Application of an improved RepVGG algorithm to DAS-Based train operation detection

Abstract Train operation monitoring based on distributed acoustic sensing (DAS) has attracted increasing attention for its potential to ensure railway safety and support intelligent transportation systems. However, accurately recognizing train vibration signals and distinguishing them from noise remains a significant challenge. This study proposes a classification-model-based framework to identify both the train direction of travel and interfering noise signals. The approach transforms orthogonally-demodulated vibration power spectrum features into image representations, where conventional image-processing techniques are first applied to extract candidate vibration regions. A deep-learning-based classifier is then employed to perform the final recognition. To improve discriminative feature modeling, we adopt the RepVGG architecture with structural re-parameterization and further integrate a lightweight Normalization-based Attention Module (NAM). This design enhances the model’s ability to capture direction-of-travel cues while suppressing redundant information, thereby reducing false positives and false negatives. The dataset is randomly divided into training and testing subsets with an 80/20 ratio. Four models including VGG16, ResNet18, RepVGG, and the proposed RepVGG-NAM are trained and evaluated on this dataset. Experimental results demonstrate that RepVGG-NAM achieves superior classification performance, attaining an accuracy of 97.45% on the test set, which corresponds to absolute improvements of 0.2%, 0.3%, and 0.2% over VGG16, ResNet18, and RepVGG, respectively. These consistent improvements verify the effectiveness of NAM in dynamically calibrating channel weights to highlight salient features. Overall, the proposed RepVGG-NAM framework provides a high-accuracy and computationally efficient solution for DAS-based train vibration signal classification, contributing to the advancement of intelligent railway monitoring systems.