7.9 Case-based reasoning in process supervision and diagnosis – selected issues

This section is devoted to brief presentation of potential applications of Case-Based Reasoning (CBR) in proces supervision and diagnosis. Case-based reasoning constitutes a selection of methodologies based on the idea of reusing existing knowledge and experience to solve new problems; the retrieved, existing knowledge must be selected and adapted to the specific case. In many situations concerning of complex systems and processes, not fully understanded and having no precise numerical models this kind of "working solution" may provide a practical, engineering approach to resolve problems and tasks of supervision.

An extended list of existing applications of CBR and description of existing CBR tools cane be found in literature; among other, the following positions provide vast lists and descriptions......In the following subsections we shall present only a brief outline of the topic related to process supervision and diagnosis.

 Potential applications of CBR in tasks related to supervision   The potential applications of CBR can be roughly classified into two basic classes of problems; these are:
 
  • classification problems,
  • synthesis problems
    •


    The first type of problems include tasks such as situation assessment, diagnosis, prediction, support of process control and planning. The second type of tasks include design, planning and configuration. Both of the groups can find potential applications in supervision.

    Recal that the generic cycle of CBR operates according to the following scheme:
     

  • retrieve a set of cases similar to the one under consideration; no direct matching is necessary, the selected cases may only partially resemble the current problem,
  • select the best one, e.g. closest to the current one with respect to some similarity measure, or satisfying some predefined conditions,
  • adapt the selected case so that it solves the current problem,
  • retain the solution for future use.
    •


    In classificatio tasks, the main point consists in efficient storage and retrieval mechanism. The basic problem is to classify the new case so that it can be compared to existing one. Efficient classificationn is based on flexible pattern matching procedures and indexing schemes.

    In synthesis tasks the main difficulty consists in construction of the design or plan. The main problem is to develop a working basic plan or design and then, using the recorded case knowledge, adjust it to specific conditions. Such tasks are more difficult that classification since they require skills based on creativity.

    In both types of tasks CBR can be combined with other AI and other techniques. The most frequently used include various types of search methods, information retrieval techniques from databases (especially soft matching mechanisms, e.g. fuzzy ones), various statistical methods (especially discriminant analysis), rule-based system methodology, machine learning techniques (especially inductive classification algorithms), neural networks technology and different knowledge representation formalism.

    Basic potential applications of CBR approach in supervision include, among other, the following specific tasks:
     

  • static situation recognition: the current state of the system may form quite a complex pattern, composed of numerical and symbolic data, while certain specific situations of interest can sometimes hardly be matched by strict formal description; in such a case instead of simple generalization checking, the current state can be classified with use of existing classified cases. Example applications include detection of potential abnormal situations, detection of dangerous situations, detection of faults, assesment of arriving at certain stage of a multi-stage process (especially in batch processes), etc.
  • dynamic situation recognition: as above, but the situation assessment may require monitoring and anlysis of a sequence of states over some period of time and anlysis of direction and degree of change of certain parameters. CBR can be especially practical in case of processes when such assesment requires extensive expertise in the domain, since the number of possible trajectories grows obviously fast with a number of states possible to occur and length of the observed sequence,
  • prediction of events/situations: this kind of application is an extension of the dynamic situation assessment; if a particular sequence of states is observed and classified, it can be used to retrieve typical "extrapolations" from existing similar cases. Typical applications may include prediction of dangerous situations, prediction of output (especially characterised in a qualitaive way) and its characteristics, and alarm generation,
  • selection of process control algorithm: again, for a recognosed specific situation, particular control algorithm may be required; suc an algorith may be selected on the base of similarity to existing retrieved cases and, if necessary, adapted/modified to fit the current case,
  • diagnosis: in fact, this kind of application is similar to situation recignition; complete diagnostic process may require more sophisticated infetence, specific for the current case, so as to provide detailed specification of the current diagnosis (e.g. which elements are out of order, the type of malfunctioning, all potential diagnoses, etc.); disgnostic process may also require methods for diagnoses verification and development of repair schemes,
  • control synthesis and planning: this kind of application consists in synthesis of control algorith (e.g. plan of actions) for achieving a new, specified goal or arriving at desired final state or following specific trajectory over some period of time. This kind of CBR application belong to synthesis tasks, and may be supported with search techniques and plan generation methods,
  • model adaptation: in a hierarchical control system of complex processes the control may be synthesized with use of explicit system model; if the model changes significantly over time, classical identification methods based on adjustment of parameter may become insufficient; in such a case selection of model structure and components may be based on retrieving similar behaviour pattern and corresponding model.
    •


    Note that in all cases successful application of the CBR methodology depends on existing recorded cases sufficiently similar to the one currently analised. The advantage of application of CBR in supervision consist in that case recording may be considered as inherent activity completing process monitoring, signal-to-symbol transformation and knowledge abstraction. Thus during the plant operation the base of cases can be constructde as a result of automatics or semi-automatic process. However, positive past experience must be present explicitly; in the other case the system would be able to suggest what situations and controls should be avoided and not to provide constructive solutions.

    The main advantages of using CBR technology in process supervision include:
     

  • recording successful cases for immediate reuse; this include elimination of human errors and assuring consistent, stable operation over long time horizon,
  • preserving and redistributing the best know-how coming from selected domain experts,
  • transferring expertise from skilled specialists to the novice,
  • building a common, corporate bank of technological knowledge,
  • eventually, supporting discovering knowledge from examples.
    •


    In general, application of CBR technology can be considered as a crucial issue for improving reliability, safety and quality of supervision. However, CBR itself should not be considerd as an alternative to other knowledge-based technologies; it constitutes a valuable element to be composed with existing domain knowledge and technology so that the resulting gain is an effect of Synergism rather than competitivity.

     Selected applications of CBR in supervision related tasks   In this sebsection a brief listing of a sellection of existing successful applications of CBR tools and techniques for taks related to supervision is presented. Some most interesting ones include the following:
     
  • Clavier - A CBR system for autoclave configuration at Lockheed. Production of certain composite materials requires following a precise schedule of temperature and preassure change over a period of 6-7 hours. The process is carried out with use of autoclave and the process is monitored with use of thermocouples attached to the part. The process operator can make small adjustments of the temperature and pressure so as to follow the desired process pattern and assure the requested final quality. For a single cycle many parts of different shape and size are placed together inside the autoclave. The design of correct layout is an art rather than science, and requires much experience and expertise. The operators used to rely on former successful layouts. As the parts changes over time, the operators had to select a successful layout matching as closely as possible the current selection of parts and adapt it to the specific situation. The CLAVIER system was developed in the period 1987-1990 to assist the operators in the task of layout design. It performs tasks such as case retrieval and adaptation using appropriate representation of layout case. This kind of application can be classified as decision or control support, and from the point of view of AI methodologies it classification and specialized limited planning support.
  • Wayland - a CBR advisory system on the setup of aluminium pressure in die-casting machines at Kaye Presteigne. The pressure die design problem consists in selecting appropriate process parameters for for the process of injecting molten metal into the mold. Machine settings are critical factor for the product quality, but they are also compromise between quality, cost, die life period, etc. As there are no generally applicable procedures for calculating the parameters, they must be determined based on the existing experience and partial qualitative and quantitative knowledge. The Wayland system was designed and implemented to automatically identify past dies with similar characteristics, adapt the die settings to cover for the existing differences and validate the solution with certain limits. The case representation is based on numerical values of predefined attributes, while case adaptation involves both use of numerical formulas (when accessible) and rule-based system for condition evaluation and warning generation.
  • CASELine - assistance in aircraft Boeing 747-400 fault diagnosis and repair at British Airways. Even temporal elimination of an aircraft from service may cause very large cost to the owner airlines. Continuous monitoring of airplaines during flight allows for immediate detection of faults and other abnormally functioning elements. Such faults can be detected automatically or by the pilots. They are they transferred by radio to the staff engineers who have to identify the cause and take decisions about repair. CASELine is a CBR system for case retrieval and identification of fault, as well as for assisting them in determine the procedureds having the highest likelyhood of success.
  • CASSIOPÉE - a CBR system for maintenance and diagnosis of the CFM 56-3 jet power units for Being 737 airplans. The system was designed to cover knowledge about failures which happend over long period of time with the specific type of the engine. The case base contains 23 000 cases which are characterised with about 80 attributes. The system performs case retrieval, inductive learning of decision trees, situation-dependent requests for additional information, connection to illustrated catalogue with 25 000 pictures, and other. The main goal of application was to significantly reduce the diagosis and repair time by assisting the engineers with quick access to large body of similar cases.
  • Process control for Naheola Mill. The Mill products are: tissue, board and processed pulp. As the input material softwood and hardwood trees pulverised into cheaps are used. The cheaps are fed into digesting batch process. The batch process is very costly and the output product is sensible to process parameters; it can be erratic at quite a high rate. A great deal of of effort was devoted to automate the monitoring process and achieve high level of computerised decision support so as to optimise the production. The CBR approach was applied at one of the production steps that involved the operation of Pulp Dryer. The goal was to maximise the continuous production and minimise the rate of change of the Kamyr Digester process. Around two hundred inputs were monitored and thousands of cases were recorded. The outcome is a decision whether to shut the process or not; it provides also a fuzzy alerting system providing early warning to the operator if something goes wrong.
    •


    A number of other atechnical pplications of CBR is further reported in the literature.