Scholarlink

<p>In this work, a hybrid LSTM-PINN fatigue life prediction model is proposed, integrating long short-term memory (LSTM) and physics-informed neural networks (PINN). Without processing the load path data, the raw strain data obtained from experiments, combined with the material’s mechanical properties, are utilized as inputs, while fatigue life is designated as the output. Three physical constraints are introduced: (1) a positive correlation exists between fatigue life and yield strength, (2) a positive correlation exists between fatigue life and tensile strength, and (3) fatigue life is capped at 1 × 10<sup>7</sup> cycles. The model’s performance is validated using a comprehensive dataset and compared to other machine learning models. The results demonstrate that the proposed model achieves superior predictive performance, with an <i>R</i><sup><i>2</i></sup> value of 0.930 and an <i>RMSE</i> of 0.210. Nearly, 85% of the predictions on the test set fall within the twice scatter band. In addition, the research shows that LSTM effectively captures relevant information from the loading path. Integrating physical constraints enhances predictive accuracy, accelerates model convergence, and reduces overfitting during training while aligning with the underlying data patterns.</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 multiaxial fatigue life prediction method based on LSTM-PINN

  • 期刊论文

  • 发表日期：

    2025年12月27日

查看全文
  • Meng Wang,
  • Hu Chen,
  • Enze Zhu,
  • Shi Shi,
  • Pujiang Huang,
  • Xingbo Fang,
  • Xiaohui Wei,
  • Hong Nie

摘要

In this work, a hybrid LSTM-PINN fatigue life prediction model is proposed, integrating long short-term memory (LSTM) and physics-informed neural networks (PINN). Without processing the load path data, the raw strain data obtained from experiments, combined with the material’s mechanical properties, are utilized as inputs, while fatigue life is designated as the output. Three physical constraints are introduced: (1) a positive correlation exists between fatigue life and yield strength, (2) a positive correlation exists between fatigue life and tensile strength, and (3) fatigue life is capped at 1 × 107 cycles. The model’s performance is validated using a comprehensive dataset and compared to other machine learning models. The results demonstrate that the proposed model achieves superior predictive performance, with an R2 value of 0.930 and an RMSE of 0.210. Nearly, 85% of the predictions on the test set fall within the twice scatter band. In addition, the research shows that LSTM effectively captures relevant information from the loading path. Integrating physical constraints enhances predictive accuracy, accelerates model convergence, and reduces overfitting during training while aligning with the underlying data patterns.