Fault Diagnosis Method Of Rolling Bearing Based On Prototype Domain Adaptation Of Less Labeled Data

As a key part of rotating machinery,rolling bearing is widely used in industry and other related fields.However,due to its complex working environment,the data distribution between different loads is different,and it is difficult to obtain effective data.Traditional fault diagnosis methods have some defects.Therefore,it is of great significance,using a small amount of effective data to build a fault diagnosis model under variable load and realize intelligent fault diagnosis.A rolling bearing fault diagnosis model based on less labeled data is proposed by studying deep neural network and prototype domain adaptation method.The proposed model can complete the task of fault diagnosis of rolling bearing under variable load.A rolling bearing fault diagnosis method based on prototype domain adaptation is proposed to solve the problem that the traditional intelligent diagnosis method need a lot of labeled data for model training and have low generalization.Firstly,one-dimensional vibration signals are transformed into two-dimensional time-spectrum images by wavelet transform.299 unlabeled two-dimensional spectrum images with only 1 labeled data for each class under a certain load are selected as source domain data,and the unlabeled two-dimensional spectrum images under different loads from the source domain are selected as target domain data.Deep network is used to extract features from source domain data and target domain data.At the same time,prototype domain adaptation is used to narrow the differences between the two domains.In order to improve the generalization performance of the model,a meta-acon activation function is proposed to change the original activation function into an adaptive activation function.Experimental results show that the proposed diagnosis model has high diagnosis performance and generalization performance.Although the above method has good results,the accuracy still needs to be improved.Therefore,so as to improve the accuracy,an improved deep neural network structure is further proposed.This network structure can extract more effective features and provide a basis for subsequent domain adaptation.Firstly,the residual block structure of the deep neural network is improved,and the input of the3×3 convolution layer in the residual block structure is divided into four parts connected in the form of residual,which can improve the receptive field range and feature processing ability of the neural network.Secondly,in order to improve the network’s attention to effective features,an attention mechanism is added after each residual block.The features of source domain and target domain are extracted from the improved network,and domain adaptation is carried out to construct the rolling bearing fault diagnosis model.It is verified by various experiments that the improved method has better processing ability for features,and has better performance in accuracy,generalization and unbalanced data processing.