機器學習應用於橋梁沖刷監測系統之研究

Machine Learning for Bridge Scour Monitoring System

張世昇(Shih-Sheng Chang)

指導教授 ： 陳俊杉

國立臺灣大學/工學院/土木工程學研究所/碩士(2018年)

https://doi.org/10.6342/NTU201802257

摘要

鑒於台灣處於火山地震帶，地形以山地為主，地勢陡峭、河川水流湍急而導致橋墩沖刷問題十分嚴重，本研究期望藉由瞭解強降雨事件對橋樑沖刷的影響，預測沖刷深度並提早發出預警，達到橋梁之安全使用的目標。
過往的沖刷深度預測多以經驗公式推估，在準確度上常有很大的誤差，而現今有許多橋墩沖刷之監測技術可以即時性並準確地量測沖刷深度，因此可以收集到非常大量的水位、流速及沖刷深度資料。本研究透過支撐向量機(Support Vector Machine)、隨機森林(Random Forest)、深度學習(Deep Learning)等人工智慧(Artificial Intelligence)的方法，並以國家地震工程研究中心在濁水溪流域的名竹大橋、中沙大橋及自強大橋所監測的資料庫做為模型之訓練樣本和測試資料，建立水位、流速和沖刷深度之預測模型，並探討水位、流速、作用時間與沖刷深度之關聯性。
現地資料分析結果顯示水位和流速與沖刷深度關聯性極高，機器學習之預測模型明確表現作用時間對沖刷深度預測準確度的影響，並憑藉機器學習模型高度非線性的特性反應作用時間的影響以提升預測的精準度，本研究針對不同座橋梁的特性提出合適的預測架構及分析流程，而研究結果顯示利用深度類神經網路配合水位和流速兩種參數以1分鐘作用時間所建立的模型結果為佳，其在中沙大橋的平均預測精準度可以達到83%，自強大橋則為63%，雖然預測模型對於沖刷歷程無法100%精準預測，但每一場強降雨事件所導致的沖刷深度峰值皆可以準確地預測，並達到預警之作用。

關鍵字

橋墩沖刷 ； 支撐向量分類法 ； 隨機森林 ； 深度學習

並列摘要

With special characteristics of river, such as steep slope and rapid flow, bridge scour is an important safety issue in Taiwan. To guarantee the stability of bridge structure, ensure serviceability and give warning in advance during hazard-prone rainfall, the relationship between heavy rainfall events and local scour around piers needs to be resolved and is the focal point of this thesis.
In the past, empirical models were used to predict scour depth. Nowadays, different kinds of real-time monitoring techniques open the gateway to measure field data instantly and precisely. It is thus impetus to develop a predictive model of scour depth based on these data.
In this thesis, the dataset acquired from National Center for Research on Earthquake Engineering, including field data of Zhuoshui River at Mingzhu Bridge, Zhongsha Bridge and Ziqiang Bridge, is used to develop the predictive model of scour depth. Three machine learning methods, including Support Vector Machine, Random Forest, and Deep Learning, are adapted.
We find that time effect affects prediction accuracy significantly, and this issue can be resolved by highly nonlinear characteristics of machine learning. Besides, it is verified that scour depth has high correlation with water level and flow rate. A specific method of choosing the prediction time frame based on the characteristic of bridge and the corresponding analysis procedure are proposed. The feasibility of the proposed method are discussed.

並列關鍵字

Local Scour around Pier ； Support Vector Machine ； Random Forest ； Deep Learning

參考文獻

[1] Shirole, A. M., & Holt, R. C. (1991). Planning for a comprehensive bridge safety assurance program. Transportation Research Record, 1290, 39-50.

[2] Melville, B. W. (1984). Live-bed Scour at Bridge Piers. Journal of Hydraulic Engineering, 110(9), 1234-1247.

[3] Sheppard, D. M., & Miller, W. (2006). Live-Bed Local Pier Scour Experiments. Journal of Hydraulic Engineering, 132(7), 635-642.

[4] Raudkivi, A. J., & Ettema, R. (1983). Clear-Water Scour at Cylindrical Piers. Journal of Hydraulic Engineering, 109(3), 338-350.

[5] Graf, W. H., & Istiarto, I. (2002). Flow pattern in the scour hole around a cylinder. Journal of Hydraulic Research, 40(1), 13-20.