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

Petroleum and chemical industries occupy a very important position in the national economy,but safety accidents frequently occur in the process of chemical production,causing great damage to personnel,property and environment.Therefore,the development of a reliable process monitoring system is of great significance to help operators deal with abnormal conditions in the production process,avoid serious safety accidents caused by the further development of abnormal conditions,and ensure the smooth operation of the production process.Although the fault diagnosis technology for chemical process has been produced and developed for more than 40 years,many industrial application problems still need to be solved before the fault diagnosis method can be fully popularized in the chemical production process.One of the main problems is that the process data is highly nonlinear and noise signal interferes with process monitoring.Traditional fault diagnosis methods are not adaptable and self-learning ability is weak in chemical process.Common pattern recognition methods are prone to over-fitting and gradient dispersion during model training.With the advent of the big data era and the continuous updating and iteration of computer software and hardware,the amount of training data available and the amount of computer computation are increasing.These two reasons have led to the rapid development of artificial intelligence represented by deep learning(DL)in the past decade.So far,deep learning method has been widely used in many fields such as computer vision,natural language processing,and has achieved unprecedented success.Deep learning models have good self-learning ability and show great advantages in dealing with non-linear problems.Therefore,this paper proposes to apply two deep learning models to chemical process fault diagnosis,and the application results in the experimental simulation platform demonstrate their effectiveness.Firstly,this paper proposes a fault diagnosis method for chemical process based on deep belief network(DBN).The method firstly carries out unsupervised learning through several stacked restricted boltzmann machine(RBM)to extract effective information of input data types,and then uses back propagation(BP)algorithm to fine-tune the whole network model with tag data as the output target.Finally,the model is successfully applied to Tennessee Eastman(TE)process,and the optimal parameter setting in the test process is given by means of step-by-step optimization test of parameter combination.In addition,this model can learn from the newly generated data samples and realize the continuous updating of the diagnostic system.Then,aiming at the problem that noise signals such as random errors and the like commonly exist in the data in the real chemical production process,which are easy to affect the fault diagnosis results,this paper proposes a method based on lifting wavelet-deep belief network(LW-DBN)on the basis of DBN model.The method firstly decomposes the original data by lifting wavelet,then reconstructs the obtained low frequency signals to obtain ideal data and applies the ideal data to the training of DBN model.In addition,this paper proposes a local noise reduction method to process online data one by one,and the application in TE process has achieved excellent results.Finally,considering that both the traditional method and the method proposed in this paper use process data discretely,and there is a high correlation between the time series data generated in the chemical production process in the local time domain.Due to its ability to extract local features from data,this paper combines the highly correlated features between the convolution neural network(CNN)model and the time series data in chemical process,and proposes a chemical process fault diagnosis model based on dynamic convolution neural network(DCNN).The model first sorts the monitoring data into two-dimensional samples according to the time series,then extracts its local features by the way of local connection between convolution kernel and input data,then uses pooling operation to reduce the dimension of data and save effective information,convolution operation and pooling operation alternate,and finally classifies through a full connection layer.The application results in TE process show that the fault diagnosis method based on DCNN can still show good diagnosis performance under less training times.