使用UFN類神經網路與啟發式搜尋策略進行剪力構架模型修正

Abstract

工程上模型修正(model updating)指的是修改數值模型中的物理參數使得計算結果與量測資料之間的差距盡可能最小化，修正後的模型可用來評估現況或用於即時需求。在結構工程方面，結構模型修正通常是先根據量測到的結構震動資訊計算取得模態資訊，再修正結構模型中的結構參數使得數值產生的模態資訊與量測取得的模態資訊盡可能地接近，換句話說，結構模型修正算是一種工程最佳化問題。演化式計算是處理結構模型修正問題的好方法，然而由於演化式計算屬於全域式搜尋，搜尋最佳解經常需要非常多的計算時間，在探索搜尋空間階段會經常找到不重要的資料點位。因此本研究打算以區域搜尋的方式尋找最佳解，利用非監督式類神經網路預測區域最佳解，再啟發式移動搜尋區域，預測新的最佳解，如此反覆迭代搜尋全域最佳解。整體計算效率會比演化式計算在全域搜尋空間內搜尋更好。經過數值模擬分析，對於剪力房屋構架在假設質量已知不變且不考慮結構阻尼系統的情況下，本研究方法求出的結構勁度矩陣，準確度可以達到與真實結構物的勁度矩陣平均誤差僅0.71%。為了提高本研究的實用性，不完整模態資訊與量測不確定性等問題都會被考量。

In engineering, model updating aims to reconstruct the system properties of a numerical model to fit the system properties of a real system by updating physical parameters of the numerical model. The updated model can be used to estimate the real system, or be provided for online requirements. In structural engineering, model updating generally is to update the structural parameters of a structural model for minimizing the difference between computed modal parameters and real modal parameters. The real modal parameters can be achieved by calculating the measured vibration data of real structures. That is, updating structural model is a kind of structural optimization problem. Evolutionary computation is a feasible means for solving the model updating problem in structural engineering. However, evolutionary computation is a global search method. The large computational time is a heavy burden because it often explores the unimportant data points. Therefore, this research intend to use a local search method. An unsupervised fuzzy neural network is adopted to predict the local optimum of the searching region; then heuristically moving the searching region to predict a new optimum until finding the global optimum. The computing efficiency will be better than the evolutionary computation that search the global optimum in the whole region. To consider the practicality, this research will take into account the factors such as incomplete modes and uncertainty in measure.