Research On Rolling Beariiig Fault Diagnosis Optimization Method Based On Deep Learning And Transfer Learning

Rolling bearings are widely used in rotating machinery,and their failure will lead to productivity loss and higher operating costs.Therefore,the fault diagnosis of the bearing is essential to ensure high-performance transmission.Recently,various fields have successfully developed data-driven methods for bearing fault diagnosis.However,these methods still have great limitations,such as training such models when the training data and test data have the same distribution.This greatly limits the wide application of datadriven models,especially fault diagnosis models based on neural networks.Because when the operating environment and conditions change,the above methods often fail to detect faults.In order to solve this problem,this paper studies a novel method of fault diagnosis based on deep learning pre-training model,and develops a high-performance model for fault diagnosis of bearings and gearboxes.First,for the single dimension of the original signal,time-frequency imaging is used to fuse multi-channel sensor signals,and the deep convolutional neural network’s ability to capture image features is used to implement a deep learning-based fault diagnosis method.Secondly,in view of the insufficient ability to capture time-frequency features of many current diagnosis models based on deep learning,and the cumbersome and inefficient parameter training,a diagnosis idea based on pre-training models has been developed.The pre-trained model is used to migrate the lower-level module parameter weight preset network to open up the high-level The level module adjusts the adaptability of the network to specific tasks,not only speeds up the convergence speed of the deep convolutional neural network,but also improves the fault recognition rate for rolling bearings,thereby improving the efficiency of fault diagnosis.This method also reduces the degree of dependence of fault diagnosis models on manual feature selection.By comparing with a variety of machine learning-based diagnosis methods,it is found that this method can be well applied to data samples with fewer labels,and passed Case Western Reserve.The effectiveness of the method is verified on the University bearing experimental platform and the Southeast University Drivetrain Dynamic Simulator experimental platform.Finally,in view of the problem that the number of layers of the deep convolutional neural network is too complex,the amount of parameters is huge,and the accuracy gain is gradually decreasing,a model scaling idea for the rolling bearing diagnostic network model based on deep learning is developed.Starting from the depth(layers),width(channels)and resolution(resolutions),the deep network structure is optimized using mobile inverted bottleneck convolution(MBConv).It not only reduces the number of parameters by dozens of times,speeds up the convergence speed,improves the accuracy of rolling bearing fault detection,and significantly reduces the scale of the deep network and improves the scalability of the model.The validity verification was completed on the bearing experiment platform of Case Western Reserve University and the Drivetrain Dynamic Simulator experiment platform of Southeast University.