Gap Measurement Method for Railway Switch Machines Based on the Fusion of Deep Vision and Geometric Features

Area of Science:

• Railway engineering
• Computer vision
• Machine learning

Background:

• Accurate measurement of railway switch gap is crucial for turnout locking status.
• Existing visual methods lack stability and sub-pixel precision in harsh industrial environments (oil, vibration, complex lighting).

Purpose of the Study:

• To develop a robust and precise vision-based method for railway switch gap measurement.
• To address the limitations of current methods in complex operating conditions.

Main Methods:

• A fusion of vision and geometric features (G-VFM) method was proposed.
• Utilized a confidence-aware YOLOv8 for gap localization and an improved U-Net for edge map extraction.
• Employed an R34-Fusion network for feature fusion and regression, with weighted Huber loss and piecewise linear calibration for error correction.

Main Results:

• Achieved a Mean Absolute Error (MAE) of 0.0076 mm and a maximum error of 0.0193 mm.
• Attained a 100% pass rate within a 0.02 mm industrial tolerance.
• Demonstrated an end-to-end inference time of 52.23 ms, balancing precision and efficiency.

Conclusions:

• The G-VFM method offers high-precision gap measurement for railway switch machines under experimental conditions.
• The approach provides a feasible solution for vision-based status monitoring, showing stable performance across various complex scenarios.
• Future work may require active lighting or multi-sensor fusion for extremely low-light conditions to ensure reliability.