Research On Robust Fault Detection Method Based On ICA And Its Application In TE Process

Modern complex chemical processes are generally flammable,explosive and biologically toxic.If a fault occurs,the product quality will be reduced,and a very serious production safety accident will be caused.Therefore,in order to ensure the safety of the chemical process,fault detection and diagnosis techniques should be used to monitor it in real time.Because modern computer systems record large amounts of chemical process data through sensors,data-driven fault detection and diagnosis techniques have been extensively studied.Because there are a lot of non-gaussian data in the actual chemical process,the Independent Component Analysis(ICA)method can extract the non-gaussian features that represent the system and obtain a better detection effect,so this paper studies based on the ICA method.However,when ICA is applied to the fault detection task of complex chemical processes,there are some problems that affect the detection effect of the model.Therefore,this paper focuses on the problem of outliers and time-series faults problem,studies the improved model based on ICA,obtains the corresponding robust fault detection method,and conducts simulation experiments on the Tennessee Eastman(TE)chemical process.The main research work of this article is as follows:(1)Aiming at the problem that chemical process data is easy to contain outliers,which leads to the decline of ICA method detection ability,a robust fault detection method based on DPICA-SVM model is proposed.The traditional method of culling outliers based on threshold is prone to problems of excessive or insufficient culling,and may destroy the structural information of the samples.The DPICA-SVM model proposed in this paper is based on all sample modeling,which improves the robustness to outliers by decomposing preprocessing,extracting gaussian and non-gaussian features,and integrating learning of features.The experimental results show that the detection ability of this method is better than that ICA method in the case of no outlier and outlier.(2)Aiming at the problem that ICA is difficult to extract the time-series features of chemical process data,which leads to the instability of the time-series fault detection effect,a robust fault detection method based on ICA-ACVA model is proposed.The ICAACVA model proposed in this paper uses ICA to reduce the interference of data crosscorrelation,and then combines with Adaptive Canonical Variable Analysis(ACVA)to deal with auto-correlation and improve the ability to detect time-series faults by extracting time-series features.The experimental results show that this method has significantly improved the detection effect of various time-series faults.(3)The EL-ICA ensemble learning model is constructed to further improve the stability of fault detection.The existing ensemble learning methods mainly focus on ensemble the same type of ICA model,but lack of ensemble research on different improved models.This paper enhances the ability to deal with two types of problems by ensemble DPICA-SVM and ICA-ACVA two improved models with different ICA partial solutions,namely the EL-ICA ensemble learning model.The experimental results show that this model can deal with both outliers and time-series faults problem,and it has stronger detection ability and better effect than these single improved models.