摘要
为了解决"基于卡尔曼滤波的神经网络算法"由于目标模型不确定性而出现的预测信息不准确,甚至发散的问题以及由于传感器误差而造成的估计误差偏大导致跟踪失效的问题,提出将强跟踪滤波(STF)应用于人工神经网络算法中,以神经网络中各层连接权值构成STF滤波的状态向量,引入时变渐消因子,强迫残差具有正交性或近似正交性,以克服上述问题。实验仿真证明,改进后的算法提高了网络训练速度、滤波精度、数值稳定性以及对目标的跟踪性能。
Algorithm for feedforward neural network based on Kalman filter has some problems, the prediction information is inaccurate,even emanative because of the target model's inaccuracy, and the larger estimate error makes tracking - disable because of sensors' error. Aiming at these shortcomings,a learning algorithm based on Strong Tracking Filtering (STF) filter is proposed for training a neural network,it regards all the weight values as the states. The fading factor is introduced,and residuals are forced to have orthogonality or approximately orthogonality to solve these problems. Simulation results show that the new algorithm improves rapidity of network's convergence,data's accuracy, stability and target tracking performance.
出处
《现代电子技术》
2009年第2期59-62,共4页
Modern Electronics Technique