Scholarlink

<p>Adhesive joints are widely used in industries such as aerospace and automotive due to their lightweight and high mechanical performance. However, weak adhesion remains a significant issue affecting the structural integrity of these joints. Current detection methods of weak adhesion rely on destructive testing, which limits the widespread use of adhesive primary structures. This study proposes a novel nondestructive testing (NDT) technique to detect, evaluate the intensity, and localize weak adhesion in single lap joints (SLJs) using lamb waves (LWs) and machine learning (ML). The aim is to develop a ML-based pipeline capable of identifying weak adhesion with high accuracy and sensitivity, based on data from simulated and experimental SLJ samples. The proposed technique integrates LW data with convolutional neural networks (CNNs) in a ML pipeline for weak adhesion detection in SLJs. The use of a large simulated dataset combined with transfer learning allows for effective adaptation to experimental conditions, improving both the detection and localization of damage. This approach offers a significant advancement over traditional destructive testing techniques. The pipeline begins with the generation of simulated LW time-series data for SLJs with varying adhesion levels, damage locations, and sizes. After preprocessing, the data are input into a CNN, which is initially trained on synthetic data. Transfer learning is employed to fine-tune the model using a small experimental dataset. The final trained model is then applied to detect weak adhesion, estimate its intensity, and localize the damage. The proposed pipeline demonstrated high performance in both simulated and experimental datasets: regarding detection, the algorithm achieved over 95.3% accuracy in identifying damage from simulated data and near 100% detection of damaged cases in experimental data; for intensity estimation, the algorithm showed an average loss of approximately 45 MPa for weak adhesion intensity in experimental validation, with an average error of about 140 MPa and a best-case error of just near 3.6 MPa; in terms of localization, the average localization error was approximately 8&#xa0;mm in the synthetic validation dataset; with respect to flexibility, the methodology is adaptable to different damage characteristics, such as existence, intensity, and localization, without requiring substantial modifications. Summing up, this study presents a novel NDT approach using ML and LW data that significantly improves the detection, evaluation, and localization of weak adhesion in adhesive joints. Its high accuracy and adaptability have the potential to enhance structural health monitoring, ensuring the safety and durability of bonded structures in critical industries.</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

Detecting and assessing weak adhesion in structural single lap joints using a machine learning pipeline with lamb waves data

  • 期刊论文

  • 发表日期：

    2025年1月4日

查看全文
  • Gabriel M. F. Ramalho,
  • António M. Lopes,
  • Lucas F. M. da Silva

摘要

Adhesive joints are widely used in industries such as aerospace and automotive due to their lightweight and high mechanical performance. However, weak adhesion remains a significant issue affecting the structural integrity of these joints. Current detection methods of weak adhesion rely on destructive testing, which limits the widespread use of adhesive primary structures. This study proposes a novel nondestructive testing (NDT) technique to detect, evaluate the intensity, and localize weak adhesion in single lap joints (SLJs) using lamb waves (LWs) and machine learning (ML). The aim is to develop a ML-based pipeline capable of identifying weak adhesion with high accuracy and sensitivity, based on data from simulated and experimental SLJ samples. The proposed technique integrates LW data with convolutional neural networks (CNNs) in a ML pipeline for weak adhesion detection in SLJs. The use of a large simulated dataset combined with transfer learning allows for effective adaptation to experimental conditions, improving both the detection and localization of damage. This approach offers a significant advancement over traditional destructive testing techniques. The pipeline begins with the generation of simulated LW time-series data for SLJs with varying adhesion levels, damage locations, and sizes. After preprocessing, the data are input into a CNN, which is initially trained on synthetic data. Transfer learning is employed to fine-tune the model using a small experimental dataset. The final trained model is then applied to detect weak adhesion, estimate its intensity, and localize the damage. The proposed pipeline demonstrated high performance in both simulated and experimental datasets: regarding detection, the algorithm achieved over 95.3% accuracy in identifying damage from simulated data and near 100% detection of damaged cases in experimental data; for intensity estimation, the algorithm showed an average loss of approximately 45 MPa for weak adhesion intensity in experimental validation, with an average error of about 140 MPa and a best-case error of just near 3.6 MPa; in terms of localization, the average localization error was approximately 8 mm in the synthetic validation dataset; with respect to flexibility, the methodology is adaptable to different damage characteristics, such as existence, intensity, and localization, without requiring substantial modifications. Summing up, this study presents a novel NDT approach using ML and LW data that significantly improves the detection, evaluation, and localization of weak adhesion in adhesive joints. Its high accuracy and adaptability have the potential to enhance structural health monitoring, ensuring the safety and durability of bonded structures in critical industries.