Autonomous anomaly detection and fault diagnosis.

Abstract

With the growing complexity of dynamic control systems, the effective diagnosis of all possible failures has become increasingly difficult and time consuming. Despite the progress made tip to date, many systems rely on limited diagnostic coverage provided by a diagnostic strategy which tests only for known or anticipated failures. To circumvent these difficulties and provide more complete coverage for the detection of all possible faults, two divide and conquer approaches are developed in this thesis. The first, relies on the use of Self-Organizing network (SON) for regionalization of the system operating conditions, followed by the performance assessment module based on Time-Frequency Analysis (TFA) and Principal Component Analysis (PCA) for anomaly detection and fault isolation. And the second, bases on Growing Structure Multiple Model System (GSMMS) and localized decision making for detecting and quantifying the effects of anomalies. Several application examples are provided to demonstrate the effectiveness of the proposed approaches. The results show that local decision making is largely invariant to variations in inputs under practical assumptions and performs better at extracting degradation indicative features. GSMMS, which combines Growing Self-Organizing Networks (GSON) with efficient cooperative learning for local parameter estimation, is proposed in this thesis for identifying general nonlinear dynamic systems. The Voronoi sets defined by network naturally partition the full system operation space into smaller regions where system dynamics can be modeled locally using relatively simpler local models whose parameters can be estimated through minimization of residual sum of squares. Several interesting theoretical aspects regarding the effects of cooperative learning and topology preservation of the self-organizing network on the properties of local model parameter estimation are reported in this thesis. Our work mathematically supports the heuristic that a good learning strategy for identifying the local model parameters is to choose a neighborhood function whose effective region is initially wider and is reduced gradually during learning. This way, one can achieve a higher convergence rate at the beginning of the learning process and a smaller bias at the end of the learning process.