Gearbox Fault Diagnosis Based On Lightweight Network Model

The gearbox is an important device used to transmit the power of the motor.Its role is to transmit the speed and torque of the motor to the various parts of the machine,thus enabling the machine to work properly.Since it is subjected to huge loads and friction in operation,it often fails,resulting in decreased equipment operation efficiency and increased downtime,which has a great impact on production efficiency and cost.Therefore,how to accurately and timely diagnose the failure of gearboxes has become a common concern in industry and academia.First,the mechanism of gearbox faults and the fault diagnosis method based on deep learning are investigated.The local binary convolution layer is constructed based on the principle of local binary partten,and the method of using local binary convolution layer instead of traditional convolution layer is proposed to diagnose gearbox faults.Based on the theoretical model of the local binary convolution layer,a residual structure model with channel-by-channel convolution and local binary convolution is constructed,and the network model is pruned using a model pruning algorithm.A theoretical foundation is laid for further deep learning-based gearbox fault diagnosis.Secondly,a gearbox fault diagnosis method based on residual structure and local binary convolution(1D-RSLBCNN)is proposed to address the problem of multiple parameters and long training and detection time in traditional fault models.Different from the traditional convolutional neural network,the 1D-RSLBCNN network proposed uses the local binary convolution layer to replace the traditional convolution layer,which reduces the model parameters while accelerating the training speed and Rate of convergence.By introducing a residual structure,the saturation or decrease in accuracy caused by an increase in network depth is avoided.In order to verify the effectiveness of this method,experiments were carried out on the single fault gearbox data set of Southeast University and the self-made composite fault gearbox data set.The results show that the model has good stability and reliability.Finally,in order to further reduce the model parameters and compress the network model,a 1D-RDLBCNN based on model pruning is proposed for gearbox fault diagnosis.The proposed network structure utilizes the channel-by-channel convolution in the depth-separable convolution and the local binary convolution to form the main body of the residual network,and introduces a batch normalization layer to speed up the model fitting;in order to compress the network model,the model pruning method is introduced.The experimental results show that combining the channel-by-channel convolution and local binary convolution in depth-separable convolution not only further reduces the model parameters but also has good diagnostic accuracy;pruning the network model effectively reduces the model size.