Research On Fault Diagnosis Method Of Mechanical Parts Based On Transfer Learning

Mechanical equipment is composed of mechanical parts.As an inseparable single part,mechanical parts are an indispensable part of machinery and machine.The important influence of mechanical parts in machinery can not be underestimated.The whole process of production is continuous in modern industry,and the process equipment interact with each other,forming a unified and complete whole.In industrial production activities,once these core components fail,it will have serious consequences.Taking rotating machinery as an example,the faults caused by rolling bearings,gears,rotors and other mechanical parts account for 70% of all faults.Any fault of these mechanical parts may change the operation state of the whole mechanical equipment,resulting in equipment failure or safety accidents.Therefore,in-depth research on the state detection of mechanical parts and mastering the fault diagnosis method can effectively reduce the risk of mechanical equipment failure and improve the continuity,reliability and maintainability of mechanical parts.In order to effectively reduce the risk of mechanical equipment failure and enhance the continuity,reliability and maintainability of the normal operation of mechanical parts,this paper makes an in-depth study on the condition monitoring and fault diagnosis methods of mechanical parts(1)A method of mechanical parts fault diagnosis based on multi-channel convolution neural network is proposed.By combining multi-channel information fusion and convolution neural network technology,the fault diagnosis model of mechanical parts based on multi-channel convolution neural network is established.First,the data collected from different sensors are input into the multiple channels of one dimensional convolution neural network.Then multi-channel convolution operation is performed,and the features obtained by the pool layer are sampled and reduced.After the convolution layer and the pool layer,the whole connection layer and the Softmax classification layer are obtained.The extracted features are classified and the final fault diagnosis results are output.Finally,the case of gear fault diagnosis proves that the method is effective and practical.(2)A fault diagnosis method of mechanical parts based on transfer learning is proposed.This method is mainly aimed at the fault diagnosis method based on multi-channel CNN.The premise of its superior performance is to meet two basic conditions:(1)a large number of high-quality labeled data are used for model training to train a diagnosis model with superior performance;(2)the data to be tested and the training data must be combined They obey the same distribution and are independent.However,in the actual industrial fault diagnosis scenario,these two conditions are basically not tenable.A fault diagnosis method of mechanical parts based on transfer learning is proposed,which is used for cross domain fault diagnosis when the training data and test data do not meet the same distribution caused by different working conditions.Firstly,the source domain data and target domain data containing multi-sensor information are input into the feature extractor through multiple channels,and the depth feature of the data is extracted through the feature extractor.Then,the domain metric module is used to realize the difference degree,so that the feature has cross domain invariance.Finally,the fault classifier is used to realize the fault classification of the target domain unlabeled data.This method is verified by using gear fault diagnosis cases and bearing fault diagnosis cases under multiple working conditions.The results show that the method proposed in this chapter can effectively classify the target data set under different working conditions and one working condition has no label,and achieves good results.(3)In this paper,a mechanical parts fault diagnosis method based on ensemble transfer learning is proposed.Firstly,based on SAE and domain metrics,multiple sub learners SAEs with differences are established through different activation functions and training samples.Then,the source domain data features and target domain data features are extracted by SAEs to construct feature pool.Then,the feature sets in the feature pool are evaluated The first k feature sets with high evaluation value are selected according to the feature evaluation value,and multiple sample subsets are constructed by using the selected feature sets;finally,multiple softmax classifiers are trained by using the source domain data features in the sample subset,and the final results are obtained by integrating multiple classifiers.The results show that the performance of the proposed method is better than that of single SAE and some other deep learning and machine learning methods.It shows that the method of feature extraction using different SAEs and combining with feature selection and multi classifier integration is helpful to the accuracy of the model in the cross condition scenario.Compared with the method in Chapter 3,the proposed method is more effective This method is more suitable for the scene with high computational efficiency and strong real-time performance.