基于变分自动编码器的接触网缺陷检测方法

Defect detection method for catenary based on variational autoencoder

关键词:  机器学习  接触网  缺陷检测  变分自编码器  类间分布距离  特征空间

Key words:  machine learning  catenary  defect detection  variational autoencoder  distribution distance between classes  feature space

摘要:

接触网支撑悬挂部分是铁路接触网的关键基础设施,受弓网之间长期接触振动影响,接触网零部件易产生各种缺陷。基于接触网4C图像开展缺陷监测是运维的核心工作,直接关系铁路运输安全和可靠性。传统人工检测方法存在劳动强度大、效率低、易漏检等问题,利用图像处理和人工智能技术实现缺陷自动检测是该领域研究的热点问题。接触网零部件种类繁多且各类缺陷样本稀缺,现有依赖大量训练样本的深度学习方法难以适用。为此,提出基于变分自编码器（VAE）的接触网缺陷分类方法（DefVAE）。该方法基于同类样本在特征空间满足高斯分布的假设,利用VAE编码器输出的潜在特征确定已知缺陷样本的特征分布,通过分布空间重采样和解码生成大量缺陷数据以弥补样本不足；编码阶段引入辅助标签信息,增大潜在特征空间的类间分布距离；缺陷分类阶段采用滑动标签辅助的图像生成方法,结合重构误差提升分类精度。在开源数据集及接触网4C数据集上的对比实验和消融实验结果表明,DefVAE在开源数据集上多数指标优于基线方法,在接触网缺陷分类中具有很高的分类精度。

Abstract:

The supporting and suspension parts of the catenary are the key infrastructure of the railway catenary. However, due to the long-term contact-induced vibration between pantograph and catenary, the components of the catenary are prone to various defects. Defect monitoring based on 4C images of catenary is the key to operation and maintenance tasks, which directly relates to the safety and reliability of railway transportation. Traditional manual inspection methods face challenges such as high labor intensity, low efficiency, and a high missed detection rate. Therefore, using image processing and artificial intelligence technology to detect defects automatically is a hot issue in this research field. The components of the catenary are diverse in type, and samples of each type of defect are scarce, making the existing deep learning methods that rely on a large number of training samples difficult to apply. To overcome this problem, we proposed a classification method, named defect detection based on variational autoencoder (DefVAE) for catenary. This method was based on the assumption that samples of the same class follow a Gaussian distribution in the feature space. It utilized the potential features from the output of a variational autoencoder (VAE) to determine the feature distribution of known defect samples and generated a large amount of defect data through resampling and decoding in the distribution space to compensate for the lack of samples. During the encoding phase, we incorporated auxiliary label information to increase the inter-class distribution distance in the latent feature space. During the defect classification phase, we adopted an image generation method assisted by sliding labels and combined the reconstruction error to improve the classification accuracy. The results of comparative and ablation experiments on open-source datasets and catenary 4C datasets show that DefVAE outperforms the baseline methods in most indicators on the open-source datasets and has high classification accuracy in the classification of catenary defects.