Development of an efficient AI model for fault identification and directional AE source identification in railway steel tracks

The safety and durability of railway infrastructure rely on the timely and accurate identification of structural defects before critical failures arise. Acoustic emission (AE) sensing has become a potential real-time monitoring technique. Conventional multi-sensor identification techniques are generally not feasible for large-scale implementations. This study presents a low-cost, single-sensor methodology for identifying faults and classifying zones on railway tracks using AE signals collected from a single wideband (WD) (100–900 kHz) AE sensor, utilizing deep learning methods. The AE raw signals are transformed into time-frequency representations using a continuous wavelet transform, enabling accurate extraction of the temporal and spectral characteristics of non-stationary signals. The experimental dataset, comprising 1452 AE events from 363 distinct pencil lead break locations, has been developed to assess the proposed methodology. A new spatial zoning framework is developed to improve the practical applicability of the proposed methodology. The study presents three deep learning models for analysis: base-model artificial neural network, convolutional neural network (CNN), and hybrid CNN-LSTM. Damage-induced AE signal data recorded support the conclusion that the CNN-LSTM architecture outperformed the other models, achieving 98.7% accuracy for fault-position identification and 99.17% accuracy for zone classification. The high performance of the CNN-LSTM architecture is attributed to its ability to extract both spatial patterns from scalograms and temporal dependencies in the AE signal evolution. This study shows that hybrid CNN-LSTM architectures can achieve very high accuracy and early fault identification with a single AE sensor. This approach provides an efficient and economical solution for fault identification in railway tracks.