视觉智能Shapelet Transform驱动的SHM数据关联分析与域自适应迁移机制深度学习

深度学习是实现结构健康监测（SHM）大数据解读的有效手段之一。研究基于视觉形状的SHM数据挖掘技术，能为从数据本身揭示结构安全状态提供理论指导。目前已通过GARCH模型获得振动响应数据新息序列，并基于深度学习投影特征对裂缝图像进行识别与重构，解决结构化与非结构化数据的可视化问题，证明可进一步利用视觉智能Shapelet Transform（SF）挖掘SHM数据的内在表征机理。拟进一步开展：1.研究SF对多源异构数据形状特征的通用提取方法，揭示SHM数据Shapelet特征蕴含的物理含义；2.建立多源异构数据的关联关系及其与结构的映射规律，阐明数据融合的关键技术；3.利用域自适应迁移学习胶囊神经网络，探索标签数据到无标签数据的深度学习机制。本项目有望进一步从视觉角度认识和揭示结构响应、荷载与性能演化规律，为结构诊断与预测提供新的理论依据，为结构损伤识别算法由理论走向实际工程应用提供新途径。

Deep learning is one of the promising approaches to interpret big data of structural health monitoring (SHM). Research of big data mining technology in SHM based on visual shape can provide theoretical guidance for revealing the structural safety state of structures from the data itself. At present, GARCH model has been used to obtain Innovation series of vibration response data, and cracks images have been deep learned to extract projection features to recognize and reconstruct the damage information. It has been proved that the visual intelligent Shapelet Transform (SF) can be further utilized to mine the intrinsic representation mechanism from SHM data when the visualization problems of structured and unstructured data have been solved. The research plan is: 1) To study general extraction method for the shape features of multi-source heterogeneous data by SF, and reveal the physical meaning of the Shapelet features of SHM data; 2) Establish the association relationship of multi-source heterogeneous data and the mapping rule with structure, and clarify the key technologies of data fusion; 3) Explore the deep learning mechanism from labeled data to unlabeled data by using domain adaptive transfer learning capsule neural network. This project will further understand and reveal the evolution law of structural responses, loads and performances from the perspective of visual intelligence. It will provide a new theoretical basis for structural diagnosis and prognosis, and pave the way for structural damage identification algorithm from theoretical developments to practical applications.