以結構信息圖嵌入結合長短期記憶深度神經網路預測結構物之動力反應

Abstract

近年來深度學習的快速發展極大地擴展了其在結構工程中的應用可能性，基於深度學習方法的建築結構代理模型在過去幾年被廣泛地研究，至今已有許多研究提出不同的深度學習模型來預測特定結構物受震後的反應歷時。雖然過去的研究所提出的深度學習模型在特定結構物的反應歷時預測任務上能達到良好的準確度，但這些深度學習模型僅適用於預測特定結構物，無法同時預測其他不同結構物的反應歷時，這使得原結構物變更設計或使用者需要評估其他結構案例進行時，就需要從新蒐集資料來訓練新的代理模型。 本篇論文針對結構物反應歷時預測任務提出一全新的深度學習方法。本研究將建築結構資料以圖資料結構封裝，與地表加速度歷時資料一併作為輸入資料，並提出一基於圖嵌入網路(Graph embedding network)與長短期記憶神經網路(Long short-tern memory, LSTM)模型的融合深度學習模型架構，能夠根據輸入的結構物與地震資訊，預測出對應結構物的反應歷時。本篇研究針對結構反應歷時預測任務提出了完整的深度學習方法論，分別探討序列模型的優化演算法，以及圖嵌入網路中不同聚合函數擷取結構物特徵的能力。本篇研究針對 LSTM 演算法的計算特性提出了集裝填充序列學習策略與序列壓縮學習策略，透過實驗證實提出的學習策略能有效提升模型的預測能力與訓練效率。圖嵌入網路的部分，以圖卷積網路 (Graph convolutional network, GCN) 與圖注意力網路 (Graph attention network, GAT) 兩種不同的圖神經網路，實驗不同聚合函數的預測能力，並透過非監督式學習降維方法將高維的圖嵌入向量降為至二維空間進行資料視覺化。本研究觀察出不同建築結構的圖嵌入依據第一模態週期以及樓高有顯著的分群效果，驗證了深度學習模型內部的圖嵌入確實攜帶結構物的物理信息，同時也說明了本研究提出的深度學習架構具有優異的建築結構特徵擷取能力。

The development of deep learning in recent years has dramatically expanded its application possibilities in structural engineering. Structural surrogate models based on deep learning methods have been widely studied in the past few years, and many studies have proposed different deep learning models to predict the response history of specific structures. Although the deep learning models presented in past studies can achieve good accuracy in the task of predicting the response history of particular structures, these deep learning models are only suitable for predicting a specific structure but not for other different structures at the same time, which makes it necessary to re-collect data to train a new surrogate model when the original structure changes the design or the user needs to evaluate other structural cases. This study proposes a new deep-learning method for predicting the response history of different structures. We encapsulate the structure data in a graph and use the graph as input data along with the ground-motion data. Then we propose a fusion deep learning model architecture based on a graph embedding network and sequence model, which can predict the response history of the corresponding structure based on the input structural graph and seismic information. This study proposes a complete deep learning methodology for structural response time prediction tasks, exploring optimization algorithms for sequence models and the ability of different aggregation functions in graph embedding networks to capture structural features. This study proposes the Packing Padded Sequences (PPS) learning strategy and the Sequence Compression (SC) learning strategy for the computational characteristics of the LSTM algorithm. Through experiments, it has been proven that the proposed learning strategy can effectively improve the model’s prediction ability and training efficiency. For the part of the graph embedding network, two different graph neural networks, the Graph Convolutional Network (GCN) and the Graph Attention Network (GAT), are used to experiment on the effects of varying aggregation functions on prediction ability. For data visualization, unsupervised learning dimensionality reduction methods are used to reduce high-dimensional graph embedding vectors to two-dimensional space. We observed that the graph embedding of different building structures has a significant grouping effect according to the first modal period and the building height. This verifies that the graph embedding inside the deep learning model carries structural information. It also shows that the deep learning architecture proposed in this study can extract structural features.