Research On Ensemble Fault Diagnosis Method Of Complex Process Based On LSTM

With the rapid development of science and technology,the industrial process is becoming more and more complex and varied.Any small fault will bring huge losses.Intelligent fault diagnosis technology is an effective method for processing industrial data.It can processes the original data quickly and efficiently.Data and provide accurate diagnosis results.In order to better extract the effective information in the data,this thesis proposes a combination of long and short-term memory(LSTM)networks with 1D convolutional networks and 2D convolutional networks in view of the time series and high-dimensional nonlinear characteristics of chemical process data.The ensemble fault diagnosis method,referred to as ELSTM.First,the LSTM network is used to effectively process the long-term dependence of the time series,and the output of the hidden layer obtained contains the timing information of the original data,and then the data is processed by 1D convolution.Finally,the 2D convolution network has a strong multidimensional The ability to extract features from the data,pass the extracted feature data through the fully connected layer to obtain the classification result.The main work is as follows:First of all,for the time series features of the extracted data,choose to use a long short-term memory network,which can efficiently extract the time series features contained in the data,effectively save the time information existing between multiple time series input at one time,and successfully capture the long-term Time characteristics.Then,the 1D convolutional network is used to further strengthen the sequence features extracted by the LSTM network in time to obtain more effective time features,and to mine the time features hidden in the data;at the same time,multiple 1D convolution kernels can generate multiple times.The feature sequence increases the number of feature channels,which has an effect on the dimension of the data;after 1D convolution,a nonlinear activation function is introduced,which greatly increases the nonlinearity of the network model,and improves information integration and information interaction.Has an important impact on the forecast results.Finally,the 2D convolutional network is used to extract the spatio-temporal characteristics of the data.The data in the experiment has high-dimensional complexity.The long and short-term memory network cannot obtain all the information of the data well.The 2D convolutional network can perform the previous time series again.Feature extraction,because the long and short-term memory network reads data at a certain step size,only limited data information can be obtained at a time,therefore,it is necessary to use a 2D convolutional network to further extract features from the data sample,the feature fusion method used Concatenate effectively fuses the temporal features and spatial features,and finally obtains all the features of the sample data for fault classification.The experiment in this thesis is based on the Tennessee Eastman process for verification.The batch normalization method is added in the training process.Compared with the method without this method,the network training speed is accelerated and the model training accuracy is improved.The ELSTM model and the LSTM model and the LSTM-CNN model are compared with the TE process fault data set for comparative experiments,and the model is analyzed from the aspects of test accuracy and training accuracy.It is verified that the fault diagnosis method based on ELSTM used in this thesis is higher than other models in classification accuracy.It proves the effectiveness of this method.