Zhao Jingyu, Xie Weihua, Yu Dong, Yang Qiang, Meng Songhe, Lyu Qihui
Science and Technology on Advanced Composites in Special Environment Laboratory, Harbin Institute of Technology, Harbin 150080, China.
School of Science, Harbin Institute of Technology, Shenzhen 518055, China.
Polymers (Basel). 2023 Mar 19;15(6):1520. doi: 10.3390/polym15061520.
Intelligent composite structures with self-aware functions are preferable for future aircrafts. The real-time location of damaged areas of composites is a key step. In this study, deep transfer learning was used to achieve the real-time location of damaged areas. The sensor network obtained acoustic emission signals from different damaged areas of the aluminum alloy plate. The acoustic emission time-domain signal is transformed into the input image by continuous wavelet transform. The convolutional neural network-based model automatically localized the damaged area by extracting features from the input image. A small amount of composite acoustic emission data was used to fine-tune some network parameters of the basic model through transfer learning. This enabled the model to classify the damaged area of composites. The accuracy of the transfer learning model trained with 900 samples is 96.38%, which is comparable to the accuracy of the model trained directly with 1800 samples; the training time of the former is only 17.68% of that of the latter. The proposed method can be easily adapted to new composite structures using transfer learning and a small dataset, providing a new idea for structural health monitoring.
具有自感知功能的智能复合材料结构对未来飞机而言更具优势。复合材料损伤区域的实时定位是关键一步。在本研究中，采用深度迁移学习来实现损伤区域的实时定位。传感器网络从铝合金板的不同损伤区域获取声发射信号。通过连续小波变换将声发射时域信号转换为输入图像。基于卷积神经网络的模型通过从输入图像中提取特征来自动定位损伤区域。利用少量复合材料声发射数据通过迁移学习对基础模型的一些网络参数进行微调。这使得该模型能够对复合材料的损伤区域进行分类。用900个样本训练的迁移学习模型的准确率为96.38%，与直接用1800个样本训练的模型准确率相当；前者的训练时间仅为后者的17.68%。所提出的方法利用迁移学习和小数据集能够轻松适应新的复合材料结构，为结构健康监测提供了新思路。
