Research On Industrial Process Fault Detection Method Based On Independent Component Analysis

The contribution of the industrial processes to the national economy is very important.Because most of the industrial equipments work in harsh working environment,the probability of failure is greatly increased.It is rather significant for using of advanced fault detection technologies to detect the occurrence of fault in time and prevent further deterioration of the consequences.Compared with other methods of fault detection,fault detection method based on data driven can eliminate the problems of establishing of mathematical model or knowledge and rules of monitoring plants.This method can directly extract feature information form large amount of sensor data,and it also has good applicability and reliability.Data-driven industrial process fault detection has become a hot spot of research.Independent component analysis(ICA)is a classical method of data driven method.ICA can extract independent components hidden behind observation data and construct high-order statistics,which is more sensitive to fault information,this has attracted the attention of many researchers.However,the ICA method still has many deficiencies,such as: the robustness of the method is not ensured;optimization method easily falls into local minimum;the problem of selection of the threshold and poor performance for minor faults.This paper is devoted to solving the existing problems in the research of ICA method.The main contents include:(1)A method based on the combination of BBO algorithm and ICA is proposed,which is called BBO-ICA,to improve the robustness and accuracy of the traditional ICA algorithm.This method uses BBO instead of the classical Newton iteration method to achieve the independent component extraction.It solves the problem of traditional FastICA algorithm which is sensitive to the initial point and easily fall into the local minimum value.The effectiveness and superiority of the method is verified in the simulation platform of the DAMADICS process.(2)In order to prevent the determination of an unreasonable monitoring threshold,a support vector data description(SVDD)algorithm is used to combine with ICA: ICASVDD.The SVDD algorithm uses the super sphere contour the independent components projected into the high dimensional space,and the radius of the super sphere is used as the monitoring threshold.ICA-SVDD can overcome the problem that traditional kernel density estimation may lead to unreasonable monitoring threshold when data are too dispersed or centralized,and it also provides a new idea for threshold selection.The feasibility and superiority of the method are verified with TE process for the first time.(3)In view of the problem that the traditional ICA algorithm has a poor detection effect on small faults,a method of combining a typical variable analysis(CVA)algorithm with ICA is adopted: CVA-ICA.Firstly,typical variables are estimated by CVA and then independent components are separated by ICA.This method improves preprocessing capability of nonlinear data,which also improves the sensitivity and detection ability for minor faults.The effectiveness and superiority of the method is verified with TE process.Finally,this paper is summarized,some future research directions and problems still need to be solved are pointed out.