Model- and data-driven approaches to fault detection and isolation in complex systems

Abstract

Complex engineering systems pervade every aspect of our daily lives. The need for increased performance, safety, and reliability of these systems provide the motivation for developing system health monitoring methodologies for these systems. It is important to develop diagnosis for nonlinear systems that are robust to model uncertainties and noisy measurements. For greater applicability, it becomes essential to extend these diagnosis methods to cover hybrid behaviors, and to apply in a distributed manner to complex systems. On the other hand, the lack of sufficiently rich and complete models for these complex systems, but the availability of data collected from system operations may indicate the need to move toward data-driven diagnosis approaches. The contributions of this thesis research are primarily categorized into: 1) model-based, and 2) data-driven diagnosis. For robust model-based diagnosis, we have developed sensitivity analysis methods to quantify the performance of residuals generated for fault diagnosis in the presence of noise and uncertainty. We combine the robustness analysis with efficient residual selection algorithms to find a residual set that meets pre-specified diagnostic criteria. A second contribution of this thesis in the model-based diagnosis domain is developing two general approaches for distributed fault detection and isolation. The first method guarantees minimum communication of measurement values among subsystems. The second algorithm is computationally more efficient and it does not use the global system model for designing the local (distributed) diagnosers. However, it does not guarantee the minimum communication of measurement values. Our next contribution is developing model based methods that combine mode detection and diagnosis of faults in hybrid systems. Our approach does not require pre-enumeration of all possible modes and, therefore, it is feasible for hybrid systems with large number of switching elements. Finally, towards our contributions to data-driven diagnosis, we combine the use of unsupervised learning techniques with expert knowledge to develop an anomaly detection method to discover previously unknown faults in a complex system. Our clustering algorithm helps determine regions of nominal behavior, and by extension outliers and anomalous groups that are well-separated from the nominal clusters. We derive significant features associate for each outlier group, and then consult experts to identify and characterize these anomalies as faults in system behavior.