The fault detection of the mechanical components in railway freight cars is important to the safety of railway transportation. Owing to the small size of the mechanical components, a manual detection method has a low detection efficiency. In addition, traditional computer vision technology has difficulty detecting multiple categories of objects simultaneously. Inspired by the use of one-stage deep-learning-based object detectors, in this paper, a multi-feature fusion network (MFF-net) for the simultaneous detection of three typical mechanical component faults is proposed. By embedding three modules in the network to improve the detection effect of small mechanical component faults, the feature fusion module is used to supplement the deep semantic information of the shallow feature maps. A multi-branch dilated convolution module uses dilated convolution and multi-branch networks to obtain the fusion features of multi-scale receptive fields, and the squeeze-and-excitation block is embedded in the network to enhance the channel features. All experiments used Nvidia 1080Ti GPUs for training on the PyTorch platform. The experimental results show that the three modules used in the network all contribute to the fault detection of railway freight car mechanical components, and that the detection performance of MFF-net is better than that of most other popular SSD-based one-stage object detectors. When the input image size is 300 pixels × 300 pixels, MFF-net can achieve 0.8872 mAP and 33 frames per second. It has good robustness to complex noise environment and can realize real-time fault detection of railway freight car mechanical components.

铁路货运车辆机械部件的故障检测对铁路运输的安全至关重要。由于机械部件的尺寸小，因此手动检测方法的检测效率低。另外，传统的计算机视觉技术难以同时检测多种类别的物体。受一阶段基于深度学习的目标检测器的启发，提出了一种同时检测三个典型机械部件故障的多功能融合网络（MFF-net）。通过在网络中嵌入三个模块以提高对小机械部件故障的检测效果，特征融合模块用于补充浅层特征图的深层语义信息。多分支扩张卷积模块使用扩张卷积和多分支网络来获得多尺度接收场的融合特征，并且将挤压和激励块嵌入网络中以增强信道特征。所有实验都使用Nvidia 1080Ti GPU在PyTorch平台上进行训练。实验结果表明，该网络中使用的三个模块均有助于铁路货车机械部件的故障检测，并且MFF-net的检测性能优于大多数其他流行的基于SSD的一级目标检测器。当输入图像尺寸为300像素×300像素时，MFF-net可以达到0.8872 mAP和每秒33帧。