A Fault-Diagnosis Method for Railway Turnout Systems Based on Improved Autoencoder and Data Augmentation.

Area of Science:

• Railway engineering
• Artificial intelligence
• Signal processing

Background:

• Railway transportation security is critical, with railway turnout systems (RTS) being vital infrastructure.
• Traditional fault diagnosis methods struggle with the increasing data volume and complexity in RTS.
• Intelligent fault diagnosis is a research hotspot due to limitations in traditional approaches.

Purpose of the Study:

• To address the challenges of deep feature extraction and accurate fault identification in unbalanced datasets for RTS.
• To propose an intelligent fault diagnosis method for railway turnout systems.
• To improve the precision and generalization ability of fault diagnosis in railway transportation.

Main Methods:

• An improved autoencoder was developed for noise smoothing and deep feature extraction from RTS signals.
• Synthetic Minority Oversampling Technology (SMOTE) was employed to address unbalanced datasets by expanding fault types.
• A Softmax regression model was trained on balanced feature data for health state identification.

Main Results:

• The proposed method effectively extracts deep features and identifies fault modes in unbalanced RTS data.
• The integration of an improved autoencoder and SMOTE significantly enhances fault diagnosis capabilities.
• Experiments on a real-world railway dataset achieved an average diagnostic accuracy of 99.13%.

Conclusions:

• The developed fault diagnosis method is effective and feasible for railway turnout systems.
• The approach overcomes limitations of traditional methods in handling complex and unbalanced data.
• This intelligent system offers superior diagnostic accuracy and generalization ability for railway safety maintenance.