Research On Quantitative Diagnosis Method For Bearing Fault Based On Deep Residual Shrinkage Network And Transfer Learning

With the rapid development of Industry 4.0 and industrial Internet of Things technology,equipment condition monitoring,and fault diagnosis have entered the intelligent stage.As one of the critical components of a mechanical transmission system,bearings play an essential role in ensuring the safe and stable operation of equipment.The quantitative fault diagnosis of bearings can reveal faults’ development degree and evolution law,provide a basis for reliability assessment and life prediction of mechanical equipment,and avoid economic losses caused by insufficient or excessive maintenance.Deep learning transforms mechanical fault diagnosis from traditional feature extraction and classification to end-to-end intelligent diagnosis mode by automatically extracting fault information from vibration signals.Transfer learning is based on the correlation between fault information in vibration signals,which allows the intelligent diagnosis model to solve the problem of lacking a large amount of historical data under complex working conditions.Therefore,this paper researches a quantitative diagnosis method of bearing faults based on a deep residual shrinkage network and transfer learning technology as the support.The main research contents are as follows:(1)The quantitative diagnosis methods for rolling bearing faults are reviewed.The research on the quantitative diagnosis of rolling bearing faults conducted by related scholars in recent years is summarized into three major categories.The quantitative diagnosis method is based on double shock characteristics and dynamics modeling,shock pulse method,and big data-driven intelligent quantitative diagnosis method.Furthermore,the basic ideas,research progress,appropriate environment,and advantages and disadvantages of these three methods are introduced.(2)Aiming at the problems of inconsistent distribution of bearing vibration data under different working conditions and insufficient or excessive adaptation in the adaptive process of the source domain and target domain,a fault diagnosis and fault severity assessment method under different working conditions based on substructure optimal transport and support vector machine is proposed.By processing the substructure generation,weighting,and mapping of bearing sample data in the source and target domains,the problem of under and over-adaptation of source and target domain data in the domain-level,class-level,and sample-level matching fault diagnosis methods are avoided.Based on the substructure generation,weighting,and mapping,the support vector machine model is selected to improve the bearing fault diagnosis.The proposed method further realizes the estimation of the fault severity of the outer ring of the bearing across working conditions.(3)To address the problem that the intelligent diagnostic model is prone to overfitting when diagnosing complex data containing noise and the model does not converge in some cases,which affects its diagnostic capability,a quantitative diagnosis method for bearing faults based on an improved deep residual shrinkage network is proposed.The residual shrinkage module is constructed by adding sub-networks to the residual module to achieve automatic soft thresholding of the signal and reduce noise interference in the vibration signal.Based on the deep residual shrinkage network constructed by the residual shrinkage module,a convolutional layer with a wide convolutional kernel,Drop Block layer,Leaky Relu activation function,and Adamax optimizer are used in the network to improve the diagnostic capability of the model further.A double-fed wind turbine testbed is used to build a dataset of bearing outer ring,inner ring,and rolling body under different fault sizes.The effectiveness of the proposed improved deep residual shrinkage network method is demonstrated using this dataset.(4)To address the problem that a large amount of labeled historical data is required for the quantitative bearing fault diagnosis method based on the improved deep residual shrinkage network,a quantitative cross-condition bearing fault diagnosis method is based on the improved deep residual shrinkage network the entropy conditional domain adversarial network is proposed.While building the improved deep residual shrinkage network model to enhance the fault diagnosis capability is combined with the entropy conditional domain adversarial network.The quantitative diagnosis of bearing outer ring,inner ring,and rolling body faults under cross-conditional conditions is realized by aligning the source and target domain bearing sample data through the adversarial idea.