Research On Fault Diagnosis Method Of Rolling Bearing Under Variable Working Conditions Based On Adam-S CNN

Bearings are the core parts of high-speed train bogies,wind power gear boxes and other rotating machinery.Their health status has an important impact on operational safety.Because rolling bearings work under harsh conditions of strong load,high speed and large workload for a long time,they are prone to various failures,which may lead to catastrophic accidents.Traditional intelligent fault diagnosis methods need to manually extract features from vibration signals,such as local mean decomposition,empirical mode decomposition,HilbertHuang transform,wavelet transform,etc.Manually extracted features are input into machine learning methods to obtain diagnostic results.Common examples of these machine learning methods include K-nearest neighbor,random forest,support vector machines,etc.However,the traditional intelligent fault diagnosis method has many shortcomings.Overall,the final diagnostic performance is highly correlated with the quality of feature extraction.Therefore,it is of great significance to propose an appropriate feature extraction method for the effective diagnosis of rolling bearing faults.In view of the problems such as difficulty in feature extraction,this paper uses the convolutional neural network theory to preprocess bearing time-domain signals,design deep learning fault diagnosis models,extract vibration signal time-frequency image information,and identify bearing fault types from the perspective of bearing fault diagnosis mechanism.The focus is on bearing fault type diagnosis under variable working conditions.The main content of this paper is as follows,(1)In recent years,data-driven intelligent diagnosis methods have been widely used in the field of bearing fault diagnosis,but these methods all involve some expert experience and knowledge,and cannot accurately mine bearing fault characteristics under variable working conditions.To solve this problem,a method based on first-order differential filtering spectrum division(FDFSD)is proposed in this paper to realize fault diagnosis of bearings under varying working conditions.First,the time-domain signal is processed by first-order difference,and the frequency domain diagram is generated by frame division,windowing and Fourier transform.Then,a triangle filter is designed to re-partition its spectrum,highlighting the characteristic information and improving the signal-to-noise ratio.Finally,time spectral features are extracted to obtain the feature graph of time-frequency combination.(2)In recent years,as a very special convolutional neural network,residual network has been widely used in the field of intelligent diagnosis due to its powerful function.In recent years,as a very special convolutional neural network,residual network has been widely used in the field of intelligent diagnosis due to its powerful function.A novel dense residual network(DRNet)is proposed in this paper.The model combines the structural advantages of dense connections in dense networks with the advantages of residuals learning to prevent the gradient disappearing and network degradation caused by network deepening to carry out efficient fault diagnosis for rolling bearings.First,each Sub-block in the dense network is further processed to make it have better nonlinear representation ability to represent complex transformation,and then deeper fault characteristics are extracted.Then,the residual learning is embedded in each Sub-block of the dense network,so that the network degradation will not occur in each dense block of deepening processing.(3)Finally,an Adam-S optimization algorithm is proposed.By introducing the second order momentum of the previous gradient into the expression of the second order momentum of the current gradient,the relationship between the parameters of the two adjacent gradients in Adam algorithm is enhanced,which makes the algorithm more reliable and the gradient prediction more accurate.Without adding additional parameters,the training stability and generalization ability of the algorithm in complex environment are further improved.