利用載重試驗資料庫進行基礎承載力預測：準場址特定方法

Abstract

載重試驗在大地工程中扮演關鍵角色，其量測之極限承載力為樁基礎與淺基礎設計時的重要依據，直接影響結構物地基之穩定與安全。然而，載重試驗往往需較大作業空間，且試驗成本高昂，促使數據驅動之預測方法逐漸受到重視，以期降低時間與成本。 本研究以 Phoon and Tang (2019) 及李宜庭 (2023) 所建構之載重試驗資料庫為基礎，並採用 Ching et al. (2021) 所提出之準場址特定階層貝氏模型 (Hierarchical Bayesian Model, HBM) 進行極限承載力預測。但文獻指出，HBM 在資料稀缺或資料變異性較高時，預測結果易出現不穩定性與信賴區間過寬之現象。因此，本研究提出三項改進策略，分別為模型簡化、客製化聚類區域化 (Tailored Clustering Enabled Regionalization, TCER) 及資料變數降維 (Dimension Reduction, DR) ，以期提升模型預測之穩健性與適用性。 然而，不同基礎型式與資料情境下，並無法找到單一通用最佳模型。故本研究進一步整合多種模型，發展以最小微分熵為選擇準則之 HBM 堆疊集成學習(HBM-Based Stacking Ensemble Learning)，以動態篩選表現最佳之模型。本研究驗證此方法於四種基礎型式、共十二個資料庫中的應用成效，並以交叉驗證下之均方根誤差 (Root Mean Square Error, RMSE) 作為驗證指標，系統性分析並驗證該方法於多基礎承載力預測之可行性與優勢。

Load testing plays a critical role in geotechnical engineering, as the measured ultimate bearing capacity serves as a key reference for the design of pile and shallow foundations, directly influencing the stability and safety of structural foundations. However, load testing typically requires large working space and incurs high costs, which has driven increasing attention toward data-driven prediction methods to reduce both time and expenses. This study is based on the load test databases compiled by Phoon and Tang (2019) and Li (2023), and adopts the Quasi-Site-Specific Hierarchical Bayesian Model (HBM) proposed by Ching et al. (2021) to predict the ultimate bearing capacity. However, previous studies have indicated that HBM tends to produce unstable predictions and overly wide confidence intervals when the available data is limited or exhibits high variability. Therefore, this study proposes three improvement strategies: model simplification, Tailored Clustering Enabled Regionalization (TCER), and dimension reduction (DR), aiming to enhance the robustness and practical applicability of the model predictions. Nonetheless, there is no universally optimal model applicable across different foundation types and data conditions. To address this, this study further integrates multiple models and develops an HBM-Based Stacking Ensemble Learning approach that employs minimum differential entropy as the selection criterion to dynamically identify the best-performing model. The effectiveness of this method is validated across four foundation types and twelve independent databases, with Root Mean Square Error (RMSE) under cross-validation adopted as the performance indicator. A systematic analysis is conducted to verify the feasibility and advantages of the proposed method in multi-foundation bearing capacity prediction.