Few-shot outliers classification in high-speed railway track geometry based on self-supervised transformer

External disturbances, data transmission, sensor signal offsets and weather conditions are common sources of outliers in track geometry inspection data. The infrequent occurrence of these outliers leads to a scarcity of labeled samples, making high-accuracy few-shot classification challenging with traditional supervised learning methods. To address this, we propose a method for the few-shot classification of outliers in high-speed railway track geometry inspection data based on a self-supervised transformer. First, the self-supervised transformer pre-trains on a substantial amount of unlabeled data, enabling the model to learn and extract fundamental features and patterns from the inspection data. Next, the limited labeled outliers are used to fine-tune the model, enhancing its adaptability to the task of classifying outliers. This approach allows for the automatic identification and classification of outliers even with limited labeled data and without prior knowledge. Experimental results demonstrate that the proposed method accurately identifies and classifies outliers such as local burr, turnout gauge widening, constant section of unilateral gauge data, and abnormal distribution in track geometry inspection data, achieving a classification accuracy of up to 97.8 % and an F1-score as high as 97.9 %. This performance surpasses five supervised baselines by 4–26 % in accuracy and by 5–35 % in F1-score. Moreover, the method maintains an accuracy rate exceeding 92 % across different inspection trains and lines, demonstrating excellent generalization performance.