Few-Shot Fault Diagnosis of Railway Switch Machines Using Regularized Supervised Contrastive Meta-Learning

Area of Science:

• Railway Engineering
• Machine Learning
• Signal Processing

Background:

• Railway switch machines are critical for train operation safety.
• Traditional fault diagnosis methods struggle with scarce fault samples in few-shot scenarios.
• Limited data hinders diagnostic accuracy and generalization for railway switch machines.

Purpose of the Study:

• To address the challenge of insufficient accuracy in railway switch machine fault diagnosis under few-shot conditions.
• To propose a novel fault diagnosis method for switch machines using limited sensor data.

Main Methods:

• Proposed a regularized supervised contrastive meta-learning (RSCML) fault diagnosis method.
• Utilized tri-axial vibration signals transformed into axis-wise STFT spectrograms.
• Employed channel expansion and attention enhancement for feature learning within a contrastive ANIL framework.

Main Results:

• Achieved maximum accuracy of 99.73% on 3-way and 5-way few-shot tasks.
• Obtained an F1-score up to 99.72% in few-shot scenarios.
• Demonstrated 93.08% accuracy and 92.84% F1-score in cross-category generalization experiments.

Conclusions:

• The RSCML method shows superior classification performance for switch machine fault diagnosis with limited samples.
• The proposed method exhibits stronger generalization to unseen fault categories, enhancing robustness.
• This approach holds significant potential for improving railway safety through effective fault diagnosis.