钢轨波磨快速识别是轨道交通运维部门制定养修计划的基础，使用车厢内部（车内）噪声数据进行钢轨波磨快速识别具有显著优势。为解决车内噪声标签数据无法大量获取的问题及少标签样本通用学习的困难，提出了一种包含“预训练+精调”两阶段的钢轨波磨自监督识别模型。首先，获取了地铁列车运行时大量车内噪声数据，并利用小波降噪与小波变换技术建立了车内噪声时频图数据集。之后，预训练阶段采用降噪卷积自编码器对大量无标签车内噪声时频图进行表征学习，获取预训练最优参数；在精调阶段利用“编码器+分类器”对少量有标签的车内噪声时频图进行监督学习。最后，开展了模型实验及钢轨波磨识别评估。结果表明：所提出的钢轨波磨自监督识别模型能够自动有效抽取时频图主要特征，模型针对“无波磨”、“30~100 mm波磨”、“100~300 mm波磨”和“30~100 mm+100~300 mm波磨”的检测精度分别为95.62 %、96.46 %、92.38 %和90.37 %。

Rapid identification of rail corrugation is essential for railway maintenance departments to plan maintenance activities. Using carriage inner noise data for quick identification of rail corrugation has significant advantages. To address the challenge of limited labeled carriage inner noise data and the difficulty of general learning from few labeled samples, a two-stage self-supervised rail corrugation recognition model based on "pretraining + fine-tuning" is proposed. First, a large amount of carriage inner noise data from operating subway trains was collected, and a time-frequency diagram dataset was created using wavelet denoising and wavelet transform techniques. During the pretraining stage, a denoising convolutional autoencoder was used to perform representation learning on numerous unlabeled time-frequency diagrams, and optimal pretraining parameters were obtained. In the fine-tuning stage, supervised learning was conducted on a small set of labeled time-frequency diagrams using an "encoder + classifier" approach. Finally, model experiments were carried out, and rail corrugation recognition was evaluated. The results show that the proposed model effectively extracts key features from the time-frequency diagrams. The detection accuracies for "no corrugation," "30-100 mm corrugation," "100-300 mm corrugation," and "30-100 mm + 100-300 mm mixed corrugation" are 95.62%, 96.46%, 92.38%, and 90.37%, respectively.