Railway Track Inspection Using Deep Learning Based on Audio to Spectrogram Conversion: An on-the-Fly Approach.

Area of Science:

• Railway Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Periodic railway track inspection is crucial for preventing accidents caused by structural and geometrical defects.
• Current manual acoustic inspection methods in Pakistan are labor-intensive, error-prone, and require expert domain knowledge.

Purpose of the Study:

• To integrate deep learning models with acoustic-based systems for improved railway track fault detection.
• To enhance the performance and reliability of track inspection systems, thereby reducing train accidents.

Main Methods:

• Implementation of Convolutional Neural Network (CNN) 1D, CNN 2D, and Long Short-Term Memory (LSTM) models.
• On-the-fly spectrogram generation as a deep learning model layer for feature extraction.
• Training and testing models with audio samples of varying lengths (1.7s, 3.4s, 8.5s) and extensive data augmentation.

Main Results:

• The LSTM model, utilizing 8.5-second audio splits, achieved the highest accuracy (99.7%), precision (99.5%), recall (99.5%), and F1 score (99.5%).
• The proposed deep learning approach demonstrated superior performance compared to traditional methods.

Conclusions:

• Deep learning models, particularly LSTM, offer a highly accurate and efficient solution for automated railway track inspection.
• Integrating AI with acoustic analysis can significantly improve railway safety by detecting faults more reliably.