Scholarlink

Introduction <p>Vibration data collected by in-service vehicle measurement systems provide valuable insights for assessing track quality and detecting faults early. Due to the large volume of data, machine learning methods have gained significant interest in processing and analysis. However, existing publications apply diverse strategies that require systematic classification and discussion.</p> Purpose <p>This article proposes a systematic mapping of&#xa0;machine learning methods for&#xa0;fault detection in track monitoring,&#xa0;using vibration data collected from in-service vehicles.</p> Method <p>Publications were collected from reliable databases following&#xa0;PICOC (Population, Intervention, Comparison, Outcome, and Context) criteria.&#xa0;The review explores training data sources, prediction strategies, preprocessing, model accuracy, and deployment.</p> Results <p>The analysis reveals that most studies rely on real-world data, followed by simulation data. Although real-world data enhances reliability, challenges like limited public datasets, extensive data processing, and label validation difficulties remain. Four primary fault prediction strategies are identified, including direct fault classification, inverse prediction, anomaly detection based on normal conditions, and primary acceleration prediction. The study also examines preprocessing techniques, highlighting the critical role of location in vibration-based monitoring. Machine learning approaches include classical, deep learning, ensemble, unsupervised, and hybrid methods.&#xa0;Notably,&#xa0;Deep learning-based methods dominate onboard monitoring applications. For simple cases, classical methods&#xa0;suffice, while advanced tasks require deep learning or ensemble methods at higher computational cost. Unsupervised techniques for dimensionality reduction and signal decomposition should&#xa0;be further explored to boost data efficiency.&#xa0;The review identifies key challenges, including real-time detection system&#xa0;deployment, multi-fault detection, and&#xa0;lightweight&#xa0;models.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Roles of Vibration-Based Machine Learning Algorithms in Railway Vehicle Monitoring for Track Condition Assessment: A Review

  • 期刊论文

  • 发表日期：

    2025年3月21日

查看全文
  • Agustinus Winarno,
  • Rienetta Ichmawati Delia Sandhy,
  • Nurhazimah Nazmi,
  • Herjuno Rizki Priatomo,
  • Addin Suwastono,
  • Giovani Ega Charisma,
  • Irfan Bahiuddin

摘要

Introduction

Vibration data collected by in-service vehicle measurement systems provide valuable insights for assessing track quality and detecting faults early. Due to the large volume of data, machine learning methods have gained significant interest in processing and analysis. However, existing publications apply diverse strategies that require systematic classification and discussion.

Purpose

This article proposes a systematic mapping of machine learning methods for fault detection in track monitoring, using vibration data collected from in-service vehicles.

Method

Publications were collected from reliable databases following PICOC (Population, Intervention, Comparison, Outcome, and Context) criteria. The review explores training data sources, prediction strategies, preprocessing, model accuracy, and deployment.

Results

The analysis reveals that most studies rely on real-world data, followed by simulation data. Although real-world data enhances reliability, challenges like limited public datasets, extensive data processing, and label validation difficulties remain. Four primary fault prediction strategies are identified, including direct fault classification, inverse prediction, anomaly detection based on normal conditions, and primary acceleration prediction. The study also examines preprocessing techniques, highlighting the critical role of location in vibration-based monitoring. Machine learning approaches include classical, deep learning, ensemble, unsupervised, and hybrid methods. Notably, Deep learning-based methods dominate onboard monitoring applications. For simple cases, classical methods suffice, while advanced tasks require deep learning or ensemble methods at higher computational cost. Unsupervised techniques for dimensionality reduction and signal decomposition should be further explored to boost data efficiency. The review identifies key challenges, including real-time detection system deployment, multi-fault detection, and lightweight models.