Scholarlink

<p>To improve surface quality and enhance machining efficiency in intelligent manufacturing, the implementation of intelligent tool condition monitoring is essential. However, existing intelligent monitoring methods often struggle with spectrally aliased noise prevalent in harsh industrial environments and suffer from the shallow integration of physical knowledge, leading to inadequate prediction accuracy. To tackle these challenges, this paper introduces a physics-informed prediction method combined with deep adaptive denoising. First, a signal processing module based on the adaptive spectral block is employed to perform intelligent denoising and enhance critical frequency components of the cutting signals through learnable frequency-domain filtering. Then, a feature extractor based on the deep residual shrinking network achieves deep noise suppression and feature extraction in the feature domain through an intrinsic adaptive soft-thresholding mechanism. Meanwhile, the feature extractor integrates a channel attention mechanism based on dual-pooling aggregation to emphasize salient features and incorporates a multi-scale convolutional module to efficiently capture feature patterns across various temporal scales. Subsequently, a bidirectional gated recurrent unit is adopted to model the long-range temporal correlations present in the feature sequence. Finally, the entire network is optimized within a PINN framework, which dynamically integrates the data-driven model’s fitting capability with the physical constraints of a non-linear wear accumulation model, utilizing an adaptive weighting mechanism considering the prediction error uncertainty. As revealed by experimental validation, the proposed model yields notable reductions in prediction error, with the MAE decreasing by 9.9–29.0% and the RMSE dropping by 2.0–20.1% relative to other advanced models. It further establishes a robust and reliable solution for high-precision tool wear prediction under intricate industrial conditions.</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

Physics-informed deep adaptive denoising and multi-scale feature fusion method for tool wear prediction

  • 期刊论文

  • 发表日期：

    2026年3月18日

查看全文
  • Miaoxian Guo,
  • Duofu Gong,
  • Zhiwen Huang,
  • Weicheng Guo,
  • Xiaohui Jiang,
  • Dejin Lv

摘要

To improve surface quality and enhance machining efficiency in intelligent manufacturing, the implementation of intelligent tool condition monitoring is essential. However, existing intelligent monitoring methods often struggle with spectrally aliased noise prevalent in harsh industrial environments and suffer from the shallow integration of physical knowledge, leading to inadequate prediction accuracy. To tackle these challenges, this paper introduces a physics-informed prediction method combined with deep adaptive denoising. First, a signal processing module based on the adaptive spectral block is employed to perform intelligent denoising and enhance critical frequency components of the cutting signals through learnable frequency-domain filtering. Then, a feature extractor based on the deep residual shrinking network achieves deep noise suppression and feature extraction in the feature domain through an intrinsic adaptive soft-thresholding mechanism. Meanwhile, the feature extractor integrates a channel attention mechanism based on dual-pooling aggregation to emphasize salient features and incorporates a multi-scale convolutional module to efficiently capture feature patterns across various temporal scales. Subsequently, a bidirectional gated recurrent unit is adopted to model the long-range temporal correlations present in the feature sequence. Finally, the entire network is optimized within a PINN framework, which dynamically integrates the data-driven model’s fitting capability with the physical constraints of a non-linear wear accumulation model, utilizing an adaptive weighting mechanism considering the prediction error uncertainty. As revealed by experimental validation, the proposed model yields notable reductions in prediction error, with the MAE decreasing by 9.9–29.0% and the RMSE dropping by 2.0–20.1% relative to other advanced models. It further establishes a robust and reliable solution for high-precision tool wear prediction under intricate industrial conditions.