Subspace Aided Data-Driven Fault Detection And Diagnosis In Dynamic Systems

Fault diagnosis are currently receiving considerably increasing attention in the engineering and academic research. As is well known, data-driven techniques are widely applied in the process industry for fault detection and isolation(FDI). Widely used data-driven methods such as artificial neural network(ANN), hidden markov models(HMM), support vector machine(SVM) are typical machine learning approaches developed primary from computer science. However, the multivariate analysis technique seems only efficient in dealing with dynamic processes in the steady state and at higher levels in large scale systems. In comparison, subspace identification methods(SIM),which have remarkably developed and become one of the main streams of the research in system identification, provide more efficient and powerful tools to investigate FDI issues in highly dynamic systems and control loops generally located at the process level.From the viewpoint of application, based on the subspace identification methods(SIM), the procedure from the rough data to the final implementation consists of three steps as followed: System identification, FDI system design and on-line implementation of the FDI systems. In addition, a model based FDI system contains two parts:residual generation and residual evaluation including threshold computation and decision making.The main computational tools are the QR decomposition and singular value decomposition(SVD). More accurate analysis and prediction,more precise control, and more reliable outcomes, which are conventionally obtained on the basis of cumbersome mechanism models, are expected to implement through data-driven methods. The major advantage of this method is that they can provide high FDI performance similar to the model based FDI methods but without a sophisticated system design on a fact that SIMs needn’t conduct model parameterization or nonlinear optimization as do other system identification methods, improving the efficiency and reducing the cost of fault detection.This thesis deals with data-driven fault detection for LTI systems. The basic idea is to directly identify the residual generator for fault detection instead of identifying the system’s mathematical model. By storing the input and output data, it is possible to retrieve certain subspace that are related to the system matrices of the signal generating state-space model, which complete the fault detection successfully and reduce the computational complexity of the algorithm. Secondly, by introducing the Akaike information criterion to identify system order, the paper compares the method with the traditional singular value decomposition method and analyze its advantages and disadvantages.Simulation studies on the benchmark of Tennessee Eastman process demonstrate the validity of the proposed approach.