Fault Diagnosis Method And Application Based On Deep Learning For Industrial Production Processes

Modern industrial production processes become more and more complex and their scale is constantly expanding.Fault diagnosis technology is of great significance for ensuring the safe operation of industrial production processes.The extraction of fault features is a very critical step in fault diagnosis.Traditional methods generally rely on expert knowledge and manual extraction.Deep learning,because of its deep network structure with multiple hidden layers,enables deeper mining of process data.Compared with traditional fault diagnosis methods,it can more accurately express the mapping relationship between features of the industrial production process data and fault types.Therefore,based on Deep Auto Encoder(DAE),this thesis studies and applies the fault diagnosis method on the industrial production processes which include continuous production processes and batch production processes.The main research work of this thesis is introduced as follows.(1)Based on DAE,feature extraction and visualization analysis are carried out for industrial production process data.The DAE is compared with the commonly used feature extraction methods such as Principal Component Analysis(PCA)and Kernel Principal Component Analysis(KPCA).The above two experiments demonstrate the superior feature extraction ability of DAE from both qualitative and quantitative aspects.(2)For the industrial production processes,a fault diagnosis model of the industial production processes based on DAE is established,and the influence of key parameters of the model such as the number of network layers and the number of nodes in the middle feature layer on the performance of the model is analyzed.The established fault diagnosis model is tested based on two Benchmark problems of continuous production processes and batch production processes.The test results show the effectiveness of the industrial production process fault diagnosis model based on DAE.(3)For the typical continuous production process of continuous annealing in the iron and steel industry,considering the characteristics of multi-classification and sample imbalance of the strip steel fault data in the continuous production processes,an improved ensemble SMOTE-DAE model is proposed.Experiment results based on actual production process data show that the ensemble SMOTE-DAE model has higher fault classification performance than other fault diagnosis methods.(4)For the typical batch production processes of semiconductor production,considering the characteristics of three-dimensional and batch unequal length of semiconductor production process data,an improved deep convolution autoencoder model is proposed,which uses convolutional layer to process three-dimensional data and uses spatial pyramid pooling to solve the problem of batch unequal length.Experiment results based on actual production process data show that the improved deep convolution autoencoder model has higher fault classification performance than the traditional algorithm.