Research On Fault Diagnosis Method Of Complex Process Based On Deep Learning

In recent years,the process of industrialization has been rapid,and the production system has gradually become more complicated and intelligent.In the industrial process,the influence of each part is complicated.Industrial process is a complex process,which has the characteristics of nonlinearity,non-Gaussian,multiple operating conditions,large hysteresis,strong coupling and so on.If any one of the parts fails,the production will not continue.Based on the characteristics of the complex process,the single fault diagnosis method has certain limitations.With the advent of the era of big data,the amount of data generated by complex processes has expanded geometrically,and the fault diagnosis method of deep learning has been widely used due to its powerful batch processing capabilities.But according to the characteristics of the complex process,there is still a lot of research space.When analyzing complex process data,due to the collection,transmission,storage and other reasons,the collected data are missing,so that the fault diagnosis method based on deep learning cannot correctly analyze the information in the data.In the face of the above problems,in order to improve the fault diagnosis ability in the absence of complex process data,this thesis proposes a new collective fault diagnosis method based on the combination of variational auto-encoder and long-shortterm memory network(VAE-LSTM).This method first inputs the missing data into the VAE,sets the network parameters,selects the appropriate dimension of the hidden variable space z,and uses the encoding-z-decoding structure of the VAE to iteratively fill in until the error reaches the set standard or reaches the number of iterations.Output the filled data,and then input the data into the LSTM.LSTM is a variant of recurrent neural network(RNN).The concept of entry is added to the network structure to make the network learn the dynamic information of time series well.In order to improve the performance and generalization ability of LSTM network,batch normalization algorithm is introduced.The optimization algorithm of LSTM is Adam algorithm,which is fast in operation and small in memory.It is suitable for dealing with unsteady data and improving the convergence speed of network.In order to verify the effectiveness of the method for fault diagnosis,this thesis chooses to use the TEP model,and takes the root mean square error(RMSE)as the criterion for data filling similarity.By comparing the filling results of mean filling and multiple filling,it is verified based on VAE data fill the accuracy of TEP data.At the same time,a comparative experiment of mean-filled-LSTM,multiple-filled-LSTM and the proposed method is set,and the fault diagnosis accuracy curve verifies that the proposed method has high diagnostic accuracy and stability in the completion and diagnosis of missing data in complex processes.