中華民國內政部建築研究所中華民國內政部建築研究所

一、研究緣起
台灣主要建築與工程之結構系統為鋼筋混凝土(RC)結構物，火災發生後對其結構系統受火損之量化與質化調查(尋因)，與後續居安使用功能之評核(驗果)至關重要。內政部建築研究所曾於107年度進行火害現場結構材料探傷檢測系統韌體之開發，其可用於量測鋼筋混凝土結構材料之彈性波速，並根據104-106年度於實驗室與現場系列調查之數據所提出的「簡速非破壞性」之火損初勘量化評估方法，藉由正規化指標(剪-壓波速比)評估其火害相關之因果數據，如火場最高受溫分佈與柱-梁-版之混凝土強度折減率，可分別作為後續火場調查與構材火損判識。為進一步擴大既有之研究成果，導入人工智能分析，增加現場應用案例，並開發鋼筋握裹損傷判釋方法，而進行本年度之研究工作。
二、研究方法
本研究擬在前述成果基礎上，利用人工智能分析方式包括支援向量機、線性回歸、類神經網路以及決策樹等演算法進一步對成果資料分析建立火害判識方法，且透過雲端資料庫系統的開發，增加資料庫的使用團隊，以利增加相關數據之累積速度；另方面採室內試驗探討不同狀態下之火害受損深度與握裹力喪失特徵，建立運用導波於火害受損厚度與鋼筋握裹力狀態之評估方法，且藉已開發之設備功能，評估在既有硬體規格下導入導波量測分析應用的可行性。另一方面，持續增加現場的應用案例以掌握相關實務狀況與可能遭遇之問題。
三、重要發現
1.在人工智能與火害特徵調查之資料庫及資料探勘方面，因果關聯相關性顯示剪力波速、壓縮波波速、剪-壓波速比與最高溫度及強度折減比例有高度相關，將剪-壓波速比做為輸入，最高溫度與強度折減比例分別作為輸出，進行支援向量機、線性回歸、類神經網路以及決策樹M5P之預測模型建立，結果顯示在最高溫度與強度折減比例之預測模型上，採用決策樹M5P之演算法有最佳效果，其餘方法預測性較差且會有負值之不合理預測結果產生。
2.在應用導波進行握裹力傷損調查判釋方法開發上，兩階段調查策略具有可行性，第一階段採用應用雷利波之表面波分析方法可獲得混凝土剪力波速剖面影像，以10%之剪力波速折減作為特徵值可定義受損範圍，判釋此受損範圍之分布特徵若呈現自表面向下延伸則顯示遭受火損，若此範圍超越保護層，則進入第二階段判釋。第二階段判釋採用距敲擊點10公分處之接收器進行時頻分析，在時頻譜上若呈現水滴狀能量分布則為正常，若呈現類靴狀且突出段在40-100kHz範圍則顯示可能有握裹力喪失之可能性，需進行殘餘握裹力調查。
四、主要建議意見
建議一：AI人工智能端
後續須蒐集更多數量與變數之內業數據，須包含正常強度混凝土(f_c^'=210、f_c^'=280 及 f_c^'=350)以及更多階級之混凝土火害最高溫度(T_max=200°C~T_max=800°C，每50°C階層漸進)
針對WEKA軟體可以使用更多預測模型進行測測法之推估，尋找更適合之預測法，並了解其程式碼方便後續修改。
針對大數據分析可以使用非監督式學習。使其預測法內自行尋找其特徵值與佔比權重，可使推估值更貼近現實情況。在此需求下應盡快建立資料庫雲端平台以及相對應之韌體設備，以有效增加有效之數據量。
建議二：握裹力傷損之導波量測判識方法
導波檢測之判識方法已建立，但仍屬定性之方法，建議後續可持續開發定量之評估方法，並進行實場測試以掌握在實場應用中可能面臨的問題。
第一階段判識採用剪力波速影像圖受既有反算方法限制而使得深度解析度僅1公分，未來可針對此深度上解析度的提升進行進一步的改善研究。
建議三：導波量測系統建置
本研究對導波技術導入既有設備進行之評估結果顯示整合具有可行性，建議未來可將其組建並發展儀控軟件，搭配網路連接可有效發揮以建置之雲端資料庫系統之功用。

Concerning Reinforced Concrete (RC) structure is the main system for buildings and constructions in Taiwan currently, the quality/quantity evaluation of fire-induced damage/degradation in RC structure is important for living safety. To establish a new working system which can be used to measure the fire-damage degree rapidly for both brittle and ductile of reinforced concrete structure. This study investigated the full-field fire damage characteristics of concrete and the interface bonding between concrete and steel using ultrasonic pulse (UP) and guide wave. The obtained results were used to construct the fire damage investigation firmware. It should be available to detect the degree of reduction of RC structural performance as well as maximum temperature. The obtained data can be provided to fire-prevention department and structure engineers to rebuild fire field and define structure recovery method respectively.
In this study, artificial intelligence analysis is used, subjoin more field application cases and establish the method of steel bond damage interpretation.
In terms of databases and data mining for artificial intelligence and fire hazard investigation, Causal correlation shows that shear wave velocity, compression wave velocity and wave-velocity ratio are highly correlated with the maximum temperature and strength reduction ratio The wave-velocity ratio is taken as the input, and the maximum temperature and strength reduction ratio are taken as the output respectively, Support vector machine, linear regression, neural network and decision tree m5p prediction model are established. The results show that the support vector machine algorithm is the best in the prediction model of the maximum temperature and strength reduction ratio. The other methods have poor predictability and negative and unreasonable prediction results.
The two-stage investigation strategy is feasible in the application of guided wave to the investigation and interpretation of bond damage, in the first stage, the shear wave velocity profile image of concrete can be obtained by using the surface wave analysis method of Rayleigh wave. With 10% shear wave velocity reduction as the characteristic value, the damage range can be defined. It can be judged that if the distribution characteristics of the damage range extend downward from the surface, it will show fire damage. If the range exceeds the concrete cover, it will enter the second stage. In the second stage, the time-frequency analysis is carried out. In the time-frequency spectrum, it is normal if there is a water drop like energy distribution; if there is a shoe like shape and the protruding section is in the range of 40-100kHz, it shows the possibility of loss of bond stress. Therefore, it is necessary to investigate the residual bond stress.
For immediate strategy:
The current evaluation standard is the identification of correlation coefficient, mean absolute error and root mean square error. In the future, it is suggested to integrate them into a composite index. In addition, the identification method of guided wave detection has been established, but it is still a qualitative method. It is suggested that the quantitative evaluation method should be developed in the future, and the field test should be carried out to grasp the possible problems in the field application.
For long-term strategy:
Unsupervised learning can be used for big data analysis. Making the prediction method to find its own characteristic value and proportion weight can make the estimation closer to the reality.