Research On Ensemble Techniques Of Fault Detection And Diagnosis

Fault detection and diagnosis (FDD) is an important means to ensure industrial safety, reliability, product quality and effectiveness. Traditional single fault detection and diagnosis techniques are only suitable for handling some simple systems, and are difficult to be directly extended to complex multi-variable, dynamic, stochastic industrial processes with time-varying, nonlinear, non-Gaussian, multi-models, fed batch characteristics. For these reasons, ensemble fault detection and diagnosis techniques, which combine several single methods, in particular, ensemble real-time fault detection and diagnosis techniques have become research hotspots and are full of challenge in recent years.To improve the accuracy, real-time and rapidity of ensemble fault detection and diagnosis, this thesis based on summary and analysis of existing research work at home and abroad, pays much attention on three key technologies of the ensemble real-time fault detection and diagnosis:data real-time de-noising technique based on lifting wavelet transform, incremental fault feature extraction technique, fault classification and clustering technique based on online machine learning. The main tasks of this thesis are:1) to do research on general structure of ensemble real-time fault detection and diagnosis;2) to present some improved real-time ensemble fault detection and diagnosis methods based on real-time data;3) to put forward some improved fast ensemble approaches of fault detection and diagnosis based on historical data.A typical structure of ensemble real-time fault detection and diagnosis, which is composed of "lift wavelet de-noising+incremental feature extraction+online fault classification and clustering" is presented in this thesis.To reduce the impact of noise on the performance of ensemble real-time fault detection and diagnosis, a real-time de-noising method which based on lifting wavelet bivariate threshold is proposed.To improve the accuracy, real-time and rapidity of ensemble fault detection, three ensemble fault detection approaches are proposed. Specifically, they are the improved real-time fault detection methods based on lifting wavelet and moving window PCA (LW-MWPCA) for stochastic time-varying systems, lifting wavelet and dynamic kernel PCA (LW-DKPCA) for dynamic nonlinear systems, and lifting wavelet and multiway PCA (LW-MPCA) for batch processes.To improve the accuracy, real-time and rapidity of ensemble fault diagnosis, three improved ensemble real-time fault diagnosis methods that are based on lifting wavelet and adaptive recursive least squares support vector machine (LW-ARLSSVM), or based on lifting wavelet and incremental probabilistic neural network (LW-IPNN), or based on lifting wavelet and incremental clustering (LW-ICLUSTER) are proposed for real-time data. In addition, based on real-time noise reduction by lifting wavelet, two improved ensemble fast fault diagnosis approaches including fast independent component analysis and adaptive recursive support vector machine (LW-FICA-ARLSSVM), fast independent component analysis and incremental probabilistic neural network (LW-FICA-IPNN) are presented for non-Gaussian processes.Theoretical analysis and experimental studies have shown that the performance of ensemble fault detection and diagnosis approaches proposed in this thesis are superior to that of single type methods of fault detection and diagnosis. This thesis provides a theoretical basis for the application of ensemble real-time fault detection and diagnosis techniques.