Scholarlink

<p>Accurate prediction of the remaining useful life of water filter cartridges is critical for ensuring water quality in both domestic and industrial applications. Conventional methods based on physical or chemical measurements are often time-consuming, costly, and reliant on expert knowledge. While data-driven time series models present a viable alternative, they frequently fail to capture long-range dependencies in complex, high-dimensional sensor data, limiting their feature extraction capability. To address these bottlenecks, this paper proposes a novel triple-fusion architecture that integrates Temporal Convolutional Networks, a Gated Attention Mechanism, and an LSTM module. The primary contributions are threefold: (1) Employing dilated causal convolutions to mitigate gradient vanishing and enable ultra-deep temporal feature extraction; (2) Introducing a gated attention mechanism during feature decoding to adaptively focus on critical time steps by dynamically weighting features, thereby enhancing modeling sensitivity; (3) Designing an HTCA-LSTM cell to improve the capture of long-term dependencies and high-dimensional complexity of key features. Experimental results demonstrate the model’s superior performance across prediction horizons of 96, 172, and 720 steps, with average reductions in MAE and MSE of 6.84–11.51%, and , respectively. This validates its efficacy in modeling long-term dependencies and extracting complex features, showcasing significant generalization ability and application potential for RUL prediction.</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

A hybrid temporal convolutional attention model for water filter remaining useful life prediction

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月4日

查看全文
  • Jia Chen,
  • Yaqiong Yang,
  • Liang Su

摘要

Accurate prediction of the remaining useful life of water filter cartridges is critical for ensuring water quality in both domestic and industrial applications. Conventional methods based on physical or chemical measurements are often time-consuming, costly, and reliant on expert knowledge. While data-driven time series models present a viable alternative, they frequently fail to capture long-range dependencies in complex, high-dimensional sensor data, limiting their feature extraction capability. To address these bottlenecks, this paper proposes a novel triple-fusion architecture that integrates Temporal Convolutional Networks, a Gated Attention Mechanism, and an LSTM module. The primary contributions are threefold: (1) Employing dilated causal convolutions to mitigate gradient vanishing and enable ultra-deep temporal feature extraction; (2) Introducing a gated attention mechanism during feature decoding to adaptively focus on critical time steps by dynamically weighting features, thereby enhancing modeling sensitivity; (3) Designing an HTCA-LSTM cell to improve the capture of long-term dependencies and high-dimensional complexity of key features. Experimental results demonstrate the model’s superior performance across prediction horizons of 96, 172, and 720 steps, with average reductions in MAE and MSE of 6.84–11.51%, and , respectively. This validates its efficacy in modeling long-term dependencies and extracting complex features, showcasing significant generalization ability and application potential for RUL prediction.