Comparative deep learning architectures for railway train detection using accelerometer

Abstract

Railway level crossing accidents remain a critical safety concern worldwide, necessitating reliable train detection systems particularly at unguarded crossings. This study presents comparative evaluations from four deep learning (DL) architectures convolutional neural network (2D-CNN), long short-term memory (LSTM), MobileNetV2, and hybrid CNN-LSTM of railway train detection using triaxial accelerometer sensor data. Vibration signals acquired from piezoelectric sensors mounted on railway infrastructure were transformed into spectrogram representations and classified into four proximity categories: idle, far-distance (>400 m), near distance (200-400 m), and train passing (<200 m). The experimental results demonstrate that the hybrid CNN-LSTM architecture achieves highest classification performance with 98.2% accuracy, 98.0% F1-score, and convergence to 95% accuracy within 68 training epochs, outperforming CNN-2D (96.9%), MobileNetV2 (92.3%), and LSTM (89.5%). The hybrid model effectively combines spatial feature extraction through convolutional layers with temporal dependency modeling via recurrent components, achieving signal detection at distances exceeding 400 meters with signal-to noise ratios above 30 dB. These findings provide empirical evidence for optimal DL model selection in accelerometer-based railway safety monitoring systems deployed on resource-constrained edge devices.