Janos Gertler

Survey of model-based failure detection and isolation in complex plants

Techniques to detect and isolate failures in complex technological systems, such as sensor biases, actuator malfunctions, leaks, and equipment deterioration are surveyed. The methods are based on analytical redundancy afforded by a mathematical model of the system. The main components of such techniques are residual generation using the model, signature generation by statistical testing, and signature analysis. Model-structural conditions for failure isolation are introduced together with transformation methods to implement them. Sensitivity and robustness considerations are presented, and a design framework based on model redundancy is proposed.<<ETX>>

A new structural framework for parity equation-based failure detection and isolation

Abstract The paper describes a new framework for developing parity equations that prevent incorrect isolation decisions under marginal size failures in a decision process that tests each residual independently. Test thresholds that take the noise conditions into account are set high to reduce the occurrence of false alarms while maintaining the algorithms ability to detect and isolate larger failures. The method is applicable to additive failures on the measured input and output variables and to additive plant disturbances. A transformation algorithm provides a multitude of models that satisfy the isolability requirements. A search procedure utilizing this model redundancy integrates model robustness considerations into the design.

Fault detection and isolation using parity relations

Abstract The design of dynamic parity (consistency) relations, for the detection and isolation of faults, is described. Both additive and multiplicative (parametric) faults are considered. Various isolation schemes are discussed and decoupling from disturbances and certain model errors is included. Links to diagnostic observers and parameter estimation are pointed out.

Optimal residual decoupling for robust fault diagnosis

his paper deals with residual generation for the diagnosis of faults in the presence of disturbances. The emphasis is on modelling errors, represented as multiplicative disturbances, and on parametric faults. These are both characterized as discrepancies in a set of underlying parameters. The residuals are obtained using parity equations. To address the situation when the number of uncertain parameters is too high to allow perfect decoupling, two approximate decoupling methods are introduced. One utilizes rank reduction of the model-error/fault entry matrix via singular value decomposition. The other minimizes a least squares performance index, formulated on the residuals, under a set of equality constraints. It is shown that, by the appropriate construction of the entry matrix or of the performance index and the constraints, a broad variety of structured and directional residual strategies can be implemented.

Analytical Redundancy Methods in Fault Detection and Isolation - Survey and Synthesis

Abstract The best known residual generation methods in model based fault detection and isolation, including parity equations, diagnostic observers and Kalman filtering, are presented in a consistent framework. The discussion is organized along two residual enhancement concepts, namely structured and fixed direction residual sets. It is shown that, once the design objectives are selected, parity equation and observer based designs lead to equivalent residual generators. Robustness in the face of modelling errors is addressed and partially robust residual generator algorithms based on multiple model variants and on partial parameter insensitivity are reviewed.

Model-based on-board fault detection and diagnosis for automotive engines☆

Abstract This paper desribes an algorithm developed for the online detection and diagnosis of faults in automobile engine sensors and actuators, using the on-board microcomputer. The algorithm is based on the structured parity equation methodology. The parity equations are derived from an engine model having linear dynamics and static nonlinearities, obtained by identification from simualtion experiments.

Generating directional residuals with dynamic parity relations

Abstract It is shown how diagnostic residuals, which exhibit directional properties at all times in response to an arbitrary mix of input and output faults, can be generated using dynamic parity relations. The parity relations are applied directly to the observables of the monitored plant; the design relies on the plants dynamic input-output model. This residual generator can be made computationally polynomial (moving average) by including the invariant zero polynomial of the fault system in the specified fault responses. The polynomial design yields moving average noise transfers as well. White noise transfer and disturbance decoupling are achieved by extending the response specification. The parity relation approach is compared with the traditional detection filter design, and is shown to be more straightforward and have milder existence conditions; if subjected to the same specification, the two approaches yield identical residual generators.

Sensor and actuator fault isolation by structured partial PCA with nonlinear extensions

Abstract Partial PCA based on principal component analysis (PCA) with ideas borrowed from parity relations is a useful method in fault isolation (J. Gertler, W. Li, Y. Huang, T.J. McAvoy, Isolation enhanced principal component analysis, AIChE Journal 45(2) (1999) 323–334). By performing PCA on subsets of variables, a set of structured residuals can be obtained in the same way as structured parity relations. The structured residuals are utilized in composing an isolation scheme for sensor and actuator faults, according to a properly designed incidence matrix. To overcome the limitations of PCA, nonlinear approaches based on generalized PCA (GPCA) and nonlinear PCA (NPCA) are proposed. The nonlinear methods are demonstrated on an artificial 2×2 system while simulation studies on the Tennessee Eastman process illustrate the linear method and some extensions.

Principal Component Analysis and Parity Relations - A Strong Duality

Abstract It is shown that the standard PCA is a subspace representation of the primary set of parity relations. It is possible then to define partial PCAs, as subspace representations of transformed parity relations. These inherit the structural properties of the parity relations, in that they are selectively sensitive to subsets of faults. With this, it is possible to design an incidence matrix for a set of such partial PCAs, resulting in a structure with the same fault isolation properties as a structured set of parity relations. While these properties are being deduced from parity relation features, the actual analysis is performed entirely in the PCA framework.

Augmented Models for Statistical Fault Isolation in Complex Dynamic Systems

The equation error approach to fault isolation implies the statistical testing of balance equation errors. In this paper, some substantial extensions to the existing methodology are proposed, including - generalized linear dynamic models - the concept of statistical isolability - the idea of and an algorithm for model augmenting - fault sensitivity analysis and filtering