Sun Yajie, Zhou Xiaowen, Long Qian, Jian Mingchun
School of Computer Science, Nanjing University of Information Science & Technology, Nanjing 210044, China.
Materials (Basel). 2025 May 1;18(9):2081. doi: 10.3390/ma18092081.
To address the challenges of feature extraction from time-domain signals and imprecise damage localization in conventional plate structure damage identification methods, this study proposes an innovative damage localization approach integrating Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks. The one-dimensional signal data extracted from the aluminum plate is converted into a two-dimensional grayscale image, leveraging the advantages of CNN to accurately extract the information features of the damaged image on the aluminum plate. These extracted features are subsequently channeled into the LSTM network and the unique forgetting and memory mechanisms inherent in LSTM are employed to integrate the input feature information through a three-layer LSTM network, which is then fed into a fully connected layer for regression prediction. This method is not only an innovative application of joint deep learning methods in the field of damage detection but also accurately predicts the coordinates of the damage location, effectively overcoming the limitations of traditional damage localization methods. To validate the effectiveness of our proposed method, experiments were conducted on aluminum plates. The results demonstrate that our method shows strong performance in accurately localizing artificial damage on aluminum plates.
为应对传统板结构损伤识别方法中从时域信号提取特征以及损伤定位不精确的挑战，本研究提出一种将卷积神经网络（CNN）和长短期记忆（LSTM）网络相结合的创新损伤定位方法。从铝板提取的一维信号数据被转换为二维灰度图像，利用CNN的优势准确提取铝板上损伤图像的信息特征。随后将这些提取的特征导入LSTM网络，并运用LSTM固有的独特遗忘和记忆机制，通过三层LSTM网络整合输入的特征信息，然后将其输入到全连接层进行回归预测。该方法不仅是联合深度学习方法在损伤检测领域的创新应用，而且能准确预测损伤位置的坐标，有效克服了传统损伤定位方法的局限性。为验证所提方法的有效性，在铝板上进行了实验。结果表明，我们的方法在准确识别铝板上的人工损伤方面表现出强大性能。
