Identification Of Abnormal Working Conditions In Distillation Process Based On Deep Learning Networks

Identifying abnormal working conditions for distillation process benefits preventing and controlling distillation accident at the beginning,which also is one of key support technologies in keeping distillation process operate with states of “safety,stability,long term,full capacity,and optimum”.As a relatively novel method for data analysis,deep learning network can extract features from data automatically,without relying on manual experience.Therefore,it is necessary to use the methods based on deep learning network to achieve automatic diagnosis for abnormal conditions in distillation process,reducing the risk of manual misjudgment.This paper first constructs a Generative Adversarial Networks(GAN)to preprocess the abnormal data using Convolutional Neural Networks(CNN)as a generator and Deep Auto-Encoders(DAE)as a discriminator.The impact of GAN hyper-parameters on reconstruction result is discussed with similarity between the reconstructed data set and real data set as the measurement index.Then,a fault identification model combining Spearman's Rank Correlation Coefficient(SRCC)and Deep Brief Networks(DBN)is constructed.Finally,three different feature variable selection methods including SRCC,Pearson Correlation Coefficient(PCC)and Mutual Information(MI)are compared for the impact of fault identification models on training time and diagnostic accuracy.An industrial ethylene thermal separation distillation process is applied to test the proposed GAN and SRCC-DBN models,with the model verified by the the Tennessee Eastman(TE)process at first.The results show that the method based on GAN and SRCC-DBN models proposed in this paper can effectively achieve the purpose of identifying abnormal conditions in distillation process,with the accuracy rate not less than 94%.