列车转向架故障光学图像在线检测方法研究

Railway vehicle condition monitoring and fault diagnosis technology is of great significance to the transportation safety and efficiency. The train bogie failure, such as surface damage, parts missing, components loosing, usually leads to abnormal location and shape of key components, which can be significantly characterized by optical images. The traditional fault detection methods are inefficient, and the accuracy tends to be greatly affected by the illumination. To overcome these shortcomings, this project manages to develop on-line train bogie health detecting technique based on the image processing theories. Firstly, considering the effects which brought by the low light source intensity at the bottom of the train, and nonuniform illumination with compensatory light, the multi-scale Retinex algorithm and gray stretching techniques are investigated for developing the image enhancement methods, accordingly, the image containing more useful information can be obtained. Secondly, because the abundant edge features about the various components in the train can be obtained from the images, the image edge feature extraction method that combining the multi-scale analysis and mathematical morphology theories is developed for more accurate and higher definition edge features. Finally, aiming at the comparative analysis between the operation status image and normal status image for the train fault detection, the similarity characterization method for the edge features is studied by combining the Shape Context (SC) algorithm and principal component analysis (PCA) method, and as a result, the operation status of the train bogie can be effectively detected. This work will provide fundamental basis, and shall develop a new technique for train bogie on-line fault detection.

铁路车辆状态监测与故障诊断对提高运输安全及效率意义重大。对于列车转向架表面破损、零件脱落、部件松动类多发故障，传统检测方法效率低，且准确率易受光照条件影响。本项目以提高检测准确率与效率为目的，结合此类故障表现为部件位置、形状发生改变，易于通过光学图像表征的特点，引入与改进图像处理方法，开展转向架故障在线检测方法研究。首先针对列车底部光源强度弱、补偿光源光照不均匀的问题，研究基于多尺度Retinex算法、灰度拉伸算法等的图像增强方法，获得更为显著的图像信息；之后结合转向架部件繁多，图像边缘特征丰富的特点，研究基于多尺度与数学形态学理论的图像边缘特征提取方法，获取更为精细、准确的边缘特征；最后针对故障检测须对比分析运行状态、正常状态图像的需求，研究结合SC匹配算法与主分量分析方法的图像边缘特征相似度表征方法，实现转向架状态的自动判别。本项目研究成果将为列车转向架故障在线检测提供理论和技术支持。

铁路车辆状态监测与故障诊断对提高运输安全及效率意义重大，本项目针对列车转向架表面破损、零件脱落、部件松动类多发故障，开展了转向架故障图像在线检测方法研究，重点研究了转向架图像边缘特征提取方法，在光照条件不佳情况下的转向架图像增强方法，并研究了基于多种描述子的边缘特征相似度表征、形状匹配、目标识别方法，研究结果为列车转向架故障在线检测提供一定的理论和技术支持。具体研究成果包括：（1）在研究了经典一阶、二阶微分算子的基础上，确立了各种算子在列车转向架图像边缘提取的优缺点，肯定了多尺度小波边缘检测方法在转向架图像连续性边缘提取方面的优点；提出了一种针对弱光照条件下，基于多尺度Retinex、数学形态学的列车转向架边缘检测方法；提出了一种以像素强度和梯度的稀疏特性为双重约束的，能更好保留边缘特征的图像平滑方法。（2）在总结图像增强算法以及图像质量评估算法的基础上，开展了一系列的列车转向架图像增强算法研究；针对转向架图像亮度不均匀问题，提出了一种基于引导滤波的Retinex图像增强算法；针对转向架图像亮度较暗问题，提出了一种采用梯度稀疏和最小平方约束的细节成分增强方法；此外，还针对局部严重过曝光图像散失纹理及颜色的问题，研究了一种过曝光图像的纹理及颜色修复算法。（3）基于降维思想，提出了一种可以提高图像形状匹配和目标识别的速度以及抗噪能力的PCA-SC图像匹配和目标识别方法；为了提高图像识别的准确率和鲁棒性，降低计算复杂度，提出了一种基于弹性匹配的目标识别算法；提出了一种根据不同边缘角点特征具有不同价值尺度的分层描述的边缘轮廓描述算法；提出一种能提高对边界噪声的鲁棒性，同时兼顾形状匹配算法的检索精度和运算效率的离散曲线演化形状匹配算法；提出了一种基于弦角轮廓特征的兼顾形状匹配检索率和运算效率的形状描述算法；提出了一种基于同底三角形面积的描述方法，并采用动态规划算法进行相似度匹配。