類神經網路在脈衝式結構主動控制之應用

Abstract

本研究提出結合類神經網路的可調性(Adaptive)發展出一個新的主動控制法則。此法則之 控制力的作用方式為脈衝式主動控制(Active Pulse Control)，其精神在於減低地震力作 用下結構物反應之大量累積的效應，亦即在下一個量測到的資料進來前，施加一個微小時 段的脈衝式主動控制力，目的是要將目前已量測到的結構物反應抵消掉，而在這一個時段 內的地震力仍允許其造成結構物的反應，於是從整個歷時來看結構物的反應在每一個時段 都是從零開始，不斷歸零的結果使得結構反應大幅的下降了。因此不需要去預測地震力的 大小，而且因為每一脈衝式的控制力施加時間很短，在量測到資料後仍留有一小段時間可 供主動控制力的計算和準備，於是時間延遲的問題便可以獲得解決。結構物在經歷地震後 會產生些許破壞而造成結構系統參數的改變，但是破壞點與破壞程度的不可知，使得古典 控制理論所要求的結構系統參數矩陣不易隨之修改，但藉由類神經網路的可調性，在地震 過後將量測到整個歷時的地震力和主動控制力，配合量測到的結構反應當作檢核，進行權 值的學習及修改以符合破壞後的結構物特性，如此下次地震來襲時所施加的主動控制力就 能更適切。本研究最後以實際例子的模擬分析，可以發現脈衝式主動控制在減低結構物在 地震力作用下反應大量累積的效果，以及類神經網路可調性應用於結構主動控制的可行性 。 In this work, a new active neural network structural control model is develope d to control the civil engineering structures under seismic loadings. The stra tegy of the developed control model is to reduce the structural cumulative re sponses during earthquakes with active pulse control force. The effect of puls es is assumed to be postponed to the time that is asmall interval before the n ext sampling time so that the control force can be calculated in time and prep ared for applied. The problem of time delay was circumvented in the proposed c ontrol model. The parameters, such as damping and stiffness, of civil engineer ing structures will be changed, if it is damaged, after subjected to earthquak es. These parameters of structures under traditional control theory are diffic ult to be modified due to the several unknowns, such as damage of elements and degrees. By employing the property of adaptive in neural networks, a network c an be retrained with the detected structural responses as the desired output d ata. Then, these data are compared with the response of real structures. As a result, the more suitable control forces will be applied to thedamaged structu res during next earthquakes with a proper seismic response.From the illustrati ve examples, it is shown that the effect of reducing a larger cumulative struc tural responses under the proposed active pulse control model. Moreover, the p racticability of using the adaptive active neural network structural control m odel is also demonstrated in this research.