Birgit Köppen-Seliger

Current Developments in the Theory of FDI

Abstract In this paper current developments in the FDI theory are reviewed. Attention is focused upon the analytical approaches that make use of quantitative models, knowledge based approaches using qualitative models as well as approaches using computational intelligence techniques. Besides the description of different fault diagnosis methods, a number of prospective research topics are outlined.

New developments using AI in fault diagnosis

Abstract This paper is intended to give a survey on the state of the art of model-based fault diagnosis for dynamic processes employing artificial intelligence approaches. Emphasis is placed upon the use of fuzzy models for residual generation and fuzzy logic for residual evaluation. By the suggestion of a knowledge-based observer-like concept for residual generation, the basic idea of a novel observer concept, the so-called “knowledge observer”, is introduced. The neural-network approach for residual generation and evaluation is outlined as well.

Fuzzy logic and neural network applications to fault diagnosis

Abstract This contribution gives a survey on the state of the art in artificial intelligence applications to model-based diagnosis for dynamic processes. Emphasis is placed on residual generation and residual evaluation employing fuzzy logic. Particularly for residual generation, a novel observer concept, the so-called knowledge observer, is introduced. An artificial neural network approach for residual generation and evaluation is outlined as well.

Fault detection and isolation in technical processes with neural networks

In this paper a neural network based fault detection and isolation concept for supervision of technical processes is introduced. Three different possibilities to employ neural networks for fault diagnosis are discussed. In most existing schemes the steps of residual generation and residual evaluation are performed on the basis of analytical process knowledge. Here neural networks are proposed for these tasks. First of all a neural network can be used instead of a mathematical model for residual generation, secondly another neural network can be trained to perform the classification task for residual evaluation and therefore fault isolation. A third possibility is a one-step diagnosis (OSD), where one neural network is trained to directly detect and isolate possible faults from the available measurements without the need for prior generation of intermediate signals as residuals. Results from the application of a restricted Coulomb energy neural network (RCE) to the residual evaluation and alternatively to the one-step diagnosis at an industrial actuator benchmark problem are presented.

Modelling for fault detection and isolation versus modelling for control

The goal of this paper is to emphasize both the particularities of models needed for model-based fault detection and isolation (FDI) and the differences with respect to the models used in control. Of special interest is the question of complexity. This depends basically on the given situation such as the kind of plant, the kind and number of faults to be detected, the demands for fault isolation, robustness and the measurements available. However, in contrast to the wide-spread opinion that models for FDI have always to be more complex than those for control, the paper shows that diagnostic models for controllable and observable plants comprise only a partial description of the input/output model and are therefore less complex than those for control. This issue is discussed in terms of different model-based FDI approaches — analytical, data- and knowledge-based. As for the analytical approaches the necessary order of the diagnostic model is that of the transfer operator from the fault vector to the system output.

Neural Networks in Model-Based Fault Diagnosis

Abstract In this contribution neural networks form the basis for model-based fault diagnosis schemes. Three different possibilities to employ neural networks for supervision tasks are discussed. In most existing schemes the steps of residual generation and residual evaluation are performed on the basis of analytical process knowledge. Here, neural networks are proposed for these tasks. First of all a neural network can be Used instead of a mathematical model for residual generation, secondly a different neural network can be trained to perform the classification task for residual evaluation and thus isolate the probable fault cause. A third possibility is an one-step diagnosis (OSD), where one neural network is trained to directly detect and isolate possible faults from the available measurements without the need for prior generation of intermediate signals such as residuals. Results from the application of a modified Radial-Basis-Function (RBF) Neural Network to the residual generation and the Restricted Coulomb Energy (RCE) Neural Network to the one-step diagnosis at an industrial actuator benchmark problem are presented.

Modelling for Fault Detection and Isolation versus Modelling for Control

The main purpose of this paper is to emphasize the particularities of models needed for model-based fault detection and isolation (FDI) in contrast to the models used for control. Of special interest is the question of complexity of the model, which is of great importance for the practical implementation. This, of course, depends basically on the given situation such as the kind of plant, the measurements, the kind and number of faults to be detected and the demands for fault isolation and robustness. However, the paper shows that diagnostic models, in contrast to the wide-spread opinion that those have always to be more complex than the functional models for control, may be even less complex, because they are restricted to only those parts of the system in which the faults occur. The issue of model complexity is discussed in terms of different model-based FDI approaches analytical, knowledge-based and data-based. The ideas are illustrated in a case study, where several types of model-based FDI techniques are compared with the same plant, the amira three tank system.

Residual evaluation for fault detection and isolation with RCE neural networks

In this paper a neural network based residual evaluation concept for fault diagnosis is introduced. Based on the idea of emphasizing the evaluation part of a diagnostic concept in contrast to most existing schemes a restricted Coulomb energy neural network (RCE) is employed to classify residuals coming from a standard parameter estimation procedure. The classification aims at the detection and isolation of different faults in the process under supervision. The developed scheme is applied to an industrial actuator benchmark problem which was especially designed for the purpose of comparison of different fault diagnosis techniques. The presented results prove the capability of the presented scheme.

MAGIC – IFATIS: EC-Research Projects

Abstract In this contribution two European research projects funded by the European Commission in the framework of the IST programme are introduced. The goal is to present the basic ideas and concepts of the projects “Multi-Agents-based Diagnostic Data Acquisition and Management in Complex Systems (MAGIC)” and “Intelligent Fault Tolerant Control in Integrated Systems (IFATIS)” which both aim at the solution of diagnostic problems but in different respects. While MAGIC is based on a multi-agents-based architecture which integrates all levels in a diagnostic scheme, IFATIS deals with hierarchically structured fault-tolerant control for integrated systems. A wide range of industrial applications is expected for both projects.

A fault diagnosis concept for a high-pressure-preheater line

This contribution describes a combined analytical/fuzzy model-based fault diagnosis concept which has been applied to the high-pressure-preheater line of a power plant. The key idea is to divide the whole system under supervision into several subsystems and to employ all available analytical knowledge to generate residuals for each subsystem seperately. Thereby the necessary observers for fault detection are of a handable size. For fault isolation, the residual evaluation is done by applying qualitative knowledge about the fault effects on the residuals and about the interaction of the different subsystems. Here a rule base is evaluated using fuzzy logic. With this method a complex system can be supervised without the need for a complex analytical model of the whole system. Furthermore, the presentation of the fault isolation results leaves the final decision about a fault alarm to the human operator. Results from a power plant prove the successful application of the proposed supervision concept.