Data-driven Complex Industrial Processes Monitoring And Fault Diagnosis

Modern industrial processes trend to be more scaled,more complicated and more refined with the rapid development of national economy.Once accidents occur in complex industrial processes,they will cause huge economic losses and even cause human casualties.Therefore,the reliability and safety of industrial production processes need to be improved urgently.It is an important way to improve the reliability of the system and reduce the risk of accidents to monitor the operation status of the industrial production process in real time,find out the possible fault state in time and diagnose the causes of the fault.For different data characteristics of complex industrial processes,the nonlinear and time-varying processes data are modeled,this paper focus on complex industrial processes monitoring and fault diagnosis methods.The main research contents are as follows:(1)The traditional multivariate statistical process monitoring methods are difficult to adapt to the time-varying characteristics of complex industrial processes,which can easily result in high false alarm rates in process monitoring or failure to timely alarm.This paper presents an adaptive sliding window recursive sparse principal component analysis method for the online process monitoring of time-varying industrial processes.Firstly,the feature information of normal process data space is extracted by sliding window,and the sparse principal component analysis is applied to the current window block matrix to construct the sparse principal component analysis-based process fault monitoring model.Then,the forgetting factor is adjusted in real time according to the similarities of adjacent windows to update the sliding window size adaptively,so that the sparse principal component process monitoring model can effectively track the time-varying process.Finally,the sparse load matrix of the sliding window is renewed recursively to update the process monitoring model dynamically.Process monitoring results of the numerical simulation system and the Tennessee-Eastman process show that the proposed method can effectively improve the fault detection accuracy and adapt to the online process monitoring of long process industries with time-varying processes.(2)Aiming at the common problems of the low accuracy for fault diagnosis based on multivariate statistical methods in complex industrial processes,a fault diagnosis method based on stacked sparse denoise autoencoder and bayesian network is proposed.This method effectively uses the sparse denoise autoencoder to robustly extract the feature information of nonlinear processes and implements fault identification by stacking the sparse denoise autoencoder with the softmax classifier model,which effectively improves the accuracy of fault classification.On the basis of fault identification,the contribution graph method is used to isolate fault variables with the highest contribution rates and make as the evidence nodes for faults in the bayesian network.The root cause of fault,as well as fault propagation pathway,are diagnosed by updating the bayesian network with evidence,which can effectively achieve fault isolation.Numerical simulation experiments and the TE process show that the proposed method can effectively improve the accuracy of fault identification for nonlinear process and can effectively achieve fault isolation.(3)Based on the actual engineering requirements,the proposed theoretical algorithm should be oriented to the application.The process monitoring and fault diagnosis software system is developed in this paper.In the test of TE process monitoring and fault diagnosis,good application effect was achieved,which provides strong technical support for intelligent automatic monitoring and safe production of industrial processes.