International Journal of Science, Engineering and TechnologyAuthors: Asawari Gawade, Pranita Jadhav, Dhananjay Rode, Prof.Bangar A.P, Prof.Bhosale S.B.
Abstract: Indian Railways is one of the largest rail networks in the world. The structure of the rail track is crucial for passenger safety and freight reliability. Traditional inspection techniques involving visual surveys, ultrasonic flaw detectors and hammer tests are labor-intensive, episodic and miss early-stage degradation. This study presents a comprehensive AI-based framework which involves monitoring acoustic waves through the use of real-time stress-wave acquisition which leads to defect classification with the usage of deep-learning and predictive rail-wear analysis. A mobile inspection unit that is built using an ESP32 microcontroller continuously samples vibration signals from piezoelectric and MEMS-based at 8 kHz. Moreover, it performs on-edge pre-processing of that data (which include pre-emphasis filtering, Hamming windowing, FFT and Mel-Frequency Cepstral Coefficient extraction). After that, it transmits feature vectors to the cloud server through Wi-Fi using the REST API. A CNN+LSTM scheme classifies the defect into normal, surface crack, internal flaw, joint defect and severe wear. The stacked-LSTM regressor forecasts cumulative wear at a 180-day horizon and uses historical sensor data along with outdoor temperature and traffic load data for its predictions. Field trials conducted on a 2.4 km test track produced 12,500 labelled samples. The proposed model achieved an overall classification accuracy of 97.2 %, precision of 96.5 %, recall of 96.9 % and F1-score of 96.7 %. The model outperforms other benchmarks which includes SVM (86.3 %), Random Forest (89.7 %), KNN (82.1 %), CNN-only (93.4 %) and LSTM-only (94.6 %). The wear-prediction module achieved a Mean Absolute Error (MAE) of 0.18 mm in comparison of 0.61 mm for ARIMA and 0.84 mm for linear regression. A web dashboard made using HTML, CSS and JavaScript visualises live track condition, GPS coordinates, vibration spectra, alert logs and wear-forecast curves. The system facilitates condition-based predictive maintenance, reduces the life cycle cost by approximately 28% and presents low-power, scalable solution for next generation railway-health monitoring while reducing dependency on manual inspection.
