Scholarlink

<p>To address the difficulty in effectively segmenting wear states and the low identification accuracy in DC railway relay contact mechanisms, this paper proposes a wear state identification method that combines temporal convolutional autoencoder-physically constrained binary segmentation (TCAE-PCBS) and bidirectional long short-term memory-multi-head attention (BiLSTM-MHA) algorithms. Firstly, the importance of feature parameters is estimated using the random forest algorithm, and low-importance parameters are removed to construct an optimal input subset. Then, based on the TCAE-PCBS algorithm, temporal degradation information of feature parameters is extracted and integrated to construct one-dimensional degradation indicator data, which better conforms to the exponential degradation model. By applying physical constraints, the model’s computational efficiency and wear state segmentation performance are effectively improved. Ultimately, the BiLSTM-MHA identification model with automatic feature weight allocation is constructed, effectively suppressing interference from non-critical features, thereby significantly enhancing feature extraction capabilities and identification accuracy. The model’s interpretability is further enhanced using the SHAP method. Compared with the other five models, the BiLSTM-MHA model achieves superior state identification performance, providing a theoretical basis for state identification of DC railway relay contact mechanisms.</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

Research on the identification of wear states in DC railway relay contact mechanisms

  • 期刊论文

  • 发表日期：

    2026年7月6日

查看全文
  • Yankai Li,
  • Shuxin Liu,
  • Chaojian Xing

摘要

To address the difficulty in effectively segmenting wear states and the low identification accuracy in DC railway relay contact mechanisms, this paper proposes a wear state identification method that combines temporal convolutional autoencoder-physically constrained binary segmentation (TCAE-PCBS) and bidirectional long short-term memory-multi-head attention (BiLSTM-MHA) algorithms. Firstly, the importance of feature parameters is estimated using the random forest algorithm, and low-importance parameters are removed to construct an optimal input subset. Then, based on the TCAE-PCBS algorithm, temporal degradation information of feature parameters is extracted and integrated to construct one-dimensional degradation indicator data, which better conforms to the exponential degradation model. By applying physical constraints, the model’s computational efficiency and wear state segmentation performance are effectively improved. Ultimately, the BiLSTM-MHA identification model with automatic feature weight allocation is constructed, effectively suppressing interference from non-critical features, thereby significantly enhancing feature extraction capabilities and identification accuracy. The model’s interpretability is further enhanced using the SHAP method. Compared with the other five models, the BiLSTM-MHA model achieves superior state identification performance, providing a theoretical basis for state identification of DC railway relay contact mechanisms.