Research On Fault Diagnosis Of Chemical Process Based On Deep Neural Network

In the field of modern chemical industry,not only the system structure is becoming more and more complex,but also the chemical process operations are interlinked.Therefore,how to quickly and accurately diagnose the types of faults during the operation of industrial systems has become a huge challenge.In the face of large-scale data with high-dimensional,time-varying,nonlinear and other characteristics in industrial systems,many traditional fault diagnosis techniques are difficult to adaptively extract a large amount of effective feature information from chemical process data,and do not consider process data.The time series feature problem.Based on neural network technology,this paper aims to efficiently and accurately extract chemical industry process fault types,and conduct in-depth research on process data sets under chemical conditions.(1)From the perspective of neural network structure,an end-to-end model fusion feature learning method based on deep bidirectional gated recurrent unit(MCNNDBi GRU)is proposed,which can be used for chemical process fault diagnosis through experiments.Firstly,by analyzing the different representation capabilities of spatial feature information and semantic information brought by convolutional neural networks at different scales,a feature-aligned multi-scale feature extraction model(MCNN)is designed.Secondly,in order to better extract the time series features in the process data,a deep bidirectional mechanism is proposed and applied to the recurrent neural network.This mechanism makes the recurrent neural network not only present the forward processing input features from the past to the future,but also the reverse processing from the future to the past.Synthesizing these feature processing improves the diagnostic performance of the network model.Finally,in order to verify the effective diagnostic accuracy of the proposed model for chemical fault diagnosis,simulation experiments were carried out on Tennessee-Eastman(TE)process and chemical coking furnace,and compared with several conventional network models.Experiments not only demonstrate the effectiveness of the model,but also confirm that the model is superior to other conventional neural networks in both diagnostic accuracy and feature robustness.(2)Starting from the optimization and improvement of the model algorithm,a deep network model for chemical process fault diagnosis named improved WDCNN is proposed.The model diagnosis process is divided into three parts.First,the data is initially extracted through WDCNN.The first-layer wide convolution kernel in WDCNN will automatically remove the features that are not helpful for diagnosis of the original signal.The subsequent convolutional network layers are set with multiple layers of small convolution kernels.The purpose is to increase the representational power of the model.Secondly,a two-layer LSTM neural network is set in the middle part of the model,in order to extract the time series features in the original data.Finally,the traditional Softmax algorithm is replaced by the XGBoost classification algorithm,which optimizes the robustness and feature learning ability of the model in the face of complex chemical working conditions.Compared with the previously proposed model,the model is much simplified.The total number of neurons in the entire model is far less than the previously proposed model,the training time is greatly reduced,and the effect is much improved than before,which fully demonstrates the effectiveness of this method.