Scholarlink

<p>Drilling of unidirectional carbon fiber reinforced polymer sheets often leads to damage around the hole, compromising the mechanical integrity of composite structures. To address this challenge, a deep learning framework was developed to predict drilling-induced damage images using process parameters as inputs. A convolutional autoencoder (CAE) was first employed to augment the limited experimental dataset by generating synthetic grayscale damage images. Subsequently, a multi-layer perceptron (MLP) decoder model was trained to predict damage images based on spindle speed and feed rate. Three CAE architectures were evaluated, with CAE Type I achieving the lowest reconstruction error in the damage area, with an error of 10.14% and an R<sup>2</sup> value of 0.9862. Four MLP-Decoder models were tested using different combinations of original and CAE-generated images. The model trained with both original and CAE Type I images (MLP-Decoder Type B) showed the highest prediction accuracy, with an MSE of 1.13 and a predicted damage area of 36.92 mm<sup>2</sup>, which is closer to the validation data. Comparative analysis against experimentally validated images demonstrated that the proposed framework can effectively estimate drilling damage 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

Estimation of Damaged Surface Images in Drilled Unidirectional Carbon Fiber Reinforced Polymer Sheets Using Convolutional Autoencoder and Multi-Layer Perceptron Decoder

  • 期刊论文

  • 发表日期：

    2025年9月23日

查看全文
  • Eddy Kurniawan,
  • Yong Chan Hur,
  • Ji Hoon Kim

摘要

Drilling of unidirectional carbon fiber reinforced polymer sheets often leads to damage around the hole, compromising the mechanical integrity of composite structures. To address this challenge, a deep learning framework was developed to predict drilling-induced damage images using process parameters as inputs. A convolutional autoencoder (CAE) was first employed to augment the limited experimental dataset by generating synthetic grayscale damage images. Subsequently, a multi-layer perceptron (MLP) decoder model was trained to predict damage images based on spindle speed and feed rate. Three CAE architectures were evaluated, with CAE Type I achieving the lowest reconstruction error in the damage area, with an error of 10.14% and an R2 value of 0.9862. Four MLP-Decoder models were tested using different combinations of original and CAE-generated images. The model trained with both original and CAE Type I images (MLP-Decoder Type B) showed the highest prediction accuracy, with an MSE of 1.13 and a predicted damage area of 36.92 mm2, which is closer to the validation data. Comparative analysis against experimentally validated images demonstrated that the proposed framework can effectively estimate drilling damage patterns.