Research On Dynamic Principle Component Analysis-Based Adaptive Fault Diagnosis Methods

With increasing demands for efficiency and product quality and progressing integration of automatic control systems in high-cost and safety-critical industrial processes, the field of supervision (or monitoring), fault detection and fault diagnosis plays an important role. The considered subjects are also known as fault detection and isolation (FDI) or fault detection and diagnosis (FDD), which are also called fault diagnosis for short.From the data-driven based fault diagnosis methods, the main researcher of this paper is dynamic principle component analysis algorithm based offline and online adaptive fault diagnosis process. The main work and contribution of this paper is as follows:(1) describes the status quo of the fault diagnosis, studies the classification of fault diagnosis methods from the viewpoint of qualitative analysis and quantitative analysis, studies the DPCA algorithm as well as its development progress, and in detail summarizes and analyzes the cause and effect of the dynamic of industrial process data (the auto-relation of the data).(2) studies the multi-dimensional wavelet denoising algorithm and adaptive principle component analysis algorithm, including Recursive Principle Component.Analysis (RPCA) algorithm, Moving Window Principle Component Analysis (MWPCA) algorithm and Exponentially Weighted Principle Component Analysis (EWPCA) algorithm.(3) proposes a fault diagnosis method based on wavelet denoising and DPCA algorithm. The method uses wavelet denoising to process augmented matrix data, improves the computational efficiency of the DPCA. Thought the simulation and study on the typical faults of Tennessee-Eastman Process (TEP) as well as the strip breaking fault in steel rolling process, verifies the effectiveness of the proposed method.(4) proposes an adaptive fault diagnosis method based on DPCA and MWPCA algorithm. The method introduces the ideas of RPCA algorithm, gives the simplified recursive formula of the correlation matrix, simplifies the original4times rank-one modification to a2times rank-one modification, shorts the computational time, improves the online update rate, and saves the hardware storage space. Thought the simulation and study on the typical faults of Tennessee-Eastman Process (TEP), verifies the effectiveness of the proposed method.