Assessment of Continuous Pavement Deflection Measuring Technologies (2013)

Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

12 This chapter summarizes the results of the literature review, survey of practice, and selection of the most promising tech- nologies. It also provides a comprehensive assessment of these technologies, including repeatability, comparability, and operational characteristics. Catalogue of Deflection Measuring Devices Using this definition, several devices can be removed from consideration. Static impulse loading devices (e.g., FWD variations and Dynaplaque) are capable of sample intervals of less than 300 mm (1 ft) but must be stationary to record measurements. The FWD was thus removed from consider- ation as a continuous measurement device, but it is used as a reference device for field testing for the continuous deflec- tion devices. Plate loading devices, certain rolling wheel load devices (Benkelman beam, Dehlen curvature meter), vibrat- ing load devices (e.g., Dynaflect), and the Flexigraphe laser were also omitted from consideration. The devices consid- ered are listed in Table 3.1. There are also several devices that, although the vehicle is nonstationary while testing, keep the measurement equip- ment stationary while sampling. Because this does not co- incide with the definition of continuous used for this report, these devices were also omitted from further investigation. These include the traveling deflectometer (6 to 11 m [20 to 36 ft] spacing), Deflectograph variations (3 to 10 m [10 to 33 ft] spacing), and the Curviametre (5 m [16 ft] spacing). These devices are listed in Table 3.2. Devices that met the definition of a continuous deflection device were the Portancemetre, the moving FWD, the Mea- suring Ball, the traffic speed deflectometer (TSD), the rolling dynamic deflectometer (RDD), the rolling wheel deflectome- ter (RWD), the airfield rolling weight deflectometer (ARWD), the road deflection tester (RDT), and the image deflection measurement (IDM) device. These are presented in Table 3.3. Among the devices listed in Table 3.3, the TSD, RWD, RDT, ARWD, RDD, and IDM can apply loads similar to that of a truck, but only the first three are capable of surveying with- out the need for traffic control (each capable of surveying at 70 km/h [45 mph] or faster). The Portancemetre, the Measur- ing Ball, and the RDD operate at walking pace and are based on a vibrating wheel whose acceleration is doubly integrated to produce deflections. The TSD, RWD, RDT, and ARWD use laser measurements to determine the pavement deflection or deflection slope. Finally, the IDM, which uses image analysis methods to determine pavement deflections under loading, is still in the early stages of development. Descriptions of each device are presented below. Overview of the Most Promising Devices Portancemetre The Portancemetre continuously measures the bearing capacity of a road. A 10-kN (2.2-kips) test wheel is mounted on a specific trailer using a retractable axle (Figure 3.1). A system comprising a hydraulically unbalanced mass makes the wheel vibrate at a 35 Hz frequency providing an additional 6 kN (1.3 kips) load- ing. The instrumentation allows the measurement of the ver- tical acceleration components of the vibrating and suspended masses. Double integration of the vertical acceleration signal determines the vertical load applied to the ground and the cor- responding deflection. This method allows the measurement of the rigidity of the structure. Since the vibrating wheel is pulled at a slow speed (3 to 4 km/h [2 to 2.5 mph]), measurements are taken every 30 mm (1.2 in.), and peak deflection is normally reported at 1-m (3.3-ft) intervals. Measuring Ball and Moving FWD The Measuring Ball is a vibrating steel wheel mounted in a two-wheel, one-axle trailer towed by a car at about 5 km/h C h a p t e r 3 Analysis and Findings

13 Table 3.1. Static Measurement Devices Generic Name Equipment Type Model Equipment Characteristics Nominal Load Speed While Testing (km/h) Status as of 2010 Impulse loading device Falling weight deflectometer Falling weight deflectometer Automated impulse load 0 Production model Heavy weight deflectometer (HWD) 0 Production model Light weight deflectometer (LWD) 0 Production model Loadman LWD 0 Production model Dynaplaque Impulse generator developed by LCPC for foundation assessment 0 Production model Rolling wheel load na Benkelman beam 0 Production model Dehlen curvature meter a.k.a. South African curvature meter 0 Production model Plate load test Plate load test NA Static load applied by circular plate; primarily for foundation assessment 0 Production model The Thumper NA Plate test affixed to van 0 Production model Vibrating load device Dynaflect Dynaflect Oscillating load applied through steel wheels 0 Production model Schwinger Swiss version of Dynaflect 0 Production model Road Rater Similar to Dynaflect but higher loading 0 Production model Laser-based device Flexigraphe laser NA Laser and photocell 0 Production model Note: na = not applicable; NA = not available. Table 3.2. Moving Measurement Vehicles with Stationary Measurement Apparatus Generic Name Equipment Type Model Equipment Characteristics Nominal Load Speed While Testing (km/h) Status as of 2010 Rolling wheel load na Traveling deflectometer California vehicle with 2 Benkelman beams attached 1–1.5 Former production model Deflectograph Deflectograph (original model) Automated Benkelman beams developed by LCPC, France 2.4 Production model Double-beam deflectograph LaCroix deflectograph developed for use on rigid pavements with both beam arms on same side 2.4 Production model Pavement deflection data logging machine (PDDL) U.K. version of LaCroix deflectograph 2.5 Production model Deflecto Variation on LaCroix deflectograph 3.5 Production model PASE (pavement strength evaluator) Australian version of LaCroix deflectograph 4 Production model Deflectolab Australian version of LaCroix deflectograph 5 Production model Flash Updated, faster LaCroix deflectograph by LCPC 3–10 Production model na Curviametre Geophones mounted on chain stationary on pavement measuring vertical deflection 18 Production model Note: na = not applicable.

14 (3 mph). The vertical vibration of the wheel is measured by means of an accelerometer mounted at the wheel hub. The measurement principle is based on the idea that the stiff- ness of the ground will cause an acceleration at the wheel. The acceleration is processed in a computer housed in the towing vehicle, and the relationship between the highest Table 3.3. Moving Measurement Vehicles with Nonstationary Measurement Apparatus Generic Name Equipment Type Model Equipment Characteristics Nominal Load Speed While Testing (km/h) Status as of 2010 Vibrating mass loading na Portancemetre Vertical accelerations measured from steel wheel vibrations developed by LCPC, France 3.6 Production model Moving FWD Developed by KUAB, Sweden 30 Early prototype Measuring Ball Similar to Portancemetre 5 Early prototype Rolling dynamic deflectometer Vibrating load applied through coated steel wheels developed in Texas 5 Prototype Rolling wheel load na Airfield rolling weight deflectometer Loaded and unloaded longitudinal profiles measured by lasers; developed by QuestUSA for U.S. Air Force 35 Decommissioned prototype na Road deflection tester Loaded and unloaded transverse profiles measured by lasers; developed by VTI, Sweden 70 Prototype na Rolling wheel deflectometer Loaded and unloaded longitudinal profiles; developed by ARA for FHWA Up to 80 Prototype na Traffic speed deflectometer Doppler laser sensors measuring vertical pavement velocity; developed by Green- wood, Denmark 60–80 Prototype Image deflection measurement na IDM device Developed by LCPC, France, using structured light pattern; tested in laboratory and statically on test track 4 Early prototype Note: na = not applicable. Figure 3.1. The Portancemetre. acceleration peak and the resulting sinusoidal acceleration signal is calculated. The result is a measure of the relative stiffness of the ground and is expressed in terms of a scale from 0 to 150. The peak load generated by the vibration is not known but is likely to be significantly less than a typical heavy vehicle wheel load.

15 Through private correspondence it has been found that a device was developed by KUAB of Sweden. This device was referred to as a moving FWD and measured, without stopping, one test point every 15 m (50 ft) at 30 km/h (19 mph). The device did not measure with the same quality as an FWD, but more like a deflectograph or Benkelman beam. It was not developed further into a commercial model because it was not as accurate as the FWD, and it no longer exists. Rolling Dynamic Deflectometer The RDD is a heavy truck weighing about 200 kN (45 kips) that surveys at 4.8 km/h (3 mph). It carries a servo-hydraulic vibrator capable of producing dynamic loads up to 310 kN (70 kips) in the frequency range of 5 to 100 Hz super- imposed on a static load that can be selected within the range of 65 to 180 kN (15 to 40 kips). The load is transmitted to the road using two sets of dual wheels mounted side by side on separate axles with a spacing of 1,180 mm (4 ft) between them, meaning they are rolling inside of the road wheels of the truck (Figure 3.2). Deflections are measured by means of accelerometers mounted between further sets of dual wheels rolling between the loaded wheel sets and isolated from the dynamic system. The deflections are obtained by the double integration of the acceleration signal. Using an accelerometer, however, means that only deflections caused by the dynamic load variations can be detected. Traffic Speed Deflectometer The TSD (Figure 3.3) is an articulated truck with a rear axle load of 100 kN (22 kips), which, in the model evaluated, uses four Doppler lasers mounted on a servo-hydraulic beam to record the deflection velocity of a loaded pavement. Three Doppler lasers are positioned such that they measure deflec- tion velocity at a range of distances in front of the rear axle: 100, 200, and 300 mm (4, 8, and 12 in.); and 100, 300, and 750 mm (4, 12, and 30 in.) in the two present prototypes. The fourth sensor, acting as a reference laser, is positioned 3.6 m (12 ft) in front of the rear axle largely outside the deflection bowl. The beam on which the lasers are mounted moves up and down in opposition to the movement of the trailer in order to keep the lasers at constant height from the pavement surface. To prevent thermal distortion of the steel measurement beam, a climate control system main- tains the trailer temperature at a constant 20°C (68°F). Data is recorded at a survey speed of 70 km/h (45 mph) at a rate of 1000 Hz, that is, a 20-mm (0.8-in.) spacing of the raw mea- surements. These results are usually reported as averaged over 10 m (33 ft). Figure 3.3. Two TSD devices at the Transport Research Laboratory (TRL) test track and a computer-generated schematic. Source: Arora et al., 2006. Figure 3.2. Rolling dynamic deflectometer.

16 Rolling Wheel Deflectometer The RWD (Figure 3.4) is based on the spatially coincident method for measuring pavement deflections. Three lasers located in front of the dual tires (away from the applied load and, therefore, deflection bowl) are used to measure the unloaded pavement surface, and a fourth laser (addi- tional lasers have been added in a newer version, as dis- cussed later in the report) located between the dual tires and just behind the rear axle measures the deflected pavement surface. Deflection is calculated by comparing “spatially coincident” scans as the RWD moves forward. The RWD applies a 40-kN (9-kips) load through 2 wheels spaced 330 mm (13 in.) apart and surveys at speeds up to 80 km/h (50 mph). The deflection profile is obtained by subtracting the profile of the deflected shape from that of the unde- flected shape measured in the same location. The RWD surveys with a 2 kHz sampling rate, that is, every 11 mm (0.5 in.) at 80 km/h, and averages the deflection values over longer sections, typically 160 m (0.1 mi), to produce a single deflection measurement. Airfield Rolling Weight Deflectometer The ARWD was designed to measure runway deflections under a wheel load of 40 kN (9 kips) moving at a speed of 35 km/h (20 mph). The ARWD was developed by Quest Inte- grated, Inc. in the 1990s for the U.S. Air Force. The equipment was updated in the mid 2000s with the help of Dynatest after the equipment was transferred to the U.S. Army for repair. The equipment was decommissioned in 2010 by the Air Force after failed attempts to repair the system. The sensors were returned to Quest. The device uses four sensors to estimate the deflection due to an applied wheel load. The ARWD places one sen- sor near the load wheel and three sensors ahead of it in line with the first sensor and beyond the deflection bowl (Figure 3.5). Distances to the pavement surface are measured by the first three sensors and then again by the second, third, and fourth sensors. The measurements are timed so that they are spatially coincident. The sensors are placed 2.74 m (9 ft) apart based on the idea that the deflection bowl in most pavements at highway speeds is generally less than 2.74 m (9 ft) in radius. This implies that the beam in which the sen- sors are mounted must be greater than 8.22 m (27 ft) long. However, the deflections of the beam tend to cause signifi- cant errors, which are magnified in computations. To over- come these limitations, the ARWD uses a laser beam that is Figure 3.4. Rolling wheel deflectometer and its measurement principle. Measurement Methodology Source: FHWA, 2009. Figure 3.5. Airfield rolling weight deflectometer.

17 passed inside the physical beam as a reference to measure the deflection of the physical beam and makes corrections for this deflection in the computations. This process overcomes the problem of thermal expansion and vibrational bending of the beam. Road Deflection Tester The RDT (Figure 3.6) consists of a truck that has been retro- fitted with two arrays of laser range finders, each consisting of 20 sensors arranged in a line transverse to the direction of travel. The first array is positioned 2.5 m (8 ft) behind the front wheels, and the second array is placed 0.5 m (1.6 ft) behind the rear wheels. Thus, the distance between the two arrays is approximately 4 m (13 ft). The first array measures the transverse deflection profile largely outside of the deflection basin; the second measures the deflection profile near the center of the deflection basin. The truck has two weights of 4 kN (1 kip), which can be moved back and forth. During testing, these weights are moved to the rear of the truck. The weights are moved back to the front of the truck during transportation for better weight distri- bution. The engine of the truck is also placed in the rear, and together with the weights it can produce a force of 40 to 70 kN (9 to 15.7 kips) on the rear axle. The sampling hardware operates at a sampling frequency of 1 kHz, and at a speed of 70 km/h (45 mph). Samples are stored every 20 mm (0.8 in.), but are normally reported at 50-m (165-ft) intervals. Image-Based Deflection Measurement Device This equipment has been recently developed by the LCPC Nantes and LRPC Strasbourg in France using the projection of a structured pattern on the road surface. A camera cap- tures the surface; and software analyzes the pattern defor- mation, thereby measuring the pavement deflection. The technique has been checked in the laboratory, in static tests, and with a load moving at 4 km/h (2.5 mph). The latter tests were carried out on the LCPC circular acceler- ated loading facility at Nantes (Figure 3.7). Development is ongoing to turn this device into a robust operational mea- surement tool. Summary of Promising Devices Several continuous deflection devices exist that can measure when constantly moving and can collect data at intervals of approximately 300 mm (1 ft) or smaller using load levels typi- cal of truck loading (i.e., 40 to 50 kN [9 to 11 kips] per wheel or load assembly). These include the following three main types of devices: • Laser-based devices that measure the deflection below a moving truck load—including the TSD, RWD, RDT, and ARWD; • Devices that apply a vibratory load—including the Por- tancemetre, the Measuring Ball, and the RDD; and • The IDM device, which uses image analysis methods to determine pavement deflections under loading; this repre- sents a very promising technology, but it is still in the early stages of development. Only the devices in the first group are currently capable of surveying without the need for traffic control. The vibra- tory devices operate at walking pace, and the IDM was still being tested in a stationary mode at the time of the evaluation. Figure 3.6. Swedish road tester. Figure 3.7. IDM prototype being tested at LCPC’s pavement accelerated load facility.

18 Survey of State DOt practices and Needs The demand and the potential value of continuous deflec- tion devices were evaluated through a survey of state and provincial DOTs. The two-stage survey included questions to assess technical needs and also aimed to determine the value assigned by the agencies to the collected data. Stage I: Web Survey For the web survey, a list of potential participants provided by National Cooperative Highway Research Program (NCHRP) for a similar project was complemented with personal contact from the project team. The final list comprised 63 potential participants. A commercial web-based survey application, SurveyMonkey, was used. The web survey link was sent to 56 of the potential participants because seven had the survey service blocked. Of the 56 recipients, 44 completed the sur- vey, resulting in a response rate of 79%. Practices and Uses of Deflection Testing Thirty-five of the survey respondents (84%) replied their agency uses pavement deflection testing. As presented in Table 3.4, most deflection testing is performed exclusively in-house (77%), while only 9% of deflection testing is exclu- sively outsourced. A number of agencies (five, or 14%) rely on both in-house collection and outsourcing. The main uses of the deflection testing are as follows: • To determine the subgrade modulus or bearing capacity (97%) for flexible pavements; and • To evaluate the joint/crack transfer efficiency (59%) for rigid pavements. Deflection testing is performed mainly at the project level, but several agencies indicated that they are also doing so at the network level. All respondents use static deflection testing devices (FWD), which probably explains why more testing is performed at the project level rather than at the network level. These results are consistent with those reported by Flintsch and McGhee (2009). The average spacing between FWD tests is 133 m (437 ft) and 396 m (1,297 ft) for project and network level, respectively (Table 3.5). A number of respondents suggested it would be very valu- able to them to have a continuous deflection device, most importantly for network-level data collection. The main desired uses for the continuous device would be to do the following: • Determine subgrade modulus (65%); • Calculate overlay thickness (65%); and • Select the most appropriate type of rehabilitation (50%). The main concerns for the adoption of a continuous deflec- tion device, voiced by the respondents, were safety, accu- racy, and cost and savings. The average costs reported for project- and network-level deflection measurements were variable ($28 to $3,000 and $10 to $790, respectively). These numbers were primarily based on best estimates, with some respondents reporting typical agency values. The average current cost for network-level data collection was estimated to be around $100/km ($167/mi), based almost exclusively on rough estimates. Pavement Rehabilitation Design Applications Whether rehabilitation design procedures are based on empirical or mechanistic–empirical (ME) methods, deflec- tion measurements data can provide valuable information. This is supported by the survey responses, which indi- cate that 85% (29 out of 34 respondents; 10 respondents skipped the question) of agencies incorporate deflection testing into their pavement rehabilitation design procedure. The main uses of deflection measurements to support their pavement rehabilitation design procedures include subgrade modulus determination and overlay thickness determination Table 3.4. Summary of Responses to Question: How Does Your Agency Currently Collect Pavement Deflection Data (Please Check All that Apply)? Answer Option Response Percentage Response Count In-house collection 77.1% 27 Outsourced 8.6% 3 Both 14.3% 5 Note: This question was answered by 35 respondents and skipped by nine. Table 3.5. Summary of Responses to Question: If Your Agency Routinely Uses Deflection Testing, Please Indicate an Estimate of the Typical Testing Spacing in Feet Answer Option Response Average Response Count Project (ft) 437 24 Network (ft) 1,297 14 Note: This question was answered by 25 respondents and skipped by 19.

19 (65% of responses each), followed by determination of type of pavement rehabilitation (50% of responses). In most cases (91% of responses), agencies use deflection testing results to determine multiple parameters (two or more). Most of the respondent agencies still rely on an empiri- cal pavement design methodology, mainly the AASHTO 1993 methodology or some modification of it (26 of 34 respondents or 76%). Although several agencies (16 out of 34 respondents or 47%) also use mechanistic–empirical design methods, only four agencies (12%) were exclusively using an ME design procedure. Table 3.6 summarizes the main engineering parameters that survey respondents would like to derive from the deflection measurements. Flexible pavements’ subgrade structural bearing capacity was the most frequently men- tioned parameter, followed by deflection values and layer moduli (in bold). Pavement Management Applications Perhaps the primary benefit from a continuous deflection measuring device is its ability to provide an overall assess- ment of the structural condition of the pavement network. Deflection test results can be incorporated into an agency’s pavement management system (PMS) to support main- tenance and rehabilitation strategy scoping and resource allocation decisions, among other asset management busi- ness functions. Although most of the respondent agencies (93%) have implemented a PMS, only five incorporate the results of deflection testing into their PMS. The dollar amount that agencies are willing to pay to obtain continu- ous deflection is in the same range as the amount they cur- rently pay for FWD measurements, around $6 to $125/km ($10 to $200/mi). Stage II: Follow-Up Interviews Using results of the survey, the research team identified a subset of states to interview. A more detailed question- naire was prepared, and interviews were conducted over the phone with the following nine state DOTs: Arizona, Florida, Indiana, Kansas, Montana, New Hampshire, New Mexico, Oregon, and Virginia. This list includes states that use network-level deflection testing in their PMS (Arizona, Florida, Indiana, Kansas, and Virginia), as well as states that have some experience with a continuous deflection device, mainly the RWD (Indiana, Kansas, New Hampshire, Oregon, and Virginia). Questions were divided into three categories: (1) desired capabilities and applications of a continuous mea- suring device, (2) use of deflection data within the PMS, and (3) experience with the RWD. Desired Uses and Capabilities The responses indicate that most respondents envision using a continuous deflection device for network-level data col- lection. Within this framework, speed is perceived as the most critical characteristic even if it means sacrificing some accuracy, as long as results are comparable to static deflec- tion measurements, such as with the FWD. However, a few respondents indicated a desire to obtain a deflection basin, area parameter, or some other parameter that can be used to assess the structural capacity of the various pavement lay- ers or detect hidden problems (e.g., stripping) in some of the undersurface layers. Ground-penetrating radar (GPR) to determine layer thicknesses was reported as a desired fea- ture that could be easily added to the system. Other desir- able characteristics, which would facilitate the adoption of the technology, include: (1) ease of operation, (2) availabil- ity of fast data processing software and service support, and Table 3.6. Summary of Responses to Question: What Are the Key Engineering Parameters that You Would Wish to Derive from Deflection Testing? Key Engineering Parameter Number of Respondents Percentage Breakdown by Pavement Type Flexible Composite JCP CRCP Subgrade bearing capacity 31 86% 31 16 14 6 Deflection values 26 72% 26 14 12 5 Layer moduli 25 69% 25 12 8 3 Joint/crack transfer efficiency 19 53% 2 8 19 4 In-service structural number 18 50% 18 10 6 2 Deflection basin area 18 50% 17 7 7 1 Pavement remaining service life 14 39% 14 10 8 4 Depth to bedrock 6 17% 6 3 2 1 Note: JCP = jointed concrete pavement; CRCP = continuously reinforced concrete pavement.

20 (3) data format compatibility with the current agency data- base structures. The primary application of the continuous deflection device at the network level would be to do the following: • Help identify “weak” (i.e., structurally deficient) areas that can then be investigated further at the project level; • Provide network-level data to calculate a “structural health index” that can be incorporated into a PMS; and • Differentiate sections that may be good candidates for preservation (good structural capacity) from those that would likely require a heavier treatment (showing struc- tural deficiencies). A desired application at the project level would be to provide input for rehabilitation pavement design (e.g., input to the Mechanistic–Empirical Pavement Design Guide (MEPDG)/ DARWin-ME or other overlay thickness design method). Other desired applications mentioned include determination of long-term trend in structural capacity, overall evaluation of bounded layers (e.g., detecting stripping), and calculation of remaining service life. Important parameters that users indicated should be considered in evaluation of the equipment include speed (safety), repeatability, accuracy (and feasibility of estab- lishing correlations with existing technologies, such as the FWD), equipment cost, ease of operation, customer service (availability of service and maintenance), ease of use of the data collected, availability of software for interpretation of results, reliability, size of the vehicle, relevance of the information (e.g., use in MEPDG/DARWin-ME), and past experience. Current PMS Uses The parameters that have been used by DOTs that currently use FWD data in a PMS include the effective structural num- ber (SN) and layer moduli for flexible pavements, and AREA and k-value for rigid pavements (Virginia). Results of deflec- tion testing have also been used to compute a structural index (Indiana) and as part of a decision tree for project scoping (Indiana and Virginia), as well as for pavement overlay design and pavement deterioration monitoring (Virginia). The Kansas DOT uses the FWD’s center and last deflections to make remaining-life calculations. Some reasons cited for not incorporating results of deflection testing into the PMS include cost associated with data collection, technical issues such as software and programming, and organizational issues related the agen- cy’s structure (i.e., one division does the data collection and another runs the PMS; planning division versus maintenance division). Experience with Existing Continuous Deflection Measuring Equipment A number of interviewed state DOTs (Indiana, Kansas, New Hampshire, Oregon, and Virginia) have had some experi- ence with the RWD, mostly through FHWA-sponsored dem- onstration projects. In general, the representatives found RWD test results to be repeatable, successful in identifying problem areas, and generally well correlated with FWD test results (except in the case of Virginia). The main data col- lected included maximum deflections every 0.1 mi (tempera- ture corrected) and location, along with speed in some cases. Some reports provided by state DOT representatives also included repeated test and correlations with the FWD; these are discussed in detail in the following sections. In addition to the demonstration projects, the Kansas DOT independently contracted measurements on one seg- ment of a four-lane rural interstate highway, I-70. Testing was conducted as a screening tool to detect potential hid- den problems along the highway corridor. Although no sur- prises were found, the assessment was that the equipment performed well. Summary of User Needs The majority of agencies perform at least some deflection testing using the FWD. Most testing is performed to support project-level decisions, and only a small number of agen- cies (five) have incorporated deflection data into their PMS. Potential users in general agree that the main advantage of a continuous measuring device would be for supporting network-level decisions. The assessment of user needs sug- gests the following: • Important parameters that users indicate should be considered in the evaluation of the equipment include: speed (safety), repeatability, accuracy (and feasibility of establishing correlations with existing technologies, such as FWD), equipment cost, ease of operation, customer service (availability of service and maintenance), ease of use of the data collected, availability of software for interpretation of the results, reliability, size of the vehi- cle, relevance of the information (e.g., use in MEPDG), and past experience. • While the responses to the initial survey suggested users would like to be able to collect continuous pavement response data to support project-level decisions, the follow-up interviews showed that respondents under- stand the current limitation of the technologies and agree that network-level applications are more likely in the near future. Furthermore, respondents agreed on the need of pavement structural data to support network-level PMS decisions.

21 • At the network level, the primary application of the continuous deflection device would be to (1) help iden- tify “weak” (or structurally deficient) areas that can be then investigated further at the project level; (2) provide network-level data to calculate a “structural health index” that could be incorporated into a PMS; and (3) differenti- ate sections that may be good candidates for preservation (good structural capacity) from those that would likely require a heavier treatment (showing structural deteriora- tion and deficiencies). Selection of Candidate Devices A more detailed summary highlighting the current knowl- edge of the capabilities of the continuous deflection devices is presented in Tables 3.7, 3.8, and 3.9. These tables provide a somewhat subjective evaluation of the various technologies in the following broad categories: • Measurement capability; • Types of pavements suitable for measuring; • Sampling rate; • Accuracy; • Operating conditions; • Development status; • Available interpretation methods for different types of applications; and • Extent of usage for different applications. It should be noted that these tables have not been updated since they were produced for Phase I of this study in early 2010. The FWD is also included because it was adopted as the reference device by which to evaluate the comparability of the continuous deflection measurement devices. Although the authors understand that this technology has limitations, it represents the most common mobile deflection measur- ing device available worldwide with some degree of stan- dardization. Thus, the FWD can be used as a reasonable reference in both U.S. and European assessments of the equipment. Based on information collected in the literature review and summarized in Tables 3.7, 3.8, and 3.9, the research team identified two devices as the most promising to deliver the information needed by users under operating conditions com- patible with SHRP 2 objectives. These devices are the rolling wheel deflectometer (RWD) and the traffic speed deflec- tometer (TSD). As indicated in the Catalogue of Selection Measuring Devices, static loading devices in which the vehicle is non- stationary while testing but keeps the actual measurement equipment stationary while sampling (e.g., the deflecto- graph) were not considered because they did not conform to the adopted definition of a continuous deflection device. The device was defined as a deflection measuring device constantly moving that can collect data at intervals of approximately 300 mm (1 ft) or smaller using load levels typical of truck load- ing (i.e., 40 to 50 kN [9 to 11 kips] per wheel or load assembly). Devices that met the definition of a continuous deflection device were evaluated in detail in Tables 3.7, 3.8, and 3.9. The Portancemetre and the Measuring Ball were eliminated from consideration because they do not apply loads similar to that of a heavy vehicle. The Portancemetre measures the response of a pavement under an oscillating load with an average value of 10 kN (2.2 kips) and an amplitude of 6 kN (1.3 kips) at a speed of 3 to 4 km/h (2 to 2.5 mph). Similarly, the Measuring Ball is towed by a car at about 5 km/h (3 mph) and applies a load significantly lower than a typical heavy vehicle wheel load. Since the magnitude of the load is small and the speed slow, these devices are primarily used for quality checks on unpaved sur- faces. The RDT and ARWD were also eliminated from fur- ther consideration because the existing prototypes have been decommissioned or reassigned to other uses. The IDM system appears to be very promising; however, an operational prototype is not yet available and, at pres- ent, does not meet the survey speed requirement. The most recent published information on the IDM device described trials in which the measuring device was stationary and only the loaded wheel was moving, at just 4 km/h. The pre- liminary trial concentrated on measuring the bowl shape in the area of maximum change, partly for convenience and partly because associated modeling and analysis has suggested that such information is a useful supplement to maximum deflection when assessing pavement condition. In an ongoing project, developers hope, by mounting a ver- sion of the system on a heavy truck and with the aid of a fast camera, to measure deflection in a continuous fash- ion. Benefits of this system are the potential for continuous measurements and the potential to measure across joints if the geometry of measurement close to the loaded wheel can be resolved. The RDD was also originally identified as a good candi- date, especially for measurements on concrete pavements; however, it was not selected for further evaluation because the user needs survey indicated that the majority of the users would prefer a device that can measure at traffic speed, that is, at least at 55 km/h (35 mph). The original machine operated at around 1.5 km/h (1 mph); it has recently been updated to operate at 5 km/h (3 mph). This is still far short of operating at a speed that does not require traffic control on busy roads. Nevertheless, the combination of a suitable frequency (gener- ally between 5 and 100 Hz) and relatively low survey speed enables the assessment of the deflection response at the joints in concrete pavement that provides valuable guidance as to required rehabilitation measures.

22 Table 3.7. Detailed Device Evaluation Table Short Description Measurement Device FWD (Denmark) IDM (France) Portancemetre (France) Measuring Ball (Sweden) Measurement capability Measures nominal peak deflection or equivalent Yes Unknown Yes Unknown Measures nominal deflection bowl shape or equivalent Yes Unknown No Unknown Transverse measurement position 1 (mid-vehicle) 1 1 (mid-vehicle) 1 Type of load Impulse on 300-mm diameter plate Unknown Oscillating at 35 Hz on 1-m (3-ft) diameter wheel Unknown Peak load level 50 kN (11 kip) Unknown 10 ± 6 kN (2.2 ± 1.3 kip) Unknown Pavement type suitability Fully flexible Yes Unknown No Unknown Rigid Yes Unknown No Unknown Granular Yes Unknown Yes Unknown Sampling rate Normal sampling rate na Unknown 30 mm (1.2 in.) at 3 km/h Unknown Normal reporting rate na Unknown 1 m (3 ft) Unknown Accuracy Repeatability Very good Unknown Moderatea Unknown Comparability Good (many devices) Unknown Moderatea (several devicesa) Unknown Relation to other devices na Unknown Unknown Unknown Operating conditions Typical survey speed 0 4 km/h (2.5 mph) 3 km/h (2 mph) 5 km/h (3 mph) Surface conditions Any Drya Anya Unknown Traffic management required Yes Yes Yes Yes Status Production/development status Production model Laboratory prototype Production model Unknown % commercially developed 100 10 100 Unknown Available interpretation methods for pavement types Flexible pavements Yes No No Unknown Rigid pavements Yes No No Unknown Granular pavements Yes No Yes Unknown Use Distance surveyed Unknown 0 Unknown Unknown Surveying Flexible/rigid/composite pavements Unknown 0 Unknown Unknown Screening structurally deficient sections Yes Unknown No Unknown Defining rehabilitation strategies Yes Unknown No Unknown Designing rehabilitation treatments Yes Unknown No Unknown Note: na = not applicable. a Author’s estimate, which could not be verified by available information at this stage.

23 Table 3.8. Detailed Device Evaluation Table Short Description Measurement Device RDD (United States) Moving FWD (Sweden) ARWD (United States) Measurement capability Measures nominal peak deflection or equivalent Yes Yesa Yes Measures nominal deflection bowl shape or equivalent Yes (4 points) Unknown No Transverse measurement position 1 (mid-vehicle) 1 1 (mid-vehicle) Type of load Oscillating at 30 Hza on 300-mma (12-in.) diameter wheel Impulsea load every 15 m at 30 km/h Fixed dual wheel assembly Peak load level 55 ± 20 kN (12 ± 4.4 kip) on rigid pavement Unknown 40 kN (9 kip) Pavement type suitability Fully flexible Yes Yes Yes Rigid Yes No No Granular Yes Yes No Sampling rate Normal sampling rate 13 mm at 5 km/h (0.5 in.) Unknown 3 m at 35 km/h (9 ft) Normal reporting rate 0.6–0.9 m (2–3 ft) 15 m at 30 km/h (50 ft) 25 m (90 ft) Accuracy Repeatability Gooda Unknown Unknown Comparability na (Only one device) Unknown na (Only one device) Relation to other devices Strong with FWD Unknown Unknown Operating conditions Typical survey speed 1.5–5 km/h (1–3 mph) 30 km/h (19 mph) 35 km/h (22 mph) Surface conditions Anya Unknown Dry Traffic management required Yes Yes Probably Status Production/development status Working prototype Decommissioned prototype Decommissioned prototype % commercially developed 80 Unknown 60 Interpretation methods available for . . . Flexible pavements Yesa Unknown No Rigid pavements Yesa Unknown No Granular pavements No Unknown No Use Distance surveyed Unknown Unknown Unknown Flexible/rigid/composite pavements surveyed Unknown Unknown Unknown Screening structurally deficient sections Yesa Unknown Unknown Defining rehabilitation strategies Ya Unknown Unknown Designing rehabilitation treatments Y Unknown Unknown Note: na = not applicable. a Author’s estimate, which could not be verified by available information at this stage.

24 Table 3.9. Detailed Device Evaluation Table Short Description Measurement Device RWD (United States) RDT (Sweden) TSD (Denmark) Measurement capability Measures nominal peak deflection or equivalent Yes Yes Yesa Measures nominal deflec- tion bowl shape or equivalent No Yes Yesa (3 points) Transverse measurement position 1 wheelpath 2 wheelpaths 1 wheelpath Type of load Fixed dual wheel assembly Fixed dual wheel assembly Fixed dual wheel assembly Peak load level 40 kN (9 kip) 40–70 kN (9–16 kip) 50 kN (11 kip) Pavement type suitability Fully flexible Yes Yes Yes Rigid No No Maybe Granular No No No Sampling rate Normal sampling rate 11 mm at 80 km/h (0.4 in.) 20 mm at 70 km/h (0.8 in.) 20 mm at 70 km/h (0.8 in.) Normal reporting rate 30 m (100 ft) 50 m (165 ft) 10 m (30 ft) Accuracy Repeatability Gooda Moderateb Goodb Comparability na (only one device) na (only one device) Goodb (two devices) Relation to other devices Strong with FWD Poor with FWD and deflectograph Strong with FWD and deflectograph Operating conditions Typical survey speed 80 km/h (50 mph) 70 km/h (45 mph) 70 km/h (45 mph) Surface conditions Dry Dry Dry Traffic management required No No No Status Production/development status Working prototype Prototype reassigned to other uses Two working prototypes % commercially developed 80a 60 95 Interpretation methods available for pavement types Flexible pavements Yesb No Yesb Rigid pavements No No No Granular pavements No No No Use Distance surveyed 12,000+ km (7,500+ mi) Unknown 30,000+ km (19,000+ mi) Flexible/rigid/composite pavements surveyed Mostly Flexible/composite Unknown Mostly Flexible/composite Screening structurally deficient sections Yes No Yesb Defining rehabilitation strategies Yesb No Yesb Designing rehabilitation treatments No No No Note: na = not applicable. a Measures vertical deflection velocity, which is converted to deflection slope by dividing by the horizontal vehicle velocity. Three deflection slopes enable maximum deflection and part of the bowl shape to be estimated. b Author’s estimate, which could not be verified by available information at this stage.

25 Detailed Description of the Selected equipment This section expands on the characteristics of the selected devices and provides a preliminary assessment of their tech- nical capabilities based on the information collected from trials and demonstration projects from the United States and Europe. Both selected devices, the RWD and TSD, conduct measurements (deflection or deflection velocity of a loaded pavement, respectively) under a truck axle and at speeds close to that of traveling traffic. By contrast, the technology currently in use, the FWD, applies an impulse load (using known weights that are dropped from specific heights onto a load plate) and measures the response (deflections) to those loads at seven to nine surface locations, starting at the center of the loading plate and extending radially up to 1.8 m (72 in.) from the load plate center. The deflection basin at each test location is indicative of the stiffness of the underlying pavement structure and given the various sensors, allows a quantification of the structural capacity of various lay- ers or layer types. A production rate of approximately 4 lane-km (2.5 lane-mi) per day is typical, assuming testing at an interval of 23 m (75 ft). This testing scheme is best suited for project- level testing. Agencies employing FWD for network-level FWD test protocol may adjust this test spacing to achieve a produc- tion rate of approximately 32 to 40 lane-km (20 to 25 lane-mi) per day, assuming an interval of 320 m (0.2 mi) (Diefenderfer, 2010). Rolling Wheel Deflectometer (RWD) Detailed Description The RWD has been developed by Applied Research Associates, Inc. (ARA), with support from the FHWA. The RWD system is mounted within a custom-designed 17-m (53-ft) semitrailer. The measured deflection is the response from one-half of an 80-kN (18-kip) single-axle load traveling at normal traffic speeds. In a previously tested version, an aluminum reference bar, suspended beneath the trailer, contained four laser sensors to measure the distance to the pavement surface (Figure 3.8). The RWD uses a spatially coincident methodology for measur- ing pavement deflection. This method was originally developed by the Transport and Road Research Laboratory (TRRL) and implemented by Dr. Milton Harr at Purdue University. In the evaluated model, three lasers are used to measure the distance to the unloaded pavement surface (i.e., forward of and outside the deflection basin), and a fourth laser, located between the dual tires and just behind the rear axle, measures the distance to the deflected pavement surface. The deflec- tion is calculated by comparing the laser scans profile as the RWD moves forward. The beam uses a curved extension to pass under and between the dual tires, placing the rearmost laser approximately 150 mm (6 in.) rear of the axle centerline and 178 mm (7 in.) above the roadway surface. The wheels are spaced a safe distance from the laser and beam using cus- tom lugs spacers (Steele and Vavrik, 2006). The equipment has recently been retrofitted to enclose the laser sensors and an additional sensor has been added, as discussed in New Improvements to the Evaluated Equipment in Chapter 4. At 89 km/h (55 mph), the RWD’s 2-kHz lasers take readings approximately every 11 mm (0.5 in.), resulting in extremely large data sets. The average deflection every 160 m (0.1 mi) is typically reported; this averaging helps reduce scatter and file size. Figure 3.9 shows an example of the data collected during the demonstration in Virginia. Each square represent the average deflection for each 160 m (0.1 mi); the continu- ous line represents a 1.6-km (1-mi) moving average. The data is generally filtered to eliminate outliers due to bridges, Source: Diefenderfer, 2010; and FHWA, 2009. Figure 3.8. RWD during testing in Virginia and close-up of laser sensor placed between dual tires.

26 sudden changes in speed, and so forth, before the analy- sis is performed. Additional details of the RWD deflection measurement process are presented elsewhere (ARA, 2005a; Steele and Hall, 2005). Recent upgrades to the RWD include improved laser sensors that are located within a temperature- controlled housing. The result of these improvements on comparison testing with FWD is ongoing and unknown at this time. The RWD technology can test approximately 320 to 480 km (200 to 300 lane-mi) per day. A potential benefit of the RWD is that the load, loading mechanism, and loading rate of the RWD are thought to match more closely the actual dynamic effects on pavements caused by vehicle loading. In addition, the RWD testing is conducted at or near highway speeds with limited or no traffic control requirements and minimal interruption to the highway users. However, the RWD does not currently allow for some of the structural capacity analysis offered by the FWD. In its current state of development, it is anticipated that the RWD could be used to prescreen the pavement network to identify areas that might require additional and more detailed study at the project level using traditional techniques such as the FWD, or to identify segments that could be good candi- dates for pavement preservation. The FHWA has sponsored RWD demonstration projects throughout the United States. Testing has been conducted in coordination with at least 16 U.S. state highway agencies, including those in California, Colorado, Connecticut, Indiana, Iowa, Kansas, Kentucky, Minnesota, New Hampshire, New Jersey, New Mexico, Ohio, Oregon, Texas, Virginia, and West Virginia; on a federal road under the jurisdiction of the fed- eral lands (Natchez Trace); and on several test tracks including the National Center for Asphalt Technology (NCAT), Virginia Smart Road, and MnRoad. According to documented test reports, the total mile- age tested exceeded 11,300 km (7,000 lane-mi). Each state agency self-developed its test plan. Several of these tests included FWD measurements on the same sections; how- ever, not all were conducted at the same time as the RWD measurements were taken. A few of the demonstrations included multiple runs to assess the repeatability of the device. One of the earliest RWD test reports was authored by Arora et al. (2006) and described testing in Texas in 2004. The RWD was used to test approximately 425 km (264 lane-mi) (a vari- ety of state routes with five repeat runs per roadway) with some companion FWD testing. FWD deflection values ranged from approximately 100 to 1,300 microns (4 to 50 mils). The authors stated that the RWD testing was repeatable, based on visual observation of the plotted deflection results. In dis- cussing a relationship of RWD to FWD deflection results, the authors stated that some relationship exists although the data “shows some scatter especially at smaller deflection values.” The authors suggested that lower deflection values might be measurable only at lower speeds. Gedafa et al. (2008) reported on RWD testing of 333 km (207 mi) of non-interstate highway in Kansas in 2006. The results of the RWD testing were compared to FWD testing that was conducted from 1998 to 2006. RWD testing was per- formed at 89 km/h (55 mph) with deflection readings aver- aged every 160 m (0.1 mi). The FWD data were collected at 5 to 10 points per mile. The average FWD center deflec- tion value ranged from approximately 0.13 to 0.45 mm (5 to 18 mils) (40 kN [9 kips] load). The results showed that the RWD deflection reading and the FWD center deflection value were statistically similar based on a significant difference test statistic. A linear regression analysis was also performed that 0 2 4 6 8 10 12 14 105 110 115 120 125 130 135 140 R W D D ef le ct io n (m ils ) Mile Marker RWD Deflection (0.1 mile average) Moving Average Sectioning (based on last resurfacing) Figure 3.9. Example of deflection data collected on eastbound I-64 in Virginia in 2006.

27 showed a strong correlation between FWD and RWD deflec- tion readings. Virginia reported RWD testing on portions of two inter- state routes and a loop consisting of primary rural highways in 2005 (Diefenderfer, 2010). All RWD testing was done at or near the prevailing traffic speed. Companion FWD test- ing was conducted in 2006 on the two interstate test sections. The FWD deflection values ranged from 0.08 to 0.38 mm (3 to 15 mils), with a majority less than 0.2 mm (8 mils) (40 kN [9 kips] load). The two interstate test sections com- prised hot-mix asphalt (HMA) (200–300 mm [8–12 in.]) over compacted aggregate and HMA (100–150 mm [4–6 in.]) over CRCP (200 mm [8 in.]). Statistical testing of RWD repeat- ability was performed by use of a non-paired t-test assuming equal variances. The results showed that for 8 of 15 trials on interstate highways and all non-interstate test sections, the RWD data were repeatable. A poor linear correlation was found between the RWD and FWD measurements (adjusted R2 values less than 0.2). However, the FWD measurements were taken several months after the RWD measurements and only on interstate sections with relatively low and uniform deflections. The results suggested that the deflection value may be influenced by surface texture as the standard devia- tion varied approximately at locations where the HMA surface mixture also varied. Traffic Speed Deflectometer (TSD) Detailed Description The TSD (Figure 3.10) is mounted on an articulated truck with a rear axle load of 100 kN (22 kips), which, in the model evaluated, uses four Doppler lasers mounted on a servo- hydraulic beam to record the vertical deflection velocity of a pavement as it is loaded by one of the dual wheel axles. Three Doppler lasers are positioned such that they measure deflection velocity at a range of distances in front of the rear axle. The fourth sensor is positioned 3.6 m (12 ft) in front of the rear axle, largely outside the deflection bowl, and acts as a reference laser. The beam on which the lasers are mounted moves up and down in opposition to the movement of the trailer in order to keep the lasers at a constant height from the pavement surface. To prevent thermal distortion of the steel measurement beam, a climate control system maintains the trailer temperature at a constant 20°C (68°F). Two proto- types had been developed at the time of the evaluation by the manufacturer, Greenwood Engineering A/S of Denmark. One is owned and operated by the Danish Road Institute (DRI); the other is owned by the U.K. Highways Agency (HA) and operated on their behalf by the U.K. Transport Research Laboratory (TRL). Newer production devices have incorpo- rated more Doppler laser sensors. The lasers are mounted at a small angle to measure the hori- zontal vehicle velocity, the vertical and horizontal vehicle sus- pension velocity, and the vertical pavement deflection velocity. Due to its location midway between the loaded trailer axle and the rear axle of the tractor unit, the reference laser is expected to measure very little vertical pavement deflection velocity, and its response can therefore be used to remove the unwanted sig- nals from the three measurement lasers. When accurately cali- brated, the TSD produces measurements of deflection velocity that depend on driving speed. To remove this dependence, the deflection velocity is divided by the instantaneous survey speed to give a measurement of deflection slope, as illustrated in Fig- ure 3.10. Deflection velocity is measured in mm/s while survey speed is measured in m/s; therefore, deflection slope measure- ments are given in units of mm/m (Ferne et al., 2009b). The DRI Machine The DRI and Greenwood jointly developed the TSD, initially called the high speed deflectograph (HSD), and have pub- lished a number of papers on this work. An early independent Figure 3.10. Computer render of the TSD, with the TSD in operation (inset).

28 evaluation by the Laboratoire Central des Ponts et Chaussées (LCPC) in 2003 (Simonin et al., 2005) showed that even though the early prototype had limitations, it demonstrated good repeatability in the short term and a good degree of correlation with the maximum deflection recorded by other devices such as the FWD and the deflectograph. Other DRI publications confirm some aspects of this work when assessing a developed version of this device that they currently own and operate. The DRI (Figure 3.11) also have practical experience operating the device on their network, having covered their main network from 2005 to 2007. Based on this experience, Baltzer (2009) reported the DRI daily survey coverage for the device of around 170 to 225 km (105 to 140 mi). The Highways Agency TSD The second prototype TSD currently in operation is owned by the U.K. Highways Agency (HA) and operated by the U.K. Transport Research Laboratory (TRL). The TRL has reported its development and performance in recent confer- ence papers (Ferne et al., 2009a and 2009b). An example of typical survey results expressed over 10-m (33-ft) and 100-m (1/16-mi) lengths are provided in Figure 3.12. Equipment Status At the time of completion of the Phase I evaluation (Febru- ary 2010), the RWD was a working prototype, with only one such prototype in existence. This prototype has been recently upgraded by adding an additional laser sensor and provid- ing temperature control to the beam that supports the lasers. There were also two working TSDs with new models (with more sensors) under construction. At the time of the cur- rent report, two additional TSDs have been constructed, for agencies in Italy and Poland. A fifth device is currently under construction for the South African Highway Administration. The latter three devices incorporate improvements to the Figure 3.11. DRI device measuring side by side with the U.K. TSD at TRL in 2008. 0.000 0.100 0.200 0.300 0.400 0.500 0.600 0.700 0.800 0.900 1.000 0 500 1000 1500 2000 2500 3000 3500 4000 D efl ec ti on S lo pe [m m /m ] Chainage [m] 10m averages 100m averages Figure 3.12. TSD deflection slope at 10 m (33 ft) and 100 m (330 ft) means against distance.

29 earlier two prototypes, such as additional velocity sensors and improved calibration facilities. The website of the manu- facturer, Greenwood Engineering A/S, currently designates the TSD as a production model and gives details about some features added to the equipment. Available Data Interpretation Methods Fully developed methods of data interpretation are not avail- able specifically for either device. In principle, the RWD deflection should be usable as an input for any procedure that requires only a maximum deflection response as its pavement response input. It has been proposed that the velocity mea- surements from the Danish device configuration can be used to produce surface curvature index values that are akin to those measured by an FWD. Therefore, they should be viable as input for procedures that require only surface curvature index (SCI) as a pavement response input. Greenwood Engineering has developed a method to inter- pret the deflection velocity from the TSD by using a beam on elastic foundation approach. The model builds a full deflection basin using a two-parameter function and information from three deflection slope measurements. The model proposed for the deflection basin is given below (Krarup et al., 2006): d x A B Cos Bx Sin Bx e Bx( ) = − ( )+ ( )( ) − 2 3 1( . ) Where: d(x) = deflection at any point within the basin, x = distance of deflection from center of load, and A and B = constants to be optimized. The implications of the model are that the deflection slope directly under the load is zero. This can be seen directly by differentiating the model with respect to the variable x, and evaluating it at zero. In the United Kingdom, the main method for interpreting pavement deflection response uses the maximum deflection measured by a slow-moving deflectograph to estimate resid- ual lives and strengthening requirements. Research in the United Kingdom, reported earlier in this section, has shown that equivalent deflectograph values can be estimated from TSD measurements, thus providing an approximate interpre- tation methodology for the English strategic road network. Routine network surveys of this network started in Novem- ber 2009, and the measurements are being converted to one of four structural condition categories before being stored in the Highways Agency Pavement Management System for use by the agents responsible for the various parts of the network to assist them with their management of the network. The above information refers to just flexible pavements. For rigid and unpaved roads, there is as yet no explicit interpretation method for either device although recent research has suggested that the TSD equipment may have a role in the preliminary evaluation of the joint condition of rigid pavements. Use Table 3.10 and Table 3.11 summarize the status of survey cov- erage for each type of device as of February 2010. For the RWD, most of this testing was conducted on flexible pave- ments with a total survey length of more than 12,100 km (7,500 mi). For the TSD, close to 100% is of flexible construc- tion with a total surveyed length of more than 21,000 km (13,000 mi) in the United Kingdom. The Danish device has covered well over 10,000 km (6,500 mi) in Denmark. In 2010 the device was commissioned to cover 20,000 km (12,500 mi) of the road networks in two Australian states, as described by Baltzer et al. (2010). To date, little data have been explicitly used for specific pavement management activities, so it is not possible to determine the appropriate use of data such as screening structurally defi- cient sections, defining rehabilitation strategies, or designing rehabilitation treatments. phase I assessment The devices were further evaluated to determine their capa- bilities based on existing data found in the literature review and obtained from interviews with DOT officials. Both can- didate devices have been used in pilot projects over multiple locations, and evaluation of accuracy and repeatability has been conducted and reported. This section presents the past research conducted on the devices. Accuracy Equipment accuracy has many interpretations, whether considering individual measurement accuracy or the over- all accuracy of the device. Therefore, accuracy is considered under a number of factors: choice of averaging length, short- term repeatability, long-term repeatability, effect of external variables, comparability, and comparison with other deflec- tion measures. The term “short-term repeatability” indicates that the surveys have been repeated as quickly as possible in order to minimize the effect of external environmental con- ditions such as temperature changes on the results. When assessing long-term repeatability, the surveys were carried out over a period of several days or even weeks, so the results could potentially include the external effects. For each factor, the capability of the two devices is considered on the basis of available information collected in Phase I of this study and is detailed in this section. This section presents preliminary

30 values based on a limited dataset collected for previous stud- ies; a more rigorous calculation based on a larger dataset is given later. Choice of Averaging Length The RWD demonstration at the Eastern Federal Lands Highway Division (EFLHD) included a comparison using 160- and 32-m (0.1- and 0.02-mi) intervals for averaging the results. Figure 3.13 illustrates the effect of a shorter averaging interval on one of the tested sections. The figure suggests that decreasing the sample unit length does not significantly affect the overall trend (or the mean deflection for the overall sec- tion), but it does increase variability over the section length. On the basis of these limited results, the manufacturer has cautioned against using a sampling interval that is too small to reduce random error sufficiently (ARA, 2005b). Table 3.11. Lane Lengths Surveyed by U.K. HA TSD by Survey Type (September 2005 to February 2010) Approximate Length Surveyed Survey Type km mi European continent 350 220 TRL track 1,300 800 Local roads 10,400 6,500 Scottish road network 800 500 English trunk road and motorway network 8,200 5,100 Total 21,050 13,100 Table 3.10. Summary of Tests Conducted in the U.S. as of February 2010 Location Date Lane-mi FWD Data Availability FWD Sampling Frequency Repeat Runs Road Functional Class Louisiana 2009 NA Good NA No NA Kansas 2008 466 Good 0.1 No U.S. and state New Mexico September 2008 443 Good 0.1 No U.S. Colorado October 2008 230 Partial 0.1 No Int., U.S., and state New Hampshire July 2007 712 NA NA No Int., U.S., and state Connecticut September 2007 204 NA NA No Int., U.S., and state Kansas July–August 2006 506 Good 0.1 Research sites U.S. and state Iowa July 2006 278 Good 0.1 No Int., U.S., and state Oregon June–July 2006 579 Partial 0.1 No Int., U.S., and state California June–July 2006 685 NA NA Research sites Int., U.S., and state Virginia October 2005 488 Partial 0.1 3 interstate, 2 primary Int., U.S., and state New Jersey October 2005 803 Partial Varied No Int., U.S., and state Minnesota September 2005 NAa Partial 0.1 MnRoad sites U.S., state, and county Kentucky–Ohio–West Virginia September 2005 437 Good in OH 0.1 No Int., U.S., and state Indiana September 2004 688 NA NA Yes U.S. and state Natchez Trace November 2004 800+ NA NA No U.S. park service NCAT July 2005 NAb NA NA Yes Test track Texas July 2003 264 Good NAc Yes 38 test sections; U.S. and state routes Total 7,583+ Note: NA = not available; Int. = Interstate. a Testing on county roads and MnRoad facility; mileage not recorded. b Testing at varying speeds on 1.3-mi test track; mileage not recorded. c FWD, MDD, and RDD testing on specific spots (see FHWA, 2009).

31 The TSD collects raw data at around 1000 Hz, but there is significant random noise in this raw signal. Even when averaged over a 0.1-m (4-in.) length, this noise is noticeable, as illustrated by the black line in Figure 3.14. Also shown in this figure are 1-m (40-in.), 10-m (33-ft), and 100-m (330-ft) contiguous averages. This site is generally of a very variable and weak composite construction with corresponding very variable deflections. The figure illustrates how some features of the true deflection profile are probably suppressed as the averaging length increases from 1 m to 100 m (3.3 ft to 333 ft). Therefore, in the United Kingdom, it has been decided to store results at 1-m (3.3-ft) intervals and generally report results as 10-m (33-ft) averages. From chainage (distance) 215 m to 250 m (705 ft to 820 ft) the construction changes to a rigid concrete construction, which has a relatively low and uniform deflection response. This is demonstrated in Figure 3.15, which Source: ARA, 2005b. Figure 3.13. Effect of sample unit length on RWD deflections. 0 0.5 1 1.5 2 2.5 0 50 100 150 200 250 300 D efl ec ti on Sl op e [m m /m ] Chainage [m] Slope Vs. Chainage, TRL Large Loop (Long Straight) 23-Jun-09, P100 0.1m 1m 10m 100m Figure 3.14. TSD deflection slope profile for Transport Research Laboratory track with various averaging lengths.

32 shows a 50-m (165-ft) section of Figure 3.14 covering from 200 m to 250 m (655 ft to 705 ft). This exaggerated scale sug- gests that on weak composite pavement even a 10-m average length hides some true deflection variations. This is discussed further in the following repeatability section. Short-Term Repeatability Several of the RWD demonstration projects included multiple runs. Figure 3.16 shows the results of conducting multiple runs at the MnRoad test facility. RWD deflections are averaged over 15-m (50-ft) intervals. Figure 3.16a shows 10 repeat passes on the inner lane of the low-volume road loop. This loop included 11 asphalt concrete (AC) test sections with different pavement structures. The sections included 4 cells (Nos. 27 through 30) in very poor condition and one cell (No. 31) that had been recently overlaid and was in excellent condition. These conditions were reflected in the deflection profile. The repeatability standard deviations considering the individual 160-m (0.1-mi) segments ranged from about 25 microns (1 mil) for the section recently overlaid to approximately 100 microns (4 mils) for the cells in poor condition. Fig- ure 3.16b presents three repeated runs on the outer (driving) lane of the mainline experiment, which included AC test cells of variable ages and AC layer thicknesses ranging from 100 to 380 mm (6 to 15 in.). Deflections were very uniform within the majority of cells, with standard deviations typically rang- ing from 50 to 75 microns (2 to 3 mils) (ARA, 2006). In general, the various evaluations showed relatively good repeatability that seemed to be appropriate for network- level analysis. On the other hand, Diefenderfer (2010) con- ducted statistical testing of RWD repeatability by use of a non-paired t-test assuming equal variances and the results showed that the RWD data were repeatable for only 8 of 15 trials. Of the non-interstate test sections, 100% of the trials were found to be repeatable. This raised some ques- tions about the applicability of the system for detailed (e.g., project-level) evaluations, especially in areas where low deflection ranges are expected. Figure 3.17 shows an exam- ple of three repeated runs on a stretch of interstate highway in Virginia. The repeatability standard deviation for the average 0.1-mi segments is shown at the bottom of the chart; the average standard deviation was 20 microns (0.79 mils), or 17% of the mean deflection. However, the repeatability standard deviation for the average values for the entire tested sections showed good repeatability (Table 3.12). For the U.K. Highways Agency TSD, Ferne et al. (2009b) reported the results of testing conducted to investigate the effect of testing speed. Measurements were taken on the TRL track over a range of speeds, and the results showed that as the speed increased, a slightly lower value of deflection slope 0 0.5 1 1.5 2 2.5 200 205 210 215 220 225 230 235 240 245 250 D efl ec ti on S lo pe [m m /m ] Chainage [m] Slope Vs. Chainage, TRL Large Loop (Long Straight) 23-Jun-09, P100 0.1m 1m 10m 100m Figure 3.15. TSD deflection slope profile for TRL track with various averaging lengths: rigid section and transition.

33 (a) Low-volume road (10 passes) (b) Mainline experiment (3 passes) Source: ARA, 2006. Figure 3.16. RWD deflections at 50-ft intervals at MnRoad. was recorded. This being the case, the testing speeds used during further tests were strictly controlled to enable repeat- able results to be obtained. Figure 3.18 shows a sample of the results of 6 runs on a 440-m (0.25-mi) length of the TRL track, which had mainly a composite pavement but included a 50-m (165-ft) length of jointed concrete at a nominal speed of 70 km/h (45 mph). The data showed reasonable short- term repeatability, with a relatively low standard deviation despite the relatively wide range of deflection slopes measured (i.e., changing by a factor of over seven). Both the LCPC assessment of the first DRI prototype and TRL’s assessment of the HA TSD suggest that the level of repeatability is not particularly dependent on the mean level of the slope. Therefore, in this section of the report they are given in absolute, not proportional, terms. The consistency of the latest version of the HA TSD has been assessed on a small number of U.K. roads. Results of these tests in terms of the standard deviation of the mean values of each of five runs of various lengths have been sum- marized in Table 3.13 for the P100 and P300 TSD sensors.

34 This good level of short-term repeatability of the TSD that is achievable under controlled conditions can also be observed graphically. Figure 3.19 shows a 20-m (66-ft) sam- ple length of the TRL track with the TSD P100 sensor results calculated at 1-m (3-ft) intervals plotted against distance for all five repeat runs. The repeated identification of weak spots at the same location (i.e., stations 187 to 188 m [613 to 617 ft] and 197 to 198 m [646 to 650 ft]) is clearly seen. Long-Term Repeatability Figure 3.20 shows a sample of five runs recorded over 5 months (September 2009 to February 2010) on 4 km (2.5 mi) of a U.K. site, which is of flexible composite construction, with a nomi- nal testing speed of 70 km/h (45 mph). The data shown has been averaged into 100-m (330-ft) lengths so that the change in deflection slope is more visible. Table 3.14 shows that the standard deviations of the mean values of each of the five runs are very similar to those in Table 3.13, meaning that repeatability apparently changed little when assessed over longer periods of time. This suggests that changes in pavement temperature have only a small effect on the measured slope as surface temperature changed from 4°C to 19°C (40°F to 66°F) during these surveys. This is not unexpected as the pavement is of flexible composite construction. In the United Kingdom, deflection surveys on strong flexible composite pavements are left uncorrected for pavement temperature. Figure 3.21 shows a 500-m selected section of the same U.K. Site B as in Figure 3.20 but with 10-m (33-ft) averaging used. Although the runs were performed over 5 months, all five surveys identify the weaker section in the same location, that is, from 2,350 m to 2,400 m (7,710 ft to 7,874 ft). Comparability The comparability of the RWD cannot be assessed because only one such device has been produced. Even with two devices, an assessment of true device reproducibility, such as with the TSD, is not possible. However, some limited com- parisons have been made but not published. One such com- parison was made in September 2008 in the United Kingdom. 0 2 4 6 8 10 12 14 16 18 20 0 1 2 3 4 5 6 7 8 9 Distance (miles) D ef le ct io n (m ils ) Pass 1 (P1) Pass 2 (P2) Pass 3 (P3) Std Deviation (0.1 mi) Correlation Matrix P1 P2 P3 P1 1.00 0.69 0.63 P2 0.69 1.00 0.65 P3 0.63 0.65 1.00 Figure 3.17. Repeated runs of the RWD on I-64 in Virginia. Table 3.12. Summary Statistics for Average Section RWD Deflections of I-64 and I-81 Highway Average (mils) Repeatability Std. Dev. (mils) Average (microns) Repeatability Std. Dev. (microns) Eastbound I-64 4.53 0.28 115.1 7.3 Westbound I-64 4.7 0.26 119.4 6.6 Northbound I-81 7.77 0.14 197.4 3.5 Southbound I-81 5.08 0.59 129.0 15.0

35 Figure 3.22 illustrates the consistency between these two devices when operating on the same 11-km (7-mi) length of varying construction and deflection response, bearing in mind that the two devices measure in different wheelpaths. Figure 3.23 illustrates differences between the wheel- paths as revealed by surveys conducted by a slow-speed deflectograph that records peak deflections in both wheel- paths at the same time. Comparison of the two figures con- firms that any differences between the two devices are likely explained by the different deflection responses of the two wheelpaths. Comparison with Other Deflection Measures Investigations have been conducted comparing the RWD and TSD to other deflection measuring equipment, in particular the FWD. However, since the FWD and rolling wheel devices load the pavement in different ways, the relationship between them will not necessarily be one of equality. Several of the RWD demonstrations included FWD mea- surements on at least some sections; however, not all were conducted at the same time that the RWD measurements were obtained. Figure 3.24 presents examples of section- level comparisons between RWD and FWD maximum deflections. In general, the RWD reports collected during follow-up interviews suggest that the average results of the RWD deflection measurements (normalized to a standard temperature) correlate relatively well with the average max- imum FWD deflection when aggregated by homogeneous sections. The example from Texas (Figure 3.24d) suggests 0 0.25 0.5 0.75 1 1.25 1.5 1.75 0 50 100 150 200 250 300 350 400 Concrete 450 Sl op e [m m /m ] Chainage [m] Std Deviation (10m) Source: Ferne et al., 2009a. Figure 3.18. Repeatability of deflection slope at 70 km/h (45 mph) on the TRL track. Table 3.13. Repeatability Standard Deviation of TSD for Five Runs in Terms of TSD Slope for Short-Term Repeatability Site Overall Length Sensor Repeatability Standard Deviation (mm/m) Averaging Length 10 m (33 ft) 100 m (330 ft) 160.9 m (1/10 mi) TRL track 291 m/ 0.2 mi P100 0.071 0.046 0.040 P300 0.053 0.038 0.034 U.K. Site A 1,080 m/ 0.7 mi P100 0.037 0.012 0.010 P300 0.037 0.013 0.011 U.K. Site B 3,871 m/ 2.4 mi P100 0.054 0.025 0.023 P300 0.071 0.052 0.051

36 0.000 0.500 1.000 1.500 2.000 2.500 3.000 180 182 184 186 188 190 192 194 196 198 200 D efl ec ti on S lo pe [m m /m ] Chainage [m] Slope Vs. Chainage, TRL Large Loop (Long Straight) 23-Jun-09, P100 (1m data) T120090623001 T120090623002 T120090623003 T120090623007 T120090623008 Figure 3.19. Selected 20-m (66-ft) length of TSD slope data as 1 m (3.3 ft) means on TRL track. 0.000 0.100 0.200 0.300 0.400 0.500 0.600 0.700 0 500 1000 1500 2000 2500 3000 3500 4000 D efl ec ti on S lo pe [m m /m ] Chainage [m] Slope Vs. Chainage, A33 SB, P300 100m Averages 10-Feb-10 09-Nov-09 06-Nov-09 09-Oct-09 18-Sep-09 Figure 3.20. Long-term repeatability of deflection slope (P100) at 70 km/h (45 mph) on U.K. Site B, 100-m (330-ft or 0.06-mi) averages.

37 that the correlation is better on sections with high deflec- tions (having “weak” structural capacity). This is expected because a wider range of the dependent variable increases the correlation coefficient, as is discussed later in this report. Additional analysis was performed on the New Mexico data that were provided for this project and is presented in the comparability section of the report. The data were col- lected on U.S. Route 550 in New Mexico and were provided by ARA for this project. Many comparisons have been made between the TSD and other deflection measuring devices. The early independent evaluation by the LCPC in 2003 (Simonin et al., 2005) of the first Danish Research Institute (DRI) prototype showed a strong correlation (R2 = 0.86) between the slope measured by the DRI TSD and the peak central deflection measured by an FWD over a range of sites in France. Comparisons in the United Kingdom between the HA TSD and FWD measurements have been less common to date because the main emphasis has been on comparison with the deflectograph, the prime deflection measuring device used in the United Kingdom. However, some comparisons of deflection profiles on specific sites have been made. For example, Figure 3.25 shows a comparison between an FWD central deflection profile at 2-m (6.6-ft) intervals compared with a TSD deflection slope profile averaged over the same intervals on a 400-m (1,300-ft) section of the TRL track. The pavement structure includes both weak flexible compos- ite materials and rigid concrete. Similarities in the shapes of the two profiles are very encouraging despite the 4-year inter- val between the surveys. It should be noted, however, that the vertical scales of the two parameters are relatively arbitrary and have been adjusted to approximately align the two pro- files vertically. In the United Kingdom, extensive comparisons have been made between the TSD slope and peak deflection measured by a U.K. deflectograph. Figure 3.26 illustrates the average relationship, together with 95% confidence limits, between deflectograph (DFG) values and TSD slope values for the P300 sensor, which is located 300 mm (1 ft) from the center Table 3.14. Repeatability Standard Deviation of TSD for Five Runs in Terms of TSD Slope for Long-Term Repeatability Site Overall Length Sensor Repeatability Standard Deviation of TSD Slope (mm/m) Averaging Length 10 m (33 ft) 100 m (330 ft) 160.9 m (0.1 mi) U.K. Site B 3871 m (2.4 mi) P100 0.065 0.040 0.039 P300 0.063 0.038 0.038 0.000 0.100 0.200 0.300 0.400 0.500 0.600 0.700 0.800 0.900 1.000 2200 2250 2300 2350 2400 2450 2500 2550 2600 2650 2700 D efl ec ti on S lo pe [m m /m ] Chainage [m] Slope Vs. Chainage, A33 SB, P100 10m Averages 10-Feb-10 09-Nov-09 06-Nov-09 09-Oct-09 18-Sep-09 Figure 3.21. TSD slope data, as 10 m (33 ft) means, for five repeat runs over a 5-month period on U.K. Site B.

38 of the load. The analysis covered almost 5,000 10-m (3.3-ft) segments on a wide range of U.K. roads. Phase II Assessment The data obtained in Phase I and collected in the field trials were analyzed to evaluate the repeatability of the TSD and RWD and the comparability of both devices by comparison with the FWD. In the Detailed Description of the Selected Equip- ment section, results of different studies that evaluated RWD (conducted in the United States) and TSD (conducted in Europe) were analyzed. It was noted in this analysis that repeatability and comparability were not uniformly defined across all those studies. Measures and methodologies used to evaluate the devices included correlation, regression analysis, standard deviation, and in some cases subjective visual inspec- tion of plots. Most studies also suggested that data averaging length affected repeatability. For example, ARA recommends that RWD results be averaged over 160 m (0.1 mi) but, in the United Kingdom, TSD test results are stored at 1-m (3.3-ft) averages and reported at 10-m (33-ft) averages. This section first discusses, evaluates, and highlights some of the draw- backs associated with the use of correlation and regression analysis to evaluate repeatability and comparability. Then, repeatability and comparability analysis based on the limits of agreement (LOA) method suggested by Bland and Altman (1986) is recommended and used to evaluate the continuous deflection devices. A method of evaluating repeatability from one run is also presented and compared to the method based on the LOA. Finally, the use of smoothing splines as a tool to Figure 3.22. Comparison of Highways Agency TSD and Danish Research Institute TSD on a major U.K. road (70-km/h [45-mph], 50-m [165-ft] averages). Figure 3.23. Deflectograph data for a major U.K. road at 50-m (165-ft) averages (nearside indicates outside wheelpath, offside indicates inside wheelpath).

39 remove the noise from TSD deflection slope measurements is investigated. This smoothing splines denoising methodol- ogy shows potential to improve the frequency at which useful information can be obtained (i.e., data averaging distance). In this report, repeatability (comparability) is defined as the 95% confidence interval of the difference between repeated measurements (difference between measurements of TSD and FWD or RWD and FWD). Correlation, cross-correlation, and regression are widely used to evaluate repeatability and com- parability in many pavement engineering applications such as profile or friction measurements. Regression Analysis For regression, the following example uses computer-generated data that simulate repeated measurements. Because the correct answer is known, it illustrates how regression analysis can lead to wrong conclusions. This argument about regression analysis follows closely the one presented by Bland and Altman (2003). The reason it is included in this report is because the use of cor- relation and regression is so pervasive in the pavement field that their shortcomings (as will be illustrated) are often ignored. The example supposes the true value of any measurement at 600 different locations (for example, pavement deflection) is known to be a sinusoidal wave varying between a minimum of 4 and a maximum of 6 units (Figure 3.27). Repeated measure- ments, m1 and m2, are obtained using an instrument that is known to produce measurements that are contaminated with Gaussian (from a normal distribution) noise with mean zero and standard deviation of 0.5 units (Figure 3.28). Since the relationship between m1 and m2 is known to be m1 = 1.0m2 + 0.0 (i.e., the line of equality), it is desirable that an appropriate sta- tistical analysis can suggest with some confidence this one-to- one relationship. Figure 3.29 shows m2 versus m1 with the true relationship (line of equality) and the regression line. The slope of the (a) Kansas (b) Minnesota (c) Iowa (d) Texas FWD Deflection (mils) R W D D ef le ct io n (m ils ) FWD Deflection (mils) R W D D ef le ct io n (m ils ) R W D D ef le ct io n (m ils ) FWD Deflection (mils)FWD Deflection (mils) R W D D ef le ct io n (m ils ) Source: FHWA, 2009. Figure 3.24. Examples of RWD versus FWD comparisons.

40 0.0 0.5 1.0 1.5 2.0 2.5 0 100 200 300 400 500 600 700 800 900 1000 0 50 100 150 200 250 300 350 400 TS D De fle cti on S lo pe [m m /m ] FW D De fle cti on [ μm ] Chainage [m] TRL Track Long Straight. FWD 2m spacings (17-Oct-2005), TSD P100 2m averaging (23-Jun-2009) FWD D1 TSD P100 Figure 3.25. Comparison of HA TSD slope and FWD central deflection profiles on flexible composite pavement. 0 100 200 300 400 500 600 700 800 0 0.2 0.4 0.6 0.8 1 1.2 P300 [mm/m] D FG [µ m ] Figure 3.26. Deflectograph versus P300 with 95% confidence intervals.

41 0 10 20 30 40 50 60 4 4.5 5 5.5 6 Location Tr ue V al ue Figure 3.27. Sinusoidal function example for regression analysis. 0 10 20 30 40 50 60 2 3 4 5 6 7 8 Location M ea su re m en t m1 m2 Figure 3.28. Sinusoidal signal with added Gaussian noise. 2 3 4 5 6 7 8 2 3 4 5 6 7 8 m1 m 2 Measurements Orthogonal Regression Line of Equality Regression m2 vs m1 Regression m1 vs m2 Figure 3.29. Limitation of linear regression when errors are present in the regressor. The relationship between m1 and m2 should be the line of equality.

42 regression line is 0.69, which is different from 1.0, and the intercept is 1.55, which is different from zero. The 95% (a = 0.05) confidence interval on the slope is [0.63; 0.75], while the 95% confidence interval on the intercept is [1.08; 2.03]. Both slope and intercept are statistically different from 1.0 and 0, respectively, suggesting that the relationship between the two measurements does not follow the line of equality (which we know it does). Therefore, ordinary regression analysis is leading to the wrong conclusion. The cause for the failure of regression analysis in this case is the presence of error in the regressors (errors in m1), which violates the con- ditions of the Gauss-Markov theorem. This violation leads to the least-squares regression coefficients to be biased (Myers, 1990). The irony of this is that the more measurements that are obtained, the narrower the confidence interval on the biased slope, which strengthens the wrong conclusion that the relationship between the two measurements does not fol- low the equality line. The alternative to linear regression when errors are pres- ent in both variables is the total least-squares regression (Van Huffel and Wandewalle, 1991). When errors in both vari- ables have the same variance, total least-squares regression is equivalent to orthogonal regression. The difference is that while ordinary least squares minimizes the squared distance from the dependent variable to the fitted function, orthogo- nal regression minimizes the square of the perpendicular dis- tance to the fitted function. The orthogonal regression for m1 and m2 is presented in Figure 3.29. The slope of the orthogo- nal regression line is 0.96, which is very close to 1.0 (the 95% confidence interval is [0.92; 1.00]). More information on this procedure can be found in Leng et al. (2007). Another way to look at this example is using the relation- ship m2 = 0.69m1 + 1.6, calculate m1 = 1.45m2 - 2.32. Since there is no specific reason to do the regression with m1 as the x-variable, it could be done with m2 as the x-variable. In this case, the relationship m1 = 0.66m2 + 1.7 is obtained, which is different from m1 = 1.45m2 - 2.32. The two regressions, using m1 or m2 as the x-variable, are presented in Figure 3.29. The relationship between m1 and m2 is not the same in each case. There is no reason to favor the use of m1 as the x-variable to the alternative of using m2 as the x-variable. This clearly illustrates the inadequacies of linear regression to evaluate the repeatability of a given device. Data Analysis Using Correlation The drawbacks of using correlation are similar to the draw- backs of regression analysis (although they are not com- pletely the same). Here, instead of using artificial data, the actual repeated TSD slope measurements obtained on differ- ent pavement sections are used to illustrate the drawbacks of correlation. Correlation measures have been extensively used to eval- uate repeatability or “accuracy”—with respect to FWD— of measures of continuous deflection data. This use of correlation is also prevalent in the analysis of pavement profile and friction data. However, the use of correlation can be very misleading, as discussed by Bland and Altman (1986, 2003). Correlation does not give agreement between repeated measures. For example, two measures that vary exactly by any factor give a correlation of 1 (or -1, if the factor is negative). A measuring device that gives repeated measurements that can vary by some factor is not one that is described as repeatable. Another drawback of correlation is that it depends on the range of the true mea- surement; the wider the range, the greater the correlation. In the extreme case, a pavement that is perfectly homo- geneous (i.e., strength is constant) will practically result in a zero correlation no matter how repeatable the device is. This is because the calculated correlation in this case is that of the error terms, which are randomly uncorre- lated. Correlation should therefore be used with caution when evaluating repeatability. This is not to say that cor- relation should never be used. For example, the proposed method of taking differences is not applicable when com- paring devices that measure two different physical quanti- ties (such as TSD and FWD). In this case, unless the two measurements can be converted to the same quantity, cor- relation (or for that matter, linear regression) might be a better choice. The average correlations between the different repeated TSD measurements obtained in this study for each section are presented in Figure 3.30. The correlations are not the same for the different sections. Interpreting the correla- tion as a measure of repeatability would give significantly different repeatability results depending on the tested sec- tion. As seen in Figure 3.30, for an averaging distance of 1 m, the correlation varies from under 0.10 to almost 0.90. Which correlation value in this range gives the repeatability of the device? The tested sections had a significant effect on the correlation. As expected, sections with low correla- tions are those that had low variation in the measured slope, and sections with high correlations are those that had high variation in the measured slope. For example, section F1 resulted in a significantly higher correlation than did all the other sections; especially for sensor 100 and 1-m averaging length. It can be concluded that correlation is a good indi- cator of the variability in the pavement section rather than in the repeatability of the device. This sentiment was some- what echoed, in more technical terms, by Bland and Alt- man (2003): “[T]he correlation coefficient is a measure of the information content of the measurement.” This clearly shows how correlation can lead to false conclusions when evaluating a device.

43 Figure 3.30. Correlation coefficient versus averaging length for three TSD sensors.

44 Figure 3.31 shows two different flexible pavement sections, one with high variation in the deflection slope (UK_F1) and another with low and uniform deflection slope (UK_F5). The correlation between repeated runs is significantly different for each section; UK_F1 had a high correlation (close to 0.9), while UK_F5 had a much lower correlation (less than 0.5). However, the measurements’ noise levels are comparable, as can be observed from visual inspection of the plots. Another observation is that the correlation varies with the distance between the sensor and the applied load. This results from the fact that sensors closer to the loaded area measure higher slopes, which increases the correlation. As a sum- mary, correlation depends on the tested pavement (or range of measurements), the instrument location (again, partly caused by different range of measurements), and averaging length. In many cases, these factors have a much more signifi- cant effect on the correlation than does the effect of errors in the measurements. However, a device repeatability measure should be, as much as possible, independent of the tested pavement. Not having this independence can lead to significantly different opinions about the suitability of the device. For example, somebody evaluating the device on the F1 section would be very pleased with the performance based on the correlation and somebody evaluating the device on the C2 section would be very disappointed in the device. The repeatability measure adopted in this study, in contrast, gives comparable results and is therefore much less affected by the tested pavement section. Repeatability The definition of repeatability given by the British Standard Institution (1979) was adopted in this report. It is defined as “the value below which the difference between two single test results . . . may be expected to lie with a specified probability.” The specified probability was set at 95%, and repeatability was calculated using the procedure suggested in a series of papers by Bland and Altman (Altman and Bland, 1983; Bland and Altman, 1986, 2003, 2007). The main idea is to estimate the standard deviation of the difference between repeated mea- surements from the same device (repeatability) or difference between measurements from two different devices (compa- rability) and construct the 95% confidence interval using 1.96sd, where sd is the standard deviation of the difference. In the case of comparability, this 95% confidence interval is referred to as the limits of agreement (LOA) between the two devices. In their procedure, Bland and Altman also specified calculating the bias between two different devices, while for the same device they incorporated this bias in the repeatabil- ity measure. This report shows the results of incorporating or not incorporating the bias in the repeatability measure. For all practical purposes, the two methods resulted in the same repeatability because the bias was negligible. For repeatability 0 500 1000 1500 2000 2500 3000 0 2 4 6 Distance (m) Sl op e (m m /m ) 0 500 1000 1500 2000 0 2 4 6 Distance (m) Sl op e (m m /m ) 0 500 1000 1500 2000 2500 3000 0 2 4 6 Distance (m) Sl op e (m m /m ) 0 500 1000 1500 2000 0 2 4 6 Distance (m) Sl op e (m m /m ) (c) (d)(b) (a) Figure 3.31. Measured deflection slope on two different flexible pavement sections: (a) first run on Section UK_F5, (b) second run on Section UK_F5, (c) first run on Section UK_F1, and (d) second run on Section UK_F1.

45 measures between the FWD and TSD or the FWD and RWD, the bias was not incorporated. TSD Repeatability The analysis of TSD repeatability was performed for mea- surements averaged over 1-m, 10-m, and 100-m distances. Five runs were obtained for each test section (except for Sec- tion F1, which had four runs, and for Sections F3 and R2, each which had three runs) resulting in five sets of slope mea- sures at each location. For repeatability (or comparability) analysis, it is impor- tant to check whether measurement repeatability depends on the actual measurement level (in other words, the measure- ment error depends on the actual measurement). In the case of two repeated measures, Altman and Bland (1983) suggested plotting x1 - x2 against (x1 + x2)/2. For the case where three or more repeated measures were obtained, the plot shows the standard deviation of the measurements at each loca- tion against the average of the measurements at each loca- tion (both plots give the same qualitative view). Figure 3.32 shows the standard deviation as a function of average slope for a relatively strong flexible pavement test section labeled F3 for slope measurements averaged over a 1-m interval. The figure suggests the standard deviation (and therefore equip- ment repeatability) is independent of the measured slope in the range of measurements. The observation was consistent for all other tested sections except for the flexible section labeled F1 shown in Figure 3.33, and the rigid section labeled R2. Both sections included R² = 2E-05 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 St an da rd D ev ia ti on Mean Reading Scatter of Standard Deviation; Site F3 - 1m Averaging - Sensor 100 Figure 3.32. Scatter of standard deviation versus mean for low average readings. R² = 0.3572 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 -0.2 0.8 1.8 2.8 3.8 4.8 5.8 6.8 7.8 St an da rd D ev ia ti on Mean Scatter of Standard Deviation; Site F1 - 1m Averaging - Sensor 100 Figure 3.33. Scatter of standard deviation versus mean for high average readings.

46 relatively very weak spots, with significant deterioration. In Figure 3.33, the standard deviation is dependent on the asso- ciated measured slope. A possible explanation for this depen- dence could be that there are two main sources of variation in the slope measurement. The first source is due to the error from the sensors, vehicle dynamics, and any other factors that can affect the TSD (temperature, moisture, etc.). This source of variability is expected in most cases to be independent of the associated measurement. The second source of variabil- ity is due to the spatial variability in the pavement strength. Weaker pavements tend to have more distress factors, such as cracking, which result in a greater spatial variability on the strength and, therefore, the deflection measurements. It is important to note that this source of variability is not caused by the device. Repeatability is closely related to the estimation of error in measurements obtained from a device. This error is com- posed of variance and bias. Therefore, repeatability can be defined by either the variance or the error (variance and bias). The advantage of defining repeatability in terms of variance and keeping bias as a separate measure is that causes of bias can often be identified and corrected. For example, in measurements on flexible pavement, bias can be the result of a temperature difference between repeated tests. More com- monly, bias is caused by different operational characteris- tics or the equipment getting out of calibration. Table 3.15 Table 3.15. Consistency of TSD Measurements over Repeated Runs: Incorporating Bias in Standard Deviation Road Section Data Averaging (mm/m) Sensor 100 Sensor 300 Sensor 756 Average Reading (mm/m) Std. Dev. (mm/m) cov Average Reading (mm/m) Std. Dev. (mm/m) cov Average Reading (mm/m) Std. Dev. (mm/m) cov F1 1 1.0468 0.2661 0.2542 0.6198 0.1918 0.3095 0.3169 0.144 0.4544 10 0.1054 0.1007 0.0717 0.1157 0.0383 0.1209 100 0.0625 0.0597 0.038 0.0613 0.0146 0.0461 F3 1 0.3226 0.1305 0.4045 0.2202 0.1376 0.6249 0.2279 0.1281 0.5621 10 0.0407 0.1262 0.0415 0.1885 0.0412 0.1808 100 0.0155 0.048 0.0139 0.0631 0.0137 0.0601 F5 1 0.4356 0.1166 0.2677 0.3691 0.1123 0.3043 0.3767 0.1096 0.2909 10 0.0401 0.0921 0.0383 0.1038 0.0367 0.0974 100 0.0147 0.0337 0.0128 0.0347 0.0111 0.0295 F6 1 0.4176 0.1254 0.3003 0.3252 0.121 0.3721 0.4316 0.1169 0.2709 10 0.0406 0.0972 0.0382 0.1175 0.0373 0.0864 100 0.0139 0.0333 0.0124 0.0381 0.0119 0.0276 C1 1 0.4786 0.1466 0.3063 0.3779 0.1398 0.3699 0.4386 0.1355 0.3089 10 0.0519 0.1084 0.0489 0.1294 0.0498 0.1135 100 0.0306 0.0639 0.0292 0.0773 0.0276 0.0629 C2 1 0.3469 0.1222 0.3523 0.2811 0.1226 0.4361 0.2949 0.1146 0.3886 10 0.0408 0.1176 0.0398 0.1416 0.039 0.1322 100 0.0142 0.0409 0.0146 0.0519 0.0143 0.0485 R2 1 0.5859 0.2072 0.3537 0.3972 0.1671 0.4207 0.3511 0.1217 0.3466 10 0.0643 0.1098 0.0558 0.1405 0.0512 0.1457 100 na na na na na na R3 1 0.343 0.1483 0.4322 0.2356 0.1378 0.5846 0.3441 0.1419 0.4122 10 0.0432 0.1258 0.0404 0.1716 0.0433 0.1257 100 0.0153 0.0446 0.0136 0.0579 0.0148 0.0431 Note: cov = coefficient of variation. F = flexible; C = composite; R = rigid (road types); and na = not applicable.

47 presents results with the bias incorporated in the standard deviation calculation. Table 3.16 presents results with the bias taken out of the standard deviation. A comparison of the results presented in both tables sug- gests that the effect of bias in the obtained measurements is negligible. This is confirmed by the bias results presented in Table 3.16, which shows the bias to be small compared to the standard deviation. However, the effect of the bias becomes more significant for measurements averaged over longer distances (for example, when averaging over 100 m [330 ft] compared to 1 m [3.3 ft]). This is because averaging over longer distances reduces the variance while keeping the bias constant. Therefore, the relative effect of the bias becomes more significant. This effect was observed in the section labeled C1. In this case, incorporating the bias into the standard deviation of measurements averaged over a 100-m (330-ft) distance resulted in a standard deviation that is twice as large as the standard deviation calculated by not incorporating the bias (compare Table 3.15 and Table 3.16 for C1). Figure 3.34 shows that one of the runs has a significant systematic bias and is shifted up compared to the others. Note that the effect of the bias is much less pronounced for averaging distances of 1 m (3.3 ft) and 10 m (33 ft) for which the standard deviation Table 3.16. Repeatability of TSD Measurements over Repeated Runs: Excluding Bias from Standard Deviation Road Section Averaging Length (m) Sensor 100 Sensor 300 Sensor 756 Std. Dev. (mm/m) Bias (mm/m) Rep.a (mm/m) Std. Dev. (mm/m) Bias (mm/m) Rep.a (mm/m) Std. Dev. (mm/m) Bias (mm/m) Rep.a (mm/m) F1 1 0.2649 0.0328 0.5192 0.1771 0.0169 0.3471 0.1333 0.0054 0.2613 10 0.1024 0.2007 0.0653 0.128 0.0353 0.0629 100 0.0583 0.1143 0.0335 0.0657 0.0257 F3 1 0.1304 0.0068 0.2556 0.1376 0.0051 0.2697 0.1280 0.0055 0.2509 10 0.0404 0.0792 0.0414 0.0811 0.0410 0.0804 100 0.0149 0.0292 0.0137 0.0269 0.0132 0.0259 F5 1 0.1165 0.0067 0.2283 0.1122 0.0056 0.2199 0.1096 0.0026 0.2148 10 0.0397 0.0778 0.0381 0.0647 0.0367 0.0719 100 0.0136 0.0267 0.0122 0.0239 0.0111 0.0218 F6 1 0.1254 0.0044 0.2283 0.1209 0.0035 0.2370 0.1169 0.0035 0.2291 10 0.0405 0.0794 0.0381 0.0747 0.0372 0.0729 100 0.0136 0.0267 0.0122 0.0239 0.0117 0.0229 R2 1 0.2069 0.0161 0.405 0.1666 0.0186 0.3265 0.1217 0.0033 0.2385 10 0.0634 0.1243 0.0541 0.1060 0.0513 0.1005 100 na na na na na na na na na R3 1 0.1482 0.0080 0.2905 0.1377 0.0062 0.2699 0.1420 0.0064 0.2780 10 0.0427 0.0837 0.0403 0.0790 0.0431 0.0845 100 0.0142 0.0278 0.01333 0.0261 0.0145 0.0284 C1 1 0.1430 0.0334 0.2803 0.1375 0.0262 0.2695 0.1332 0.0267 0.2611 10 0.0446 0.0874 0.0421 0.0825 0.0432 0.0847 100 0.0153 0.0300 0.0155 0.0304 0.0123 0.0241 C2 1 0.1221 0.0070 0.2393 0.1224 0.0103 0.2399 0.1145 0.0061 0.2244 10 0.0404 0.0792 0.0389 0.0762 0.0386 0.0757 100 0.0129 0.0253 0.0121 0.0237 0.0132 0.0259 Note: F = flexible; C = composite; R = rigid (road types); and na = not applicable. a Rep. = repeatability or 1.96 * std. dev.

48 of C1 is similar to that of the other sections. In this analysis, of all tested sections, C1 was the only one that had a statisti- cally significant bias. The results presented in Table 3.15 and Table 3.16 were obtained using the method of analysis of variance (ANOVA). The procedure can be illustrated using two repeated mea- surements. For the TSD, each slope measurement consists of the actual slope and an error term. This can be expressed as follows: y s eij ij ij= + ( . )3 2 where yij = TSD slope measurement at location i for run j, sij = actual (unknown) slope at location i during run j, and eij = error in TSD slope measurement at location i for run j. The first two runs for F5 are shown in Figure 3.35a. Figure 3.35b shows the difference between the two runs. The mean of this difference is an estimate of the bias between the two runs, while the variance of the difference represents the sum of the error variances for each run (Bland and Altman, 1986). Assuming the variance for each run is the same, the mea- surement variance can be estimated by dividing the variance of the difference by two. For example, taking the difference between runs 1 and 2, the bias and variances can be calculated as follows: Bias y y N i i i N = − = ∑ 1 2 1 3 3( . ) σd i i i N Bias y y N 2 1 2 2 1 1 3 4= − −( )[ ] −= ∑ ( . . )a σ2 1 2 2 1 1 2 1 3 4= − −( )[ ] −= ∑ Bias y y N i i i N ( . . )b where N = total number of data points per run, sd2 = variance of the differences, and s2 = variance of the measurements. If both bias and variance are used to estimate the repeat- ability (see Bland and Altman, 1986), the mean squared dif- ference (MSD) can be calculated as follows: MSD y y N i i i N = −( )[ ] = ∑1 2 3 5 1 2 2 1 ( . ) Slope Vs. Chainage, UK_C1 (REA-S), P100 100m Averages Figure 3.34. Measured slope averaged over 100-m length for Site C1.

49 The factor ½ in the calculation of the variance (or MSD) reflects that the measurement variance is half the variance of the difference. The calculated variances (in the case of Equa- tion 3.3), using the difference between the first run and each of the remaining 4 runs for F5 for the 1-m (3.3-ft) averag- ing distance, were 0.0141, 0.0132, 0.0130, and 0.0135 mm/m. Levene’s test of equal variance with a = 0.05 showed that the variances are equal. Because the calculated variances using the difference between different pairs of runs are equal, the TSD error variance can be estimated as the average of the calculated variances. The TSD error standard deviation can then be calcu- lated as the square root of the variance. The resulting standard deviation calculated for F5 and 1 m (3.3 ft) TSD measurement averaging was 0.1165 mm/m. This procedure is essentially implemented in MATLAB in the function called “anova2.” Calculation of the repeatability coefficient was performed using the 1.96sd estimate of confidence interval. This esti- mate was used after the error distribution was found to fol- low a normal distribution (using the Anderson–Darling test for normality). The repeatability calculated for all sections is presented in Table 3.16. RWD Repeatability Analysis A similar analysis (which separates the bias and variance) was conducted for the RWD using the data obtained in Phase I for Virginia. Although the device has been improved, this analysis permits the establishment of a baseline repeatability value. The analysis of repeatability was performed using three RWD repeated measurements averaged over 160 m (0.1 mi) col- lected over a distance of 32 km (20 mi) on I-64 for both east- bound and westbound directions. Figure 3.36 shows the test results for the eastbound direction. The difference between the first and second runs is shown in Figure 3.37. A test of normality showed that the differences between the two measurements are normally distributed. The cal- culated standard deviations of the difference (sd) between Runs 1 and 2, Runs 1 and 3, and Runs 2 and 3 are 24.9, 27.3, and 27.2 µm (0.98, 1.07, and 1.07 mils), respectively, for the eastbound direction, and 26.1, 24.3, and 24.8 µm (1.03, 0.96, and 0.98 mils), respectively, for the westbound direction. Dif- ferences between the standard deviations were found not to be statistically significant at the 0.05 significance level using Levene’s test of equal variances. Therefore, the error defined as the standard deviation of RWD measurements (s) was cal- culated as 18.2 µm (0.72 mils). This standard deviation was computed by dividing sd by the square root of 2. The bias between the runs is defined as the mean of the difference. For the eastbound direction, the biases between Runs 1 and 2, Runs 1 and 3, and Runs 2 and 3 are 3.93, 18.4, and 14.5 µm (0.15, 0.72, and 0.57 mils), respectively. For the westbound direction, the biases between Runs 1 and 2, Runs 1 and 3, and Runs 2 and 3 are 14.2, 6.5, and -7.7 µm (0.56, 0.26, and -0.30 mils), respectively. All these biases were found to 0 500 1000 1500 2000 2500 3000 3500 -0.2 0 0.2 0.4 0.6 0.8 1 1.2 Sl op e (m m/ m) 0 500 1000 1500 2000 2500 3000 3500 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 Distance (m) Sl op e D iff er en ce (m m/ m) Run1 Run2 (a) (b) Figure 3.35. Comparison of P100 measurements of Runs 1 and 2 for Site F5 (a), and difference between the two runs (b).

50 0 5 10 15 20 25 30 35 60 40 80 100 120 140 160 180 200 220 Distance (km) D ef le ct io n (µ m ) First Run Second Run Third Run Figure 3.36. RWD deflections on eastbound I-64 in Virginia. 0 5 10 15 20 25 30 35 -60 -40 -20 0 20 40 60 80 100 Distance (km) D ef le ct io n D iff er en ce (µ m) Figure 3.37. Difference of measurements versus average measurement. be statistically significant at the 0.05 significance level. Since the distances covered are relatively long, the biases could be due to differences in the testing conditions (e.g., pavement temperature). An average bias of 10.9 µm (0.43 mils) was cal- culated as the average of the absolute values of the biases. Finally, the repeatability of RWD was calculated as 1.96sd, which is equal to 50.4 µm (1.98 mils). The results of RWD repeatability analysis are presented in Table 3.17. Summary of Repeatability Evaluation For the TSD, except for the two sites UK_F1 and UK_R2, the evaluated LOA (repeatability) was the same for all tested sites and for the three sensors (see Table 3.16). As a function of averaging length, the LOA roughly decreased (lower LOA means higher repeatability) by a factor of L L2 1 going from the averaging length L1 to the averaging length L2. For example, going from L1 = 1 m (3.3 ft) to L2 = 100 m (33 ft), the repeatability roughly decreases by a factor of 100 1 10= . This occurs when measurement errors are uncorrelated and therefore suggests that measurement errors are uncorrelated. Furthermore, it was found that there is no significant bias (systematic error) between repeated runs (except for one run on Section C1). For Sites UK_F1 and UK_R2, the repeatability was found to depend on the actual deflection slope measure- ment. Higher measurements resulted in larger error standard deviation and therefore larger LOA (lower repeatability). A possible reason for this is that weaker sections, which result in

51 higher measurements, are much less homogeneous because of the presence of distresses (either on the surface or in the hidden layer). This results in larger LOA. For the RWD, the analysis was limited to three runs per- formed on a single section. The effect of averaging length could not be evaluated as measurements were already averaged over 160-m (0.1-mi) sections. Furthermore, there was a significant bias between the three repeated runs. The repeatability of RWD for the section evaluated was 50.4 µm (1.98 mils). Comparability Comparability is defined similarly to repeatability and is the level of agreement between the two devices (TSD and FWD or RWD and FWD). The difference is that comparability compares measurements from two different devices, whereas repeatabil- ity compares measurements from the same device. The FWD is used as a reference point because it has become the de facto standard for structural evaluation of pavement. Many engineer- ing parameters and properties are obtained from FWD testing. Most state DOTs have acquired enough experience to be able to effectively interpret and use FWD test results. Therefore, the FWD can be used as the reference for evaluating any deflection measuring device (or structural capacity measuring device). Because the TSD does not directly measure pavement deflec- tions, the analysis first investigates the correlation between TSD slope and FWD deflection. To evaluate comparability, measurements need to be converted to the same physical quantity. The physical quantities that can be obtained from both devices are the surface curvature index (SCI) and the base damage index (BDI). The first step in the comparison is to temperature correct FWD deflections to the TSD test temperature. The reason for correcting FWD deflections and not the TSD or both devices is because the temperature- correction procedure for the TSD is still under development. Temperature Correcting the FWD for TSD Sites FWD deflection values were first corrected to the temperature at the mid-depth of the asphalt layer during TSD testing. The estimated temperatures near the mid-depth of the pavement (using the U.K. methodology) during FWD testing and TSD testing were included in the files. The method adopted was to correct the center deflection of the FWD first using the proce- dure described by the FHWA (1998). The correction assumes that deflections 900 mm (36 in.) away from the center of the load plate are not significantly affected by temperature. Thus, using the deflections at this point, along with a measured temperature, a value for asphalt stiffness, and the pavement thickness, the center deflection of the FWD can be corrected to a reference temperature. The latitude of the test site was used to account for climatic differences. The corrected center deflection can be found with the fol- lowing equations: D D TAF0 0 3 6Corrected Measured=  ( . ) where TAF = temperature adjustment factor calculated according to Equation 3.7. TAF D D T T = +( ) +( ) 36 36 3 36 36 Delta Delta ref meas @ @ ( . )7 where D36 = deflection at 900 mm (36 in.) from the center of the load plate in µm, Delta36@Tref = basin factor calculated at the reference temperature, and Delta36@Tmeas = basin shape factor calculated at the measured temperature. The basin shape factor is found by the following equation: log . . log . logDelta Theta36 3 05 1 13 0 502( ) = − ( )+ac ( ) ( ) − ( ) ( ) ( ) + log . log log . D T D 36 0 00487 36 0 Theta 00677 3 8T ac( ) ( ) ( )log log ( . )Theta where ac = total thickness of the HMA in mm, Theta = latitude of the pavement section, and T = temperature at middepth of the HMA in degrees Celsius. Table 3.17. Repeatability Analysis of the RWD Westbound Eastbound 1 and 2 1 and 3 2 and 3 1 and 2 1 and 3 2 and 3 Standard Deviation 24.9 27.3 27.2 26.1 24.3 24.8 Bias 14.2 6.5 –7.7 3.93 18.4 14.5 Repeatability 51.8 49.1 50.4 Note: Units in table in µm.

52 After the center deflections were corrected, the deflections between 0 mm and 900 mm (0 in. and 36 in.) were also adjusted for temperature. The adjustment factor was assumed to vary linearly along the distance from the applied from a maximum calculated using Equation 3.5 at 0 to 0 at 900 mm away from the applied load. Figure 3.38 illustrates how the correction factor can be obtained. The calculation can be performed according to the following equation: D x D x D D D D ( ) = ( )−( ) −( )Corrected Corrected36 0 36 0 36 3 936−( ) +D D ( . ) where D(x)Corrected = temperature-corrected deflection at location x, D(x) = measured deflection at location x, D0Corrected = temperature-corrected center deflection, and D0 = measured center deflection. Comparison of TSD Slope and FWD Deflection The TSD and FWD measure two different quantities; therefore, measurements obtained from the devices cannot be directly compared. To make a comparison, a relationship between measurements from the two devices needs to be obtained. The simplest relationship is the linear one. In this case, correla- tion is used to evaluate the strength of the linear relationship. Table 3.18 shows the correlation between TSD measurements averaged over 10 m and selected FWD measurements (D0 and D300) at 10-m (33-ft) intervals, obtained at six different sites (UK_R2, UK_R3, UK_F1, UK_F5, UK_C3, and UK_F3). The correlation ranges from very good (~0.95) for Site UK_R2 to relatively poor (~0.27) for Site UK_F3. One main drawback of correlation is that it is significantly affected by the range of measurements: the wider the range of measurements, the better the correlation. This is illustrated for the six sites in Figure 3.39 and Figure 3.40. Measurements collected on Site UK_F3 were gathered over a much smaller range than those collected on the other sites, which explains why the correlation for Site UK_F3 is lower than that of the other sites. While correlation values suggest a relatively good linear relationship between FWD deflections and TSD slope, the figures show that this relationship is pavement-type specific, with flexible and composite pavements (F1, F3, F5, and C3) exhibiting the same relationship, and rigid pavements (R2 and R3) exhibiting a different relationship (Figure 3.39 and Figure 3.40); that is, no single linear (or any function) model 0 2 4 6 8 10 12 14 0 200 400 600 800 1000 D efl ec ti on S lo pe (m m /m ) Distance from Center of Load Plate (mm) A B C D Figure 3.38. Temperature-correction technique for deflections between D0 and D36. Table 3.18. Correlation Between TSD Slope and FWD Deflection UK_R2 UK_R3 UK_F1 UK_F5 UK_C3 UK_F3 TSD100 and D0 0.9492 0.2798 0.8289 0.7252 0.8007 0.3716 TSD300 and D0 0.9514 0.3721 0.8260 0.7164 0.7942 0.2189 TSD100 and D1 0.9420 0.2839 0.7984 0.6941 0.7941 0.3338 TSD300 and D1 0.9448 0.4002 0.8223 0.6865 0.7948 0.1906

53 can adequately represent the relationship between FWD deflections and TSD slope for the two pavement categories. Comparability Between TSD and FWD Test Results To evaluate the comparability between the TSD and the FWD, the two measured quantities need to be converted to the same physical quantity. The surface curvature index (SCI) and the base damage index (BDI) were chosen because they can be calculated from both the TSD and FWD measurements. The calculated SCI is SCI300, which is defined as D0 - D300, and the BDI is defined as D300 - D600. For the FWD, D0, D300, and D600 are directly measured and calculation of the SCI is straightforward. In the following section, we present the methodology used to calculate both the SCI and the BDI from TSD slope measurements. Converting TSD Slope to SCI or BDI The TSD measures the slope of the deflection bowl that results from the truck traveling over the pavement. Therefore, Figure 3.39. TSD P100 slope versus FWD D0 deflection. 0 500 1000 1500 2000 2500 0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000 FWD D0 (µm) T SD P 10 0 (µ m /m ) 0 100 200 300 400 500 0 200 400 600 800 1000 UK-R2 UK-R3 UK-F1 UK-F5 UK-C3 UK-F3 Linear Fit R Sites Linear Fit F and C Sites Figure 3.40. TSD P300 slope versus FWD D300 deflection. 0 200 400 600 800 1000 1200 1400 1600-500 0 500 1000 1500 2000 2500 3000 FWD D300 (μm) TS D P3 00 ( m/ m) UK-R2 UK-R3 UK-F1 UK-F5 UK-C3 UK-F3 Linear Fit R sites Linear Fit F and C Sites 0 50 100 150 200 250 300 350 400-200 0 200 400 600 800 0

54 deflection can, in principle, be obtained from the TSD slope by integration if a sufficiently detailed representation of the full deflection slope bowl is available. Integration is specified only up to a constant value; therefore, the deflection can- not be recovered without a reference deflection measure- ment. The difference between two deflection readings can be obtained; the constant cancels out, which gives the SCI or BDI. The relationship between slope, deflections, and SCI is presented in Equation 3.10. s x Dx D b D a a b ( ) = ( )− ( ) =∫ SCI ( . )3 10 where s(x) = slope at location x, and D(x) = deflection at location x. Recent results by Thyagarajan et al. (2011) and Krarup et al. (2006) suggest that SCI values, especially SCI300, correlate very well with tensile strain at the bottom of the asphalt layer. Krarup et al. (2006) used a functional representation of the deflection and slope presented in Equation 3.11 and Equa- tion 3.12 to perform the integration: D x A B Bx Bx Bx( ) = − ( )+ ( )[ ] −( ) 2 3 11cos sin exp ( . ) s x D x A Bx Bx( ) = ′( ) = ( ) −( )sin exp ( . )3 12 where A and B = experimentally determined constants. Equations 3.11 and 3.12, which were developed for a sen- sor configuration of 100, 200, and 300 mm (3.94, 7.87, and 11.8 inches) from the applied wheel load, were investigated to fit measured TSD slope. The results showed that the equations are not suitable to use in the sensor configura- tion tested (sensors at 100, 300, and 756 mm [3.94, 11.81, and 29.8 in.] from the load). Therefore, numerical inte- gration was adopted (trapezoidal rule) to calculate the SCI and BDI. To calculate the SCI and BDI from TSD measurements, the TSD slope was integrated numerically. Slope measure- ments were obtained at 100, 300, and 756 mm (3.94, 11.81, and 29.8 in.) from the applied load. The integration was per- formed using the trapezoidal rule. To calculate SCI300, the integration interval is [0 mm; 300 mm] ([0 in.; 11.81 in.]). In this interval, TSD slope measurements were obtained at 100 and 300 mm (3.94 and 11.81 in.) from the applied wheel load and an assumption was needed on how the slope varies between 0 and 100 m (0 and 3.94 in.). Two assumptions illus- trated in Figure 3.41 were investigated. The first assumption sets the slope to vary linearly throughout the interval from 0 to 300 m (0 and 11.81 in.). This assumption is represented by the red line between 0 and 100 mm (0 and 3.94 in.) and the blue line between 100 and 300 mm (3.94 and 11.81 in.). The integration in this case results in the area comprising the red and blue areas. The second, more realistic assumption is to set the slope at 0 mm from the load equal to zero. This assumption is valid if the load is uniformly (or approximately uniformly) applied over a specific area rather than being a point load and if viscoelastic effects are neglected. The results of both calculations are presented in Figure 3.42, which sug- gests the second assumption (slope at 0 equals 0) works bet- ter because the calculated SCI values (from FWD and TSD) 0 50 100 150 200 250 3000 100 200 300 400 500 600 Sensor Location (mm) Sl op e ( m /m ) Note: The SCI300 is the shaded area under the piecewise linear curve. Figure 3.41. Effect of assuming the slope at 0 mm to be equal to zero on the calculated SCI300.

55 better follow the line of equality. The second assumption was thus adopted in the analysis. Figure 3.42 also shows the results of the ordinary and orthogonal regressions between the SCI calculations based on the FWD and TSD measurements. Since both variables contain errors, orthogonal (or total least-squares) regression is more appropriate, as shown in the figure. The orthogonal regression line has a slope of 0.90, which is closer to the line of equality than the one obtained using the least-squares regression with FWD measurements as the independent variable, which had a slope of 0.79. SCI Comparisons Figure 3.43 shows the calculated SCI300 with the assumption of slope at 0 equal to 0 while Figure 3.44 shows the calcula- tion of BDI (also known as SCI450, which is defined here as D300 - D600). The main advantage of comparing the indices -200 0 200 400 600 800 1000-200 0 200 400 600 800 1000 1200 1400 FWD SCI300 ( m) TS D S CI 30 0 ( m ) Slope (0) 0 Slope (0) = 0 Linear Regression Orthogonal Regression Equality Line Figure 3.42. Comparison of SCI300 calculated by using various assumptions on the slope. -200 0 200 400 600 800 10000 100 200 300 400 500 600 700 800 900 1000 FWD SCI300 ( m) TS D S CI 30 0 ( m ) UK-R2 UK-R3 UK-F1 Equality Line UK-F5 UK-C3 UK-F3 -20 0 20 40 60 80 1000 50 100 150 200 Figure 3.43. Comparison of SCI300 obtained from the TSD and FWD testing.

56 is that both quantities have the same dimensions. The first important observation is that there is no obvious indication that the relationship is not the same for all tested pavement types (Figure 3.43 and Figure 3.44). The plots also show that the results of the calculated SCI300 and BDI are comparable for both the FWD and TSD; this is reflected by the close prox- imity of the observed measurements to the equality line. The limits of agreement (LOA) method was then used to assess the comparability between the two devices, as is presented in the next section. Comparability and LOA of SCI and BDI Measurements The plots presented in Figure 3.43 and Figure 3.44 suggest a relatively good agreement between both devices; however, to compare measurements obtained from the two devices, one would need to evaluate how well they agree with each other in a quantifiable way. This is usually a judgment call, but there are some parameters that provide guidance for this problem. Generally, a discrepancy in measurement between the two devices will be accepted if it is less than a certain limit, usually referred to as the confidence limit. Confidence limits should be narrow enough to expect that the result will not affect a decision that is based on it. To evaluate this dependence, the standard deviation of the difference as a function of the SCI measurement was calcu- lated. This was done with the method presented by Davidian and Carrol (1987). The procedure consists of fitting a regression line with the y variable taken as the difference in the SCI and the x variable taken as the average SCI. This regression line is an estimate of the bias (as a function of average SCI) between the two methods. The variance can be estimated by using the squared residuals obtained from the regression analysis. This is done by (again) performing regression on the squared residuals. The result of this regression is presented in Figure 3.45. These results correspond to the square root of the squared residuals, and as such, the fitted lines represent the standard deviation variation as a function of average SCI. The reason for not plotting the squared residual is because taking the square root results in a clearer visual repre- sentation. The LOA presented in Figure 3.46 can be calculated by using the derived standard deviation and constructing the 95% confidence interval around the bias. As recommended by Bland and Altman (1983), the plot of the difference of the SCI (FWD - TSD) versus the average SCI ([FWD + TSD]/2) is the first step to evaluate the LOA. This plot is presented in Figure 3.46 for both SCI300 and BDI (referred to both as SCI). The figure suggests that in both cases, the difference depends on the SCI measurements. Thus, data transformations were investigated to possibly remove this dependence. If transformation is not successful, the repeat- ability can be defined as a function of the size of measurement (Bland and Altman, 1983). A logarithmic transformation, as well as a coefficient of variation transformation (difference divided by average), on the data did not result in the removal of this association. The repeatability was therefore defined as a function of the size of measurements. 0 100 200 300 400 500 600 700 800 9000 100 200 300 400 500 600 700 800 900 FWD BDI ( m) TS D B D I ( m ) UK-R2 UK-R3 UK-F1 Equality Line UK-F5 UK-C3 UK-F3 0 20 40 60 80 100 0 50 100 150 200 Figure 3.44. Comparison of BDI obtained from the TSD and FWD testing.

57 Figure 3.45 and Figure 3.46 present two sets of results each. The first trend line (red dashed line, STDEV) and confidence interval used all the measurements to estimate the bias and standard deviation. However, only seven measurements were obtained for SCI values above 500 µm (19.7 mils), and the results of the analysis performed by using all measurements are highly influenced by those seven measurements. The same analysis was performed with only the set of points that resulted in an average SCI below 500 µm (19.7 mils). As can be seen in Figure 3.45 and Figure 3.46, the results of both analyses are essentially the same for average SCI values below 250 µm (9.84 mils), and differences increase with increasing average SCI values. To interpret the results presented in Figure 3.46, the example average SCI of 300 µm (11.8 mils) can be exam- ined. In this case, the SCI calculated from the FWD is expected to be (on average) 30 µm (1.18 mils) lower than the SCI calculated from the TSD, with values ranging from as much as 205 µm (8.07 mils) lower to 175 µm (6.89 mils) higher (for a range of 380 µm [15.0 mils]) expected to occur 95% of the time. For an average SCI of 100 µm, the values computed using the FWD measurements are expected to be (on average) 50 µm (1.97 mils) lower than the SCI calculated from the TSD, with values rang- ing from as much as 115 µm (4.53 mils) lower to 15 µm (0.59 mils) higher (for a range of 130 µm [5.12 mils]), 0 100 200 300 400 500 600 700 800 900 10000 50 100 150 200 250 300 350 400 450 Average SCI ( m) SC I R es id ua ls (μ m ) Residuals STDEV (SCI<500) STDEV Figure 3.45. Plot of absolute value on residual and the regression on the residual squares. 0 100 200 300 400 500 600 700 800 900 1000-400 -200 0 200 400 600 800 1000 Average SCI ( m) D iff er en ce S CI FW D - SC I T SD ( m ) Difference Bias 95% CI 95% CI Bias (SCI<500) 95% CI (SCI<500) 95% CI (SCI<500) Figure 3.46. Plot of difference of SCI versus average SCI with bias and limits of agreement.

58 expected to occur 95% of the time. To put these two numbers into perspective, the repeatability of TSD SCI300 measure- ments for Sites UK_F1 and UK_F5 was calculated as 51 µm (2.01 mils) (range of 102 µm [4.02 mils]) for UK_F1 and 25 µm (0.98 mils) (range of 50 µm [1.97 mils]) for UK_F5. Comparability Between RWD and FWD Comparison of RWD and FWD testing performed in New Mexico, which was provided by ARA for this project, is presented in Figure 3.47. In this case, RWD and FWD mea- surements were obtained at 160-m (0.1-mi) intervals. The plot of the difference, FWD – RWD, versus the average is pre- sented in Figure 3.48. The figure suggests the measurement error depends on the associated measurement. Both loga- rithmic and power transformations of the data were investi- gated but did not result in removal of the error dependence with the associated measurement. In the end, the normalized difference calculated as (x1 - x2)/(0.5x1 + 0.5x2) was found to be the error measurement parameter that was the least 0 20 40 60 80 100 120 140 160 180 2000 100 200 300 400 500 600 700 Distance (km) D ef le ct io n s ( m ) RWD FWD Figure 3.47. FWD and RWD measurements obtained from New Mexico. 0 100 200 300 400 500 600 700-500 -400 -300 -200 -100 0 100 200 300 400 Average RWD and FWD Measurements ( m) D iff er en ce FW D - R W D ( m ) Figure 3.48. Difference versus average measurement.

59 dependent on the associated measurement. Figure 3.49 shows the normalized difference as a function of the measurement. The dependence on the associated measurement is not com- pletely removed; however, it is appreciably lower than the association in Figure 3.48. The average normalized difference of the difference (FWD - RWD) is -0.2328 and the standard deviation is 0.3265. Figure 3.50 shows a distribution of the normalized difference. The Anderson–Darling test for normality showed that this distribution does not follow the normal distribution. The 95% bootstrap confidence interval of the normalized dif- ference was calculated as [-0.8642; 0.4286]. In comparison, the 95% confidence interval using normal assumptions is [-0.8538; 0.3882]. The two intervals are close; however, the wider calculated bootstrap interval reflects the fact that the distribution in Figure 3.50 has a heavy tail. Direct Comparison of RWD Deflections and FWD Deflections A comparison was made directly between the RWD data and the FWD data collected. Figure 3.51 presents a more detailed analysis of the data collected on U.S. Route 550 in New Mexico. The figure shows the entire length measured with the RWD (Figure 3.51a) and FWD (Figure 3.51b) divided into homogeneous sections. Each plot displays the average and characteristic deflections (upper 95% confidence limit) for 0 100 200 300 400 500 600 700-1.5 -1 -0.5 0 0.5 1 1.5 Average Measurement of FWD and RWD ( m) N or m al iz ed D iff er en ce Difference Linear Fit Figure 3.49. Normalized difference versus average measurement. -1.5 -1 -0.5 0 0.5 10 10 20 30 40 50 60 70 80 90 Normalized Difference Co u n t 1.5 Figure 3.50. Distribution of normalized difference.

60 (b) FWD measurements Deflection Data Uniform Sections 95% Upper Confidence Limit std = 2.86 std = 4.81 std = 3.15 std = 3.10 std = 1.71 std = 1.62 std = 2.46 0 20 40 60 80 100 120 Distance (miles) 0 5 10 15 20 25 30 D ef le ct io n (m ils ) 0 20 40 60 80 100 120 0 5 10 15 20 25 30 Distance (miles) D ef le ct io n (m ils ) Deflection Data Uniform Sections 95% Upper Confidence Limit std = 1.03 std = 2.25 std = 1.58 std = 5.11 std = 3.07 std = 2.81 std = 1.47 (a) RWD measurements 0 20 40 60 80 100 1200 5 10 15 20 25 30 Distance (miles) D ef le ct io n (m ils ) FWD Data RWD Data Uniform FWD Uniform RWD Correlation = 0.5993 (c) Comparison of 0.1-mi averaged measurements and section averages (d) Individual 0.1-mi segment comparisons FWD Deflections (mils) 0 10 20 30 0 10 20 30 R W D D ef le ct io ns (m ils ) Measurements Line of Equality Correlation = 0.5993 5 10 15 20 5 10 15 20 FWD Deflection (mils) R W D D ef le ct io n (m ils ) y = 1.1*x + 1.7 Uniform Section Measurements Linear Trend Line of Equality Correlation = 0.9132 (e) Homogeneous section comparisons Figure 3.51. Analysis of New Mexico data.

61 each section. The sections were segmented visually, and the average values are compared in Figure 3.51c. Figure 3.51d and e compare the individual 160-m (0.1-mi) averages and the homogeneous section averages, respectively. The correlation between the individual measurements is rela- tively weak (R = 0.60, R2 = 0.36), but it improves significantly when the data are aggregated in homogeneous sections. The RWD deflections are consistently higher than the FWD ones, but the correlation is good (R = 0.91, R2 = 0.83). Summary of Comparability To evaluate the comparability between the TSD and FWD, the TSD slope measurements and the FWD deflection measurements were converted to the SCI and BDI. This was performed because the FWD and TSD measure different physical quantities—deflection and deflection slope, respec- tively. Both these quantities can be used to calculate the SCI and BDI, which makes comparability evaluation between the two devices feasible. The LOA between the TSD and FWD was found to depend on the average measurement of the two devices (see Figure 3.45 and Figure 3.46). Furthermore, there is a bias between the TSD and FWD, SCI, and BDI measurements that depends on the average measurement (see Figure 3.46). To interpret the results presented in Figure 3.46, the example of average SCI of 300 µm (11.8 mils) can be examined. In this case, the SCI calculated from the FWD is expected to be (on average) 30 µm (1.18 mils) lower than the SCI calculated from the TSD, with values ranging from as much as 205 µm (8.07 mils) lower to 175 µm (6.89 mils) higher (for a range of 380 µm [15.0 mils]) expected to occur 95% of the time. For an average SCI of 100 µm (3.94 mils), the value computed using the FWD measurements are expected to be (on average) 50 µm lower than the SCI calculated from the TSD, with values ranging from as much as 115 µm (4.53 mils) lower to 15 µm (0.59 mils) higher (for a range of 130 µm [5.12 mils]) expected to occur 95% of the time. It is important to point out that the large range in the LOA between the two devices does not imply that the TSD fails to give accurate measurements. The LOA between the two devices depends on the repeatability of the TSD, the repeatability of the FWD, and the comparability between the two devices. The repeatability of the FWD was not evaluated. The repeatability between the FWD and RWD was also found to depend on the average deflection measurement. However, the cov was found to be relatively uniform across all measurement values and was therefore used as a measure of repeatability. The average cov of the difference (FWD - RWD) was calculated as -0.2328 (or -23.28%). The 95% interval of the cov was calculated as [-0.8642; 0.4286] or, in percent, [-86.42%; 42.86%]. This means that on average, FWD deflection measurements are 23.28% lower than RWD deflection measurements, and in 95% of the cases, FWD deflection measurements will be between -86.42% lower to 42.86% higher than RWD deflection measurements. This difference seems to be high however, and as in the case of the TSD comparison with FWD, the repeatability of the FWD was not evaluated. Therefore, it was not possible to quantify how much of this large range resulted from repeatability of the FWD. TSD Repeatability from Single Measurement Run The calculation of repeatability requires at least two runs of the TSD repeated on the same pavement section. Ideally, these runs should be performed under the same conditions. For example, for flexible pavement sections, the test temperature should, as much as possible, be the same. This section presents a method to evaluate TSD repeatability from measurements obtained from a single run. Such a method can be very use- ful in cases where repeated runs performed under the same conditions are not practically or economically feasible. This probably includes the majority of applications of continuous deflection devices, specifically the TSD. Difference Sequence Method for Standard Deviation Estimation Difference sequence methods (DSMs) arose from the need to estimate the error standard deviation in nonparametric regres- sion models. The calculated standard deviation can be used, among other things, for the computation of confidence bands or the optimal choice of smoothing parameter (Munk et al., 2005; Brown and Levine, 2007). The general procedure is sum- marized in Hall et al. (1990), who introduced difference-based estimators of arbitrary order r using a difference sequence di of real numbers as follows: di i r = = ∑ 0 3 13 0 ( . ) di i r 2 0 1 3 14 = ∑ = ( . ) ˆ ( . )σ = −( )    +== − ∑∑1 3 151 0 2 1n r d yj j j r i n r It can easily be verified that, except for the case of r = 1, differ- ent combinations of the values di can be used and still satisfy Equations 3.13 and 3.14. Here, the results of the best estimator for the investigated data (Katicha et al., 2012), which was the second order estimator (r = 2), known as the Gasser estimator

62 (Gasser et al., 1986), are presented. This estimator for equally spaced observations is given by the following equation: ˆ ( .σ = −( ) − +  + + = − ∑2 3 2 1 2 1 2 31 2 2 1 2 n y y yi i i i n 16) Calculation of Standard Deviation Table 3.19 and Table 3.20 present the results of standard deviation estimation for the measurements averaged over a 1-m (3.3-ft) length by using the Gasser estimator for P100 and P300, respectively. These are compared with the standard deviations obtained by using repeated runs (presented in the TSD Repeatability section). In general, the two estimates of the standard deviation agree very well, which suggests that the DSM can be used to obtain the standard deviation from a single run. Note that in the DSM, the bias is not accounted for (because there is only one run investigated at a time, the bias cannot be evaluated). Effect of Sampling Frequency on Accuracy of Standard Deviation Estimation Although the results presented in the previous sections show that the standard deviation estimated from a single run agrees Table 3.20. Standard Deviation Calculated with Second Order Difference Sequences for Sensor P300 Slope Standard Deviation (mm/m) Run Average Difference MethodSite 1 2 3 4 5 F1 0.2053 0.2057 0.2027 0.2096 na 0.2058 0.1918 F3 0.1267 0.1363 0.1245 na na 0.1293 0.1376 F5 0.1137 0.1153 0.1071 0.1114 0.1077 0.1111 0.1123 F6 0.1117 0.1156 0.1215 0.1164 0.1181 0.1167 0.1210 C1 0.1491 0.1325 0.1426 0.1357 0.1467 0.1415 0.1398 C2 0.1199 0.1207 0.1201 0.1212 0.1263 0.1217 0.1226 R2 0.1674 0.1617 0.1482 na na 0.1593 0.1671 R3 0.1418 0.1550 0.1358 0.1473 0.1344 0.1431 0.1378 Note: na = not applicable. Table 3.19. Standard Deviation Calculated with Second Order Difference Sequences for Sensor P100 Site Slope Standard Deviation (mm/m) Run Average Difference Method1 2 3 4 5 F1 0.2706 0.2623 0.2647 0.2750 na 0.2682 0.2661 F3 0.1260 0.1360 0.1226 na na 0.1283 0.1305 F5 0.1118 0.1173 0.1096 0.1152 0.1136 0.1135 0.1166 F6 0.1181 0.1253 0.1233 0.1226 0.1213 0.1221 0.1254 C1 0.1503 0.1429 0.1510 0.1357 0.1489 0.1459 0.1466 C2 0.1214 0.1184 0.1186 0.1172 0.1241 0.1200 0.1222 R2 0.2151 0.2087 0.1921 na na 0.2055 0.2072 R3 0.1566 0.1739 0.1501 0.1571 0.1601 0.1598 0.1483 Note: na = not applicable.

63 with the standard deviation estimated by using repeated measurements, there are some limitations to the application of the DSM. For the method to be successful (which depends on how accurate the estimation of standard deviation needs to be), the variation in the true measurement (deflection slope, in this case) profile should happen relatively smoothly. In practical terms, for the most part of the profile, the differ- ence of the true deflection between adjacent points should be small compared to the noise level in the measurements. In a pavement structure, the deflection measured at a given location is influenced by the pavement properties within a certain distance, say of radius R, of that location. The radius R depends on the pavement structure (number of layers, thicknesses, and so forth) and the mechanics that govern the deformation of the pavement (e.g., multilayer analysis). Results presented in the previous section suggest that, for all practical purposes, to obtain a good estimate of the standard deviation from a single run, a distance of 1 m (3.3 ft) is within the radius of influence R for the tested pavements. To illustrate the effect of measurement distance, the stan- dard deviation for the case of P100 measurements averaged over 10 m and 100 m (33 ft and 330 ft) were calculated by using the DSM. The results of this calculation, along with the 1-m (3.3-ft) results, are presented in Table 3.21. The table shows that the calculated standard deviation for 10-m (33-ft) averaging derived with the DSM does not agree with the calculated standard deviation derived with repeated measurements. The DSM produces wrong estimates of the standard deviation for deflection averaged over a 10-m (33-ft) section, probably because the average values can be significantly different from one point to another. The more variation there is in a tested section, the worse the estimate of the DSM. This is confirmed by the fact that Section F1 and Section R2, which had the most variation, resulted in the worse estimate of the standard deviation (see Table 3.21 and Figure 3.52). As expected, the estimate is also worse for longer aver- aging distance; for example, the results for 100-m (330-ft) averaging are very different except for Site C2, which is a relatively uniform site. This shows one limitation of differ- ence sequence methods. However, the standard deviation for different averaging distances can still be estimated from the standard deviation for 1-m (3.3-ft) averaging using the following formula: σ σ n n = 1 3 17( . ) where n = section length in meters, sn = the standard deviation for the sections of length n, and s1 = the standard deviation for 1-m averaging. Denoising and Data Aggregation The purpose of continuous deflection devices is to measure a physical characteristic of the pavement (e.g., deflection in the case of RWD and deflection slope in the case of TSD) at specific locations. Measurements are contaminated with noise; from these noisy measurements, the objective is to make inferences about the expected value of the true physical quantity being measured with a confidence interval on that expected value. For TSD measurements, this can be interpreted as finding out how the deflection slope varies as a function of the measurement location (i.e., along the roadway). This Table 3.21. Calculation of Standard Deviation for Measurements Averaged over Distances (mm/m) 1 m (3.3 ft) 10 m (33 ft) 100 m (330 ft) Site Difference Sequence Repeated Runs Difference Sequence Repeated Runs Difference Sequence Repeated Runs F1 0.2682 0.2611 0.2548 0.1054 0.3504 0.0625 F3 0.1282 0.1305 0.0436 0.0407 0.0256 0.0155 F5 0.1135 0.1166 0.0578 0.0401 0.0544 0.0147 F6 0.1223 0.1254 0.0448 0.0406 0.0227 0.0139 C1 0.1458 0.1466 0.0659 0.0446 0.0423 0.0153 C2 0.1200 0.1222 0.0409 0.0408 0.0168 0.0142 R2 0.2053 0.2072 0.1766 0.0643 na na R3 0.1596 0.1483 0.0500 0.0432 0.0179 0.0153 Note: na = not applicable.

64 can be done by using nonparametric regression. Common regression analysis, which is extensively used by engineers, is parametric regression where a (parametric) model (e.g., linear model) is postulated to represent the observed behav- ior and model parameters are obtained using a specified criterion (e.g., least squares, maximum likelihood). In some cases, such as for continuous deflection measurements, the form of the regression curve is not known. In such cases nonparametric regression, which is not restricted to a given form (such as linear and exponential), can be used. Essen- tially, the collected data are used to infer on the regression function. The three most common methods of nonparametric regres- sion are kernel regression, smoothing spline regression, and least-squares spline regression. The familiar moving average falls under kernel regression. From a practical perspective, all three methods give very similar results, and selecting a particu- lar one is often the result of individual preference and ease of use. In this report, the method of smoothing splines is used for its simplicity and ease of implementation. The main ques- tion in smoothing spline regression is how much smooth- ing should be performed. A number of objective methods have been developed to answer this question. These methods consist of optimizing a parameter that controls the trade-off between smoothness and adherence to the measurements (i.e., controls variance and bias). The method can be formu- lated using the following model for the TSD deflection slope measurements: y f xi i i= ( )+ ε ( . )3 18 where yi = TSD deflection slope measurements, i = 1, 2, . . . , n (the number of measurements), xi ∈ [0, 1], f(xi) = true deflection slope (which is not known and is to be estimated), and ei = i.i.d random variables with mean zero and known or unknown variance s2. The smoothing spline method consists of finding the func- tion g that minimizes the following formula: 1 3 19 2 1 2 0 1 n g x y g u dui i i n m( )−( ) + ( )( ) = ( )∑ ∫λ ( . ) where g(m) is the mth derivative of g. In this report, m is taken as m = 2 (the typical value used for m). In this case, the solution to the minimization problem, g is a cubic spline. The parameter l (the only parameter that needs to be determined) is a smoothing parameter that controls the 0 500 1000 1500 2000 0 2 4 6 a 0 500 1000 1500 2000 2500 3000 -0.2 0 0.2 0.4 0.6 b 0 500 1000 1500 2000 2500 3000 0 0.5 1 c 0 500 1000 1500 2000 2500 3000 3500 0 0.2 0.4 0.6 0.8 d 0 1000 2000 3000 0 0.5 1 1.5 e 0 500 1000 1500 2000 2500 3000 0 0.2 0.4 0.6 0.8 f 0 200 400 600 800 1000 1200 0 2 4 Distance (m) Sl op e (m m/ m) g 0 200 400 600 800 1000 1200 1400 0 0.5 1 h Figure 3.52. TSD slope measurements for Sensor P100 for U.K. sites: (a) F1; (b) F3; (c) F5; (d) F6; (e) C1; (f) C2; (g) R2; and (h) R3.

65 trade-off between the “roughness” of the solution and the fidelity to the measured data. A popular method to estimate l, which is used in this report, is the generalized cross validation (GCV) criterion first suggested by Craven and Wahba (1979). Illustrative Simple Examples with a Sinusoidal Function Two simple examples of spline smoothing using a sinusoidal function are first presented. The purpose of these examples is to give an intuitive feeling on the performance of the method. The example compares synthetic measurements with relatively low noise levels (s = 1) and relatively high noise levels (s = 5) of a sinusoidal function given by the following equation: y x= ×( )+sin . ( . )2 0 1 3 3 20pi Measurements are obtained for x between 0 and 60 and 0.01 spacing (units are not important in the example; however, these can be meters). Figure 3.53a shows the results for the case of low noise levels (s = 1). From the measured signal, the underlying signal is presumed to be fluctuating (such as a sinusoidal). The smoothing spline estimate (dashed red line) is very close to the actual sinusoidal signal (continuous blue line). Figure 3.53b shows the measured signal for the case of high noise levels (s = 5). In this case, visual inspection of the plot may not by itself suggest that the true signal is fluctuat- ing. However, even for the level of noise in the signal, the performance of the smoothing spline in estimating the true function is still very reasonable. Application to TSD Measurements In this section, the application of the smoothing spline is demonstrated on three sets of TSD deflection slope measure- ments from the P100 sensor on Sites UK_F3, UK_F5, and UK_F1. These were selected because of the different range of TSD slope variation in each site (see Figure 3.52). UK_F3 exhibited very low variation in the measured TSD deflection slope, UK_F5 exhibited some variation in the measured TSD deflection slope, and UK_F1 exhibited the most variation in the measured TSD deflection slope. Figure 3.54 shows the measured first run on Site UK_F3, with the smoothing spline regression function and the 95% Bayesian confidence interval (Wahba, 1983; Nychka, 1988, 1990) superimposed to the measurements and the averages over 100-m (330-ft) sections. The smoothing function gives results that are as “noise-free” as averaging results over 100-m (330-ft) sections. One important difference between the smoothing function and the 100-m (330-ft) average sections is that the smoothing function is an esti- mate of the deflection slope at 1-m (3.3-ft) intervals (com- pared to an estimate of the average of a 100-m [330-ft] section) and can therefore be used for project-level appli- cations (compared to network-level applications). Further- more, a confidence interval can be constructed around the estimate. The operation of smoothing filters the noise from the data at the cost of introducing bias. The GCV criterion finds the compromise between the bias and variance to be a trade-off. In practical terms, if more smoothing is per- formed, important features that are not spikes due to noise will be smoothed out. This is better illustrated with Section UK_F5. Figure 3.55 shows the results of smoothing measurements on Section UK_F5. In this case, an appreciable difference between the smoothed regression estimate and the results averaged over 100-m (330-ft) sections can be observed. The figure suggests that results for the 100-m (330-ft) aver- aged section are smoothing statistically significant features of the deflection slope profile. Therefore, these features are not likely due to random noise and perhaps (depending on whether they are important from an engineering perspective) should not be ignored (or smoothed out). 0 10 20 30 40 50 60-2 0 2 4 6 8 Measured Signal True Signal Estimated Signal 0 10 20 30 40 50 60-15 -10 -5 0 5 10 15 20 25 Measured Signal True Signal Estimated Signal (a) (b) Figure 3.53. Spline smoothing of sinusoidal function contaminated with (a) low and (b) high noise levels.

66 UK_F1 shows an even more extreme example (Figure 3.56). In this case, averaging over 100-m (330-ft) sections is likely to result in smoothing out significant features (both statistically and from an engineering perspective). Figure 3.56 also shows that averaging over 10-m (33-ft) sections results in smooth- ing out statistically significant features. This was expected, as UK_F1 shows significant variation in the measured deflec- tion slope. In this case, the noise in the signal is relatively small compared to the signal itself. Therefore, smoothing has a much more significant effect on changing the signal than on reducing the level of noise and should be minimal. Again, how much to smooth is controlled by the GCV criterion, which finds a compromise between bias and variance. Summary The results presented in this section suggest that the smooth- ing (or averaging) deflection test results should not be set to a constant value, but rather controlled by the actual deflec- tion profile. Using an “optimal” smoothing may help improve the capabilities of continuous deflection devices to be used for project-level applications. Furthermore, the smoothing spline analysis can be used to identify features that vary with distances as small as 1 m within a certain level of confidence (using the confidence interval as shown in Figure 3.54). Although the limited (in terms of number of sections and section length) data used are only for the TSD, a similar type of analysis should be feasible for the RWD if more closely spaced data are provided. However, closely spaced (raw) RWD measurements were not available for this project and this type of analysis could not be carried out for this device. Operational Characteristics Within this study it has not been possible to examine in detail the operational characteristics of the equipment being assessed. Although external factors such as temperature, road geometry, road profile, texture profile, moisture, acceleration, deceleration, and so forth, may have been recorded during the surveys, it is not possible to control these factors and, therefore, not easy to 0 500 1000 1500 2000 2500 3000 -0.4 -0.2 0 0.2 0.4 0.6 0.8 Distance (m) D ef le ct io n Sl op e (m m/ m) Measured Smoothed Upper 95% CI Lower 95% CI 100 m average Figure 3.54. Smoothing slope deflection measurements of site UK_F3. 0 500 1000 1500 2000 2500 3000 -0.2 0 0.2 0.4 0.6 0.8 1 Distance (m) D ef le ct io n Sl op e (m m/ m) Measured Smoothed 100 m average Figure 3.55. Smoothing slope deflection measurements of Site UK_F5.

67 assess their effect on the measurements. Control of survey speed is possible, but sometimes other uncontrollable effects confound the effort. The following sections discuss these issues together with the measuring capability of current and future devices on the typical range of pavement types encountered on a network. Operating Conditions Survey Speed The RWD is normally operated as close as possible to 80 km/h (50 mph). As yet, no method has been developed for converting surveys at other speeds to the standard operating speed. At present the TSD operates as close as possible to 70 km/h (45 mph) on divided highways, but a range of 60 to 80 km/h (40 to 50 mph) is considered acceptable. For two- lane roadways in the United Kingdom, because of the speed limit, the standard operating speed is 60 km/h (40 mph). The effect of survey speed has yet to be investigated fully in either device. The deflection response of a pavement will be influenced by a number of factors, which include the speed of the loading wheel and the composition of the pavement. In order to ascertain the extent of the effect, any experiment will need to be very carefully controlled in terms of operating conditions; and the results are likely to vary with the proper- ties, particularly the viscoelastic properties, of the pavement layers. This kind of control was not possible within the scope of the project. However, a limited examination based on available data has been carried out for both devices. The California RWD demonstration included repeated runs at various speeds on flexible and rigid pavements. Fig- ure 3.57 shows 160-m (0.1-mi) RWD results on two of the tested sections, with flexible and rigid pavements. The figures suggest that RWD deflections are relatively insensitive to truck speed for the speed ranges investigated (50 to 110 km/h [30 to 70 mph]). Since the tests were conducted at different times, some differences for the asphalt pavement could be due to variations in temperature. Repeat runs have been carried out with the HA TSD at different speeds. A small change with increasing speed has been shown, but data collected so far have been insufficient to develop a correction procedure to a reference speed. Fig- ure 3.58 illustrates this by presenting the TSD P100 slope profiles as 10-m (33-ft) means for a 1.5-km (1-mi) section with flexible pavement at 60, 70, and 80 km/h (40, 45, and 50 mph). An analysis of network TSD measurements collected in the United Kingdom during 2010 and the early part of 2011 over a wide range of speeds, some outside the current recommended limits, is shown in Figure 3.59 in the form of a distribution plot. This suggests a reduction in response with increasing speed. However, there are many other confounding effects present in this data, so it can be taken only as an indication of the likely effect of survey speed. The figure shows a two-dimensional ver- sion of a three-dimensional plot, in which the lines are con- tour lines containing given proportions of the total dataset. For example, 20% of the data were collected at speeds of around 61 km/h (38 mph) with slopes from 0.18 to 0.37 mm/m. The other concentration of data is just below 70 km/h (44 mph). These were the two target speeds in the surveys, but other speeds were covered for various practical reasons. Road Geometry and Profile The specific effect of road geometry has yet to be investigated for either device. In particular, measurements on curves are a potential issue, especially with the RWD because of the way it compares the same texture profile from the deflected sensors and the undeflected or reference sensor. The two sensors may not follow the same trajectory while measuring on curves. A superficial examination of the effect of road curvature has been carried out in the United Kingdom by examining the TSD 0 500 1000 1500 2000 0 1 2 3 4 5 6 7 Distance (m) D ef le ct io n Sl op e (m m/ m) Measured Smoothed 100 m average 10 m average 1200 1220 1240 1260 1280 1300 0 1 2 3 4 5 6 7 Figure 3.56. Smoothing slope deflection measurements of Site UK_F1.

68 Figure 3.57. Example of repeated RWD runs at different speeds. (a) AC test section on eastbound SR-16, Mileposts 14 to 15.5 (b) JPCP test section on southbound I-505, Exits 24 to 28 Source: ARA, 2007c. network data collected during 2010 and 2011. This examination showed no obvious relationships between longitudinal profiles, gradient, transversal slope, or curvature. The data, with respect to longitudinal profile, are shown in the form of a contour map of the distribution of almost 10,000 lane-km (6,250 lane-mi) of data of deflection slope versus 3-m (10-ft) longitudinal profile variance in Figure 3.60. Surface Characteristics Research with the U.K. TSD has shown that surface type can influence the response of the velocity sensors. In particular, new binder-rich surfaces can cause faulty operation of the velocity sensors on the TSD but normal measuring performance returns after a few months of trafficking as the surfacing becomes less

69 0.000 0.100 0.200 0.300 0.400 0.500 0.600 0.700 0.800 0.900 1.000 1500 1700 1900 2100 2300 2500 2700 2900 D e ec tio n Sl op e [m m /m ] Chainage [m] 80km/h 70km/h 60km/h Figure 3.58. TSD slope profiles for 10-m (3.3-ft) means on a U.K. road section at various survey speeds. Figure 3.59. Effect of survey speed on TSD deflection slope. Note: The blue area highlights the overall trend of the reduced deflection slope with increased velocity.

70 reflective. Otherwise, the equipment functions reliably on a wide range of surfaces, including rigid concrete. This reliability has been illustrated by surveying on a length of road with various types of road surfacing and in various conditions. This work established that the optics perform better on lighter surfaces, such as jointed concrete, and least well on new bituminous surfaces. This conclusion can be seen in Figure 3.61, where the data rate of the four Doppler lasers is shown for one particular survey along a 13-km (8.1-mi) length of a U.K. motorway. The figure shows changes in data rate (the rate at which laser signals are successfully returned from the road surface) over lengths of the jointed concrete and a length of freshly laid thin surfacing in comparison to the rate returned from the old bituminous surface. This phenomenon is likely due to optical properties of the different surfaces or surface condition. Although it is not fully understood, it is important to identify such features because they help to define the capabilities of the technology, and such information will probably be employed in the quality assurance of TSD data in the future. Information on the effect of such parameters on RWD measurements is not available but is likely to be similar. Moisture Both devices use laser-based noncontact sensors that fail to measure correctly when the road is damp or wet. The laser reflection is degraded by the water on the surface. Dynamic Loading The specific effect of dynamic loading has yet to be investigated for either device. At present, neither the RWD nor any existing TSDs have the capability to routinely measure dynamic load- ing on the measuring load. The Danish Road Institute has been exploring this possibility with BASt in Germany using strain gauges mounted on the rear axle of the trailer. The latest TSD, currently being constructed for the South African government, is said to include this technology. Some have argued that because the equipment has a suspen- sion typical of other trucks, any additional deflection response caused by additional dynamic loads represents what normally occurs at that point on the road and so provides a representative estimate of the structural condition at that location. However, knowledge of the dynamic load would provide a more complete understanding of pavement behavior. Acceleration and Deceleration Longitudinal vehicle acceleration and deceleration are likely to affect the accuracy of deflection measurements, so oper- ating limits for these parameters have been developed for the U.K. HA TSD. When operating at normal survey speed, neither device requires traffic management. However, some form of traffic management will be necessary for operation at slower speeds, as well as when operating in nonstandard locations, such as in a U.K. motorway outer lane where trucks are not normally permitted. Acquisition and Operation Costs Only the TSD is available commercially, at a cost between $2 million and $2.5 million depending on the number of sensors requested. Although detailed operation and main- tenance costs were not obtained for this study, experience in the United Kingdom suggests that the cost of operating Figure 3.60. Effect of longitudinal profile variance on TSD deflection slope.

71 the TSD, not including the capital or maintenance costs, is approximately $25 to $40 per mile. Adding the process- ing cost is estimated to bring this to approximately $75 to $90 per mile. Measurement Capability Bowl Shape Detail The current RWD is designed to measure just the vertical deflection close to the wheels in the outside wheelpath. This should be closely related to the maximum deflection response of the pavement, although the location of this maximum will vary depending on the composition of the pavement and the survey speed, due to any viscoelastic properties of the pave- ment materials. The RWD as evaluated in this report did not have the capability to measure the full deflection bowl, or a sampled representation of it, as provided by the multiple sen- sors fitted to most FWDs. However, recent modifications have added an additional sensor to the RWD at a second position farther away from the rear axle to provide such information (see New Development to the Evaluated Equipment section). At present the two fully operational versions of the TSD measure the vertical velocity of the pavement response to the dual wheel assembly loading at three offsets in front of the rear axle, which is then converted to deflection slope. In the case of the Danish device, these offsets are 100 mm (4 in.), 200 mm (8 in.), and 300 mm (12 in.). In the case of the U.K. device, they are 100 mm (4 in.), 300 mm (12 in.), and 756 mm (30 in.). The deflection sensors respond to velocity (not displacement), which are converted to deflection slope, and therefore cannot directly provide either the full deflection or the maximum value. However as discussed later, the mea- sured slopes closest to the axle, P100 and P300, have shown a strong relationship to the peak deflections measured by other devices. The manufacturer, Greenwood A/S, has developed an approximate relationship that can be fitted to the three offset slope measurements and thus enable estimation of the surface curvature index of the pavement surface under load. At present it is not possible to estimate the full deflection bowl from the current sensor configurations, but a device with more sensors was delivered in 2010. Both devices currently measure in just one wheelpath in between the two loaded wheels, which are mounted at a slightly wider spacing than are standard truck dual wheels to enable room for the measurement sensors. At present, both measure in the nearside (outside) wheelpath closest to the pavement outer edge in the countries in which they operate. However, measurement in both wheelpaths should be feasible with suit- able modifications and at an additional cost. The latest TSD being constructed for the South African government is reported to have the capability of measuring in both wheelpaths. At present the RWD operates with a 40-kN (9-kip) dual wheel assembly load and the TSD operates with a 50-kN (11-kip) load. Other loads can be employed relatively easily. Sampling and Reporting Intervals Both devices sample the raw measurements frequently, the RWD at around 2,000 Hz, equivalent to around 11 mm at 80 km/h (0.4 in. at 50 mph), and the TSD at around 1,000 Hz, equivalent to around 22 mm at 80 km/h (0.8 in. at 50 mph). However, there is much noise in the raw signal, so results are normally reported over much longer lengths. Some examples of different sampling lengths on the results are presented in a later section. Published material on the RWD suggests that the device is suitable for measuring only on flexible and compos- ite pavements, normally presenting results at 160-m (0.1-mi) Figure 3.61. Effect of road surfacing type on data rate from each TSD laser (p/s 5 number of successfully returned laser pulses per second). P100 P200 P300 Pref Events 0 200 400 600 800 1000 1200 0 2000 4000 6000 8000 10000 12000 Chainage [m] D at a R at e [p /s] Exposed Concrete New bituminous surface

72 TSD deflection response Height laser response (5m transverse joint slots) Figure 3.62. Longitudinal profile and deflection slope results from TSD survey on jointed rigid pavement. -2 0 2 4 6 8 10 0 50 100 150 200 1 3 5 7 9 11 13 15 17 19 Sl op e 10 0 (m m/ m) FW D L T( %) Chainage (m) FWD LT % Bend TSD10K Good Moderate PoorPoor PoorPoor Figure 3.63. Raw TSD slope measured at 10 km/h (6 mph) and FWD joint load transfer efficiencies against location.

73 intervals. The position with the TSD is similar except that results are normally presented at 10-m (33-ft) intervals. However, some surveys have suggested that shorter length structural vari- ations can be distinguished if the variability is sufficient. This has been discussed further in mathematical terms and examples given earlier in the Denoising and Data Aggregation section. Joint Load Transfer Efficiency Limited measurements have been made on an unreinforced jointed rigid pavement on a research track at a survey speed of just 10 km/h (6 mph). Figure 3.62 shows some results over a 90-m (300-ft) length. The red lower profile shows the longitudinal height profile with spikes indicating the 5-m slab joints. The upper blue trace shows the response of one TSD sensor with significant spikes at chainages of 2,827 and 2,859, suggesting poor joint condition. In Figure 3.63, the raw deflection slope profile of another section is shown in comparison to load transfer efficiencies assessed with an FWD. This preliminary investigation suggests that the TSD may have potential to assess the transfer efficiency at joints, but further work is needed to make this a practical routine proposition. No information is yet available on the applica- bility of either device to unsurfaced granular pavements.