Scholarlink

<p>The fundamental goal of prognostics is to predict a system or component’s remaining useful life (RUL), or the span of time until it fails to operate as intended, which is crucial for reducing downtime. Recently, various deep learning approaches have been used for machine prognostics. The sequential models like RNN, LSTM, GRU, and Transformer models can effectively manage implicit contexts defined within the data, such as temporal contexts. Further, the attention mechanism enables the model to focus on different parts of the input sequence. However, such models may struggle with explicitly incorporating external contextual information such as machine operating conditions. In this paper, a novel recurrent neural network (RNN)-based architecture is proposed for RUL prediction that explicitly embeds contextual information into the model. Context influences both the input feature weighting and the computation of the attention vector, resulting in a context-aware attention mechanism. This not only improves prediction performance but also provides interpretability by highlighting the most influential features and parts of input sequences. The proposed architecture is evaluated using two widely adopted benchmark datasets: the NASA C-MPASS Turbofan Engine Degradation Simulation dataset and the Bearing dataset (IEEE PHM 2012 Challenge). Results show that incorporating context significantly improves performance over baseline and state-of-the-art methods. The model achieves up to 18.90% and 59.10% reductions in RMSE and Score on FD002, and 6.86% and 53.88% on FD004 (C-MPASS). On the Bearings dataset, it yields an 84.54% improvement in RMSE and 54.98% in Score over the best existing model. Notably, these gains are achieved using a lightweight model architecture with minimal parameters, demonstrating both effectiveness and efficiency.</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

Recurrent neural network with contextual attention for prognostics

  • 期刊论文

  • 发表日期：

    2025年9月30日

查看全文
  • Rashmi Dutta Baruah,
  • Mario Munoz-Organero

摘要

The fundamental goal of prognostics is to predict a system or component’s remaining useful life (RUL), or the span of time until it fails to operate as intended, which is crucial for reducing downtime. Recently, various deep learning approaches have been used for machine prognostics. The sequential models like RNN, LSTM, GRU, and Transformer models can effectively manage implicit contexts defined within the data, such as temporal contexts. Further, the attention mechanism enables the model to focus on different parts of the input sequence. However, such models may struggle with explicitly incorporating external contextual information such as machine operating conditions. In this paper, a novel recurrent neural network (RNN)-based architecture is proposed for RUL prediction that explicitly embeds contextual information into the model. Context influences both the input feature weighting and the computation of the attention vector, resulting in a context-aware attention mechanism. This not only improves prediction performance but also provides interpretability by highlighting the most influential features and parts of input sequences. The proposed architecture is evaluated using two widely adopted benchmark datasets: the NASA C-MPASS Turbofan Engine Degradation Simulation dataset and the Bearing dataset (IEEE PHM 2012 Challenge). Results show that incorporating context significantly improves performance over baseline and state-of-the-art methods. The model achieves up to 18.90% and 59.10% reductions in RMSE and Score on FD002, and 6.86% and 53.88% on FD004 (C-MPASS). On the Bearings dataset, it yields an 84.54% improvement in RMSE and 54.98% in Score over the best existing model. Notably, these gains are achieved using a lightweight model architecture with minimal parameters, demonstrating both effectiveness and efficiency.