Mechanical Fault Diagnosis Based On Multi-source Domain Adaptation

Mechanical fault diagnosis is an effective method to ensure the high accuracy and reliability of equipment.Intelligent mechanical fault diagnosis has attracted the attention of researchers because it improves the speed and accuracy of diagnosis as well as reduces the cost of manpower and equipment maintenance.In practical engineering applications,due to the difference of data distribution under different working conditions,the trained model is not suitable for data under new working condition and the diagnostic accuracy is reduced.Based on transfer learning,this thesis studies the problem of mechanical fault diagnosis with multiple data sources under different working conditions,which is of great significance for fault diagnosis in practical applications.This thesis first introduces the current data-driven fault diagnosis methods,summarizes and analyzes the advantages and disadvantages of these methods.Based on machine learning and transfer learning,a Multi-source Unsupervised Domain Adaptation Network(MUDAN)algorithm that adapts to single or multiple data sources is proposed,and a mechanical fault diagnosis method based on MUDAN is designed.Aiming at the problem of different data distribution under different working conditions,combined with domain adversarial,a multi-source cross-domain discriminator and its corresponding loss function that adapt to single or multiple data sources are designed,which effectively improves the diagnostic accuracy and robustness of the method.Secondly,the transfer tasks of mechanical fault diagnosis in real scenarios is designed,and the experimental comparison rules are discussed in this thesis.Experiments are carried out on two bearing data sets,and the experiments results were analyzed in detail.The experimental results confirm that the proposed mechanical fault diagnosis method based on MUDAN can effectively perform domain adaptation under multiple sources,and achieve good diagnosis results.Under the same basic network framework,this method is compared with other single source and multiple sources domain adaptive methods,and it is found that the fault diagnosis method based on MUDAN has better accuracy and robustness.This method can effectively diagnosis the target domain data without the prior knowledge of experts and target domain labels,and has high fault diagnosis accuracy.