Fault Detection, Diagnosis and Fault Tolerance Approaches in Dynamic Systems based on Black-Box Models

• Main Author: Brito Palma, Luis
• Publication Date: 2006
• Language: eng
• Source: Repositórios Científicos de Acesso Aberto de Portugal (RCAAP)
• Download full: http://hdl.handle.net/10362/66049
• topic: fault detection; fault diagnosis; fault tolerance; fault tolerant control; black-box models; dynamic systems; Domínio/Área Científica::Engenharia e Tecnologia::Engenharia Eletrotécnica, Eletrónica e Informática

In this dissertation new contributions to the research area of fault detection and diagnosis in dynamic systems are presented. The main research effort has been done on the development of new on-line model-based fault detection and diagnosis (FDD) approaches based on blackbox models (linear ARX models, and neural nonlinear ARX models). From a theoretical point of view a white-box model is more desirable to perform the FDD tasks, but in most cases it is very hard, or even impossible, to obtain. When the systems are complex, or difficult to model, modelling based on black-box models is usually a good and often the only alternative. The performance of the system identification methods plays a crucial role in the FDD methods proposed. Great research efforts have been made on the development of linear and nonlinear FDD approaches to detect and diagnose multiplicative (parametric) faults, since most of the past research work has been done focused on additive faults on sensors and actuators. The main pre-requisites for the FDD methods developed are: a) the on-line application in a real-time environment for systems under closed-loop control; b) the algorithms must be implemented in discrete time, and the plants are systems in continuous time; c) a two or three dimensional space for visualization and interpretation of the fault symptoms. An engineering and pragmatic view of FDD approaches has been followed, and some new theoretical contributions are presented in this dissertation. The fault tolerance problem and the fault tolerant control (FTC) have been investigated, and some ideas of the new FDD approaches have been incorporated in the FTC context. One of the main ideas underlying the research done in this work is to detect and diagnose faults occurring in continuous time systems via the analysis of the effect on the parameters of the discrete time black-box ARX models or associated features. In the FDD methods proposed, models for nominal operation and models for each faulty situation are constructed in off-line operation, and used a posteriori in on-line operation. The state of the art and some background concepts used for the research come from many scientific areas. The main concepts related to data mining, multivariate statistics (principal component analysis, PCA), linear and nonlinear dynamic systems, black-box models, system identification, fault detection and diagnosis (FDD), pattern recognition and discriminant analysis, and fault tolerant control (FTC), are briefly described. A sliding window version of the principal components regression algorithm, termed SW-PCR, is proposed for parameter estimation. The sliding window parameter estimation algorithms are most appropriate for fault detection and diagnosis than the recursive algorithms. For linear SISO systems, a new fault detection and diagnosis approach based on dynamic features (static gain and bandwidth) of ARX models is proposed, using a pattern classification approach based on neural nonlinear discriminant analysis (NNLDA). A new approach for fault detection (FDE) is proposed based on the application of the PCA method to the parameter space of ARX models; this allows a dimensional reduction, and the definition of thresholds based on multivariate statistics. This FDE method has been combined with a fault diagnosis (FDG) method based on an influence matrix (IMX). This combined FDD method (PCA & IMX) is suitable to deal with SISO or MIMO linear systems. Most of the research on the fault detection and diagnosis area has been done for linear systems. Few investigations exist in the FDD approaches for nonlinear systems. In this work, two new nonlinear approaches to FDD are proposed that are appropriate to SISO or MISO systems. A new architecture for a neural recurrent output predictor (NROP) is proposed, incorporating an embedded neural parallel model, an external feedback and an adjustable gain (design parameter). A new fault detection and diagnosis (FDD) approach for nonlinear systems is proposed based on a bank of neural recurrent output predictors (NROPs). Each neural NROP predictor is tuned to a specific fault. Also, a new FDD method based on the application of neural nonlinear PCA to ARX model parameters is proposed, combined with a pattern classification approach based on neural nonlinear discriminant analysis. In order to evaluate the performance of the proposed FDD methodologies, many experiments have been done using simulation models and a real setup. All the algorithms have been developed in discrete time, except the process models. The process models considered for the validation and tests of the FDD approaches are: a) a first order linear SISO system; b) a second order SISO model of a DC motor; c) a MIMO system model, the three-tank benchmark. A real nonlinear DC motor setup has also been used. A fault tolerant control (FTC) approach has been proposed to solve the typical reconfiguration problem formulated for the three-tank benchmark. This FTC approach incorporates the FDD method based on a bank of NROP predictors, and on an adaptive optimal linear quadratic Gaussian controller.