Defect Detection of Primary Coil Spring of Heavy-Haul Locomotive by Dynamic Adaptive Parallel Fusion Residual Network

The damage of primary coil spring in heavy-haul locomotive poses a significant risk to the safety of railway transportation. However, the strong background noise caused by the complicated running environment usually drowns out the defect features of coil spring in the locomotive vibration response, which makes it challenging to detect the spring defect states in time. Considering the advantages of the efficient channel attention (ECA) and residual network (ResNet) in data mining, this article reports a novel dynamic adaptive parallel fusion residual network (DAPFRNet) for accurately detecting the defect degree of the coil spring of heavy-haul locomotive. The DAPFRNet begins by integrating a convolutional neural network (CNN) with ECA to autonomously extract sensitive features from the raw bogie frame acceleration signals. Then, a multiscale parallel residual network is constructed to deeply mine the multiscale latent features from the extracted information, which can effectively improve the feature learning capability by dynamically adjusting the scale of branches. Finally, an adaptive feature fusion module is designed to accurately integrate the branch features by enhancing the relationship awareness among distinct parallel branches, and thus achieves intelligent grading assessment of coil spring defects. Both dynamics simulations and field tests are carried out to show that the proposed method can effectively identify different degrees of defects in coil spring under different running speeds and damage locations.