Application Of Machine Learning In Process Industrial Fault Diagnosis

In recent years,chemical enterprises have shown a trend of integration and largescale production in the production process.With the rapid popularization of distributed control systems(DCS)in the overall industrial field,the production process of the chemical industry has become more and more automated,fault detection and diagnosis(FDD)technology is also playing an increasingly important role for the safe production of chemical companies.With the rapid progress of communication technology and Internet of Things technology recently,FDD based on the data-driven is a hot research field.However,some fault diagnosis models have not strong generalization ability and they are not suitable for all kinds of sensor signal data in the chemical process,the fault detection rate is not high.Therefore,how to improve the practicability of the models is an important issue in the current research.To solve the problem,this paper studies the artificial neural network model in the field of machine learning and applies it to complex fault diagnosis problems in the process industry.The works and conclusions of research are as follows:(1)Firstly,the convolutional neural network(CNN)is studied in this paper,the one-dimensional convolutional neural network(1D-CNN)model can directly process signal samples without changing the one-dimensional characteristics of the data,which may be more suitable for processing such signal data.Therefore,a new1D-CNN architecture is proposed for FDD.The network architectures,including convolutional layers,pooling layers,fully connected layers,and various parameters are optimized in the proposed method.(2)In addition,this paper continues to study the Long Short-Term Memory(LSTM)and Gated Recurrent Unit(GRU)models of the Recurrent Neural Network(RNN)model.By analyzing the internal structure of them and the gate mechanism for processing data,it can be found that they have good learning capabilities for nongraphic data,for example,they can effectively process long-sequence data.(3)Several models proposed in this article are used in Tennessee Eastman process(TE process)to assess the performance of the model.The TE process is a standard process system engineering simulation platform specially used to evaluate process monitoring methods.To comprehensively reflect the performance of the model,three evaluation indexes are selected in this study: accuracy,F1-score,and fault detection rate(FDR).All the three proposed models can achieve good performance,1D-CNN model achieves 92.6%,92.6% and 92.6%,LSTM model achieves 92.2%,92.1% and92.1%,and GRU model achieves 95%,95% and 95%,respectively.The experimental results indicate that compared with other diagnostic methods,three models can remarkably improve the detection and classification for various working conditions in the TE process.