【摘要】：為實現(xiàn)對機載設(shè)備工作狀態(tài)的在線狀態(tài)預測,提出了一種稀疏核增量超限學習機(ELM)算法。針對核在線學習中核矩陣膨脹問題,基于瞬時信息測量提出了一個融合構(gòu)造與修剪策略的兩步稀疏化方法。通過在構(gòu)造階段最小化字典冗余,在修剪階段最大化字典元素的瞬時條件自信息量,選擇一個具有固定記憶規(guī)模的稀疏字典。針對基于核的增量超限學習機核權(quán)重更新問題,提出改進的減樣學習算法,其可以實現(xiàn)字典中任一個核函數(shù)刪除后剩余核函數(shù)Gram矩陣的逆矩陣的前向遞推更新。通過對某型飛機發(fā)動機的狀態(tài)預測,在預測數(shù)據(jù)長度等于20的條件下,本文提出的算法將預測的整體平均誤差率下降到2.18%,相比于3種流形的核超限學習機在線算法,預測精度分別提升了0.72%、0.14%和0.13%。
[Abstract]:In order to predict the working state of airborne equipment, a sparse kernel incremental learning machine (ELM) algorithm is proposed. In order to solve the problem of kernel matrix expansion in nuclear online learning, a two-step thinning method based on instantaneous information measurement is proposed, which combines construction and pruning strategy. By minimizing dictionary redundancy in the construction stage and maximizing the instantaneous conditional self-information of dictionary elements in the pruning stage, a sparse dictionary with fixed memory scale is selected. In order to solve the problem of kernel weight updating of incremental overlimited learning machines based on kernel, an improved sample reduction learning algorithm is proposed, which can update the inverse matrix of the residual kernel function Gram matrix after the deletion of any kernel function in the dictionary. Through the state prediction of a certain aircraft engine, under the condition that the predicted data length is equal to 20, the overall average error rate of the prediction is reduced to 2.18%. Compared with the on-line kernel overlimited learning machine algorithm of three manifolds, the prediction accuracy is improved by 0.72%, 0.14% and 0.13%, respectively.
【作者單位】： 海軍航空工程學院科研部;
【基金】：國家自然科學基金(61571454)~~
【分類號】：TP181;V267

【相似文獻】

相關(guān)博士學位論文 前1條

1 雷達;基于智能學習模型的民航發(fā)動機健康狀態(tài)預測研究[D];哈爾濱工業(yè)大學;2013年

，

本文編號：2509738