Gear Box Fault Diagnosis Based on Neural Network with Multi-Scale Weighted Fusion

Aiming at the problem that fault feature of gearbox is difficult to exactly detect under strong background noise,a novel design method of multi-scale weighed feature fusion block (MSWFFB) was proposed. Based on this, a two-dimensional convolutional neural network (CNN) was constructed for the fault diagnosis of gearbox. In the network, features are extracted using convolutional kernels of different sizes, and the multi-scale feature information of the gearbox fault signal was extracted from different receptive field. Importing the depthwise over-parameterized convolution (Do-Conv) and the dual attention network (DANet), the individual channels of the characteristic signal are weighted and fused in the process, DANet recalibrates the multi-scale features and allocates attention weights to reduce the influence of redundant features. The experimental results show that the proposed network can diagnose the fault of the gearbox under different noise conditions, and the diagnostic accuracy can reach 98.5% under the original signal, and still has an accuracy of 91.2% under the background of strong noise. The network are carried out to prove the feasibility of the method, which is conducive to the research and development and practical application of intelligent fault diagnosis.