Fault Detection And Diagnosis Based On Kernel Forecastable Component Analysis

The security has gained great attention as the scale and complexity of industrial processes. As a result, data-driven based fault detection and diagnosis technique is bring forward.Forecastable Component Analysis(ForeCA) is a new dimension reduction technique and feature extraction method for multivariate time series. It discovers the potential forecastable information structure from large amounts of data utilizing a distinctive model-independent measure. ForeCA employs variable autocorrelation to extract dynamic time series characteristics and measures forecastability of observed variables. Since industrial processes are strongly dynamical of time series, ForeCA is an appropriate approach for fault detection and diagnosis.Based on the above considerations, the paper introduces ForeCA into fault detection and diagnosis. Specifically, the main contents of this dissertation are as follows:(1) KForeCA is proposed by combining ForeCA and kernel trick. KForeCA extends ForeCA to nonlinear field and is more suitable for real-world problems.(2) Fault detection based on KForeCA is modeled. Cumulative forecastable component contribution ratio is proposed. Then two statistics are constructed: the L2 statistic and the SPE statistic. Simulation results on the Tennessee Eastman(TE) process illustrate the effectiveness of the proposed method.(3) Small shift detection model based on MCUSUM-KForeCA is established. We analyze the influence of step size of MCUSUM to the effect of fault detection and compare this approach with the one based on KForeCA. Simulation results show that the proposed method is suitable for small shift detection.(4) KForeCA-SVM model is etablished for fault diagnosis. The diagnosis results of the approach are compared with SVM and KPCA-SVM under both conditions of single fault and multiple faults. Experimental results reflect excellent performance of the method for fault diagnosis of industrial processes.