Scholarlink

<p>To address the limitations of traditional metro equipment maintenance, such as heavy reliance on manual experience, poor mobility, difficulty assessing equipment health, and the inability to meet actual operation and maintenance needs, this paper proposes a health prediction method for metro platform screen door (PSD) system based on an XGBoost-BiGRU-Attention model optimized with the Lookahead optimizer. Using the PSD system’s fault records and control principles, data were collected for 14 characteristic parameters. The XGBoost algorithm was applied to rank these variables in order of importance, and the top-ranking characteristic parameters were selected as inputs to the prediction model. An attention mechanism was then integrated into a Bidirectional Gated Recurrent Unit (BiGRU) network using the Lookahead optimizer to improve prediction performance. Results show that, compared with a standard BiGRU model, the proposed XGBoost-BiGRU-Attention model reduces Root Mean Square Error (RMSE) by 52.4%, Mean Absolute Error (MAE) by 57.1%, and Mean Absolute Percentage Error (MAPE) by 57.0%, achieving a coefficient of determination of 0.972. These findings indicate that the proposed model offers strong practicality and feasibility for health management of metro equipment.</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

Health prediction of metro platform screen door system based on XGBoost-BiGRU-attention model

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月26日

查看全文
  • Zijun Zhu,
  • Shaohu Tang,
  • Liang Zhang,
  • Hailin Kang,
  • HongKang Song,
  • Pengyu Li

摘要

To address the limitations of traditional metro equipment maintenance, such as heavy reliance on manual experience, poor mobility, difficulty assessing equipment health, and the inability to meet actual operation and maintenance needs, this paper proposes a health prediction method for metro platform screen door (PSD) system based on an XGBoost-BiGRU-Attention model optimized with the Lookahead optimizer. Using the PSD system’s fault records and control principles, data were collected for 14 characteristic parameters. The XGBoost algorithm was applied to rank these variables in order of importance, and the top-ranking characteristic parameters were selected as inputs to the prediction model. An attention mechanism was then integrated into a Bidirectional Gated Recurrent Unit (BiGRU) network using the Lookahead optimizer to improve prediction performance. Results show that, compared with a standard BiGRU model, the proposed XGBoost-BiGRU-Attention model reduces Root Mean Square Error (RMSE) by 52.4%, Mean Absolute Error (MAE) by 57.1%, and Mean Absolute Percentage Error (MAPE) by 57.0%, achieving a coefficient of determination of 0.972. These findings indicate that the proposed model offers strong practicality and feasibility for health management of metro equipment.