基于三级级联架构的接触网定位管开口销缺陷检测

Detection of split pins defect in catenary positioning tube based on three-level cascade architecture

针对高铁接触网定位管开口销在列车长期运行振动中容易松脱并且松脱样本数量匮乏的问题,本文提出一种基于深度卷积生成对抗网络(DCGAN),扩充缺陷样本集后,再训练卷积神经网络(CNN)检测开口销缺陷的三级级联架构。该架构首先采用中心点法提取训练需要的相同规格开口销图像。然后通过改进的DCGAN生成模拟缺陷样本,并搭建轻量级CNN网络对生成的模拟缺陷样本进行筛选。最后将添加了模拟缺陷样本的扩充缺陷样本集与正样本集输入优化后的VGG16卷积神经网络中,以训练分类模型,检测开口销缺陷。实验结果表明,本文所提方法检测接触网定位管开口销缺陷的准确率高达99%。

The split pins of high-speed railway catenary positioning tube are easy to be loosed in the long-term running vibration of the train. However, the number of loose samples is scarce. To solve these problems, this study proposes a three-level cascade architecture expand the defect samples based on deep convolutional generative adversarial network(DCGAN). Then, the convolutional neural network(CNN) is trained to detect split pins defect. Firstly, according to the central point method, the same size image of split pins for training is extracted. Then, DCGAN is used to generate simulated defect samples and a lightweight CNN network is formulated to screen the generated samples. Finally, the extended defect sample set and the positive sample set are utilized to train the detection model on the adjusted VGG16 convolutional neural network. In this way, the defective pins defect state detection can be realized. Experimental results show that the proposed method can achieve 99% accuracy in split pin defect detection of catenary positioning tube.