, 铁路货车零部件故障是造成货车事故的主要原因之一, 当前列检工作多采用“人眼识图”判断零部件故障, 无法实现故障自动识别。针对关键零部件故障自动识别的问题, 根据“先定位, 后识别”的思路提出基于深度学习和图像处理技术结合的图像检测算法。首先, 构建零部件检测数据集和零部件故障识别数据集。其次, 融合小目标检测层、卷积注意力模块, 提出改进的Small-Target-Detect-Layer and CBAM YOLOv5s (SC-YOLOv5s) , 实现多尺度关键零部件的定位、分类、裁剪。最后, 提出基于目标检测结果与先验知识的算法直接判断丢失类型故障, 改进的Mobile-NetV3分类算法实现滚动轴承甩油、锁紧板松动、摇枕弹簧断裂三种故障的自动识别, 以及图像处理方法实现交叉杆弯曲、滚动轴承前盖破损两种故障的自动识别。结果表明, SC-YOLOv5s算法的mAP@0. 5和mAP@0. 5: 0. 95分别可达99. 3%、74. 9%, 检测速度可达36. 09FPS; 改进的MobileNetV3算法对轴承甩油、锁紧板松动、摇枕弹簧断裂的识别准确率分别可达98. 63%、99. 34%、90. 21%; 图像处理法对交叉杆弯曲、滚动轴承前盖破损识别准确率分别可达95. 32%、82. 88%。

The parts failure of railway freight wagon is one of the main causes of wagon accidents. Currently, train inspection work mostly uses "human eye recognition" to determine the part failures, which cannot realize the automatic fault identification. For the problem of automatic identification of key parts failure, an image detection algorithm based on is proposed based on the combination of deep learning and image processing technology according to the idea of "first locate, then identify". First, the parts detection data set and parts fault identification data set are constructed. Then, the Small-Target-Detect-Layer and CBAM YOLOv5s (SC-YOLOv5s) are integrated with the small target detection layer and Convolutional Block Attention Module to realize the localization, classification and cropping of multi-scale key parts. Finally, an algorithm based on object detection results and prior knowledge is proposed to directly determine the type of loss fault, and the improved MobileNetV3 classification algorithm is proposed to realize the automatic identification of three kinds of faults: oil dumping of rolling bearing, loosening of locking plate, and breakage of rocking pillow spring, as well as the image processing method to realize the automatic identification of two kinds of faults: bending of crossover rod and breakage of front cover of rolling bearing. The results show that SC-YOLOv5s mAP@0. 5 and mAP@0. 5: 0. 95 can reach 99. 3% and 74. 9% and the detection speed can reach 36. 09 FPS; the improved MobileNetV3 algorithm can reach 98. 63%, 99. 34%, and 90. 21% of the recognition accuracy of bearing oil dumping, locking plate loosening, and rocking pillow spring breakage; the recognition accuracy of the image processing method for crossbar bending and rolling bearing front cover breakage can reach 95. 32% and 82. 88%, respectively.