Research On Gearbox Fault Diagnosis Method Based On Horizontal Federated Learning Framework

With the advent of the era of industrial big data,data-driven mechanical equipment fault diagnosis methods have been widely studied and applied.As a key component of mechanical equipment,gearbox has become a hot research object.If enterprises that lack sufficient high-quality fault data want to use the high-quality fault data owned by other enterprises to help them complete the training of fault diagnosis model,they need to break the problem of "data island" caused by laws,regulations and industry norms.Because of its distributed learning architecture and privacy protection,horizontal federated learning provides a new method for enterprises to solve the problem of "data island".Taking the gearbox as the research object and the horizontal federated learning as the research method,this thesis studies the gearbox fault diagnosis method under the framework of horizontal federated learning.The main work and innovations are as follows:(1)A gearbox fault diagnosis method based on horizontal federated learning is proposed and verified.Aiming at the problem of "data island" of gearbox fault data,this thesis combines the horizontal federated learning framework with the gearbox fault diagnosis model to realize the joint training of gearbox fault diagnosis model by multiple participants.The experimental data of planetary gear pitting,wear,broken teeth and normal vibration waveform under the same working condition are collected,and the fault sensitive features are manually screened and labeled to make samples.The samples are trained by FATE framework and its deep neural network algorithm components.After100 rounds of joint training,the accuracy of fault diagnosis has reached 90.50%.The feasibility of applying lateral federated learning to gearbox fault diagnosis is proved.(2)A gearbox fault diagnosis model based on densely connected convolutional networks and horizontal federated learning framework is proposed.Aiming at the disadvantage that traditional machine learning models such as DNN need to manually screen fault sensitive features,this thesis studies the method of automatically extracting fault sensitive features by using deep convolution neural network,and puts forward a densely connected convolutional networks with 63 layers network structure to ensure high accuracy and reduce model parameters.Aiming at the shortcomings of FATE framework,the horizontal federated learning framework is manually built by using Python language,and the performance of the model under the horizontal federated learning framework is studied.The collected four types of fault data are made into time domain diagram samples for training and prediction.After only 20 rounds of aggregation,the accuracy of the model reaches 93.25%,which is about 3 percentage points higher than the horizontal federated learning model based on DNN.It is proved that the model has excellent fault diagnosis ability under the framework of horizontal federated learning.(3)A method of transforming one-dimensional vibration signal into twodimensional image based on fast Fourier transform and Gramian Angular Difference Field is proposed.In view of the characteristics of convolutional neural network with image as input and the limitations of the traditional method of converting onedimensional data to two-dimensional image,this thesis improves the overall performance of horizontal federated learning model from the perspective of improving image quality.Based on the diagnosis accuracy and model convergence speed,the proposed method is compared with the fault samples processed by FFT and GADF.The experimental results show that the diagnostic accuracy of the samples obtained by using only GADF is 94.00%after 20 iterations;After 11 iterations,the accuracy of the samples obtained by using FFT alone reached 100%;The accuracy of the samples obtained by the proposed method is100% in only three iterations,which is better than the first two in convergence speed and accuracy.It is proved that the proposed method has better effect in image coding of vibration sequence data.(4)A gearbox fault diagnosis model based on case standardization is proposed.Aiming at the problem that the fault data between enterprises is not independent and identically distributed,this thesis studies the reasons for the reduction of model performance caused by non-independent and identically distributed data,and proposes to use case standardization to replace the batch standardization of the original densely connected convolutional networks,so as to reduce the impact of non-independent and identically distributed data on model aggregation speed and model performance.The experimental results show that when the data distribution of the model based on Batch Normalization is different,the model training fails,while the model based on Instance Normalization can still achieve 100% fault diagnosis accuracy after aggregation for 9times.It is proved that the proposed method can effectively improve the convergence speed and performance of the overall model under non-independent and identically distributed data.