西北工业大学学报
西北工業大學學報
서북공업대학학보
JOURNAL OF NORTHWESTERN POLYTECHNICAL UNIVERSITY
2015年
2期
320-325
,共6页
神经网络%渐消滤波%学习算法%组合导航
神經網絡%漸消濾波%學習算法%組閤導航
신경망락%점소려파%학습산법%조합도항
针对BP神经网络在训练过程中易陷入局部极小值的问题,提出一种新的基于渐消滤波的神经网络学习算法。该算法采用渐消卡尔曼滤波对神经网络的权值进行更新,有效避免了梯度下降算法产生的局部极小问题;与卡尔曼滤波相比,在渐消滤波中充分考虑了最新量测值的影响,能更合理地利用新的有效信息,从而提高了学习算法的收敛速度。学习后的网络不仅具有普通神经网络的自主学习能力,而且具有渐消滤波的最优估计性能。将提出的神经网络算法应用于SINS/BDS组合导航系统进行仿真验证。结果表明,提出的算法在逼近精度方面优于BP算法和卡尔曼滤波算法,可以有效提高神经网络的泛化能力。
針對BP神經網絡在訓練過程中易陷入跼部極小值的問題,提齣一種新的基于漸消濾波的神經網絡學習算法。該算法採用漸消卡爾曼濾波對神經網絡的權值進行更新,有效避免瞭梯度下降算法產生的跼部極小問題;與卡爾曼濾波相比,在漸消濾波中充分攷慮瞭最新量測值的影響,能更閤理地利用新的有效信息,從而提高瞭學習算法的收斂速度。學習後的網絡不僅具有普通神經網絡的自主學習能力,而且具有漸消濾波的最優估計性能。將提齣的神經網絡算法應用于SINS/BDS組閤導航繫統進行倣真驗證。結果錶明,提齣的算法在逼近精度方麵優于BP算法和卡爾曼濾波算法,可以有效提高神經網絡的汎化能力。
침대BP신경망락재훈련과정중역함입국부겁소치적문제,제출일충신적기우점소려파적신경망락학습산법。해산법채용점소잡이만려파대신경망락적권치진행경신,유효피면료제도하강산법산생적국부겁소문제;여잡이만려파상비,재점소려파중충분고필료최신량측치적영향,능경합리지이용신적유효신식,종이제고료학습산법적수렴속도。학습후적망락불부구유보통신경망락적자주학습능력,이차구유점소려파적최우고계성능。장제출적신경망락산법응용우SINS/BDS조합도항계통진행방진험증。결과표명,제출적산법재핍근정도방면우우BP산법화잡이만려파산법,가이유효제고신경망락적범화능력。
Aiming at the local minimum problem in the training process of BP neural network, this paper proposes a novel neural network learning algorithm based on the fading Kalman filtering. This algorithm avoids the problem of local minimum value because it provides the fading Kalman filtering for weights optimizing in neural network train?ing. Compared with Kalman filtering, the fading Kalman filtering gives full consideration to the effects of the newest measurements and uses the effective information more reasonably. Consequently, the convergence speed of the algo?rithm is improved. The network after training not only has the autonomous learning ability of general neural network, but also has optimal estimation performance of the fading Kalman filtering. The proposed algorithm is applied to SINS/BDS integrated navigation system. Simulation results and their analysis demonstrate preliminarily that the ap?proximation accuracy of the proposed learning algorithm is better than those of BP algorithm and Kalman filtering al?gorithm. The proposed algorithm, we believe, can improve the generalization capability of neural network effective?ly.