Technical University of Munich, Department of Computer Science
MQM Paper: [Struss 97d]

Colloquium - Applications of model based reasoning

Institute of Electrical Engineers (IEE),

Savoy Place, London

17th November 1997

Model-based Diagnosis for Industrial Applications Peter Struss Technical University of Munich Dept. of Computer Science Orleansstr. 34 D-81667 München, Germany http://www9.in.tum.de/people/struss.html OCC'M Software GmbH Gleissentalstr. 22 D-82041 Deisenhofen, Germany http://www.occm.de

Powerful Systems for Automated Diagnosis - Urgently Needed

The task of diagnosing technical systems becomes increasingly challenging. Their complexity is growing rapidly, and so is the complexity of the task to identify and remove the causes of malfunctions. This holds for all types of systems, production plants, communication networks, transportation systems, but also for consumer products. At the same time, the potential impact of malfunctions grows considerably in terms of costs entailed, threats to human health, and destruction of the environment. In many cases, the capabilities of human operators, users of such systems, maintenance staff etc. do not catch up with this development. Modern ships perfectly illustrate these three aspects: complex systems with tremendous effects in case of an accident are operated by a dramatically reduced crew with limited technical skills. As a result, powerful systems that are able to support or automatically perform the diagnostic process are urgently needed.

The Hardware Basis Is Available

For many, if not most, devices that pose serious diagnosis problems, the computing power is already available or will be in the near future. PCs are or can be used within repair shops. Moreover, many devices carry their own processors, usually for control purposes. This is not only true for large systems, such as ships or production lines, but also for small devices and consumer products like photocopiers, cars or cameras. Shrinking in terms or size and price and expansion in terms of memory and speed allows to equip a broad range of technical systems with a hardware basis for diagnosis software to run on-board. Advances in networks and telecommunication open another possibility: solutions can be found even though the subject of the trouble may be quite distant for the available diagnosis knowledge: the mechanic in the workshop can be connected to a diagnosis server via internet, or a car can transmit data to a service center and receive instructions. So, everything is prepared for the exploitation of advanced computer systems for on-board, off-board and remote diagnosis, but:

What is Missing: the Advanced Diagnosis Software

Of course, there already exist on-board diagnostics, for instance, on control units of various car subsystems, but they deliver a symptom rather than a diagnosis; fault localization and identification starts with the symptom. Also, there are systems available for certain devices, again, such as cars, that guide a technician in diagnosis based on predefined decision trees. But all existing solutions fail to meet the challenge we discussed above. They are hand crafted software programs (or decision trees) targeted to very special types of devices or even individual ones. There are no diagnostics for the diesel injection system, but for EDC7.11 for an Audi type X, Variant Y, year 1996, etc. Producing the variants of diagnosis systems for all variants of the devices under consideration is a costly process, often not feasible because of the sheer number of variants, and too costly in cases where it might be feasible in principle. As a result, a fundamental, inevitable requirement for almost all industrial applications is

a solution to the variant problem. Approaches to diagnosis that require the input of a set of pre-defined associations between symptoms (or, more generally, observations) and potential faults do not provide a solution, because such associations have to be established for each variant. Decision tree-based systems as well as rule-based diagnosis belong to this class. At a technical level, what is needed to solve the problem, is a
methodology that allows to automate the derivation of links between symptoms (observations) and faults. With the perspective of diagnosis tasks as sketched above, the scope is not only expanded in terms of types of devices and their variants, but also with respect to the scenarios in which diagnosis is performed. In particular on-board diagnosis demands for
robust and flexible solutions that cover a broad range of external conditions, symptoms and faults some of which may be difficult or impossible to anticipate (e.g. multiple faults). Again, approaches assuming the existence of symptom-fault associations based on human experience fail to address this issue: their coverage is limited to exploiting known symptoms of known faults in well-experienced devices (in order to know the associations).

The Essence of Model-based Diagnosis

The technology at model-based systems addresses these issues and provides the solid foundation for the future generation of automated diagnosis systems. What enables humans to effectively and efficiently solve diagnosis problems even when they have to deal with new devices and unanticipated symptoms and, hence, cannot simply draw upon previous cases and experience? It is their capability to apply general, principled knowledge, namely knowledge about

the (intended) behavior of the constituents of the device (e.g. the components of a circuit) and
the structure of the device, i.e. how the constituents are connected for interaction.

Based on this knowledge, the diagnosis process is guided by reasoning about how to collect more information about the device (e.g. by probing, testing and direct inspection) in order to discriminate among competing diagnosis hypotheses. Model-based diagnosis systems work exactly the same way: They do not require a pre-specified set of symptom-fault associations, which will always be incomplete, limited, and expensive to maintain. Instead, they are based on an explicit representation of the principled knowledge that allows to derive such associations automatically. The crucial modules of a model-based diagnosis system are

a library of component behavior models, capturing the intended, correct behavior and possibly the behavior under relevant faults
a description of the device structure in terms of components represented in the library and their interconnections.
an algorithm for generating diagnostic hypotheses from a comparison of the model with observations of the actual behavior.

The library captures the knowledge specific for a particular application domain, say mechatronic car subsystem or electrical circuits; its model fragments can be re-used for the class of devices in this domain. The structure description is the (only) device-specific element. It allows to automatically generate a device behavior model using elements form the library (see Fig. 1). This is, besides the observations, the input to the diagnosis algorithm which is general and domain-independent and, hence, can be reused across several application domains.

Fig. 1: Generation of Diagnostic Systems

In a nutshell, a diagnosis system of this type works as follows:

Observations of the actual behavior are entered.
The device model (of the circuit behavior) computes conclusions about system parameters and variables (observed and unobserved).
If a contradiction is detected (i.e. conflicting conclusions / observations for a parameter or variable) the set of component modules involved in it indicates which components possibly deviate form their intended behavior. This can be determined by the diagnosis system because the device model has a structure that reflects the device constituents that may incorporate faults.
Diagnosis hypotheses are generated, i.e. sets of faulty components that account for all detected contradiction (fault localization).
Suggestions for useful probes and tests are generated. This is possible because the behavior model reveals where the diagnosis hypotheses entail distinctive features of behavior. This function can be used to reduce the costs in cases whose observations are expensive and to act as a filter when the amount of information is overwhelming (see e.g. [Beschta et al. 92])
In case models of faulty components behavior are provided , the same approach (checking consistency of a model with the observations) can be used to discard particular faults and to conclude correctness of certain components if the set of modeled faults is considered complete.

These key principles are fairly basic and straightforward. Nevertheless, they provide the foundation for diagnostic systems that are truly revolutionary in two respects.

A Revolution in the Technology of Diagnosis

From the description given, one can see that model-based diagnosis overcomes serious limitations of previous approaches.

Diagnosing new devices becomes possible without available experience with them. This holds because and as long as they are assembled form components whose behavior models are contained in the model library.
Unanticipated faults and, in particular, multiple faults can be dealt with. This is based on the fact that the basic algorithm requires only the specification of models of correct behavior and, hence, makes no assumption about particular faults.
New symptoms can be exploited, since any observation that manifests a deviation from normal behavior, even if encountered never before, will be detected as a contradiction and, hence, contribute to the generation of diagnosis hypotheses.

The diagnosis procedure guarantees completeness of the results with respect to the phenomena captured by the models, the faults modeled, if any, and the observations provided. This means that model-based diagnosis provides the foundation for future diagnosis systems that can take on the challenges we discussed in the beginning. An implication of this approach that is equally important concerns the way diagnosis software is produced.

A Revolution in the Software Process of Constructing Diagnosis Systems

As we pointed our earlier, the only device-specific element is captured by the representation of the device structure. Often, it will be possible to import it from CAD systems. From this information and the elements of the domain-specific model library, a device behavior model can be derived automatically. The general diagnosis algorithm operating on this device model forms a diagnosis system tailored to the particular device which means it is constructed without having to write a single line of code:

Diagnosis systems can be generated automatically rather than having to be programmed (see Fig. 1).

Since the diagnosis framework can be bought from the shelf, the entire work of programming and knowledge-base construction is in the establishment and maintenance of the model library. This is essential for solving the variant problem discussed above and, hence, for turning a desirable goal into a feasible one from a practical and economic point of view. Even if the actual runtime system is not a model-based one for some reason, model-based diagnosis provides useful tools for supporting programming of diagnostics, for instance for validation or as a starting point for the programmer who might want to violate the completeness of the model-based results in a controlled way. Also, the model-based technology can be used to generate the input to traditional types of diagnosis tools, e.g. by generating decision trees.

Another important advantage is that the generation of diagnostics requires only a "blueprint" of the device as opposed to a physical prototype and, hence, can be done concurrently with the design process.

The Current State of the Technology

After about 20 years of research, a number of prototypes that reflect practical applications have been constructed in the past few years, including examples from devices like digital and analogue circuits, power networks, photo copiers, helicopters, and coolant systems of power plants. Moreover, recently, emerging industrial applications become visible, mainly a demonstrators. The Tiger project (ESPRIT, see [Milne et al. 94]) included diagnosis of a gas turbine fuel system. Model-based diagnosis of car subsystems (both on-board and off-board) is a target of two projects: "Vehicle Model-based Diagnosis" (VMBD, a BRITE-EURAM project, see http://www.aber.ac.uk/~vmbd/) and INDIA (a project funded by the German government, see http://lki-www.informatik.uni-hamburg.de/~india/welcome.html). NASA will equip the "New Millennium" space crafts with model-based systems as basis for increased autonomy through self-diagnosis and self-reconfiguration.

The first commercial tools are now available, such as rodon (R.O.S.E. Informatik GmbH) and RAZ'R (OCC'M Software GmbH, see http://www.occm.de). It has to be mentioned that the scope of current successful applications and tools is still restricted. Industrial processes (as opposed to devices with a clear component structure), systems containing embedded software, and occurrence of structural faults may pose problems for which research is continuing and no general of-the-shelf solution may be available. Also, providing the necessary model library is obviously a crucial and potentially extensive task. Even in cases where mathematical models are used in the engineering disciplines they may require re-structuring and transformation to be of utility for model-based diagnosis. Systems supporting or automating this model acquisition process are still limited. All this means that availability of expertise in the field of model-based systems is still important both for selecting applications that are tractable for the state of the art and for designing the solutions. MONET ("Model-based systems and qualitative reasoning", a network of excellency in the ESPRIT program, see http://MONET.aber.ac.uk/ ) is dedicated to the promotion of technology transfer in this area. In summary, industrial exploitation of model-based diagnosis technology is in its starting phase, and companies that appreciate the tremendous impact of the technology and invest in its exploitation will have a significant advantage over their competitors.

Beyond Diagnosis

It is fairly obvious that the model-based technology is relevant nor only to diagnosis. In fact, much of the knowledge captured by a domain model library is not specific to the task of diagnosis. Knowledge about the behavior of components is equally fundamental to other tasks, such as design, repair, or re-configuration of a device. Failure modes and effects analysis (FMEA) is a task which can be supported by systems that automatically analyze the impact of component failures based on a behavior model (see e.g. [Struss/Malik/Sachenbacher 96]). The same models might be used for test generation, design for diagnosability, and training purposes. The model library can be considered as a repository of corporate technological knowledge. Based on an explicit and coherent representation, this knowledge can be shared across different phases of the life cycle of a product, either by humans or by automated model-based problem solvers (see Fig. 2).

Fig. 2: The Corporate Technological Knowledge Base

The impact of fully exploiting this technology will be extremely important to

improving and speeding up communication and knowledge sharing between different work processes and the people involved,
making corporate knowledge accessible independently of persons, time, and location.

Thus, model-based systems will be a major contribution to improve the mastering of the product life cycle.

References

[Beschta et al. 93] Beschta, A., Dressler, O., Freitag H., Montag, M., Struss, P.: A model-based approach to fault localization in power transmission networks. In: Intelligent Systems Engineering, Vol. 2, No. 1 , pp. 3-14, 1993.

[Milne et al. 94] Milne, R., et al.: TIGER: Realtime Situation Assessment of Dynamic Systems. In: Intelligent Systems Engineering, Vol. 3, No. 3 , pp. 103-124, 1994.

[Struss/Malik/Sachenbacher 96] Struss, P., Malik, A., Sachenbacher, M.: Qualitative Modeling is the Key to Automated Diagnosis. 13th World Congress of IFAC, San Francisco, CA, USA, Pergamon, 1996.

More Literature

[Dressler/Struss 96] Dressler, O., Struss, P.: The Consistency-based Approach to Automated Diagnosis of Devices. In: Brewka, G. (ed.), Principles of Knowledge Representation, CSLI Publications, Stanford, ISBN 1-57586-057-0, pp. 267-311, 1996.

[Hamscher/Console/deKleer 92] Hamscher, W., Console, L., de Kleer, J. (eds.): Readings in Model-based Diagnosis, Morgan Kaufman Publishers, Mountain View, 1992.

[Struss 97] Struss P. Model-based and qualitative reasoning: An introduction. In: Annals of Mathematics and Artificial Intelligence 19 (1997) III-IV, Baltzer Science Publishers, pp. 355 - 381, 1997.

If you have comments or suggestions, email me at mauss@in.tum.de
created: Jakob Mauss, November 28, 1997, last updated: Jakob Mauss, December 4, 1998