Scholarlink

<p>Laser shock peening (LSP) monitoring based on acoustic emission (AE) technology not only needs to achieve the desired monitoring accuracy, but also faces the challenges posed by the transmission and storage of high-dimensional time-series data. Existing methods mainly consider the former singularly while ignoring the latter. To address this issue, this study proposes an autoencoder-based data feature selection and a decision tree–based data feature identification method for the task of real-time LSP-AE monitoring. Specifically, the autoencoder takes deformable convolution as the core unit, which can fully consider the global and local features in the time-varying AE signals, and guide the model to obtain more valuable feature vectors through offset calculation. The decision tree model can process the encoded features efficiently and accurately, which in turn enables real-time monitoring of the laser processing quality. The encoding of high-dimensional AE signals facilitates efficient data storage, and the encoded features are more portable and operable. The feasibility and reliability of the proposed method are verified based on LSP experiments. Compared with other methods, the proposed method can simultaneously meet the requirements of monitoring accuracy and data storage by encoding the original signal. Specifically, the original time series signal with dimension 4050 is reduced to 128 dimensions and has an optimal recognition accuracy of 98.76%.</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

Deformable convolutional autoencoder-based feature selection and recognition for acoustic emission monitoring in laser shock peening

  • 期刊论文

  • 发表日期：

    2025年3月5日

查看全文
  • Rui Qin,
  • Zhifen Zhang,
  • Jing Huang,
  • Yu Su,
  • Guangrui Wen,
  • Weifeng He,
  • Xuefeng Chen

摘要

Laser shock peening (LSP) monitoring based on acoustic emission (AE) technology not only needs to achieve the desired monitoring accuracy, but also faces the challenges posed by the transmission and storage of high-dimensional time-series data. Existing methods mainly consider the former singularly while ignoring the latter. To address this issue, this study proposes an autoencoder-based data feature selection and a decision tree–based data feature identification method for the task of real-time LSP-AE monitoring. Specifically, the autoencoder takes deformable convolution as the core unit, which can fully consider the global and local features in the time-varying AE signals, and guide the model to obtain more valuable feature vectors through offset calculation. The decision tree model can process the encoded features efficiently and accurately, which in turn enables real-time monitoring of the laser processing quality. The encoding of high-dimensional AE signals facilitates efficient data storage, and the encoded features are more portable and operable. The feasibility and reliability of the proposed method are verified based on LSP experiments. Compared with other methods, the proposed method can simultaneously meet the requirements of monitoring accuracy and data storage by encoding the original signal. Specifically, the original time series signal with dimension 4050 is reduced to 128 dimensions and has an optimal recognition accuracy of 98.76%.