基于分布式光纤传感与深度学习的结构损伤识别技术研究

For the long and large important infrastructure, increasing the accuracy and efficiency of structural damage detection and reducing the human intervention of damage feature extraction have always been the research focus in the engineering field. The structural safety monitoring based on distributed fiber optical sensing technology makes it possible to greatly improve the spatial resolution of structural damage detection. Deep convolutional neural network (DCNN) provides a new way to study the automatic extraction and classification of structural damage features. Based on the optical fiber ribbon package structure and partitioned layout for measuring points, the sensing method expecting to detect multi-parameter with high-density spacing by using the distributed weak Bragg fiber optical Bragg grating (FBG) will be studied in this research. According to the characteristics that weak FBG can sense structural multi-parameter, a structural damage localization method based on stress distribution statistical under multi-scale division will be studied; the coupling relationships among moving load, vibration, strain and stiffness based on the partial differential equation of beam motion and deflection curve equation will be revealed. Finite element simulation and model test will be taken to verify and calibrate the proposed damage detection theory which is on the basis of weak FBG structural responses regarding stress and vibration. After the study regarding data augmentation and the balance processing for the classification, the sample database will be constructed to reveal the influence of the DCNN model structure and parameters on the automatic extraction of damage features for the raw sensing responses which are expressed by different kinds of tensor inputs. The dual-control method based on loss function and prediction accuracy will be elicited, and an appropriate feature extraction model structure and the corresponding parameters selection strategy will be proposed.

提高长、大重要基础设施结构损伤识别的准确性与效率，降低损伤特征抽取人为干预，一直是工程领域研究关注的重点。基于分布式光纤传感技术的结构安全监测，可大幅提高结构损伤识别空间分辨率。深度卷积神经网络（DCNN）则为研究结构损伤特征自动抽取与分类提供了新途径。本项目基于光纤并带封装及测点分区绕行布置，研究分布式弱光栅高密度感测结构多参量传感方法。基于弱光栅多参量感知特性，研究多尺度划分下由应力分布统计定位结构损伤的方法；揭示基于梁运动偏微分方程与挠曲线方程的移动荷载、振动、应力、刚度间耦合关系。采取有限元模拟和模型试验，校验完善基于分布式弱光栅应力和振动响应的损伤识别理论。面向提出的数据增强及样本平衡化处理手段所构造的样本库，揭示原始传感响应在不同张量输入表达下，DCNN模型结构，参数对损伤特征自动抽取的影响机制。以损失函数和预测精度进行双控，提出适合的特征自动抽取模型结构及相应的参数选择策略。

针对长线路重大基础设施监测，分布式光栅传感阵列具有大容量、高密度的感知优势，可为基于分布式传感响应研究结构损伤识别的深度学习方法提供大数据支撑保障。本项目针对缩尺模型和实际重大工程的结构特点，试验研究了用于感知长线结构分布式应变/振动响应的光缆布设工艺，通过光缆结构选型及表贴与埋入工艺尝试，解决了地铁隧道、机场跑道、高速公路、大跨桥梁等不同场景下开展结构多参量分布式监测的难题。基于UBC、UCR公开数据集和工程中的实测样本，提出了针对分布式光纤传感数据的信号预处理、不平衡样本集数据增强及样本间的相似性度量方法。在有限算法复杂度和低计算开销下，针对时间序列信号，试验论证了基于Savitzky-Golay开展信号平滑去噪的可行性。基于Hilbert与深度卷积生成对抗网络以及基于信号切割与混叠处理，探索了不平衡数据集样本增强的实现方法。结合堆叠降噪自编码网络与改进DTW算法提出了度量样本信号间距离的相似度算法。面向桥梁及地下结构，设计了获取包含结构异常与损伤态样本的实验方案，并建立了基于实验方案的实测样本库。针对实测样本库对应的四种场景，构建了结构异常和损伤分类识别的深度学习模型。其中，具有自适应激活机制的13层深度卷积（CNN）神经网络模型有效实现了缩尺桥梁模型中对称位置损伤的辨识。基于两阶段训练策略的深度CNN神经网络模型通过分步骤处理二分类和多标签分类问题，实现了斜拉桥结构7类标签样本的分类辨识。基于具有注意力机制的CNN/LSTM无监督混合深度网络，提出了一种通过量化误差序列和使用ROC曲线来区分异常响应的阈值设置流程，满足了不同异常扰动占比情况下结构损伤信号的可靠识别。基于伪Hilbert扫描的样本维度转换策略，通过深度CNN自编码网络训练及t-SNE聚类评价，实现了对于轨道扣件松动损伤的准确识别。研究成果提高了针对长线路重要基础设施开展结构损伤识别的准确性与效率，有效降低了损伤特征抽取过程的人为干预，为保障长线路重大结构的服役安全提供了损伤识别与定位的新途径。