Research On Fault Diagnosis Method Of TE Process Based On Siamese LSTM

With the scale of the chemical production system becoming large,the chemical production process has becoming more and more complexity.Therefore,chemical systems are more prone to major safety accidents.In the field of chemical process fault diagnosis,in order to ensure that the chemical production process can operate safely and reliably,it is difficult to study the fault in time and accurately diagnose the fault.Due to the characteristics of high dimensionality,nonlinearity and dynamic timing of chemical process data,the traditional data-driven fault diagnosis methods are more difficult to apply in complex chemical processes.Therefore this paper proposes a fault diagnosis method on the Tennessee-Eastman(TE)process based on Long Short Term Memory network(LSTM),and proposes a fault diagnosis method based on Siamese Long Short Term Memory network(Siamese LSTM).In addition,For imbalances in the number of samples,this paper improves the method of Siamese LSTM.The main research work of this article is as follows:(1)Aiming at the fact that chemical production environment is complex and easily affected by various noises,this paper utilizes the method of wavelet threshold to reduce the noise of the original signal.After the noise reduction,the signal becomes more stable and can retain the effective characteristics of the original signal.(2)Aiming at the problems that the chemical process data has the features of high dimension,nonlinear,time series,and the problem that minor faults are difficult to diagnose due to the insignificant characteristics.This paper combines LSTM with Siamese network to propose a Siamese LSTM fault diagnosis method.LSTM which can deal the probelm of long-term dependence,and the siamese network which can amplify the small differences between similar samples.Through the experiments of fault diagnosis in the TE process,the average recall rate of the Siamese LSTM method proposed in this paper is96.98%,which is 10.81% and 3.48% higher than the commonly methods: RNN and GRU,respectively.Especially for the 3 types of tiny faults in the TE process,the fault diagnosis effect of this method has been significantly improved.It proves that the proposed method has greater advantages in fault diagnosis of TE process.(3)Aiming at the problems of imbalance data in the actual chemical process,this paper proposes a combined diagnosis method based on Synthetic Minority Oversampling Technique(SMOTE)and Siamese LSTM model.Based on the classic SMOTE method,this paper firstly improves the problem of unbalanced data distribution through data synthesis;Based on the Siamese LSTM model,this paper improves the cost function of the Siamese LSTM model to adaptively update the weight of the cost function,thereby improving the sensitivity of the network to small class samples.Experimental results show that this method can obtain a more robust diagnosis result under multiple failure scenarios and multiple unbalanced data ratios.