Study On Fault Method Of Gearbox Based On VMD-DL

With the development of modern industry,the proportion of integrated and large-scale rotating machinery in mechanical equipment is increasing.As one of the key components of modern industry,the gearbox is extremely prone to mechanical failure.Therefore,the fault diagnosis and condition monitoring of the gearbox are of great significance for the normal and smooth operation of the large rotating machinery.In the context of the increasing scale of monitoring data,deep learning algorithms are widely used with their powerful feature extraction capabilities.This thesis proposes the gearbox,SKAD-VBN model,SCDAE-DBN-DBN model,KHA-VBN,(DBN)and convolutional neural network(CNN).(1)In view of the problem that the VMD decomposition process needs to artificially set the number of decomposition layers k and the penalty factor,and the subjective factors will have a great impact on the decomposition results,the VMD optimization model — KHA-VMD model based on the krill group optimization algorithm(KHA)is proposed.The optimization model is compared with EMD and EEMD models,and with the clig as the index,the good effect of the model in noise reduction is proved through envelope spectrum analysis;(2)Aiming at the problem of poor feature learning ability of ordinary models,the fusion model — SCDAE-DBN model based on deep belief network(DBN)is proposed.SCDAE is used to resist the noise in the original signal and perform low-level feature learning,and DBN learns deep features based on low-level features.The features of the two models are complementary,further improving the feature learning ability of the model;(3)For the slow iterative convergence problem of SCDAE-DBN model,the adaptive learning rate adjustment model — ad＿SCDAE-DBN model based on reconstruction error is further proposed.The experiment proves that compared with SCDAE-DBN model,ad＿SCDAE-DBN model accelerates the convergence rate and further improves the diagnostic accuracy;(4)Aiming at the relatively low diagnostic accuracy of vibration signal by SVM,the fusion model of VMD and CNN with parameter optimization—KHA-VMD-CNN model is proposed.The learning rate was dynamically adjusted using Adam method,regularization using Dropout technique,and finally,the KHA-VMD-CNN model has higher diagnostic accuracy by comparing experiments with five models.