Scholarlink

<p>Detecting load imbalances in floating slab track (FST) isolators is critical for railway safety and operational reliability, yet remains challenging due to complex, non-stationary vibration signals and the limitations of traditional diagnostic techniques. To address these challenges, we introduce FusionPathNet, a novel deep learning framework that combines multi-scale time-frequency feature extraction, temporal dependency modeling, and uncertainty-aware adaptive data augmentation for robust FST fault diagnosis. The framework uses a customized multi-path convolutional neural network (CNN) combining Res2Net and a feature pyramid network (FPN), optimized for FST vibration signals. A bidirectional long short-term memory (Bi-LSTM) captures temporal dependencies, and an adaptive interpolation strategy enhances data variability and generalization. Raw vibration signals are transformed into Mel spectrograms to reveal subtle fault patterns across temporal and frequency domains. Experimental validation on 4,800 real-world Shanghai Metro samples shows that FusionPathNet achieves 97.28% accuracy with a 4.28% false positive rate, outperforming existing baselines. The framework provides a scalable solution for intelligent condition monitoring and predictive maintenance in urban rail systems.</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

Uncertainty-aware multi-scale fusion for floating slab track isolator load imbalance fault diagnosis

  • 期刊论文

  • 发表日期：

    2025年9月9日

查看全文
  • Shi-Ming Lian,
  • Zai-Wei Li

摘要

Detecting load imbalances in floating slab track (FST) isolators is critical for railway safety and operational reliability, yet remains challenging due to complex, non-stationary vibration signals and the limitations of traditional diagnostic techniques. To address these challenges, we introduce FusionPathNet, a novel deep learning framework that combines multi-scale time-frequency feature extraction, temporal dependency modeling, and uncertainty-aware adaptive data augmentation for robust FST fault diagnosis. The framework uses a customized multi-path convolutional neural network (CNN) combining Res2Net and a feature pyramid network (FPN), optimized for FST vibration signals. A bidirectional long short-term memory (Bi-LSTM) captures temporal dependencies, and an adaptive interpolation strategy enhances data variability and generalization. Raw vibration signals are transformed into Mel spectrograms to reveal subtle fault patterns across temporal and frequency domains. Experimental validation on 4,800 real-world Shanghai Metro samples shows that FusionPathNet achieves 97.28% accuracy with a 4.28% false positive rate, outperforming existing baselines. The framework provides a scalable solution for intelligent condition monitoring and predictive maintenance in urban rail systems.