Research On Complex Chemical Processes Fault Detection Based On Deep Neural Network

Chemical Processes safety has always been one of the most important issues in the chemical industry,fault detection and diagnosis(FDD)is the most commonly used tools for abnormal events management(AEM)of chemical processes,which provides guarantee for process safety.Due to the complexity,nonlinearity,high-noise,non-Gaussian distribution and other characteristics of the modern process,the traditional chemical process fault detection method does not show excellent diagnostic performance,especially for some minor disturbances.With the development of deep neural network(DNN)technology,many deep learning algorithms have been proposed.However,these algorithms are rarely applied to FDD of chemical processes.DNN is a powerful tool for feature learning,because it can mine the information hidden in the raw data deeper than the shallow neural network,Therefore,it is a new idea that deep learning algorithms are applied to fault detection and diagnosis of chemical process.In this paper,firstly,the severe nonlinearity of complex chemical processes was considered,and a sparse filtering feature learning based novel fault detection method for complex nonlinear chemical processes is proposed.This method uses unsupervised learning through sparse filtering,and features are adaptively learned from the raw data of the chemical processes.Then the features are input into the logistic regression model,and the process operation status is classified with a supervised manner.The method is validated by Tennessee Eastman(TE)process,and the monitoring results show that the proposed method has good diagnostic performance and can diagnose faults effectively and timely.At the same time,in order to further improve the performance of fault detection,aiming at the non-linear and high-noise characteristics of chemical process,this paper has developed a novel method of chemical process fault detection based on stacked denoising sparse autoencoder with particle swarm optimization(PSO-SDSA).The layer-by-layer training method of DNN can learn the information of nonlinear process data adaptively during feature learning.At the same time,noise reduction strategy is added to enhance the robustness of the algorithm.The fault detection model is also more accurate.The TE process was used to evaluate the diagnostic performance,the experimental results show that the proposed PSO-SDSA method has excellent fault detection performance,The average fault detection rate(FDR)can reach as high as 83.48%,and the false alarm rate(FAR)is only 0.21%.And than the FDR have greatly improved,the PSO-SDSA method can detect faults more timely and effectively.Finally,the two new methods developed in this paper are applied to the fault detection of cyclohexanone production process from a petrochemical,and the fault detection performance of each method is compared.The PSO-SDSA method is selected as the fault detection method for cyclohexanone production process,the method has excellent fault detection performance for the cyclohexanone production process,and FDR can reach as high as 93.89% and FAR as low as 2.59%,so that the faults can be more effectively detected and it ensures the stable and safe operation of the petrochemical production process.