基于多尺度全维动态卷积残差网络的齿轮箱故障诊断OA

In response to the challenges of insufficient multi-scale feature analysis capability and gradient vanishing or explosion in deep network structures for gearbox fault diagnosis,this study proposes a Multi-Scale Full-Dimensional Dynamic Convolutional Residual Network(MFDCResNet)for gearbox fault diagnosis.Firstly,the Omni-dimensional Dynamic Convolution(ODConv)was utilized to dynamically adjust the weights of convolutional kernels,thereby acquiring more comprehensive fault feature information.Secondly,an improved Pyramid Split Attention(PSA)module was introduced to embed the Efficient Channel Attention(ECA)mechanism into the PSA module,so as to more effectively extract multi-scale spatial information and cross-dimensional key features.Finally,a dual-skip connection full-dimensional dynamic residual block was designed to enhance the network's ability to differentiate feature information,thereby the recognition capability of fault features was improved.By stacking these modules,the network depth was increased,and the in-depth exploration of potential fault features in fused signals was realized.Experimental results demonstrate that the proposed model can achieve the accuracy rate of 98.4%,which validates the superiority of the proposed model and provides a novel and effective intelligent method for gearbox fault diagnosis.