Intelligent Classification Method for Rail Defects in Magnetic Flux Leakage Testing Based on Feature Selection and...

Area of Science:

• Railway Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Rail integrity is crucial for transportation safety.
• Magnetic Flux Leakage (MFL) is a key non-destructive testing method for rails.
• Current MFL defect classification faces accuracy challenges, especially with limited data.

Purpose of the Study:

• To develop an enhanced intelligent classification framework for rail MFL signals.
• To improve the accuracy and robustness of defect detection in railway inspection.
• To address the limitations of conventional methods in handling small-sample scenarios and complex defect types.

Main Methods:

• Proposed an enhanced classification framework using a Particle Swarm Optimized Radial Basis Function neural network (PSO-RBF).
• Implemented a dynamic PSO algorithm with adaptive learning factors and nonlinear inertia weight for RBF parameter optimization.
• Employed a hierarchical feature processing strategy combining mutual information selection and correlation-based dimensionality reduction.
• Incorporated adaptive model architecture adjustment for effective small-sample learning.

Main Results:

• Achieved 87.5% accuracy on artificial defects, a 17.5% absolute improvement over conventional RBF.
• Obtained macro-F1 score of 0.817 and Matthews Correlation Coefficient (MCC) of 0.733 for artificial defects.
• For real-world limited samples, reached 80% accuracy, improved minority class (spalling) F1-score by 0.25, and reduced false alarms by 50%.

Conclusions:

• The optimized PSO-RBF framework demonstrates superior capability in extracting and classifying MFL signal patterns.
• The study sets a new benchmark for industrial rail inspection, particularly in discriminating between abrasions, spalling, indentations, and shelling defects.
• The proposed methods offer a robust solution for accurate MFL-based rail defect classification, even with limited data.