基于深度卷积神经网络的稀疏信号重构与结构损伤识别方法研究

In present structural health monitoring system, long-term monitoring generates massive data and sensor performance degradation causes uncompleted data. In addition, traditional feature extraction methods have low utilization of monitoring data. Furthermore, deep learning has more powerful ability of feature learning and feature representation in the advent of big data era. This project employs deep learning theory and probability statistics theory to investigate the following problems. Firstly, the techniques of response novelty and distinguishing them are studies for the structural response novelty caused by non-structural damage. Secondly, the weighted 1 norm sparse reconstruction algorithm is studied to recovery missing structural response. Thirdly, a new structural damage detection method based on deep convolutional neural network is proposed, which can fuse both feature extraction and classification blocks into a single and compact learning body, and can adaptively extracted damage sensitive factor. Finally, the feasibility and effectiveness of the related techniques are verified through laboratory tests, numerical simulation and field tests. Through this project, some technical bottleneck problems can be solved in the community of structural health monitoring, such as the misclassifying due to non-structural damage, and the incomplete data due to single transmission and sensor fault as well as low utilization of monitoring data due to Hand-crafted features extraction and classifier selection. Consequently, some independent core techniques are also achieved, which can provide technical support for the large structural health monitoring systems.

针对健康监测系统长期监测产生的海量数据和传感器性能退化导致的数据不完备，以及现有特征提取方法对监测数据利用率低的问题。本项目考虑到深度学习理论在大数据背景下较传统机器学习具有较强的特征学习和特征表达的特性，拟采用深度学习理论结合概率统计理论，首先研究非结构损伤引起的结构响应异常剔除方法；其次研究随机信号缺失的数据重构方法，获得加权1范数的稀疏重构技术；第三，研发集无监督特征提取和分类器为一体的基于深度卷积神经网络的结构损伤检测新方法，使其完成损伤敏感因子的自适应提取与健康状况的智能诊断。最后通过实验室试验、数值模拟与现场测试，验证相关技术的可行性和有效性。通过研究，可部分解决非结构损伤引起的错误报警、传感器故障和信号传输导致的随机数据丢失、及传统特征提取对结构响应利用率低的问题。获得具有自主知识产权的核心技术,为大型结构健康监测系统真正发挥防灾减灾作用提供技术支持。

针对健康监测系统长期监测产生的海量数据和传感器性能退化导致的数据不完备，以及现有特征提取方法对监测数据利用率低的问题。本项目考虑到深度学习理论在大数据背景下较传统机器学习具有较强的特征学习和特征表达的特性，采用深度学习理论结合概率统计理论，首先研究非结构损伤引起的结构响应异常剔除方法，提出基于主成分分析第一主分均方根的结构响应异常剔除方法，同时提出了基于二维卷积神经网络的结构响应异常剔除技术；其次研究随机信号缺失的数据重构方法，提出基于自适应数据融合和小波变换的结构响应重构技术，同时获得了基于加权1范数的稀疏重构技术；第三，研发集无监督特征提取和分类器为一体的基于深度卷积神经网络的结构损伤检测新方法，使其完成损伤敏感因子的自适应提取与健康状况的智能诊断。最后通过实验室试验、数值模拟与现场测试，验证相关技术的可行性和有效性。通过研究，可部分解决非结构损伤引起的错误报警、传感器故障和信号传输导致的随机数据丢失、及传统特征提取对结构响应利用率低的问题。获得具有自主知识产权的核心技术,为大型结构健康监测系统真正发挥防灾减灾作用提供技术支持。提出的以自适应一致性数据融合和DWT-Robust ICA算法相结合的考虑传感器性能退化的结构响应异常检测方法识别七层框架数值模拟结构证明该方法具有较好的容噪性，当噪声水平达到 10%仍能很好地识别；提出的IMSPCA对各种类型的非完全故障模型下传感器性能退化采集的加速度在时域范围内无论是桁架模型还是实测的天津永和大桥采集的数据的恢复与无故障数据基本吻合，且其相对误差也小于16%，且其频谱图与无故障数据完全吻合，这就减少了故障数据所产生的伪基频的现象同时减少了时域转频域的误差；提出的改进多尺度主成分分析（IMSPCA）数据恢复方法和自适应一致性数据替换方法无论对桁架数值模型数据还是天津永和大桥实测数据均能很好地恢复性能退化传感器采集的数据，故此方法为后期的进一步损伤精确识别提供完备化数据。