類神經網路系統識別模式於結構健康檢測之研究

Abstract

傳統的類神經網路結構損壞評估方法使用類神經網路來萃取及儲存未損壞和損壞結構反應之樣本的知識。因為結構的破壞模式變化很大而且無法預測，所以使用破壞狀態及其相對應之反應所組成的樣本訓練類神經網路是不可行的方法。 類神經網路具有穩健(Robust)及容錯(Fault Tolerant)的特性，並且能夠有效地處理不確定及不完全的資料，因此類神經網路非常適合用來識別結構動力系統。識別結構動力系統之類神經網路的權值儲存著被識別之結構系統的結構特性。本文之目的即是直接或間接從識別結構系統之類神經網路的權值找出一些與結構系統參數有關的指標以作為整體結構健康檢測之用。文中提出了三個類神經網路系統識別模式－偏微分形式模式（Partial Derivative Form Model）、等效線性系統模式（Equivalent Linear System Model）、以及自由振動模式（Free Vibration Model）－來檢測結構是否健全。每一個模式都包括兩個步驟。第一個步驟是系統識別：使用神經系統識別網路（Neural System Identification Networks）識別結構系統未損壞及損壞之狀態。神經系統識別網路的輸入變數是前幾個時刻的結構反應、以及前幾個時刻和現時的外在擾動，輸出變數則是現時的結構反應。第二個步驟是結構損壞偵測：直接或間接從神經系統識別網路的權值找出有用的指標以作為結構損壞偵測之用。在偏微分形式模式、等效線性系統模式及自由振動模式中結構損壞偵測之有用指標分別為神經系統識別網路的輸出變數對輸入變數之一階偏微分值、等效線性系統之模態參數及經由訓練過的神經系統識別網路所產生之自由振動的週期和振幅。藉由比較結構未損壞和損壞狀態之指標值則可以評估結構系統變化的程度。 為了驗證本文所提之類神經網路系統識別模式於結構損壞偵測的可行性，文中列舉數值以及實驗之例子來說明。此外，我們建議未來的研究工作應該擴展到真實的結構物、研究如何決定損壞的位置及程度、研究結構特性和神經識別網路輸出變數對輸入變數之偏微分值的關係、以及發展即時結構健康檢測的方法。

Conventional artificial-neural-network-based (ANN-based) structural damage assessment methods use artificial neural networks (ANNs) to extract and store the knowledge of the patterns in the response of undamaged and damaged structure. Since the failure modes of a structure are so varied and so unpredictable, it is not feasible to train the neural network by furnishing it with pairs of failure states and corresponding diagnostic response. ANNs are robust and fault tolerant. They can also effectively deal with qualitative, uncertain, and incomplete information, thereby making them highly promising for identifying systems that are typically encountered in structural dynamics. The weights of the approximating neural network store the knowledge of the structural properties of the identified system. The objective of this research was looking for some useful indices for global structural health monitoring directly or indirectly from the weights of the approximating neural network. Herein, three ANN-based system identification models (Partial Derivative Form models, Equivalent Linear System models, and Free Vibration models) for structural health monitoring were presented. Each model comprises two steps. In the first step, system identification, Neural System Identification Networks (NSINs) are used to identify the undamaged and damaged states of a structural system. The inputs of the NSIN are previous structural responses and previous and current external excitations, and the outputs are current structural responses. In the second step, structural damage detection, some useful indices for detecting structural damage are searched directly or indirectly from the weights of the NSIN. The useful indices for structural health monitoring in Partial Derivative Form model, Equivalent Linear System model, and Free Vibration model are partial derivatives of the outputs with respect to the inputs of a NSIN, modal parameters of an equivalent linear system, and the amplitudes and periods of the free vibrations generated from a NSIN respectively. By comparing the indices of damaged state with those of undamaged state, the extent of changes can be assessed. Numerical and experimental examples were presented to demonstrate the feasibility of proposed models for structural health monitoring. Besides, further studies were suggested in the area of extending this work to realistic structures, investigating how to determine the location and extent of the damage, exploring relations between structural properties and partial derivatives of the outputs with respect to the inputs of a NSIN, and developing on-line structural health monitoring methods.