Fault Diagnosis Based On Deep Learning Subject To Missing Data

Model-based fault diagnosis can provide guarantees for the safe and efficient operation of automated systems,but it is difficult for large complex systems to establish precise physical models.Therefore,datadriven fault diagnosis techniques are more favored by industrial process anomaly monitoring experts.Deep learning is an efficient data feature extraction tool,but the accuracy of deep learning based fault diagnosis methods depends on the quantity and quality of valid data.The data missing caused by random packet loss of network transmission or inconsistent sampling frequency of multiple sensors will inevitably affect the accuracy of fault feature extraction.Thus it cannot guarantee the accuracy of the deep learning based fault diagnosis model and the effectiveness of the diagnostic method.This paper focuses on the problem that deep learning cannot extract features from incomplete samples in the case of data missing.Effective data interpolation techniques and DNN model updating techniques are proposed to solve the problem that traditional deep learning based fault diagnosis method is unavailable for missing data.The main innovations are as follows:(1)Random packet loss of network transmission can result in missing of online data.This paper first establishes a missing data imputation model of BP neural network based on historical data with complete structure,which can solve the problem that online structural incomplete sample cannot be transferred to the input of the well trained deep neural network.Then imputed online sample is a sample with complete structure.Thus on-line diagnosis can be implemented by using the imputed sample as the input of the well trained DNN fault diagnosis model.(2)Significant difference of sensor sampling frequency can result in the missing of both historical data and online data.Due to the reason that the number of historical samples with complete structure is very limited,the DNN based fault diagnosis model is not accurate enough.Therefore,it is necessary to propose a DNN fault diagnosis method with model updating.By building a BP neural network data imputation model updated with samples,the structural complete sample data set is gradually expanded,and the DNN fault diagnosis model can be gradually updated based on the imputed structural complete sample.Thus the accuracy of the DNN based fault diagnosis model in the case of data missing can be well improved.