Fault Detection Of Dynamic System Based On Subspace Identification Aided Method

Fault detection technology plays an important role in improving system safety,which has been well developed for decades.Because of the advantages of high detection rate,low computational complexity and good adaptability to dynamic systems,the subspace identification aided method has gradually attracted more attentions of scholars at home and abroad.However,there are still some problems to be solved in the current research on this method,such as:(1)The off-line training of all normal data can only get a single fixed model,which makes it insensitive to the fluctuation of local small data.(2)The threshold obtained by chi-square distribution or kernel density estimation is a fixed value,which makes it difficult to take into account both fault detection rate and false alarm rate.(3)The data is trained to obtain a linear model,which has poor detection effect for Non-linear and Multi-operation points processes.In this thesis,the above-mentioned problems are studied.The main work includes:The principle of subspace identification aided fault detection algorithm is introduced.The characteristics of several existing subspace identification methods are analyzed and compared.The subspace identification aided based on principal component analysis(SIMPCA)fault detection method is adopted to detect faults.Then the process of fault detection are given,and the detection effect is verified on the Tennessee Eastman(TE)process and Intelligent Process Control Test Facility(IPC-TF)of Wuhan University of Technology.The results show that the SIMPCA method is superior to other methods.For the shortcomings of SIMPCA in off-line training model,a subspace identification aided fault detection method based on Principal Component Analysis with moving window(MW-SIMPCA)is proposed.By selecting an appropriate length of the window,the model is updated on-line according to the time series and can be sensitive to the fluctuation of local small data.Moreover,particle swarm optimization algorithm was used to optimize the window to be the optimal window,and the effectiveness of the method is verified on the TE process.In order to overcome the shortcomings of SIMPCA method,which is difficult to take both fault detection rate and false alarm rate into account by using fixed threshold,a method combining adaptive threshold and SIMPCA is proposed.The method adopts an adaptive threshold,which comprehensively considers the fixed threshold and the statistics of the previous period of time,so that the threshold can adaptively varied with time.Through the application of the TE process,the method has higher fault detection rate and lower fault alarm rate than the case of fixed threshold.A method of combining fuzzy C-means clustering with SIMPCA is proposed: FCM-SIMPCA.FCM is used to classify training data into several clusters models.SIMPCA is used to build models for each cluster data,and the weighted average of these clusters models is used for the actual process model.This method can more accurately establish a Non-linear and Multi-operation points process model.In order to verify the fault detection effect of FCM-SIMPCA method,the simulation experiment of three-phase flow equipment platform is completed,and the results prove the effectiveness of this method.