Scholarlink

<p>Abrasive water jet (AWJ) offers the benefits of cold machining and high efficiency for difficult-to-machine materials. However, predicting the separation speed of different materials using AWJ is extremely challenging. Mechanistic models can reduce trial cuts but suffer from large errors in complex scenarios. Data-driven models capture nonlinear relationships yet lack physical interpretability and are prone to noise and temporal degradation. To address the limitations of a single model, this study develops a hybrid neural network. The framework deeply integrates mechanistic knowledge with data. It combines a Degradation-Aware Long Short-Term Memory (LSTM) with a Physics Correction Neural Network (PCNN), forming the DaLSTM-PCNN model. The framework employs a degradation-aware LSTM module to adaptively extract latent temporal features and incorporates Gaussian probabilistic modeling to establish confidence intervals for each prediction. It combines multi‐scale physical constraints to organically fuse data fitting with physical priors. This design endows the model with physical interpretability and enhances its predictive robustness. The fusion model was established on an experimental dataset. The results show that the model achieves R² = 0.99393 and a predictive MAPE of 6.45%. Validation on out-of-sample materials (organic ultra-clear glass and organic green glass) further confirmed its accuracy, yielding improvement ratios of 93.74% and 97.08% over the physics-based model. These results demonstrate that DaLSTM-PCNN not only surpasses physics-only approaches in prediction accuracy but also offers tangible industrial value by reducing trial cuts and material consumption in real waterjet cutting processes. In conclusion the method could offer a new paradigm for adaptive noise suppression and probabilistic uncertainty modeling.</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 degeneration-aware LSTM-PCNN fusion network for predicting abrasive water jet’s separation speed

  • 期刊论文

  • 发表日期：

    2025年10月24日

查看全文
  • Chenghan Huang,
  • Shijin Zhang,
  • Ming Chen,
  • Yuqiang Wu

摘要

Abrasive water jet (AWJ) offers the benefits of cold machining and high efficiency for difficult-to-machine materials. However, predicting the separation speed of different materials using AWJ is extremely challenging. Mechanistic models can reduce trial cuts but suffer from large errors in complex scenarios. Data-driven models capture nonlinear relationships yet lack physical interpretability and are prone to noise and temporal degradation. To address the limitations of a single model, this study develops a hybrid neural network. The framework deeply integrates mechanistic knowledge with data. It combines a Degradation-Aware Long Short-Term Memory (LSTM) with a Physics Correction Neural Network (PCNN), forming the DaLSTM-PCNN model. The framework employs a degradation-aware LSTM module to adaptively extract latent temporal features and incorporates Gaussian probabilistic modeling to establish confidence intervals for each prediction. It combines multi‐scale physical constraints to organically fuse data fitting with physical priors. This design endows the model with physical interpretability and enhances its predictive robustness. The fusion model was established on an experimental dataset. The results show that the model achieves R² = 0.99393 and a predictive MAPE of 6.45%. Validation on out-of-sample materials (organic ultra-clear glass and organic green glass) further confirmed its accuracy, yielding improvement ratios of 93.74% and 97.08% over the physics-based model. These results demonstrate that DaLSTM-PCNN not only surpasses physics-only approaches in prediction accuracy but also offers tangible industrial value by reducing trial cuts and material consumption in real waterjet cutting processes. In conclusion the method could offer a new paradigm for adaptive noise suppression and probabilistic uncertainty modeling.