Fault Of Rolling Bearing Based On Convolutional Capsule Network Diagnostic Research

Rolling bearings are a common component in industrial equipment,they are responsible for carrying,rotating or moving loads and reducing friction.The working structure of rolling bearings is complex,variable and fragile.These characteristics lead to a variety of bearing failures,which seriously affect the normal operation of the equipment.Fault diagnosis of rolling bearings is very critical because it can prevent the normal operation of the equipment from being seriously affected,thus avoiding the safety accidents that may be caused and the huge economic losses.Since the traditional method relies too much on the subjective manual experience of experts in the process of fault diagnosis,it is influenced by the subjective manual experience,and this method is limited in effectively extracting fault feature information,which leads to the difficulty in achieving good generalization ability when dealing with various complex faults.Although deep learning methods can combine the features of fault feature extraction and classification to automatically achieve the ability to extract representative features from the original signal data,and thus achieve the purpose of overcoming the influence of subjective experience on feature extraction.Therefore,this thesis focuses on the rolling bearing fault diagnosis methods based on convolutional capsule neural network,and the main research contents of this thesis are as follows:(1)To address the problem that the vibration signals of rolling bearings are affected by changes in operating conditions and strong environmental noise,and the fault signals are easily disturbed by the complex environment after the missing features,a rolling bearing fault diagnosis method with improved convolutional capsule neural network is proposed.The method directly inputs the time domain signal data of rolling bearings,and extracts the input time domain data fault features through asymmetric convolutional layers with different scale convolutional kernel sizes,which can also effectively reduce the number of parameters while ensuring the maximum extraction of feature information in the data;in addition,the introduction of the channel attention mechanism in this module can better extract the useful channel features and improve the feature extraction capability of this method;then,by improving the network fully connected layer in this module is improved into a capsule fully connected layer,which makes the capsule avoid the loss of feature information in space when outputting vector feature information.Experiments with variable noise,variable load and variable working conditions are conducted to verify that the method has better diagnostic performance and better generalization.(2)To address the problem that the deep learning diagnosis method is too large for the amount of prerequisite parameters to ensure the accuracy of bearing fault diagnosis after the fault signal is easily disturbed by the complex environment,a dynamic capsule network lightweight fault diagnosis method through adaptive weight sharing is proposed.The method combines dynamic convolution and capsule dynamic routing algorithms to distribute adaptive weight sharing throughout the network model.To learn the vibration signal characteristics,the convolutional weights are adaptively adjusted and shared to different convolutional layers through an attention mechanism,which can effectively reduce the computational cost of the network.In addition,dynamic routing algorithm is used to generate subcapsules with shared weights.The subcapsules extract and transform fault features into vector feature information storage to reduce feature loss.Finally,the method uses jump connect line concatenation dynamic convolution to enhance the network feature extraction capability.The effectiveness of the method is verified by noise and variable load experiments,and the method has better generalization and diagnostic effect.(3)A multi-scale residual parametric convolutional capsule network(MRCCCN)method for small-sample bearing fault diagnosis is proposed to address the problem that small-sample data are insufficient to support the training of traditional intelligent diagnosis methods and it is also difficult to extract sensitive fault features from the original signals.The method splits the single input into multi-branch inputs,and then extracts the features of the multi-branch inputs by residual parametric convolution.Then,the fused features are fed into an improved parametric capsule network to further extract fault features and further store feature information by dynamic routing.Not only the learning capability of the network is improved,but also the number of network parameters is lightened.The method is able to diagnose faults accurately under small sample conditions,and the experimental results in noisy environments show that there is still good noise immunity performance.