An Uncertainty-Aware Bayesian Deep Learning Method for Automatic Identification and Capacitance Estimation of...

Area of Science:

• Railway Engineering
• Signal Processing
• Artificial Intelligence

Background:

• Compensation capacitors are crucial for high-speed railway track circuits.
• Strong noise poses challenges for accurate detection and reliability assessment.
• Existing methods struggle with misclassification in noisy environments.

Purpose of the Study:

• To develop a robust framework for compensation capacitor detection under strong noise.
• To improve the accuracy and reliability of state recognition and capacitance estimation.
• To provide uncertainty-aware predictions and enhance safety assurance in track circuits.

Main Methods:

• A hierarchical Bayesian deep learning framework integrating multi-domain signal enhancement (time, frequency, time-frequency).
• Bidirectional Long Short-Term Memory (BiLSTM) for robust feature extraction.
• Bayesian classification/regression using Monte Carlo (MC) Dropout and Stochastic Weight Averaging Gaussian (SWAG) for uncertainty quantification.
• A rejection mechanism to filter low-confidence outputs.

Main Results:

• Achieved 97.8% state-recognition accuracy and a mean absolute error of 0.084 μF.
• Demonstrated superior performance over threshold-based, CNN, and standard BiLSTM models in NLL and ECE.
• Approached theoretical 95% interval coverage, indicating high reliability and calibration quality.

Conclusions:

• The proposed Bayesian deep learning framework significantly enhances robustness, accuracy, and reliability in compensation capacitor detection.
• It offers a viable solution for intelligent monitoring and safety assurance of critical track circuit components.
• The uncertainty-aware predictions are key to improving system safety and performance in challenging conditions.