552
Abstracts
037 Feedback Neural Nets for Supervision of Dynamic Processes B. Schenker, M. Agarwal, pp 225-230 Keen supervision of dynamic processes requires reliable prediction of measured outputs and unmeasured states. This work presents a feedback neural network structure that predicts measured outputs more accurately than the feedforward network does, and additionally provides estimates of unmeasured states as well. The superior performance of the feedback structure is demonstrated on simulation examples of a linear, a nonlinear, and a chaotic system.
038 Neural Network Approach to Fault Detection under Unsteady State Operation I. KoshlJima, K. Niida, pp 231-236 The interpretation and recognition of trend data is the key task in a diagnosis system. However, the lack of useful interpretation methods makes knowledge acquisition and its representation more difficult. This paper proposes a unified approach to fault recognition by applying artificial neural networks (ANNs). In the fault detecting system called "SkilTrecer", the ANN gives a standard measure to judge whether the operating condition is in a normal or abnormal state by learning normal trend data. An example of sutrtup operation at a chemical plant demonstrates SkilTracer's fault-detecting capabilities under unsteady state without any knowledge acquisition and representation of human activities, and dynamic modeling of the plant.
039 A Decomposition Approach to Solving Largescale Fault Diagnosis Problems with Modular Neural Networks J.A. Leonard, M.A. Kramer, pp 237- 242 Radial Basis Function Networks (RBFN) are many times faster to train than similar backpropagation networks, and in diagnostic applications, more robust and less prone to misdiagnosis of novel cases. However, for industrial-scale problems, RBFNs still require large computational resources. Two decompositions are proposed to reduce this work: a decomposition in time, and a decomposition among the fault classes. Temporal trend information is retained by integrating point-in-time ("snapshot") hypotheses, rather than presenting temporal information as a time-series input. This modular approach improves training speed, simplifies system maintenance, and allows a priori probabilities and misclassification costs to be incorporated explicitly.
040 On-line Fault Diagnosis of a Distillation Column using an Artificial Neural Network Sang Gyu Lee, Sun Won Park, pp 243-248 This paper presents the Time-Delay Neural Network (TDNN) approach for on-line fault diagnosis. The online fault diagnosis system finds the exact origin of the fault of which the symptom is propagated continuously with time. The proposed method has been applied to a pilot distillation column to show the merits and applicability of the TDNN.
041 Beyond Falcon: Industrial Applications of Knowledge Based Systems D.A. Rowan, pp 249-252 The Falcon cooperative research project was one of the
early projects to apply real-time, on-line, knowledgebased systems in the petrochemical industry. Since this initial research in the 1980s, the application of this
technology has matured and is now driven by business needs and the commercial development of the technology. Overall, two major trends have emerged since the Falcon project: 1) applications focus on solving specific problems with a high business impact ranging from process watchdogging to transition management, and 2) improved commercial tools and a standard system infrastructure, which have lowered some of the costs associated with fielding knowledge-based systems applications.
042 Robust Fault Detection for a Nuclear Reactor System: A Feasibility Study R.J. Patton, J. Chen, J.H.P. Millar, pp 253-258 This paper describes a feasibility study of a robust model-based method for nuclear reactor fault detection, designed to be insensitive to disturbances, yet highly sensitive to faults. The paper describes a robust faultdetection approach based on eigensa'ucture assignment technique. The approach has been applied to an llthorder model of a nuclear reactor system. A reducedorder model is used to approximate this system and modelling errors are considered as disturbances acting upon the model contained in the observer. A new method is used for estimating the disturbances' direction(s). Simulation results show that the scheme can detect soft or incipient faults efficiently.
043 Expert System for On-line Fault Diagnosis at Hojalata Y Lamina, S.A., Mexico C. Pfeiffer, R. Soto, M. Garcla, A. de Le6n, pp 259-262 This paper describes a methodology for alarm diagnosis in the direct reduction process (HyL), which a) learns how operators and engineers make diagnoses and b) designs an expert system based on production rules that emulates the diagnostic task. The diagnostic process has been separated into subtasks such as alarm identification, alarm filtering and trend computation. An on-line prototype of a small part of the process has been implemented using a "static shell" and "snap shots" of the process variables. The purpose is not to completely solve the problem, but to help the operators speed up their diagnostic decisions. The approach is generic and can be extended to other processes. 044 Operator Support System for Fertilizer Plant A. Mjaavatten, S. Saelid, pp 263-268 This paper describes an operator support system to be implemented in a compound fertilizer plant, to help minimize process upsets. This will be done by warning of process conditions that may lead to undesirable effects, and by assisting operators in tracing root causes of upsets. A combination of qualitative (rule-based) and quantitative (model-based) methods are used. The diagnosis is performed by a topological search from the detection point to the probable cause, directed by automatic checking of process streams for unacceptable deviations. With few exceptions, rules are independent of the specific configuration, thus minimising the work needed after process modifications.
045 On-line Supervision System for a Benzol Recovery Plant A. Kawashima, T. Sato, H. Ishiguro, S. Matsumura, Y. Naka, pp 269-274 This paper presents the development of on On-line Supervision System for a Benzol recovery plant (OSSB), which provides the guidance for optimizing control