第三十七卷第四期 (期別146) (111年)

第三十七卷第四期 (期別146) (111年)

人工智慧在結構工程之應用特刊

客座主編:林子剛

標題 「人工智慧在結構工程之應用」引言
作者 林子剛
 

      近年人工智慧(AI,artificial intelligence)正對各個領域帶來前所未有革命性的影響,土木結構工程亦朝向「智慧化」邁進。結構工程導入人工智慧,運用在大量的資料彙整、分析及尋找規律上,透過AI 能有效且快速地節省人力,可直接對接進入施工過程,讓大量常規項目規劃和設計變得快速,比起以往人工判讀更具備效能與效率。美國國家科學基金會(NSF)開發了一套名為BRAILS(大規模人工智慧建築識別)的工具,可以檢測城市建築在地震、颶風或海嘯中可能面臨的風險,此即是人工智慧結合土木結構工程的案例之一。

      結構工程期刊企盼藉由此次特刊的機會,向各界介紹國內人工智慧在結構工程應用之成果。本特刊收錄了「結合神經網路熵與卷積神經網路於結構健康診斷」、「使用機器學習和統計資訊進行異常訊號檢測和分類」、「深度學習於提升地震模擬振動臺加速度性能之研究」、「含巡檢規劃、劣化辨識、損傷量化之進階建物劣化檢測架構」及「基於深度學習之合成配電網絡建置研究」。希冀藉由這些篇章的介紹,能更深入瞭解人工智慧在結構工程之應用發展與趨勢,讓臺灣在土木結構工程專案管理實務中,嘗試建立類似的工具,進而保障建築安全。

國立陽明交通大學土木工程學系 林子剛 教授 謹誌
結構工程期刊 特刊客座主編
2022 年12 月

標題含巡檢規劃、劣化辨識、損傷量化之進階建物劣化檢測架構
作者許舜翔、許筠曼、洪昊天、林禹齊、張家銘、陳宗珷、陳俊仲
關鍵字目視檢測、損傷偵測、損傷量化、深度學習
摘要目視檢測常用於結構檢測中,以快速評估建物中構件或材料之耐用性,及尋找可能的損傷。然而,傳統目視檢測相當耗費人力及時間,鑒於以上問題,本研究開發一個進階的建物劣化檢測架構,其中功能包括巡檢規劃、即時劣化識別、損傷程度判別及記錄詳細的損傷狀況。於該架構中,首先規劃巡檢草圖,提供檢測時最有效率檢測路徑;接著本研究使用Scaled-YOLOv4,進行損傷物件偵測,該方法於大規模場域中,也能快速檢測損傷,並使用SOLOv2 模型,對混凝土裂縫位置進行像素等級之實例分割,以利更精確的裂縫量化;最後,根據檢測到的劣化嚴重性及面積大小,對構件的損傷程度進行評級。本研究以一所小學之走廊進行實際場域驗證,目的為檢測和量化混凝土構件之表面劣化。由驗證結果可見,藉由本研究之進階建物劣化檢測架構,於劣化辨識、損傷量化和檢測效率上,相較傳統建物目視檢測都獲得顯著改善。
TitleFramework of Advanced Building Inspection withRoute Planning, Defect Detection, and Damage Rating
AuthorShun-Hsiang Hsu, Ho-Tin Hung, Yun-Man Hsu, Chia-Ming Chang, Tzung-Wu Chen, Chun-Chung Chen
Keywordsvisual inspection; damage detection; damage quantification; deep learning
AbstractVisual inspection is commonly adopted for building operation, maintenance, and safety. The durability and defects of components or materials in buildings can be quickly assessed through visual inspection. However, implementations of visual inspection are substantially time-consuming, labor-intensive, and error-prone because useful auxiliary tools that can instantly highlight defects or damage locations from images are not available. Therefore, an advanced building inspection framework is developed and implemented with route planning, realtime and detailed damage recognition, and damage rating in this study. The inspection route sketching is first exploited to provide an efficient plan with significantly reduced disruption. Then, Scaled-YOLOv4 and SOLOv2 models are considered in this study to detect defects even in a large-scale field quickly and acquire pixel-level damage recognition for more precise quantification, respectively. Finally, damage levels of components are rated following the importance and numbers per unit area of the detected defects. This entire framework is also implemented and verified by the hallway of an elementary school to detect and quantify surface damage of concrete components. As seen in the results, the conventional building inspection is significantly improved by the aid of the proposed framework in terms of damage localization, damage quantification, and inspection efficiency.
標題圖神經網路於線性靜力分析之應用
作者周遠同、黎光曜 、郭柏志、張慰慈、黃尹男、陳俊杉
關鍵字 
摘要結構設計是朝向最佳化的迭代過程,目前多依賴結構工程師過往的經驗結合結構分析的結果完成。結構最佳化設計需反覆進行結構分析,如何加速結構分析則是重要的突破關鍵。本研究使用深度學習模型作為線性靜力分析之代理模型,提供快速即時且精準之結構反應預測。因著結構物之空間幾何關係與圖資料結構之間之相似性,本研究將結構物表達成圖,並以圖神經網路模型去學習結構物所受之外力與結構物反應之間的關係。在透過由結構軟體所生成包含隨機樓層數、跨數、樑柱長度,以及隨機大小的側向外施載重之結構分析資料集訓練後,圖神經網路模型不僅顯示了其在預測位移以及力具有良好之表現,模型也具有很好之泛化能力,能夠預測在訓練時從未接觸過的、更高的結構物。在特徵重要性分析中也顯示模型所學習到之特徵具有一定之物理意義。
TitleLinear Static Analysis with Graph Neural Networks
AuthorYuan-Tung Chou, Kuang-Yao Li, Po-Chih Kuo, Wei-Tze Chang, Yin-Nan Huang , Chuin-Shan Chen
Keywords 
AbstractStructural design is an iterative process for optimum selection, which relies on structural analysis results and experience from structural engineers. Since iterative structural analysis is a necessary for getting a good design, accelerating structural analysis is an important task. In this work, we adopt deep learning approaches as a surrogate model for linear static analysis, which provides fast and accurate structural response prediction. Based on the similarity between the structure’s topology and graph data structure, we represent structures as graphs and leverage graph neural networks (GNNs) to learn the relationship between given external forces and corresponding structural responses. The GNN model is trained with random-generated structures, including random story number, span number, beam-column length, and value of external forces. The results show that the GNN model has good performance on displacement and force predictions and excellent generalizability on unseen, taller structures. In addition, it is shown that based on the analysis of feature importance, the GNN model can extract important physical attributes from the input features.
標題使用機器學習和統計資訊進行異常訊號檢測和分類
作者林天郇、黃謝恭、周肇昱
關鍵字異常訊號、機器學習、模式識別網路、GoogLeNet
摘要結構健康監測(structural health monitoring,SHM)和結構完整性管理(structural integrity management,SIM)為逐漸興起的技術,為了持續檢測結構狀態並不斷追蹤結構劣化,海量的資料伴隨著不正常的量測數據正不斷的被產生,至此,檢測與分類這些帶來問題的異常訊號大多依靠人工,這不僅費時費力,工作更是乏味。在本研究中,檢測與分類異常訊號的工作將被機器學習的技術取代,並嘗試利用統計學的優勢來提高檢測與分類的正確性,模式識別網路(pattern recognition network)被採納來使用一維的輸入資料,而GoogLeNet 被引入來使用二維的輸入資料,並利用一組現地量測資料進行學習,其結果顯示兩種機器學習的技術均可有效的檢測與分類異常訊號,並且其算力需求與正確性是個權衡下的結果,因此對於兩種技術的使用情境將可視應用的需求而定。
TitleDATA ANOMALIES DETECTION AND CLASSIFICATION USING MACHINE LEARNING AND STATISTIC INFORMATION
AuthorTian-Xun Lin, Shieh-Kung Huang, Jau-Yu Chou
Keywordsdata anomaly, machine learning, pattern recognition network, GoogLeNet
AbstractStructural health monitoring (SHM) and structural integrity management (SIM) are emerging recently. To continuously track the condition and constantly detect early deterioration of the infrastructure, huge amounts of data are produced and abnormal measurement is inevitable. The corrupted data can produce a lot of problems and, generally, they are examined and classified by humans. In this study, the detection and classification are replaced by the techniques of machine learning (ML) and improved by using statistic information. The neural networks based on 1-dimensional and 2-dimensional data are studied via a field dataset collected from a long-span cable-stayed bridge. Therefore, a shallow network, called pattern recognition network, is selected to use 1-dimensional data as an input and a deep network, GoogLeNet, is selected to use 2-dimensional data. The results show that both models can detect and classify the data anomalies and the usage depends on the assigned application and the trade-off between computation and performance.
標題結合神經網路熵與卷積神經網路於結構健康診斷系統應用之研究
作者林子剛、林易廷、郭凱維
關鍵字結構健康監測、神經網路熵、卷積神經網路
摘要本研究以神經網路熵(Neural Network Entropy, NNetEn)為研究基礎,並將此熵分析方法與卷積神經網路(Convolutional Neural Network, CNN)結合,發展出一套具實用性之結構健康監測系統。為了驗證其系統之可行性,將進行七層樓鋼構架的破壞實驗,並建置與鋼構架相同結構特徵之數值模型。首先使用狀態空間法建立數值模型,模擬與鋼構架相同之十六種破壞模式,並將破壞時之各樓層加速度訊號以神經網路熵進行分析,建立熵值資料庫,再將此熵值資料庫用以訓練類神經網路模型。前期研究之熵值是藉由人為觀察後定義其閥值,藉以判斷結構是否破壞,因此為了避免人為因素的誤判及自動化判讀,本研究以可視化之 heatmap 量化熵值變化,並選用適用於影像處理的卷積神經網路分析,透過將熵值轉為圖像資料之方式不僅能夠減少模型中之參數量還能提升其運算速度。神經網路模型在訓練過程中藉由提取熵值中的破壞特徵並學習,在訓練完成後模型即能在識別輸入資料的破壞特徵後定位結構物之破壞區域。最後,透過國家地震工程研究中心之七層樓鋼構架驗證所設計之十六種破壞案例,逐例討論模型的輸出值並以混淆矩陣量化數值模擬和實驗驗證之預測結果。本研究提出的結構健康診斷系統,將新興之熵分析方法結合類神經網路,其最終驗證實驗之測試樣本結果其準確率 93.13%。
TitleApplication of Convolution Neural Network and Neural Network Entropy Algorithm for Structural Health Monitoring
AuthorTzu-Kang Lin, Yi-Ting Lin, Kai-Wei Kuo
KeywordsStructural Health Monitoring, NNetEn, Convolution Neural Network
AbstractThis study combines Neural Network Entropy (NNetEn) and Convolutional Neural Network (CNN) to develop a practical structural health monitoring system. In order to verify the feasibility of the system, the failure experiment of a seven-story steel frame has been carried out with a numerical model of the same structural characteristics as the steel frame. The state space method is used to simulate the sixteen failure modes on the steel frame, where the acceleration signals of each floor at the time of failure are analyzed by neural network entropy. An entropy database is established based on the model to train the neural network model. To avoid the misjudgment and automatic interpretation of human factors, this study uses the visualized heatmap to quantify the change of entropy value, and the convolutional neural network analysis is selected for image processing. By converting the entropy value into image data, not only the number of parameters in the model can be reduced, but its operation speed can be improved. During the training process, the neural network model extracts and learns the damage features in the entropy value. After the training is completed, the model can allocate the damage area of the structure by identifying the damage features of the input data. Finally, through the verification of 16 failure cases simulated on the seven-story steel frame of the National Center for Research on Earthquake Engineering (NCREE), the performance of the proposed SHM system is evaluated by both numerical simulation and experimental verification with confusion matrix. The SHM system proposed in this study combines the emerging entropy analysis method with a neural network. The test results of the final verification have an accuracy rate of 93.13%.
標題深度學習於提升地震模擬振動臺加速度性能之研究
作者賴魁星、陳沛清
關鍵字地震模擬振動臺、加速度控制、三參數控制器、深度學習、長短期記憶神經網路
摘要地震模擬振動臺是地震工程研究重要的實驗設備,已廣泛應用在各種結構系統的耐震性能試驗,諸如鋼筋混凝土結構、鋼結構、隔減震結構以及精密設備等,因此振動臺加速度控制之精確與否在振動臺實驗中特別重要。本研究提出使用深度學習的方式, 使用了長短期記憶(Long Short-TermMemory, LSTM)神經網路,透過大量實驗資料訓練出振動臺的外迴路控制器。本研究使用油壓驅動之大型單軸向振動臺,設計一組鋼構造試體安裝其上並進行大量的振動臺實驗,以得到LSTM 之訓練資料,再將訓練所得到的LSTM 作為前饋控制器,加裝在振動臺既有的三參數控制架構中,藉此補償振動臺加速度控制的動態響應。實驗結果證明LSTM 前饋控制器可有效地降低振動臺的加速度誤差,並減少試體與振動臺的互制現象,使振動臺之加速度性得到顯著的提升。本研究成果顯示應用深度學習於提升振動臺加速度表現的發展潛能,未來可進行更多的相關研究,以改善振動臺實驗之測試品質。
TitleImproved Acceleration Tracking Performance of Seismic Simulators using Supervised Deep Learning
 Kui-Xing Lai and Pei-Ching Chen
KeywordsSeismic shake table; acceleration control, three-variable control; deep learning, long short-term memory neural network
AbstractSeismic shake table testing has been widely used for various structural systems such as steel structures, reinforced-concrete structures, energy-dissipated and base-isolated buildings, and nonstructural components etc. Therefore, accurate replication of shake table acceleration is particularly important to these tests. In this study, supervised deep learning approach is applied as an alternative for seismic shake table control. The Long Short-Term Memory (LSTM) neural network is built for training the controller to improve acceleration performance of the shake table. A large-scale servo-hydraulic uniaxial shake table is adopted. A steel specimen is designed and fabricated for performing a large number of shake table tests. Then, the shake table testing data are used to train a feedforward controller using LSTM which is implemented close to an existing Three-Variable Control (TVC) loop. The validating experimental results prove that the acceleration tracking performance is improved compared with conventional TVC. The control-structure interaction is also suppressed. The experimental results demonstrate the proposed control scheme reduces the acceleration tracking error effectively compared with conventional TVC control. The research results also show great potential for deep learning application to seismic shake table control in the future. Keywords: Constitutive model, anisotropy, shear-slip and re-contact, mesh-sensitivity, non-proportional loading, concrete, finite element
標題基於深度學習之合成配電網絡建置研究
作者林岳鴻、林其穎
關鍵字合成配電網絡、深度學習、物體識別、地理標籤
摘要現今時代處於能源高消耗的處境且加上地球暖化帶來的災害規模日益嚴重,使人們對於公共資產管理、減少能源消耗、預測災害損失更加重視。本研究提出利用深度學習建置合成電力網路,以便未來使用其擬真模型進行電網可靠度分析。本研究利用街景圖配合物體識別及計算交集的方式定位出電線桿之地理資訊,並利用此地理資訊建立其合成電力網路。研究首先使用Mask R-CNN 與YOLOv4 物體識別模型,進行時長之條件控制訓練,並比較兩種模型的精確度得出適合此研究之方法。接續調整其參數配置,比較不同參數下對於此研究的最佳模型設置。對於區域的研究進行每條街的個別物體識別,並設計兩種排序方式進行電線桿排序,分別為經緯度以及最短路徑,利用兩種排序方式進行基於馬爾科夫隨機場(Markov random field,MRF)和基於方位線(line of bearing, LOB)與基於密度之聚類演算法(Densitybasedspatial clustering of application with noise, DBCSCAN)兩種方式計算交集,接著進行電線桿之補遺,使其更符合真實電線桿最大容許間距。最後,利用距離迭代的方式合併所有街道並刪除重複預測的電線桿點位,得出四種結果。使用四種結果與真實電線桿地理資訊進行精度比對,選擇出最適者並利用最小生成樹(minimum spanning tree, MST)建立電力網路,建立後分析並比較其性質與真實電力網絡的區別。未來可使用此模型進行改善,使其更符合真實電力網路,並使用其合成電力網絡進行電網可靠度分析、公共資產管理、災害分析等分析。
TitleSynthetic Power Distribution Network Construction Based on Deep Learning Algorithm
 Yue-Hung Lin and Chi-Ying Lin
Keywordssynthetic power distribution network, deep learning, object detection, geo-positioning
AbstractGlobal warming has caused high energy consumption and an increasing scale of disasters, which make people draw more attention to public asset management to reduce energy consumption and predict losses caused by disasters. Based on a deep learning based object detection approach, this study develops a synthetic power distribution network that can serve as an alternative to the real power distribution net-work and be used to analyze its reliability. This research uses the street view images to detect utility poles and conduct geo-positioning to locate utility poles on the map. For object detection, the Mask R-CNN and YOLOv4 are trained with controlled du-ration, and then the accuracy of the two models is compared to determine which method is suitable for this research. Second, the model’s hyperparameters are adjusted and compared to determine the best model setting for the object detection task in this study. Then the selected model is used to perform the object detection task for each street in the research region. Two sorting methods, namely, the latitude and longitude sorting method and the shortest path sorting method, are proposed to sort the poles for pole geopositioning and supplementation. With two sorting methods, pole geopositioning is conducted based on two approaches: The first is the Markov random field (MRF) approach, and the second is the line of bearing (LOB) with density-based spatial clustering of application with noise (DBSCAN). After determining the detected pole location, pole supplementation is conducted to ensure the maximum allowable distance between poles. Third, four sets of results are obtained by merging all streets and removing duplicate poles by means of distance iteration. Finally, four results are compared with the coordinates of real utility poles. The most suitable method for study region is selected to establish the synthetic power distribution network using the minimum spanning tree (MST). In the future, this model can be improved to make it more in line with the real power distribution network, and the synthetic power distribution network can be used for power grid reliability analysis, public asset management, disaster analysis, power demand-supply analysis, etc.