Research On Fault Diagnosis Method Of Rotating Machinery Based On Lightweight Network Model

When rotating machinery and equipment are running,they are often in harsh environments such as high temperature,heavy load,and impact.Over time,some parts of rotating machinery may be damaged to varying degrees.In order to ensure stable operation of machinery and equipment and prevent frequent accidents,health monitoring of key components of rotating machinery is required.For a long time,only professionals can complete the fault diagnosis of rotating machinery.With the continuous progress of diagnosis technology and the rise of artificial intelligence,intelligent diagnosis methods based on deep learning have begun to be widely used.Although deep learning methods have achieved good performance in fault diagnosis applications,complex networks and limited computing power of hardware devices restrict the rapid development of deep learning in fault diagnosis.Based on the improved lightweight network model,this thesis is applied to the intelligent diagnosis of bearings and the migration diagnosis of different working conditions.It is efficient and has good accuracy.The main research contents are as follows:(1)When using the traditional neural network method for fault diagnosis,the massive data samples make the computer bear huge load and increase the operation time.To solve this problem,a lightweight network model,Shuffle Net,is used for bearing fault diagnosis.The Shuffle Net network comprehensively uses Depthwise Separable Convolution(DSC),Group Convolution(GConv)and Channel Shuffle(CS)technologies to reduce the amount of network model parameters,which is similar to traditional networks.It also has higher accuracy,and significantly reduces the computing time and reduces the dependence on computer hardware.(2)In order to further reduce the dependence on computer hardware,a ShuffleSE network model is proposed based on the Shuffle Net V2 network.The Shuffle-SE network retains the Shuffle Net V2 network unit and adds the Squeeze-and-Excitation(SE)structure.The Shuffle Net V2 network unit improves the Shuffle Net network based on four lightweight criteria.The SE structure is also very simple,and it selectively suppresses or improves the information of different channels,which improves the network performance.Compared with network models such as Mobile Net V2,Shuffle Net V1/V2,and Res Nets,Shuffle-SE takes into account both operational efficiency and accuracy.(3)For fast fault diagnosis of rotating machinery and equipment with small samples and unlabeled samples under variable working conditions,the improved lightweight network model is embedded in the domain adaptation method of highorder moment matching.The higher-order moment matching domain adaptation method can reduce the distribution difference between training data and test data,achieve fine-grained alignment,and complete migration diagnosis.The main structure of the lightweight model adopts depthwise separable convolution and pooling layers.When the number of input and output channels is large,the depthwise separable convolution effectively reduces the amount of parameters.The maximum pooling layer is used as a downsampling layer,eliminating most of the less valuable ones.characteristic information.The lightweight network model was verified by changing working conditions on two different test benches,the computing efficiency was greatly improved,and high accuracy was achieved.