Fusing Expert Knowledge with Monitoring Data for Condition Assessment of Railway Welds.

Area of Science:

• Railway Engineering
• Condition Monitoring
• Data Fusion

Background:

• Axle Box Accelerations (ABAs) provide valuable data for railway infrastructure condition assessment.
• ABA measurements face uncertainties from noise, non-linear dynamics, and environmental factors, challenging rail weld assessment.
• Existing tools struggle with ABA data uncertainties for accurate rail weld defect detection.

Purpose of the Study:

• To refine the detection of faulty rail welds by fusing ABA data with expert feedback.
• To evaluate the effectiveness of different machine learning models in improving weld condition assessment.
• To quantify the confidence in defect predictions for railway infrastructure.

Main Methods:

• Assembled a database of expert evaluations on critical rail weld samples identified by ABA monitoring.
• Fused ABA-derived features with expert feedback using Binary Classification, Random Forest (RF), and Bayesian Logistic Regression (BLR) models.
• Evaluated model performance for defect detection and confidence quantification.

Main Results:

• RF and BLR models outperformed the Binary Classification model in weld defect detection.
• The BLR model provided probabilistic predictions, quantifying confidence in assigned labels.
• Acknowledged high uncertainty in classification due to potential ground truth label inaccuracies.

Conclusions:

• Fusing ABA data with expert feedback significantly refines rail weld condition assessment.
• BLR models offer a robust approach for improved defect detection and confidence estimation in railway monitoring.
• Continuous tracking of rail weld condition is crucial for reliable infrastructure management.