Research On Fault Diagnosis Method Based On Deep Residual Network

As industrial production becomes more and more complicated and the higher cohesion of automatic production equipment,production equipment for fault diagnosis technology is also increasing demand.The traditional methods based on knowledge and analytical model are difficult to realize the fault diagnosis of complex nonlinear systems,but the data-driven fault diagnosis method can overcome this problem proficiently,so it is widely used.Among them,the fault diagnosis method based on deep learning theory has excellent engineering application prospects because it averts relying on feature selection and background knowledge.Aiming at the inherent problems that gradient disappearance and network degradation existing in the fault diagnosis method based on deep learning,the Residual Network(Res Net)was used to process the feature images,and then the fault diagnosis was realized.Compared with traditional fault diagnosis methods,the accuracy of fault identification has been significantly improved.The main research work as follows:Based on Res Net method,a fault diagnosis model with parameter optimization and wide universality is designed.The model can automatically select the number of network layers,activation function and residual module to generate the best classifier.Among them,the paper innovatively proposes the maximum pooled residual module(MP-conv block),which well overcomes the problem that the common module is difficult to obtain the local key information of the image and retain the texture features.In engineering practice,the equipment is in normal operation in most cases,and the fault samples are hard to obtain,so the classifier training often faces the problem of small samples.In this paper,the fault diagnosis method based on Res Net is improved in the form of model pre-training and fine-tuning.Firstly,aiming at the diagnosis object,the paper collects plenty of typical fault image samples to establish the prototype of fault classifier,and then collects the fault samples of corresponding equipment in the factory to be deployed.Through the collected samples to participate in training,the classifier is fine-tuned,so as to realize the effective construction of sorting machine in the case of small specimen and promote the precision of fault diagnosis.Photovoltaic module Electroluminescent(EL)images were used as experimental specimen to test the designed method.The test results show that the proposed two improved methods have high fault diagnosis accuracy,and lay a foundation for further study of fault diagnosis methods based on deep learning.In the comparison experiment,the diagnostic accuracy of the proposed model is 2.5% higher than that of the traditional Res Net method.