Fault Prediction Of Industrial Process Data Based On LSTM Model

With the rapid development of many industrial fields,its equipment systems are becoming more and more complicated and intelligent.Traditional methods such as ’ post-maintenance ’and ’ regular maintenance ’ will still cause huge losses to the life safety of staff and mechanical equipment.The reason is that such methods cannot accurately analyze and judge the status of equipment,systems and components,and it is difficult to meet the actual needs of modern times.In order to ensure the reliable and safe operation of the equipment during use,it is of great significance to solve the problem of fault prediction based on data-driven complex industrial processes.In view of this problem,on the basis of artificial neural network,a new fault prediction model based on multivariate time series data is proposed.Firstly,the concept and characteristics of time series are described in detail,which provides the basis for the following data-driven experimental methods.By adjusting the generator of the GAN network and combining the gated recurrent neural network with the autoencoder,an artificial sample data expansion model(GRU-BEGAN)for multivariate time series is established,which improves the quality of artificial samples and provides sufficient fault sample data for multivariate time series prediction model.Secondly,according to the expanded data volume and the need for time series feature extraction,this paper uses the Deep-Bi LSTM model based on recurrent neural network.This model can not only further improve the long-term dependence of the network structure under the premise of maximizing the spatial and temporal feature extraction of the sampled data,but also combine the relevant data information of the previous moment(forward)and the latter moment(backward)in the output layer,thereby improving the prediction accuracy and achieving the effect of multi-step prediction.In the prediction model,according to the actual needs of industrial processes,the loss function W-MAE of Deep-Bi LSTM is proposed,which can improve the data points of higher value for staff while the overall accuracy of the model remains unchanged or even improved.Finally,using the data expansion model and data prediction model proposed in this paper,the data of Tennessee Eastman,KJH simulation system and thermal hydraulic system of nuclear power plant are analyzed respectively,and the data prediction experiments of known transient conditions and unknown transient conditions are completed.By combining the method of single variable process monitoring(pauta criterion),the data prediction results are monitored,and the purpose of fault prediction is finally achieved.