應用多任務機器學習於結構物梁柱斷面預測與應用

Abstract

本研究開發一套以深度學習為基礎的預測系統，用於結構設計中梁柱斷面尺寸的預測。研究中採用多任務學習（MTL）模型同時預測結構元件在兩個方向上的尺寸，並與傳統單任務模型進行比較。實驗結果顯示，MTL 模型能有效捕捉兩個方向間的相依性，並提升整體預測表現。此外，本研究針對結構元件的特性，設計了多組兼具物理意義與工程背景的輸入特徵，並提出客製化的損失函數以強化模型對尺度差異的適應能力。而本研究設計加入的特徵多符合預期表現，影響模型表現。

為增進模型可解釋性，本研究引入 SHAP（SHapley Additive exPlanations）方法進行特徵重要度分析，探討各特徵在不同預測階段對模型判斷的貢獻。透過實例驗證顯示，SHAP 分析所得的重要特徵與結構設計的工程直覺具有一致性，具實務應用價值。最後，研究亦實作出一套友善的使用者介面，整合模型預測與資料輸入流程，降低應用的門檻。

This study develops a deep learning–based prediction system for estimating the cross-sectional dimensions of structural beams and columns. A multi-task learning (MTL) model is adopted to simultaneously predict the dimensions of structural components in two orthogonal directions, and its performance is compared with that of traditional single-task models. Experimental results indicate that the MTL model effectively captures the interdependence between the two directions, leading to improved overall predictive accuracy.

To reflect the physical and engineering characteristics of structural components, this study designs multiple input features grounded in domain knowledge and proposes a customized loss function to enhance adaptability to scale variations. Most of the engineered features demonstrate expected performance to the model's predictions.

To improve interpretability, SHAP is employed for feature importance analysis, showing whether features are consistent with engineering intuition. Finally, a user-friendly interface is implemented to integrate data input and model prediction.