Inspection of railway catenary systems using machine learning with domain knowledge integration

Area of Science:

• Engineering
• Computer Science
• Artificial Intelligence

Background:

• Railway catenary system inspection is vital for operational safety and efficiency.
• Current machine learning applications face challenges like data acquisition costs.
• Accurate assessment of infrastructure condition and component inventory is crucial.

Purpose of the Study:

• To present innovative machine learning solutions leveraging domain knowledge for improved railway catenary inspection.
• To enhance inference quality using existing training data and reduce false positives.
• To optimize detection of small and difficult-to-spot components like insulators.

Main Methods:

• A two-stage approach with object clustering to extract regions of interest (ROI).
• Dynamic confidence score weighting and ROI masking for enhanced precision.
• Integration of ensemble learning methods and custom test-time augmentations (TTA).

Main Results:

• Substantial improvements in AP50, precision, recall, and F1-score metrics.
• Significant enhancement in detecting small catenary components, such as insulators.
• Achieved an F1-score improvement from 61.97% to 82.53% compared to the baseline.

Conclusions:

• Integrating domain knowledge significantly boosts machine vision inspection quality.
• Proposed methods maintain practical runtime constraints for industrial applications.
• The enhanced system ensures more reliable and efficient railway catenary inspection.