基于改进Vision Transformer的道岔故障智能诊断

Intelligent diagnosis of turnout faults based on improved Vision Transformer

道岔故障种类繁多,特征复杂,存在检测难、分类难等问题,导致故障排查效率低下,对铁路运输安全构成威胁。Vision Transformer模型在图像分类方面具有较高准确度,但是其处理的是图像块,而不是传统的像素级特征,在某些情况下可能会影响曲线局部信息的获取。针对上述情况,提出一种基于改进Vision Transformer模型的故障曲线分类算法。首先,对典型道岔故障及原因进行梳理分类,指出几种典型的道岔故障;其次,对使用道岔动作电流数据生成的图像尺寸进行调整并根据故障图像特点进行数据增强,使用ResNet网络取代原Vision Transformer模型中的故障图像分块机制进行特征提取,同时采用相对位置编码增强模型的适应性和泛化能力;最后,利用模型的多头自注意力机制,综合全局与局部信息进行分类,并得到分类权重。经过实验验证,本文道岔故障分类识别总体准确率达99.77%,各分类识别的平均精确率达99.78%,与原模型相比,在训练集和验证集上的识别精度分别提升了5.4%和2.4%。为了更好地理解模型的性能,采用Grad-CAM方法将迭代过程可视化,剖析了模型关注区域的变化过程,并在测试集上与VGG-16、DenseNet121等经典分类模型进行性能对比;通过ROC曲线评估分类效果,显示改进的模型取得更优结果。研究结果为道岔故障识别分类提供了新的理论支持,并为未来的研究提供了新的思路和方法。

The diversity and complexity of turnout failures pose challenges in detection and classification,leading to inefficient troubleshooting and threats to railway transportation safety.The Vision Transformer model demonstrates high accuracy in image classification,but it processes image blocks rather than traditional pixel�level features,which may affect the acquisition of local information in certain cases.To address these issues,a fault curve classification algorithm based on an improved Vision Transformer model was proposed.First,typical turnout failures and their causes were categorized and sorted out,identifying several typical turnout failures.Second,the image size generated from the current turnout action data was adjusted and the data were enhanced based on the characteristics of the fault images.The ResNet network was used to replace the fault image block mechanism in the original Vision Transformer model for feature extraction,and relative position encoding was adopted to enhance the model’s adaptability and generalization ability.Finally,the model’s multi-head self�attention mechanism was used to integrate global and local information for classification,and classification weights were obtained.Experimental validation shows that the overall accuracy of turnout failure classification identification reaches 99.77%,and the average recognition accuracy rate of each category reaches 99.78%.Compared with the original model,the recognition accuracy on the training set and validation set increased by 5.4%and 2.4%,respectively.To better understand the performance of the model,the Grad-CAM method was used to visualize the iterative process,analyze the change process of the model’s attention area,and compare the performance with classic classification models such as VGG-16 and DenseNet121 on the test dataset.The classification efficiency was evaluated through the ROC curve,showing that the improved model achieved better results.The research results provide new theoretical support for turnout failure identification and classification and provide new ideas and methods for future research.