Application of CNN-LSTM to Detect Excessive Signal Distortions in Cab Signaling System

Abstract

The operational integrity of Automatic Locomotive Signaling systems is paramount for railway safety. However, signal distortions occurring within the rails—primarily due to electromagnetic interference and return traction currents—often evade traditional monitoring, which relies heavily on subjective visual assessment. This paper proposes an automated classification framework based on a hybrid Convolutional Neural Network and Long Short-Term Memory (CNN-LSTM) architecture. To provide high-precision features for the model, we implement the Thomson multitaper method, which offers a robust alternative to standard windowing techniques by significantly reducing spectral variance without a severe loss of resolution. Results demonstrate that the multitaper-based CNN-LSTM architecture achieved a classification accuracy of 97.7% on synthetic signals and 92.1% on operational rail data. These results significantly outperform the baseline Hann window spectrogram (86.3%), confirming that the integration of low-variance multitaper features and temporal modeling provides a superior solution for the automated monitoring of safety-critical railway signaling systems.