應用深度學習預測RC建築之柱斷面設計

Abstract

本研究應用深度學習演算法，結合至土木實務設計中，蒐集真實設計建案作為訓練資料，用以預測柱斷面設計中的柱主鋼筋量與柱斷面積，在實際應用上只需將參數輸入至模型，即可獲得預測值，此預測值可以給予設計者在初期以及調整階段的建議值，以幫助減少設計時間，也可以用於完成設計後檢核是否有斷面設計錯誤的情況。

真實設計建案由築遠工程顧問有限公司提供，從約40棟建築之結構計算書、平面圖以及柱配筋詳圖中取得輸入參數，輸入參數區分為建築參數與柱參數，建築參數用於描述不同建築物之間的區隔，而柱參數則用於描述每根柱在不同樓層之間的差異，柱參數進一步分為兩種類型，一種與柱的斷面性質相關，另一種與柱周圍性質相關(如：樓板厚度)。透過建築參數與柱參數的組合，可以充分展現每筆柱資料獨有的特性。

模型建立上須考慮預測目標與應用方式，預測目標有兩種：柱主鋼筋量、柱斷面積；應用方式依照輸入方式區分為兩種：初始建議值、調整建議值與檢核值，故本研究須建立四個回歸模型。

模型訓練過程使用的深度學習模型包括人工神經網路ANN與遞迴類型神經網路GRU，分別探討兩種模型的表現。在參數選取上搭配皮爾森相關係數來選取輸入參數集，最後從不同面向分析並探討預測結果。

This thesis utilizes deep learning algorithms in conjunction with civil engineering design practices to predict the reinforcement area and section area in column design. Real building case are collected as training data to develop a predictive model. By inputting the relevant parameters into the model, designers can obtain predicted values, which provide suggestions during the initial and adjustment phases of the design process. These predictions assist in reducing design time and can also be used to check for significant errors after completing the design.

Input parameters are obtained from the structural calculation sheets, floor plans, and detailed column reinforcement drawings of approximately 40 buildings. The input parameters are divided into building parameters and column parameters.The building parameters are used to describe the distinctions between different buildings, while the column parameters are used to describe the variations of each column between different floors.The column parameters are further divided into two types. The first type is related to the properties of the column section, while the second type is related to the properties surrounding the column. By combining the building parameters and column parameters, the unique characteristics of each column data can be fully represented.

The model building considers both the prediction targets and their applications. The prediction targets consist of the reinforcement area and section area. The applications are divided into two types based on the input method: initial suggestion values and adjustment/checking values. Therefore, four regression models are developed in this study.

During the training process, the deep learning models used include Artificial Neural Networks (ANN) and a type of Recurrent Neural Network called Gated Recurrent Unit (GRU). Parameter selection is conducted using the Pearson correlation coefficient, which helps choose the input parameter set. Finally, the prediction results are analyzed and explored from different perspectives.