Page 109
5—
Modeling and Simulation
for the Virtual Factory
Introduction
The development of manufacturing systems often proceeds in the operating factory by trial-and-error experimentation using valuable factory resources and time, with a resulting high cost in lost manufacturing productivity due to prototyping and debugging on the factory floor. Data continue to be generated at an exponentially increasing rate, but there is little opportunity to assimilate, much less act on, the information they represent. The abundance and variety of customized and unintegrated new technological capabilities introduce problems for managers, engineers, supervisors, and operators within a factory, with which the work force may be ill-prepared to cope. Better ways to introduce new manufacturing technology onto the factory floor must be developed to help improve our manufacturing capabilities.
As noted in Chapter 4, the factory environment is complex. To develop high-level insight into what is happening in the factory environment and how factory operations might be affected by changes in that environment, manufacturing researchers have attempted to construct simulations and models of factories that factory decision makers can experiment on without disrupting actual production. This approach requires a virtual factory to be a simulation that faithfully reflects that operation in all of the dimensions relevant to human managers, from the week-long and month-long time scales that characterize today's material requirements planning (MRP) and manufacturing resources planning (MRP-II) planning and scheduling systems to the second- and minute-long time scales that characterize real-time dynamic schedulers (human or computer) of shop floor operations.
Page 110
Considerations In The Development
Of Virtual Factories
Learning from Past Problems
The idea of a virtual factory is not new. But in the past, the development and application of large coherent models of factory operations have been largely unsuccessful, even as more local simulations (e.g., those operating at the equipment level or process level) have succeeded. The models of the past have not had the content to capture the essence of factory operations, yet even these inadequate models were so difficult to produce that after great effort they typically ground down into "analysis paralysis" and were abandoned.
Possible causes for such failure include the following:
| • | The initial models were too poor in detail to provide answers that satisfied factory demands, and so the concepts were abandoned. Even small events may produce large fluctuations in factory operations, and adequate models must be capable of reflecting these subtle influences. |
| • | The simulations were too slow to provide timely answers. Timeliness is paramount in factory operations; a piece of equipment that is unexpectedly down, or is being used for another task when required to perform a function, can dis rupt a shift schedule. Data must be provided and acted on very quickly to preserve the integrity of operations. |
| • | The representations of the process were inaccurate, leading to wrong answers. Processes may be simple or complex; in any event, all the important characteristics of a process need to be accurately represented so that applying a model will supply useful answers. |
| • | The user interfaces were so complicated and/or incomprehensible that they were unusable. The most common user of information is a human being, who can be overwhelmed by either the number or the complexity of the user interfaces required to access or disseminate information. For example, senior factory managers are best able to comprehend results that are explicitly tied to financial metrics of performance; results tied to metrics relevant at lower levels in the hierarchy will be less helpful to them. |
| • | There were insufficient skilled personnel to understand and apply the models intelligently. Use of models is not inherently easy. It requires skills that enable using models and simulations, as well as understanding and analyzing the results. Learning these skills requires education and training. |
| • | There were sufficient skilled experts on the factory floor to manage operations, so that modeling and simulation were considered unnecessary. When things are going relatively smoothly, new tools such as modeling and simulation are felt to be superfluous; current skill sets are thought to be sufficient to do the job. |
Page 111
| • | Invalid input data to simulations led to incorrect results. For example, communications on the factory floor were insufficient to keep the model up to date. (Factory floor communications are used, among other things, to keep track of equipment placement on the shop floor, a factor with a big impact on work flow; if a model does not reflect changes in equipment placement (which may change on a time scale of hours), the model's output relevant to making work flow decisions may well be incorrect.) |
| • | Factories themselves constantly change (e.g., new or modified manufacturing processes may be installed), and a lack of synchronization between a model and what is actually being done on the shop floor may invalidate the model. |
| • | Simulations may have been performed at inappropriate levels of detail for addressing problems of interest and relevance to decision makers. An element of considerable importance in constructing a simulation is understanding what kinds of questions need to be answered for what purposes. |
| • | Today's simulation models are difficult to expand, and they present very difficult problems in scaling up. |
For such reasons, factory-level modeling has never succeeded in capturing the attention of senior manufacturing management the way that process and product models have. However, recent advances in information technology make the idea of realistic simulations of factory operations much more feasible than they have been in the past.
Determining the Requirements for Effective Factory Models
A virtual factory model will involve a comprehensive model or structure for integrating a set of heterogeneous and hierarchical submodels at various levels of abstraction. Each submodel will be designed for a specific purpose, but together they will operate from a common source of data or knowledge base and will be able to deal with the task at hand without expensive or time-consuming hand-tailoring of interfaces for a user's particular needs. To a very high degree, the software used to control actual factory operations will also drive the operation of the virtual factory model, although it will do so very rapidly so that simulations can be run on a timely basis.
A key consideration in the development of these tools will be accounting for the essential stochastic nature of events on the factory floor.1 Since managers
1 A Research Agenda for CIM (MSB, 1988, p. 20) recommends research on methods for representing decision objects, for example, states, actions, utilities, prior and posterior probabilities, samples, costs, and decisions; methods for mapping these to manufacturing objects; development of relevant heuristics and algorithms; and exploratory assessment of the merit of these techniques in specific domains.
Page 112
will be using virtual factory models to perform "what if" exercises to anticipate unexpected events (e.g., tool breakdown, new demands for expediting orders, the arrival of new machines, unplanned maintenance) and to indicate what actions need to be taken (e.g., which machine is to be changed, what materials are to be provided), a simulation based only on deterministic factors would correspond to one "experiment," and in a stochastic environment, there is no reason to expect that the same line configuration or schedule plan would yield the same result on any particular run. However, a suite of modeling tools that could test a given configuration or plan against thousands of experiments (and the random events in each run) might well provide a meaningful basis on which to make decisions about a configuration or plan. A second important consideration is that the development of simulation models and tools should be broad enough to encompass both the factory technologies and processes of today (which are embedded in the manufacturing infrastructure and may not be replaced for many years to come) and those of tomorrow (i.e., those technologies and processes that have not yet been deployed widely or even invented yet).
A complete simulation of even a modest factory will require the integration of numerous models and is out of reach today. However, an appropriate first step in developing a full factory model is the virtual production line, which involves simulation of individual tools and, more importantly, simulation of the integrated operation of the tools in production. A wide variety of individual, single-activity models are already in use in manufacturing. But the comprehensive integrated modeling of the manufacturing enterprise will provide new insight into the causes of and remedies for scheduling bottlenecks and new strategic options.2
An ultimate goal of the research agenda outlined here would be the creation of a demonstration platform that would compare the results of real factory operations with the results of simulated factory operations using information technology applications such as those discussed in this report. This demonstration platform would use a computer-based model of an existing factory and would compare its performance with that of a similarly equipped factory running the same product line, but using, for example, a new layout of equipment, a better scheduling system, a paperless product and process description, or fewer or more human operators. The entire factory would have to be represented in sufficient detail so that any model user, from factory manager to equipment operator, would be able to extract useful results.
There are two broad areas of need: (1) hardware and software technology to handle sophisticated graphics and data-oriented models in a useful and timely
2 A Research Agenda for CIM (MSB, 1988, pp. 13-15) recommends research on methods and technologies for process operation analysis, optimization, control, and quality assurance and (on p. 17) recommends research on methods for coupling process operation with process planning and resource allocation so that these activities can be accomplished concurrently.
Page 113
manner, and (2) representation of manufacturing expertise in models in such a way that the results of model operation satisfy manufacturing experts' needs for accurate responses.
Modeling Technology
A requirement for manufacturing models is that all levels of detail be internally consistent, since very small influences at a very local level may have a significant impact even at the highest factory level (e.g., if a piece of equipment goes down, the resulting schedule disruptions can shut down the entire factory). A factory may well demonstrate mathematically chaotic behavior, such that very small influences may have totally unpredictable outcomes. The enormous range of objects (e.g., pieces of equipment, people, raw materials, batches or lots of material, process steps, process control, assemblies of objects, a wide variety of information "objects," material-handling and robotic equipment, and so on) encompassed by even a partial manufacturing model further adds to its complexity.
In general, the complexity needed in a model to generate accurate results strongly influences the time it takes to simulate some factory condition. Thus, there is a trade-off between speed of response to provide timely answers and the level of detail required to provide "good enough" answers. For example, some simulations of semiconductor processing operations may take hours or days of computation (i.e., may take much longer than the operations take in reality); such performance may be useful for analytical purposes but not for controlling or advising for real-time operations.
If advanced modeling technology is to be useful in the future, it must not disregard the large investment that has been made in existing data systems. The client-server systems and relational databases of today, and even object-oriented databases, will be the legacy systems three and four decades hence. New technology must be able to fold these legacy systems into its representation and be able to resolve semantic differences that may exist.
In addition, new modeling technology will likely reside in an extensive distributed computing environment. Thus, fundamental issues of maintaining data integrity in a distributed computing environment need to be resolved. Administration of systems, networks, and applications is today resource-intensive; more efficient and effective methods of managing these elements will be required.
Finally, information at the appropriate level of detail must be presented to relevant decision makers and analysts. Visibility of the activities of an enterprise is a prerequisite to effective control. Research is needed regarding how to present to different users (from managers to operators) the information inherent in the business model, which can reflect an actual state, potential future states, and postmortem analysis of past states to determine what went right or wrong. People should be able to prepare presentations of various kinds of information on their
Page 114
own, rapidly, and without the use of support personnel, and to retrieve "lessons learned" and present that information at the appropriate times to users.
Specific research issues related to developing manufacturing models include improving methods and tools for:
| • | Efficiently managing large amounts of related data distributed over many machines and locations. Data may be directly linked to the computer model of the production process, automatically updating the status of the model, and it will be desirable to automate data collection to the maximum degree. |
| • | Specifying complex data relationships. Since the various scales of manufacturing are intimately interlinked, models will require large amounts of detail, probably stored in distributed object-oriented databases. However, these databases will need to be consistent across geographical sites and will span many time zones. Large servers will be needed to manage the model applications and the propagation of changes. |
| • | Easing interoperability of design process tools. Interoperability problems lie in two domains, in the meaning of the information passed between tools and in the mechanics of their transfer. To address the first issue, a complete data model of the simulation process must be created, one that will specify the information at every stage of operation. When this is coupled with a flow model of the simulation, it will provide an unambiguous definition of what information is being interchanged and when. The mechanics of information transfer require either the development of standards for passing messages between tools (e.g., the communication required between objects in an object-oriented paradigm3) or the specification of programming interface standards, or both. |
| • | Constructing formal models that reflect factory resources. Reflecting factors such as resource availability, condition or repair status, and shop floor location, such models should interact with process and product models; indeed, at sufficiently high levels of abstraction, process model components may appear as components of a resource model. Careful configuration management of these models will provide an approach to alleviating the problem of validating models for a continually changing factory environment. |
| • | Presenting data (e.g., using human interface tools). For example, model displays may have to be very large, either to enable viewing large quantities |
3 This is one stated goal of the product data exchange using STEP (standard for the exchange of product model data) project described in Chapter 3.
Page 115
| of data or to show lots of visual detail. The displays may even have to be mobile, so that they can be carried around in a factory. Display panels may be attached to a computer or be in the form of glasses used by a viewer to simultaneously view the model and the real world. | |
| • | Reducing the magnitude of the testing and validation task problems by carefully configuring the relationships between models used for control and those used for simulation. A factory control system that has been successfully tested for operation in real time could in principle be connected to a set of inputs that simulate the input it receives in real time, and could thus serve as the foundation of a (partially validated) simulation model as well. |
| • | Developing new paradigms for understanding dynamic interactions among the different aspects and components of the manufacturing enterprise. A real factory consists of many interacting elements; thus, it is reasonable to expect a high-fidelity simulation to consist of many interacting elements (models) as well. But the real factory is difficult to understand, in part because of these interactions, and so a real problem of understanding is posed by a complex multitude of interacting models. Developing new ways to understand the complexity, nature, and scope of interacting components will be a major challenge with ramifications for ensuring model fidelity and validation. |
The following are research areas within modeling technology that need more work:
| • | User interfaces. Since factory personnel in the 21st century will be interacting with many applications, it will not suffice for each application to have its own set of interfaces, no matter how good any individual one is. Much thought has to be given both to the nature of the specific interfaces and to the integration of interfaces in a system designed so as not to confuse the user. Interface tools that allow the user to filter and abstract large volumes of data will be particularly important. |
| • | Model concurrency. Models will be used to perform a variety of geographically dispersed functions over a short time scale during which the model information must be globally accurate. In addition, pieces of the model may themselves be widely distributed. As a result, a method must be devised to ensure model consistency and concurrency, perhaps for extended time periods. The accuracy of models used concurrently for different purposes is a key determinant of the benefits of using such models. |
| • | Testing and validation of model concepts. Because a major use of |
Page 116
| models is to make predictions about matters that are not intuitively obvious to decision makers, testing and validation of models and their use are very difficult. For factory operations and design alike, there are many potential "right answers" to important questions, and none of these is "provably correct" (e.g., what is the "right" schedule?). As a result, models have to be validated by being tested against understandable conditions, and in many cases, common sense must be used to judge if a model is "correct." Because models must be tested under stochastic factory conditions, which are hard to duplicate or emulate, outside the factory environment, an important area for research involves developing tools for use in both testing and validating model operation and behavior. Tools for automating sensitivity analysis in the testing of simulation models would help to overcome model validation problems inherent in a stochastic environment. | |
| • | Model evolution. Models are not static, in the sense that as users learn more or as the factory environment changes (e.g., as a result of new rules, new heuristics, changing equipment sets, and so on), models have to be updated. In addition, as new knowledge is developed, models need to be enhanced, speeded up, and made more detailed. The information that can be obtained from a model depends fundamentally on the model's capabilities, which are expected to improve with time. This is a fruitful area for research. |
| • | Chaos theory. To the extent that manufacturing is a mathematically chaotic environment, chaos theory may be able to show where critical assumptions break down and where they may lead to computationally impossible situations and totally unpredictable behavior. |
Representing And Capturing
Manufacturing Expertise
It is obvious that the modeling sophistication implied by the preceding section will demand an extraordinary degree of manufacturing expertise at all scales of operation (from shop floor operator to senior management). But a number of factors specific to the manufacturing enterprise exacerbate these demands even further:
| • | Since activities and actions at the different scales of manufacturing operation cannot be considered independently, expertise relevant to one level will often couple strongly to expertise at a different level. Since models will pass information up and down the management ladder, information exchange mechanisms must be consistent. Hence, a consistent model representation language is needed that will precisely define a model's objects and the relationships between objects at different levels of abstraction. |
Page 117
| • | The desired coupling of a model (the virtual factory) directly to a real factory demands that changes in the real factory be quickly incorporated into the model, or, conversely, that the model be used to drive changes in the real factory. For example, the movement of a piece of equipment in a model would cause a robot to relocate the corresponding real piece of equipment. Or, a change in the temperature in a model object would cause the real temperature to change in the real factory. To enable such coupling, the links to the real factory from the model need to be very robust and bidirectional. |
| • | Different "windows" into a model will have to be provided for different users, including "help"-line operators, engineers, managers, supervisors, technicians, schedulers, human resources personnel, and maintenance personnel. Each type of user will have his or her own perspective on reasons for using the model, and the model should be able to be tailored to accommodate those specific needs. |
It is likely that traditional approaches in artificial intelligence with respect to knowledge engineering will be helpful in capturing expertise, since heuristic or even analytical manufacturing information is difficult to formalize. At the same time, research in the area of capturing manufacturing expertise should not be limited solely to the artificial intelligence knowledge-engineering approach. For example, it would also be valuable to investigate how people involved in the manufacturing enterprise construct meaning when they are listening to others describe some manufacturing process or activity. Hard experience in such situations suggests that even when all participants share the same spoken and written language (English plus mathematics), it is often difficult to specify certain procedures without ambiguity.
The following describes some of the research and development needed for capturing and representing information:
| • | Gathering expertise. Although it has been possible to gather knowledge from experts in certain narrow and well-focused domains, acquisition of knowledge is still difficult. Since manufacturing is such a varied and complex activity, and since the information exists over a very wide range (from high-level factory decision making to running a specific piece of equipment), capturing the tacit knowledge of an experienced work force for later reuse represents a key challenge. |
| • | Languages for gathering and representing knowledge. An information-theoretic based language is needed in which to gather and assemble responses from experts and channel them into a machine-usable form. In addition, a good language is needed to represent the information in the model. This may or may not be the same language as that used to gather knowledge; however, it will certainly be closely related. |
Page 118
Research Areas Not Specific To Manufacturing
A number of research problems in areas not specific to manufacturing will also need to be addressed if the virtual factory is to be realized. For example:
| • | Models that must be concurrently activated and used may have to be built by geographically dispersed teams. Methods to provide for concurrent engineering and for keeping models or pieces of models synchronized will have to be developed. |
| • | Models must run at real-time speeds if they are to be used in conjunction with real-time factory operations; they must run 1,000 or 10,000 times faster than real time if they are to be used to experiment with different factory configurations. |
| • | Testing and validation of models will be difficult, since there are so many variables and combinations of objects involved. ''Provably correct" modules will have to be developed, so that large models can be constructed that are free from error. Tools to help build models will be required. |
| • | Model security, to protect a model from unwanted tampering or accidental incidents, to make the model accessible to users but protected from unwanted actions, is another area of concern. |
| • | The computational requirements for such simulations will be very high, perhaps requiring the use of supercomputers. However, traditional supercomputers have been configured for homogeneously parallel problems. The virtual factory simulation will require extensive parallel computation, but it will not be homogeneous. Therefore research is needed on new techniques to apply massive parallel processors to the simulation question. |
