Research On Mechanical Fault Diagnosis Based On Deep Learning

As an important part of rotating machinery,bearing has been widely used in modern mechanical equipment.When it breaks down,it will not only seriously affect the performance of the machine,which will reduce production efficiency,but also lead to serious accidents.Therefore,the fault diagnosis of rolling bearing is very necessary.At present,deep learning has achieved good performance in fault diagnosis.However,the high accuracy of fault diagnosis is mainly based on a general rule: the training data set and the test data set are in the same distribution and contain less noise.In practical engineering applications,rolling bearings work under complex and diverse working conditions,so that the collected data are usually unevenly distributed.This means that the data sets from different operating conditions are not in the same distribution,which will lead to poor adaptability on different working conditions.Moreover,noise is an inevitable interference in industrial field,and noise interference will reduce the accuracy of fault diagnosis.Therefore,these factors extremely limit the application of these methods in solving practical fault diagnosis problems.Therefore,the main work of this thesis is as follows:(1)Multi scale network is used to extract the features of different scales and different levels of abstraction,and then the extracted details are combined with advanced semantic features to reduce the loss of information,improve feature utilization,obtain advanced features,and finally improve network performance.At the same time,spatial attention module is added to the network to extract key information and improve the overall performance of the network.The experimental results show that the proposed network structure can achieve high accuracy and migration on the data sets of different speeds and different loads.This shows that our method has high migration ability in different distributed training data and test data.(2)Nowadays,most of the deep learning models used in fault diagnosis are based on convolutional neural network or cyclic neural network.Here we use Transformer,a new deep learning model for mechanical fault diagnosis.It makes full use of self-attention mechanism to solve the problems that convolutional neural network cannot consider long-time dependence and cyclic neural network cannot be parallelized.At the same time,knowledge distillation is used to improve model accuracy,reduce model training time and improve model generalization ability.The experimental results show that the Transformer model can be used in fault diagnosis,and has high accuracy and strong anti-noise.In order to verify the performance of the two models in this thesis,the experimental results in the same environment are compared with several research results in the field of fault diagnosis in recent years.The results show that the fault diagnosis model based on multi-scale has higher migration ability,while the fault diagnosis model based on Transformer has better noise resistance.