Research On Bearing Fault Diagnosis Based On Time Frequency Fusion And Attention Mechanism For Deep Learning

As the core component of rotating mechanical equipment,rolling bearings are widely used in professional machine tools,aerospace,wind power generation,and other fields.Its work is mostly carried out in high-temperature and high-pressure environments.If the bearing fails,it will greatly reduce the reliability of equipment operation.Therefore,it is particularly important to achieve real-time monitoring and fault diagnosis of rolling bearing operation status.With the rapid development of deep learning,signal classification using convolutional neural networks has become a hot research topic in bearing fault diagnosis.However,existing models generally have problems such as limited recognition accuracy,poor anti-noise performance,and poor model generalization ability.In this regard,this thesis mainly studies the following content:Firstly,aiming at the problem of limited recognition accuracy of bearing fault diagnosis models,a bearing fault diagnosis model based on multi-channel feature fusion convolutional neural network(MFFCNN)is proposed.This method uses signal processing technology to convert one-dimensional bearing acceleration signals into three types of two-dimensional time-frequency images,and constructs a multi-channel input network to simultaneously learn the three types of images.Experimental results based on the Case Western Reserve University(CWRU)bearing dataset show that compared to a single channel diagnostic model,MFFCNN can not only achieve accurate recognition of bearing fault categories,but also achieve 100% recognition accuracy in more difficult fault severity testing.Secondly,in order to achieve accurate diagnosis of bearings in high noise environments,the structural parameters of the MFFCNN were optimized.By adding varying degrees of Gaussian white noise to the vibration signal and testing it,parameter optimization is performed on the size and step size of the first layer convolutional kernel of the model.The experimental results show that using a super convolutional kernel of the same size as the input image can effectively reduce the interference of local outliers caused by noise signals,and the anti-noise model can achieve a 95.14% accuracy rate for fault degree recognition on a test set with a SNR(Signal to Noise Ratio)of-2d B.Thirdly,to solve the problem of poor generalization ability of the model,a multichannel feature fusion network based on attention mechanism,CBAM-MFFCNN,is proposed.By introducing an attention mechanism and performing channel and spatial weighting operations on the feature map,more efficient feature fusion and feature extraction for network input are achieved.The experimental results show that under the condition of training based on single load condition data,the recognition accuracy of the model with the addition of attention mechanism on the variable load test set is improved by 5.48%.Finally,considering the complexity of practical application scenarios,a comprehensive application test of the model was designed,including exploring the bearing fault location ability of the model,diagnostic performance on different bearing models,and model migration ability under small sample training conditions.Experimental results show that the proposed CBAM-MFFCNN model can achieve a recognition accuracy of 99.25% for bearing fault locations,and has good migration.It can achieve a recognition accuracy of 96.25% for fault categories under small sample training conditions.