Scholarlink

<p>Laser melt injection (LMI) is a novel approach for localized surface enhancement and remanufacturing upgrades of high value-added parts by the strategic selection of ceramic particles that are injected into the molten pool. However, ceramic reinforced metal matrix composites (MMCs) with complex material systems and highly variable properties exhibit complex quality states in the LMI process that cannot be described in a single dimension. This study presents promising findings aimed at bridging the knowledge gap by identifying a new approach for the cross-scale in-situ monitoring of LMI. A physical feature-assisted multi-task convolutional neural network (CNN) model is proposed to simultaneously monitor visible external cracks and invisible internal thermal damage to ceramic particles in MMCs. A prototype man‒computer interaction system is used to apply the proposed model. The performances of the multi-task model with different loss design methods, single-task model, and machine learning model are compared in terms of the accuracy of predicting surface cracks (<i>ACC</i><sub><i>C</i></sub>), percentage error in predicting thermal damage to particles (<i>MAPE</i><sub><i>R</i></sub>), and proposed inference efficiency coefficient. The results show that the proposed multi-task model reaches 1.7−1.8 times the model inference efficiency of the single-task learning model by avoiding complex parallel operations and inefficient repetitive feature extraction processes. The multi-task model with the MobileNetV3 backbone achieves optimal <i>ACCc</i> and <i>MAPE</i><sub><i>R</i></sub> values of 96.11% and 3.68%, respectively. The loss weighting design methods of dynamic weight average (DWA) and random loss weighting (RLW) can further optimize the prediction precision for both tasks without extra computation time.</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

Multi-task deep learning-empowered in-situ monitoring methodology for visible and invisible defects in laser melt injection: ceramic reinforced metal matrix composite

  • 期刊论文

  • 发表日期：

    2025年8月4日

查看全文
  • Hongmeng Xu,
  • Xixi Li,
  • Zhengchun Qian,
  • Huanbo Cheng,
  • Wenzheng Ding

摘要

Laser melt injection (LMI) is a novel approach for localized surface enhancement and remanufacturing upgrades of high value-added parts by the strategic selection of ceramic particles that are injected into the molten pool. However, ceramic reinforced metal matrix composites (MMCs) with complex material systems and highly variable properties exhibit complex quality states in the LMI process that cannot be described in a single dimension. This study presents promising findings aimed at bridging the knowledge gap by identifying a new approach for the cross-scale in-situ monitoring of LMI. A physical feature-assisted multi-task convolutional neural network (CNN) model is proposed to simultaneously monitor visible external cracks and invisible internal thermal damage to ceramic particles in MMCs. A prototype man‒computer interaction system is used to apply the proposed model. The performances of the multi-task model with different loss design methods, single-task model, and machine learning model are compared in terms of the accuracy of predicting surface cracks (ACCC), percentage error in predicting thermal damage to particles (MAPER), and proposed inference efficiency coefficient. The results show that the proposed multi-task model reaches 1.7−1.8 times the model inference efficiency of the single-task learning model by avoiding complex parallel operations and inefficient repetitive feature extraction processes. The multi-task model with the MobileNetV3 backbone achieves optimal ACCc and MAPER values of 96.11% and 3.68%, respectively. The loss weighting design methods of dynamic weight average (DWA) and random loss weighting (RLW) can further optimize the prediction precision for both tasks without extra computation time.