Research On End-to-End Bearing Fault Diagnosis Based On Deep Learning Under Complex Conditions

Bearings are the critical parts of rotating machinery.Failure of bearing may lead to severe casualties and economic losses;thus,the bearing fault diagnosis is an indispens-able part of maintaining the safe and stable operation of modern machinery.Traditional bearing fault diagnosis studies focus more on manually feature designing and feature selection,including noise reduction and signal filtering.However,these manual pro-cedures are complex and lack intelligence,and it is challenging to diagnose massive data manually.Moreover,bearing operating conditions are complex and changeable,making it more challenging to extract bearing fault features.Researches on the end-to-end bearing fault diagnosis under complex conditions still need improvement.Recently,deep learning could utilize massive data to automatically learn the bearing fault features due to its powerful feature extraction and information processing capabilities,which brings the possibilities of efficient and intelligent diagnosis under complex conditions.To this end,this thesis takes bearing vibration signals as the research object and deep neural network as the method to research end-to-end bearing fault diagnosis under noisy and different working conditions.The details of this thesis are as follows:(1)In order to achieve the end-to-end bearing fault diagnosis,considering that deep convolutional neural networks have powerful feature extraction capabilities that could replace manually feature designing and selection processes;whereas con-sidering its huge amount parameters,which would affect the computational effi-ciency;therefore a lightweight deep structure optimized convolutional neural net-work is proposed for end-to-end bearing fault type and severity diagnosis.The pro-posed network employs the local-sparse structure to replace the parameter-dense layers in the original network,thus greatly reducing the number of parameters.Ex-periments show that the proposed method achieves similar performance compared to the original method by using less than half parameters as used in the original study,thereby improving the efficiency and adaptability of fault diagnosis.(2)In order to improve the performance of bearing fault diagnosis under heavy noise conditions,considering that the recurrent neural network could capture the time-dependent characteristics of the vibration signal,and the attention mechanism could adaptively reorganize features;therefore,an end-to-end adaptive anti-noise neural network framework(AAnNet)is proposed.Therein,the convolutional neu-ral network is employed to extract features from the vibration signal,and the gated recurrent neural network is used to further process the extracted features.To fur-ther enhance the adaptability under noise conditions,an input strategy based on random sampling is proposed,and the activation function is improved.Extensive experiments based on two datasets prove that the proposed framework achieves state-of-the-art results under heavy noise conditions.(3)In order to improve the performance of bearing fault diagnosis under differ-ent working conditions,considering that labeling data under all working condi-tions is extremely time-consuming and expensive;therefore,an end-to-end multi-adversarial cross-domain neural network(MACDNet)is proposed.The proposed method takes the labeled source domain data and the unlabeled target domain data to improve the performance of the target domain.The proposed method uses a multi-adversarial strategy to automatically extract domain-invariant features from different domains without manually designed metrics and features.Therein,domain-adversarial learning is adopted to reduce the domain discrepancy between different domains,and the mini-max entropy is adopted to cluster features of the same category from different domains,thereby alleviating the class misalignment problem.In addition,the virtual adversarial training is used to strengthen the local Lipschitzness assumption,and the adaptive layers are adopted to further improve the cross-domain performance of the network.The results of the cross-load and the cross-machine experiments prove the effectiveness of the proposed method.(4)In order to explore the intrinsic characteristics of the neural network,we visualize the working principle and the process of the bearing vibration signal in AAnNet.Through network visualization,it is found that the feature extraction part has good effects on feature selection and noise suppression;the attention mechanism part can automatically select and combine the features according to different fault types and severities;similar features are firstly gathered together and then separated step by step.By visualizing MACDNet,it can be directly seen that the proposed method effectively alleviates the class misalignment problem,thereby improving the per-formance of cross-domain bearing fault diagnosis.Through network visualization,we can better understand how the neural network works,thus improving the inter-pretability of neural networks and would guide the design of neural networks.