An Online Classification Method for Fault Diagnosis of Railway Turnouts

Dongxiu Ou, Yuqing Ji, Lei Zhang, Hu LiuSensors 2020 · DOI

TL;DR

This paper introduces a novel online fault diagnosis method for railway turnouts that leverages Bayesian incremental learning and scalable recognition. This approach effectively addresses the challenges of imbalanced data and the emergence of new fault types, leading to improved accuracy and efficiency in railway maintenance.

What Problem Does This Paper Address?

Railway turnouts are critical components for railway safety and efficiency, but they are also prone to failures. Existing fault diagnosis methods often struggle with real-time updating and identifying novel fault types, especially when dealing with imbalanced datasets where normal operational data significantly outweighs fault data. This limitation hinders timely maintenance and can lead to missed critical failures. The paper aims to bridge this gap by developing a system that can learn continuously and adapt to new fault patterns, making it more practical for real-world railway operations.

Methods

The proposed method combines several key techniques:

  • Knowledge Graph-Based Feature Engineering: A knowledge graph is used to represent the structure and relationships within turnout monitoring data. This helps in extracting relevant features, such as those related to current and power consumption during different operational stages of the turnout.
  • Imbalanced Data Preprocessing: To combat the scarcity of fault data, the SMOTE (Synthetic Minority Over-sampling Technique) is employed. SMOTE generates synthetic samples for the minority (fault) class, balancing the dataset for better model training.
  • Bayesian Incremental Learning: This core component allows the model to learn from new data batches without retraining from scratch. It updates the prior and conditional probabilities of the Bayesian classifier, enabling continuous adaptation. The method uses Laplace estimation for probability calculations and a specific judgment condition to decide when to update the model.
  • Scalable Fault Recognition: To handle new, previously unseen fault types, the paper utilizes the Local Outlier Factor (LOF) algorithm. LOF is an efficient novelty detection method that identifies data points significantly different from the known clusters (normal operation or known faults). If a sample is identified as an outlier across multiple binary classification models, it's flagged as a new class.

Datasets

The experimental data used in this study comes from several railway stations of Guangzhou Railway, collected between 2016 and 2018. The paper highlights the imbalanced nature of this real-world data, with abnormal samples constituting less than 0.1% of the total. For experimental purposes, a dataset was organized with 1000 normal samples and 10 samples for each of the 11 fault types, resulting in an initial dataset of 12,000 samples after SMOTE preprocessing.

Key Insights and Ideas

  • The "trap" of imbalanced data: The paper vividly illustrates how imbalanced data can lead to misleadingly high accuracy metrics if not properly addressed. A classifier that always predicts "normal" can achieve near-perfect accuracy but completely miss critical faults.
  • Incremental learning is crucial for adaptability: The dynamic nature of railway systems, with potential for new fault types and evolving operational conditions, necessitates a learning model that can adapt over time. Incremental learning provides this adaptability without the prohibitive cost of full retraining.
  • LOF for detecting the "unknown unknown": The use of LOF for scalable fault recognition is a clever way to identify entirely new fault patterns that weren't present in the initial training data. This moves beyond simply classifying known faults to detecting novel anomalies.
  • Knowledge graphs as a structured approach to feature engineering: Representing domain knowledge in a graph structure helps in systematically identifying and organizing relevant features, which is vital for complex systems like railway turnouts.

Key Results

  • The proposed scalable fault recognition method achieved an accuracy of 99.11%.
  • The incremental Bayesian learning model demonstrated a reduction in training time of approximately 30% compared to the original model when dealing with a large dataset (12 million samples), without a significant decrease in accuracy (average improvement of 0.32%).
  • The method showed improved accuracy for all 11 fault types when compared to models without scalable fault recognition, with an average improvement of about 6.8%.

Limitations and Future Work

The authors acknowledge that the proposed solution is applicable to most railway turnout systems, with the exception of those using three-phase AC type switch machines.

Future research directions suggested include:

  • Investigating more intelligent methods to further enhance the accuracy of scalable fault recognition.
  • Improving the structure of the knowledge graph for even better feature representation and understanding.

Summary updated

BibTeX
@article{ou2020online,
  author = {Dongxiu Ou and Yuqing Ji and Lei Zhang and Hu Liu},
  title = {{An Online Classification Method for Fault Diagnosis of Railway Turnouts}},
  journal = {Sensors},
  year = {2020},
  doi = {10.3390/s20164627},
}
Figure