![]() |
|||
![]()
|
![]() |
![]() |
![]() |
3. METHODOLOGIES, TECHNIQUES, AND PRACTICESIn building an expert system, knowledge acquisition or elicitation and knowledge representation are key issues. In knowledge acquisition, conventionally, the knowledge engineer interviews a domain expert and asks him to talk through a number of examples of problems and their solution. Knowledge representation is to develop computer-based models of those aspects of a problem that are not naturally amenable to numerical representation or that can be more efficiently represented by the knowledge and procedures used by a human expert. Many different approaches, such as first-order predicate calculus, semantic networks, production systems, and frames, have been attempted in the manufacturing and engineering fields. The knowledge acquisition, representation, and inference or decision-making in a "design"-type expert system are more complex and difficult than a "diagnostic"-type expert system. Although expert systems have been applied in many fields, the acquisition of knowledge is a time-consuming task and may be a bottleneck due to the gap between experts in a field and expert system builders. Furthermore, even the experts themselves may find it difficult to formalize their expertise. The advent of neural networks has made it possible to acquire knowledge from a sort of data rather than from human experts. Expert systems with neural networks are called neural expert systems. The application areas of neural networks are divided into two main categories: general information processing and processing of sensory information (Monostori and Barschdorff, 1992). The former category concerns the "design"-type expert systems, and the latter concerns the "diagnostic" type. In the "design"-type expert systems, neural networks with learning abilities substitute for time-consuming simulation of real industrial processes. In this section, such knowledge-based systems in manufacturing and engineering fields are briefly illustrated. 3.1. INTELLIGENT PROCESS CONTROLExpert systems in intelligent process control include the computer-aided design and online adaptive adjustment of parameters of conventional feedback control algorithms such as PID control and optimum control. In a typical prototype with a self-tuning regulator (Astrom, Anton, and Arzen, 1986), the expert system was composed of several software parts. The expert system part was implemented using a production system framework, OPS4, and other parts such as user interface part, real-time control part, and action parts in rule expressions were written in Lisp. The rule base contains the production rules, which are typically described as: if <situation> then <action>. The data base in the knowledge base is the repository of facts, evidence, hypotheses, and goals. The facts would include static data concerning sensor measurement tolerances, operating thresholds, constraints on operational sequencing, etc., and parameter and state estimates based on noisy observations are included in the database as evidence and hypothesis. The inference engine repeatedly finds all the rules that are matched. It selects one of them and performs its actions. When no more rules are found, the system waits for incoming messages from the user interface part or the low-level control part. A message is either a new element to be entered into working memory or a Lisp function that should be evaluated. New memory elements may cause new rules to match and thus start the rule execution again. The total expert system can be efficiently executed under a real-time multitask operating system with mailboxes for message passing between parallel processes. Since there is no difference between data and program in Lisp programs, it is simply implemented for the user to do with the expert system and to add, delete, and edit rules online by a message of an arbitrary Lisp expression. Experience from building expert systems for real applications has shown that their power is most apparent when the problem considered is sufficiently complex. Process control problems with multiple loops under unpredictable material variations are admittedly complex. In the area of online real-time process control, use of an expert system development tool, such as G2 by Gensym Corp., where the domain knowledge can be embedded in the knowledge base, reduces the burden of plant engineers. Currently, many intelligent process control systems are being developed, utilizing neural network modeling, to monitor process outputs, to correct the control inputs, and to keep the process output within the desired range in the real process. Neural networks model the process output from the real input data by the feedforward error back propagation algorithm for learning. To search for the appropriate choice of control input, which makes the model output closer to the target output, the neural network model can be employed with an optimization technique such as the genetic algorithm. Using the neural network-based statistical process control method, the process shift due to abnormal cause can be detected. To compensate the process shift by abnormality, model modification and selection of the new control input are performed to recover the process output to the target output. 3.2. KNOWLEDGE-BASED ADAPTIVE CONTROL IN MACHININGTo enhance the effective range of present adaptive control strategies in machining, a knowledge-based adaptive control system was developed (Lingarkar, Liu, Elbestawi, and Sinha, 1990). The system can regulate the cutting force by adjusting the feed rate against unpredictable changes in the cutting process dynamics during a milling operation. The knowledge-based system is a part of the primary feedback loop, and the supervision level to monitor range violations is efficiently built into a frame-based knowledge representation scheme. A frame is composed of slots, and predicates attached to slots are used to encode the heuristics for adaptive control. These predicates behave as demons for proper functioning and maintaining the integrity of the controller. Hierarchical representation of frames allows sharing of information among frames by inheritance. Symbolic processing in Prolog and numeric algorithms written in C are combined by interface clauses within a rule written in Prolog. Although knowledge in a knowledge-based controller is application dependent, the fundamental structure of the controller can be easily applied to other manufacturing fields such as grinding, injection molding, and electrical discharge machining. For more complex controllers, where the sampling period is very small, use of parallel processors (such as INMOS transputers) or dedicated hardware to do symbolic processing is recommended, since chronological backtracking by Prolog is exponential in time.
|
![]() |
|
Use of this site is subject certain Terms & Conditions. Copyright (c) 1996-1999 EarthWeb, Inc.. All rights reserved. Reproduction in whole or in part in any form or medium without express written permission of EarthWeb is prohibited. Please read our privacy policy for details. |