Research On Fault Diagnosis Method Of Pneumatic Control Valve Based On Deep Learning

Pneumatic control valves are widely used in chemical,refining,steel,metallurgy and other large industrial scenarios.If the control valve failure,there is a high risk of significant human or property losses,causing great damage to the environment.However,the traditional manual detection of control valve consumes a lot of time and labor costs.At the same time,the use of mechanistic models and signal processing methods require staff to have rich experience and knowledge,with high staff training costs.Therefore,it is necessary to use intelligent and automated fault diagnosis methods for pneumatic control valves.With the rapid development of deep learning technology and computer storage technology,more and more researchers have started to use deep learning technology for fault diagnosis research of control valves.However,there are still some problems in this field,such as inconspicuous representation of single sensor features,insufficient training data for deep learning models and poor cross domain generalization ability of models.A dual-stream attention fusion fault diagnosis model is proposed to address the problem of inconspicuous representation of single sensor features.The model acquires fault representations of pneumatic control valves from different perspectives and extracts the fault features acquired by two different sensors separately using dual-stream networks.Secondly,a weighted fusion of the two extracted features is performed according to the contribution of different data types to the diagnosis results,thus improving the diagnostic accuracy.The experimental results show that the network can effectively improve the diagnostic accuracy compared with the direct fusion network and the fault diagnosis network using single sensor data.An improved ACGAN model is proposed to address the problem of insufficient training data required for deep learning models.The model uses a deep convolutional structure instead of the original network structure and uses the Wasserstein distance instead of the original loss function.In addition,a separate classifier is constructed to improve the stability of training and the quality of sample generation.The experimental results show that the proposed model can generate higher quality samples and improve the training accuracy of the model compared with the original model.Finally,the SSIM image quality index is introduced to further filter the samples generated by the proposed model,which further improves the diagnostic accuracy of the model.To address the problem of poor generalization of fault diagnosis models across domains,a WGDANN fault diagnosis model transfer network based on domain adversarial is proposed.First,the model is pre-trained using the source domain data to make the classifier with initial discriminative ability.Second,the source domain labeled data and the target domain unlabeled data are mixed into the WGDANN model,and the discriminator is trained against the feature extractor with the help of the weighted gradient inversion layer to improve the domain common feature extraction capability.Finally,the JMMD distance is used to narrow the joint distribution of features between domains,thus improving the classification performance of the classifier.The experimental results show that the proposed model can effectively solve the problem of insufficient accuracy in cross domain diagnosis and achieve good cross-domain diagnostic results under different gas source pressure conditions.