Research On Fault Diagnosis For Chemical Processes Based On Deep Residual Network

How to guarantee the normal operation of industrial processes has always been one of the most important research issues in various engineering disciplines.The fault diagnosis technology quickly locates the fault by diagnosing the fault,thereby clearing or isolating the fault,to reduce the loss caused by the fault,which is of great significance to the industrial process,especially the chemical processes with high risk and frequent fault occurrence.For chemical processes,due to its characteristics of coupling,non-Gaussian distribution,and nonlinearity,the diagnostic performance of traditional fault diagnosis methods such as principal component analysis and partial least squares is not good enough.And these methods require a large amount of empirical knowledge manually extracts fault features rather than automatic extraction.Therefore,it is necessary to study a fault diagnosis method with better diagnostic performance to further improve the safety and reliability of the chemical processes.Firstly,given the problems in traditional fault diagnosis methods for chemical processes,such as low efficiency of fault features extraction,low fault diagnosis rate,inability to automatically extract fault features and difficulty in processing massive process operation data,deep learning method is considered to use for fault diagnosis in chemical processes.Deep learning is a technology that has developed rapidly in recent years,it has an excellent performance in features extraction and can automatically extract fault features from the data generated by the chemical processes' operations with high efficiency,eliminating the redundancy and complexity of manually select features.Secondly,considering the training difficulty in deep learning model,a fault diagnosis method for chemical processes based on deep residual network(DRN)is proposed.DRN model is a kind of deep learning model generated based on the traditional convolutional neural network.It uses the shortcut connections of identity mapping to solve the problem of training difficulty in deep neural network and uses the new activation function of rectified linear unit and the method of batch normalization in the network,which can effectively alleviate the problem of vanishing / exploding gradients.In the face of a large number of process operation data,the DRN model also has efficient data processing ability.The Tennessee Eastman(TE)process,a typical chemical processes fault diagnosis experiment object,was used to evaluate the diagnostic performance of the proposed method.Compared with the previous TE process fault diagnosis methods based on traditional deep learning models,this method has a more excellent diagnostic performance.Finally,based on the former DRN diagnostic model,a fault diagnosis method based on improved DRN is proposed.The improved DRN model modifies the structure of the residual building blocks based on the former method,and the parameters and floating-point operations of the model are greatly reduced.Similarly,this fault diagnosis method for chemical processes based on improved DRN is applied to the TE process.The results show that compared with the former method,the training speed and diagnostic accuracy of this method are improved to a certain extent,showing more superior performance.