Application of improved YOLO algorithms for train trajectory detection in distributed acoustic sensing systems

• Transform distributed acoustic sensing spectra into waterfall images for train trajectory detection. • Employ a one-stage deep-learning detector to localize train trajectories in waterfall images. • Integrate a snake-like deformable convolution module to track elongated, slanted train-induced trajectories. • Introduce a multi-scale attention mechanism to suppress noise and enhance salient train-related features. • Achieve up to 97.7% mean average precision (mAP@50) on field data from operating railway lines. To address the challenge of automatically detecting and identifying the direction of trains from train-induced vibration signals in distributed acoustic sensing (DAS) systems, this study proposes a real-time detection framework based on improved YOLO algorithms. First, the orthogonally demodulated vibration power spectra are accumulated over time and transformed into images to represent train trajectory patterns. Then, dynamic snake convolution (DSC) is integrated into YOLOv3, YOLOv5, and YOLOv8 architectures to adaptively capture elongated trajectory features. Additionally, the efficient multi-scale attention (EMA) mechanism is introduced to suppress redundant information and enhance salient responses. A unified dataset partitioning and training strategy was used to divide the dataset into training and test subsets, and both the baseline models (YOLOv3, YOLOv5, YOLOv8) and their enhanced variants (+DSC + EMA) were trained and evaluated accordingly. Experimental results show that the proposed YOLOv3 + DSC + EMA, YOLOv5 + DSC + EMA, and YOLOv8 + DSC + EMA outperform their respective baselines in terms of mAP@50 by 2.8%, 0.6%, and 1.5%, achieving mAP@50 scores of 95.2%, 97.5%, and 97.7%, respectively. Leveraging the architectural advantages of one-stage detectors, the proposed method achieves robust and real-time detection of both train trajectories and directions. These results demonstrate that integrating DAS with improved YOLO architectures significantly enhances the detection and recognition performance of vibration signals in railway environments, offering strong support for train operation safety monitoring and intelligent maintenance.