Research On Rolling Bearing Fault Diagnosis Based On Deep Learning

With the development of intelligent industrialization,rotating machinery are widely used in various production scenarios,and rolling bearings are the core of them,so the diagnosis of rolling bearing health status is of great significance.This paper takes rolling bearings as the research object,and uses deep learning technology to conduct fault diagnosis research on the collected vibration signal data by sensors,the main research content is as follows:(1)Aiming at the problem that the characteristics of time series signal are not obvious and the feature extraction ability of conventional convolution layer is poor,a fault diagnosis method based on Short-Time Fourier Transform(STFT)and residual nested convolution network is studied.The time-frequency domain information extracted from the original signals by STFT is better than the time-domain information to reflect the fault characteristics of rolling bearings.In order to enhance the influence of low-level features on the output of conventional convolutional layer,a residual nested convolutional network is designed.Residual nested convolutional network has internal and external residual structure,and all n×n convolution kernels are replaced by 1×n and n×1convolution kernels,which can reduce model parameters.Experimental results show that the proposed fault diagnosis method is better than the other four comparison methods.(2)Aiming at the problem of inconsistent data distribution between source domain and target domain under variable working conditions,an improved capsule network fault diagnosis method under variable working conditions is studied.The capsule network can simultaneously encode the spatial information between features and the probability of feature existence,and extract the spatial domain invariant features between the source domain and target domain.The introduction of residual nested convolutional network,channel attention,and involutional network to replace the convolutional layer of the capsule network can enhance the learning of the input information.Experimental results show that the proposed fault diagnosis method can obtain 89.4% accuracy when the distribution of source domain and target domain is quite different,which is better than the comparison methods.(3)Aiming at the problems of strong noise in real scenes,large number of parameters in deep learning model,and complex design in artificial feature,a lightweight fault diagnosis method based on improved deep residual shrinkage network is studied.The method directly inputs the normalized vibration signal data into the model without manually extracting the relevant features for signal processing.In order to construct the lightweight model,all the convolutional layers are replaced with blueprint separable convolutions,and the network parameters are reduced using the grouping structure.A multi-directional collaborative attention network is designed to encode spatial information in multiple directions to obtain spatial feature vectors and spatial feature surfaces.The encoded information can effectively improve feature extraction ability through squeezing,excitation and fusion operations.The improved deep residual shrinkage network calculates the weights of different channels and sums them with feature surfaces,which helps to enhance the anti-noise performance of the model.Experimental results show that the proposed fault diagnosis method performs well on single-noise datasets and mixed-noise datasets.