基于图像识别的铁路机车转向架螺栓紧固状态检测

为通过可视化图像分析手段辅助机务检修作业人员进行铁路机车转向架螺栓紧固状态检测，提出一种基于图像识别的铁路机车转向架螺栓紧固状态检测方法。首先，使用YOLOv7算法快速定位图像中的螺栓，利用深度学习算法的强鲁棒性和泛化能力，在机务检修各场景下准确获得包含螺栓及其定位漆的螺栓目标检测结果图像。其次，将螺栓目标检测结果图像转换至YCbCr空间，结合螺栓定位漆的色彩特征，提取Cr分量图像并应用自适应分割算法，有效滤除背景像素后得到仅包含螺栓定位漆的二值化图像。最后，针对螺栓定位漆的形状、位置和角度差异，提取Hu矩特征作为螺栓定位漆状态信息的定量表征，并结合SVM建立分类模型得到最终的螺栓紧固状态检测结果。实验结果表明，该方法充分利用了铁路机车转向架螺栓的特点，在保证螺栓目标检测准确率和螺栓定位漆分割精度的情况下，在所有场景下的铁路机车转向架螺栓紧固状态查准率为92.42%,查全率为94.55%,平均正确率为93.28%。

Detection of bolt fastening state of locomotive bogie based on image recognition

To assist maintenance personnel in detecting the fastening status of railway locomotive bogie bolts through visual image analysis, a railway locomotive bogie bolt fastening status detection method based on image recognition is proposed. Firstly, the YOLOv7 algorithm is used to quickly locate bolts in the image, and the strong robustness and generalization ability of deep learning algorithm are utilized to accurately obtain bolt target detection results images including bolts and their positioning paint in various scenarios of maintenance. Secondly, the bolt target detection result image is converted into YCbCr space, combined with the color characteristics of the bolt positioning paint, the Cr component image is extracted, and an adaptive segmentation algorithm is applied to effectively filter out background pixels to obtain a binary image containing only the bolt positioning paint. Finally, based on the differences in shape, position, and angle of bolt positioning paint, Hu moment features were extracted as quantitative representations of bolt positioning paint status information, and a classification model was established using SVM to obtain the final bolt tightening status detection results. The experimental results show that this method fully utilizes the characteristics of railway locomotive bogie bolts. While ensuring the accuracy of bolt target detection and bolt positioning paint segmentation, the accuracy of bolt tightening status in railway locomotive bogie bolts in all scenarios is 92.42%, the recall rate is 94.55%, and the average accuracy rate is 93.28%.