基于改进SSD的动车组底部异常检测方法研究

Study on the Anomaly Detection Method at the Bottom of EMU Based on Improved SSD

高速行驶中的动车组(EMU)底部偶尔出现螺钉丢失、布条和纸屑等异常情况,而动车组列车底部背景结构复杂、目标尺度较小和异常样本少等因素给异常检测带来极大挑战。为提高动车组底部异常检测平均精度,提出一种基于改进SSD动车组底部异常检测算法。首先,采用残差网络Resnet-101替换VGG-16,残差网络Resnet-101比VGG-16更能细化特征图的特征;其次,引入一种新的特征融合策略,特征融合有效利用了改进后SSD模型浅层的细节信息和深层的语义信息;最后,引入注意力机制,注意力机制有助于在特征图空间中建立特征之间的关系,突出相关特征,抑制不相关信息,为异常检测提供可靠指导。试验结果表明:与FasterR-CNN、SSD和YOLOV3相比,本文算法的mAP分别提高了18.14%、5.26%和3.85%,螺钉丢失的AP值分别提高了2.76%、2.66%和1.22%,纸屑的AP分别提高了51.66%、13.11%和10.32%,本文算法提高了小尺度目标的检测精度。

Some anomalies, such as missing fastening bolts, cloth strips and paper scraps, appear occasionally at the bottom of EMU. However, the complex background at the bottom of the train, small-scale targets and few abnormal samples have brought great challenges to the detection of these abnormalities. In order to improve the mean Average Precision(mAP) of anomaly detection at the bottom of EMU, an algorithm for inspecting anomaly at the bottom of EMU based on improved SSD is proposed. Firstly, VGG-16 is replaced by Resnet-101, which can refine the features of the feature map better than VGG-16. Secondly, a new feature fusion strategy is introduced to effectively utilize the detail information of the low-level feature layer and the semantic information of the high-level feature layer of the improved SSD model. Finally, self-attention mechanism is employed to contribute to the establishment of the relationship between features in the space of feature map, which highlights useful feature information, suppresses irrelevant information, and provides reliable guidance for anomaly detection. The experimental results show that: compared with FasterR-CNN, SSD and YOLOV3, the mAP of the proposed algorithm is improved by 18.14%, 5.26% and 3.85% respectively;the AP of the proposed algorithm is improved by 2.76%, 2.66% and 1.22% respectively in detecting missing fastening bolts;the AP of the proposed algorithm is improved by 51.66%, 13.11% and 10.32% respectively in detecting paper scraps. The proposed algorithm improves the detection accuracy of small-scale targets.