Scholarlink

Purpose <p>The accurate remaining useful life (RUL) prediction is crucial for predictive maintenance of mechanical equipment. Traditional deep learning models are difficult to extract deep spatio-temporal degradation features in the case of noise interference and limited annotated data.</p> Methods <p>To address this problem, we propose a novel semi-supervised residual denoising temporal convolutional additive self-attention network (RD-TCN-ASAM). In terms of unsupervised, we introduced a residual denoising (RD) module to learn the differences between noise and original data, extracting clean data from noisy datasets. For supervised learning, we designed an additive self-attention mechanism (ASAM) for modeling long sequence data, which is integrated into a temporal convolutional network (TCN) to form the TCN-ASAM architecture. The core advantage of TCN-ASAM is to capture the deep global degradation features and strengthen the dependence between sequence features to improve model performance.</p> Results <p>The RD-TCN-ASAM adopts a two-phase semi-supervision process to significantly improve the accuracy of RUL prediction by identifying the trend of spatio-temporal degradation in data.</p> Conclusion <p>Comparative experiments on bearing and turbofan engine datasets with the state-of-the-art methods demonstrate that the proposed method excels in extracting spatio-temporal degradation features and predicting RUL under complex operating conditions and noise interference.</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 Semi-supervised Temporal Convolutional Network with Residual Denoising and Additive Self-attention Mechanism for Bearing and Turbofan Engine RUL Prediction

  • 期刊论文

  • 发表日期：

    2026年3月14日

查看全文
  • Jiaxin Ji,
  • Youming Wang,
  • Tingting Qu

摘要

Purpose

The accurate remaining useful life (RUL) prediction is crucial for predictive maintenance of mechanical equipment. Traditional deep learning models are difficult to extract deep spatio-temporal degradation features in the case of noise interference and limited annotated data.

Methods

To address this problem, we propose a novel semi-supervised residual denoising temporal convolutional additive self-attention network (RD-TCN-ASAM). In terms of unsupervised, we introduced a residual denoising (RD) module to learn the differences between noise and original data, extracting clean data from noisy datasets. For supervised learning, we designed an additive self-attention mechanism (ASAM) for modeling long sequence data, which is integrated into a temporal convolutional network (TCN) to form the TCN-ASAM architecture. The core advantage of TCN-ASAM is to capture the deep global degradation features and strengthen the dependence between sequence features to improve model performance.

Results

The RD-TCN-ASAM adopts a two-phase semi-supervision process to significantly improve the accuracy of RUL prediction by identifying the trend of spatio-temporal degradation in data.

Conclusion

Comparative experiments on bearing and turbofan engine datasets with the state-of-the-art methods demonstrate that the proposed method excels in extracting spatio-temporal degradation features and predicting RUL under complex operating conditions and noise interference.