Research On Bearing Remaining Useful Life Prediction And Fault Diagnosis Method Based On Time-Frequency Transformer

With the progress of industrial technology,the mechanization of industrial production is rapidly increasing.To ensure the long-term stable operation of mechanical equipment,it is important to implement Prognostics and Health Management for mechanical equipment.In the Prognostics and Health Management,the remaining useful life prediction and fault diagnosis are in the core position.In recent years,the rapid development of data science,computer science and other fields has laid the foundation for the development of the method of remaining life prediction and fault diagnosis of mechanical equipment based on deep learning,and made it a mainstream research direction.In this paper,based on the related literature at home and abroad,we mainly study the method of rolling bearing remaining useful life prediction and fault diagnosis based on deep learning technology.Firstly,to solve the problem that the existing deep learning-based prediction methods cannot achieve efficient training and global dependency extraction at the same time,this paper introduces the Transformer model with parallel architecture and based on the fully self-attention mechanism into the research of rolling bearing remaining life prediction,and proposes the bearing life prediction method based on the time-frequency Transformer by improving the model.The method uses time-frequency transformation to convert timedomain signals into time-frequency-domain signals,and designs adaptive position encoding layers using dilated causal convolution;by improving the encoding of the original Transformer model,it avoids the destruction of the original signal by foreign data,and enhances its feature extraction capability for time-frequency data,while realizing multiscale feature extraction,that is,the analysis of variable-length historical data.Finally,it achieves efficient and accurate prediction of the remaining useful life of bearing.Secondly,a time-frequency Transformer fault diagnosis method combined with a selfsupervised training strategy is proposed to solve the problem that it is difficult to collect fault samples with labels,insufficient training samples lead to insufficient extraction of fault features,and slow convergence and poor performance of the model in practical engineering applications.The method adopts a hierarchical construction method and uses a shifted window multi-head self-attention mechanism instead of the traditional multi-head selfattention mechanism to speed up the self-attention computation to achieve multi-scale feature extraction of the input time-frequency images and achieve the purpose of extracting more discriminative feature information from limited labeled data;meanwhile,the method learns general high-level feature representation from unlabeled data with the help of the self-supervised training strategy.At the same time,the method makes full use of the large amount of unlabeled data to improve the performance of the model and achieve accurate fault diagnosis under limited labeled data.The method in this paper is experimentally validated using PHM2012 bearing dataset,Case Western Reserve University bearing dataset,Paderborn University bearing dataset and Xi’an Jiao tong University bearing dataset,and the effects of important parameters such as sliding time window size and network depth are thoroughly studied and analyzed;meanwhile,the experimental results of this paper’s method are compared with those of commonly used methods,and the comparison results show that the comparison results show that the proposed method is superior in efficiency and accuracy.