A Comparative Study of Fault Diagnosis for Train Door System: Traditional versus Deep Learning Approaches

Area of Science:

• Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Train door systems are critical for operational safety and efficiency.
• Effective fault diagnosis is essential for maintaining train door reliability and preventing failures.
• Motor current signals offer a viable data source for diagnosing train door system faults.

Purpose of the Study:

• To compare the effectiveness of a traditional feature-based fault diagnosis method with a deep learning approach (Convolutional Neural Network - CNN).
• To evaluate the performance of both methods in classifying various fault modes in a train door system.
• To determine the most suitable method for practical implementation in real-world train operations.

Main Methods:

• A test rig was developed to simulate natural and artificial wear, inducing various fault modes in train door components.
• A traditional fault diagnosis approach involved signal segmentation, time-domain feature extraction, feature selection (Fisher's discrimination), and classification (K-nearest neighbor).
• A deep learning approach utilized a Convolutional Neural Network (CNN) to directly process raw motor current signals, bypassing manual feature engineering.

Main Results:

• The traditional method achieved good accuracy after segmenting the current signal into three velocity regimes, enhancing discrimination.
• The CNN approach demonstrated superior accuracy using the original raw signal, offering simpler implementation.
• Despite CNN's accuracy, the traditional method's feature processing capabilities are more advantageous for assessing individual fault health and monitoring progression over time in operational settings.

Conclusions:

• Both traditional and deep learning methods can diagnose train door faults using motor current signals.
• The traditional feature-based method, despite requiring more processing steps, offers superior interpretability and practical utility for ongoing health monitoring in real train operations.
• The CNN approach, while accurate and simpler to implement, lacks the detailed diagnostic insights provided by the traditional method's feature analysis.