Fault Diagnosis Method For Rolling Bearings Based On Image Coding And Deep Residual Network

Rolling bearing is a key component of mechanical devices,its working conditions are the complexity of the working conditions,increased probability of failure,its working performance and state has an important impact on the safety and reliable operation of engineering equipment.Therefore,effective methods must be adopted to ensure the safe and reliable operation of bearings.In view of the shortcomings of the traditional signal processing fault diagnosis method which relies on the experience of experts,this paper takes rolling bearings as the research object,adopts the data-driven idea,extracts the multi-input and multi-output fault modes of rolling bearings with practical application value quickly and effectively from a large amount of data in the background of industrial big data,and carries out direct data analysis and processing to realize the "end-to-end" fault diagnosis of rolling bearings.In this paper,the main research content is as follows: "The end-to-end" intelligent fault diagnosis of rolling bearings,and finally complete the effective equipment condition monitoring and fault diagnosis.The main research contents of this paper are as follows:(1)In response to the shortcomings of traditional bearing fault diagnosis methods based on signal processing techniques that rely on expert experience and manual feature selection,a fault diagnosis method based on image coding and deep residual networks is discussed.Several commonly used image coding methods are used to transform one-dimensional time-domain vibration signals into two-dimensional images,and deep learning networks are used for fault identification of bearings.The experimental comparison is verified on CWRU and locomotive bearing datasets,and the results show that the two-dimensional images are used as features input to the deep learning network for fault diagnosis with good fault diagnosis performance and without human intervention;(2)Aiming at the problem that the traditional fault diagnosis methods do not fully exploit the correlation characteristics among time series of fault signals,a new fault diagnosis model based on recursive image coding and residual network is proposed by introducing recursive graph coding technology into the field of fault diagnosis.The proposed method can encode the one-dimensional original vibration signal into a two-dimensional image with obvious features,and then use the powerful image data feature extraction ability of the residual network to achieve the accurate identification of rolling bearing faults.Experimental validation using the method on CWRU and locomotive bearing data shows that the model achieves 99.99 % and99.83 % high accuracy in bearing fault diagnosis;the method still shows excellent recognition accuracy under various conditions of data length and working condition changes;the method has higher recognition accuracy and stronger robustness compared with other common image coding methods;(3)In order to solve the problem of low diagnostic recognition accuracy caused by insufficient features of single-channel vibration signals and lack of labeling of cross-service data,a rolling bearing fault diagnosis method based on multi-sensor feature fusion and migration learning(MFF-TL)is proposed.The method uses wavelet transform to transform multiple sensor signals into time-frequency maps,then fuses the time-frequency maps into multichannel images,and finally uses a pre-trained deep residual network(Res Net)as a migration model to reduce the fault diagnosis task to a multichannel image classification problem through migration learning.The method is experimentally validated on cylindrical roller bearings and locomotive bearings,and the average diagnostic accuracy of the method reaches 99.23 % and 99.78 %.Twelve migration tasks are designed in the CWRU dataset,and the cross-service fault migration diagnosis is performed under four different loads with an average recognition accuracy of 93.12 %.The results demonstrate the superiority and scalability of the proposed method.