資料探勘技術於橋梁健康管理之實務應用研究

Application of Data Mining Techniques in Bridge Health Management

馮紹庭(Shao-Ting Fong)

指導教授 ： 呂良正

國立臺灣大學/工學院/土木工程學研究所/碩士(2020年)

https://doi.org/10.6342/NTU202001900

摘要

本研究利用 Python 程式語言，依循資料探勘技術流程並使用機器學習模型，用於預測橋梁橋面板(Deck)之健康狀態，以資料科學的角度給予橋梁管理者檢測進行時機上的建議。
資料部份主要使用美國聯邦公路總署(Federal Highway Administration, FHWA)所建立之全國橋梁清冊(National Bridge Inventory, NBI)與元件資料庫(NBI Element Data)。
本研究可分為兩部份。前半部著重於建構健康狀態預測模型，先介紹資料前處理與多年度資料整合方法，再比較隨機森林模型(Random Forest)與搭配 Entity Embeddings 前處理之深度學習模型在不同閥值下的預測結果，並自分類過程中找出重要參數。此外，亦探討投入單次檢測資料與多次資料之表現差異。
後半部則著重於預測模型之實務應用，制定資料更新規則，將原先之單次預測模型推廣至連續預測。過程中亦測試不同方案下之失誤率與節省資源率，輔助橋梁管理者擬定檢修計劃。最後，除 NBI 資料庫之橋梁基本資料外亦加入元件資料，探討元件資料是否可精進模型表現。

關鍵字

橋梁檢測 ； 資料探勘 ； 機器學習 ； 深度學習 ； NBI橋梁資料庫

並列摘要

This thesis develops a prediction model to estimate the health condition of bridge deck components through data mining approaches and machine learning algorithms, aiming to provide bridge owners with bridge inspection schedule recommendations.
National Bridge Inventory (NBI) Database and NBI Element Data from the United States’ Federal Highway Administration are used in this research.
This thesis can be classified into two parts. The first part focuses on constructing the prediction model. Firstly, we introduce our data preprocessing and multi-year data integration methods. Secondly, we test the Random Forest model and deep learning model with the Entity Embeddings layer under different thresholds, analyze their prediction results and extract important features after the classification. Lastly, we respectively use single inspection data and multiple inspection data to train the models and compare their performances.
The second part focuses on practical applications of the proposed model. We establish data update rules which enable the model to predict consecutively. Moreover, we demonstrate the model with different settings and summarize their error rate and inspection resources saving rate into a chart to assist bridge owners to formulate their maintenance plan.
On the other hand, in addition to the fundamental bridge data of the NBI database, element data is also added to discuss whether the element data can improve model performance.

並列關鍵字

Bridge inspection ； Data mining ； Machine learning ； Deep learning ； NBI database

參考文獻

AASHTO. (2010). AASHTO Bridge Element Inspection Guide Manual. In: American Association of State Highway and Transportation Officials.

Bektas, B. A. (2011). Bridge management from data to policy: Iowa State University.

Bektaş, B. A. (2017). Use of recursive partitioning to predict national bridge inventory condition ratings from national bridge elements condition data. Transportation Research Record, 2612(1), 29-38.

Box, G. E., Cox, D. R. (1964). An analysis of transformations. Journal of the Royal Statistical Society: Series B (Methodological), 26(2), 211-243.

Breiman, L. (1996). Bagging predictors. Machine learning, 24(2), 123-140.