基于LSTM-ResNet模型的时变结构损伤检测

Time-varying structural damage detection based on LSTM-ResNet model

摘要:

面向火箭结构健康监测，提出了一种基于深度学习的损伤检测方法，直接将多个通道的振动数据作为输入，并基于由长短时记忆网络LSTM（Long Short-Term Memory Networks）和残差卷积神经网络ResNet（Residual Convolutional Neural Networks）组合而成的LSTM-ResNet网络进行损伤识别。其优点在于，首先利用LSTM提取信号的时间依赖特征，减轻了由某些通道信号缺失带来的影响，再利用ResNet在不损耗特征的情况下进一步提取空间特征，提高了训练效率和损伤辨识准确性。通过充液圆筒振动放水实验模拟火箭飞行状态下的燃料消耗，并基于自主构建的数据集和公用数据集对LSTM-ResNet、LSTM、ResNet以及ResNet-LSTM网络进行了训练，训练结果表明，LSTM-ResNet组合网络无论在传感器是否存在故障的情况下都具有更好的性能，损伤检测精度更高。

Abstract:

For rocket structural health monitoring, this paper proposes a damage detection method based on deep learning. This method directly takes the vibration data of multiple channels as input, and performs damage identification based on LSTM-ResNet model composed of the long short-term memory structure (LSTM) and residual convolutional neural structure (ResNet). The advantages are that firstly, LSTM is used to extract the time-dependent features of the signal, which reduces the impact caused by the lack of some channel signals, and then ResNet is used to further extract spatial features without loss of features, which improves the training efficiency and the accuracy of damage identification. In this paper, a liquid-filled cylinder vibration and water discharge experiment is used to simulate the fuel consumption of the rocket in flight. Based on self-built data sets and public data sets, the performance of LSTM-ResNet, LSTM, ResNet and ResNet-LSTM networks are compared. The training results showed that the LSTM-ResNet model was the best. It had better performance in the case of whether the sensor was faulty, and had higher damage detection accuracy.