Research On Bearing Fault Diagnosis Method Based On Neural Network

In mechanical equipment,the bearing is an indispensable part,and its operational condition is directly related to the performance of the entire equipment.Because its actual operating environment is complex and changeable,it is very prone to breakdowns.If failed bearings are found and replaced in time,it may prevent the abnormal state of the entire equipment and avoid serious economic losses.In extreme cases,it may even endanger the safety of personnel.Therefore,research on bearing fault diagnosis is a meaningful subject.In recent years,with the rapid progress of deep learning technology,a variety of neural networks have been widely used in the fault diagnosis of vibration data,and have a considerable application prospect.In contrast,the traditional fault diagnosis methods include three steps: feature extraction,feature selection and pattern classification.However,these methods often rely on expert experience and professional knowledge,introducing human interference that results in high uncertainty in the diagnosis results,making it challenging to meet practical needs.Additionally,some existing fault diagnosis methods still have certain limitations.For example,the traditional single network model lacks sufficient distinguishability in feature extraction,and its anti-interference ability to noise is weak,which makes it perform poorly in terms of fault diagnosis accuracy.With the continuous development and optimization of neural network models,the complexity and number of layers in these models have been gradually increasing.As a result,there has been a significant rise in training time.Additionally,the black-box nature of neural network models often leads to poor interpretation of diagnosis results,making it challenging for them to gain trust in the machinery industry.These factors act as obstacles to the widespread adoption and application of this method.To address these issues,this thesis proposes a bearing fault diagnosis method based on neural networks,focusing on the following key aspects.1.A new data enhancement method,called "overlapping downsampling," is proposed to increase the number of data samples while reducing the amount of data per sample,effectively reducing the time required for neural network training.In this method,the original vibration signal is cut by a fixed-size sliding window,and each cut sample is processed by downsampling,reducing the amount of data of each sample.The experimental results show that this method effectively increases the number of samples,improves the accuracy of fault diagnosis,and shortens the training time of the network.2.A new bearing fault diagnosis method based on multi-scale convolution neural networks is proposed to improve the learning ability of features.In this method,the one-dimensional vibration signal processed by overlapping downsampling is used as input,and the convolution check signals of different sizes of multiple channels are used for feature extraction.The extracted features are then fused by element-by-element product,and bearing fault diagnosis is achieved by the normalized exponential function.The experimental results show that this method has good feature extraction ability and can effectively improve the accuracy and robustness of fault diagnosis.3.A rolling bearing fault diagnosis method based on convolution neural network-bidirectional gated cycle units based on an attention module is proposed.The introduction of an attention mechanism eliminates the influence of redundant features on fault diagnosis,effectively improving the adaptive ability of fault diagnosis.Unlike the traditional multi-scale bearing fault diagnosis method,this method uses the attention mechanism to weight and fuse the extracted features and further extract time series features from the signal through the bi-directional gated cycle unit.Finally,fault classification is achieved by the normalized exponential function.The experimental results show that the method performs well on different data sets and has a certain ability to overcome the interference of redundant features,improving the accuracy,anti-noise,and self-adaptability of fault diagnosis.