【作者】 张洪涛；

【导师】 崔平远； 马培军；

【作者基本信息】 哈尔滨工业大学 ， 飞行器设计， 2007， 博士

【摘要】 近年来,多传感器信息融合理论及应用技术研究得到了快速的发展,已经成为一个重要的研究领域。其中,神经网络、模糊推理、Kalman滤波、粒子滤波等信息融合算法一直是国内外研究的热点。针对不同的实际应用需求,将这些方法和理论有机融合,是目前研究的一种趋势。本文在分析、研究目前主要融合算法的基础上,对BP(Back propagation Network)神经网络、Kalman滤波和粒子滤波等多传感器信息融合算法进行了改进和完善。首先,针对BP网络学习算法易陷入局部极小值、收敛速度慢的缺陷,提出了基于DFP(Davidon-Fletcher-Powell)和信赖域的BP网络学习算法,该方法兼顾了DFP在算法收敛速度、处理大残量问题以及信赖域法在保证全局最优性和解决数值不稳定问题方面的优越性,由于DFP算法是最优化理论中一类典型的拟牛顿法,具有超线性收敛速度。因此,采用DFP算法代替传统的梯度下降法进行BP网络学习,既保证了高效搜索和超线性的收敛速度,又减少了计算量,并且适于处理大残量问题,同时,在BP学习中引入信赖域法,又确保了算法的全局收敛性和稳定性,使算法更为实用。其次,针对大规模样本聚类的时间复杂度过高和聚类结果对经验参数设置的依赖性过强的问题,本文提出一种基于密度网格可变参数的自扩展聚类算法(PASCDU,a Parameter-Adjustable Self-expanded Clustering algorithm based on Density Units),算法先将数据空间分割成相同大小的网格单元,再将样本归一化后映射到相应的网格单元中,然后从指定密度较大的网格单元向周围扩展,直到其平均密度达到指定的下限或可扩展的聚类边界为止。聚类过程中,通过下限密度和均值密度来限制聚类间的过度扩展,如果有效样本的比率低于阈值,则自适应调整扩展密度并重新聚类。然后,将其与模糊神经网络有机融合,提出了基于该聚类算法的模糊神经网络学习算法,利用聚类算法从样本中提取规则数目,并将其应用于模糊神经网络建模,缩短了模糊神经网络的运算时间,提高了整个系统的效率。第三,考虑到虽然粒子滤波通过蒙特卡罗模拟来实现递推贝叶斯估计,在非线性非高斯系统中体现了良好的特性;但粒子滤波存在粒子退化现象的缺陷,针对这一问题,提出了一种新的重要性重采样算法,即分区重采样算法,其主要思想是根据多项式重采样与分层重采样算法的特点,把随机数区间划分成若干个区,每个区内的随机数任意排列,而区与区之间按升序排列。与目前常用的其它重采样算法相比,该方法提高了粒子滤波的平均性能。最后,针对基于Kalman滤波的神经网络学习算法适于含有较大噪声的样本的学习,但不能进行批处理操作、学习精度也不是很高的问题,本文提出了基于Kalman滤波的改进的BP网络学习算法,其主要思想是:首先在时间更新部分,通过其他学习算法给出预测权值的更新量,然后利用这个结果修改Kalman滤波增益表达式,获得了一套新的时间更新和测量更新计算公式,缓解了维数灾难和大量计算性问题,增强了神经网络的鲁棒性和抗干扰能力,同时采用批处理方式进行神经网络的学习,大大提高了神经网络的学习效率。仿真实验结果验证了本文提出的上述几类算法的有效性和可行性。更多还原

【Abstract】 In recent years, multi-sensor information fusion theory and its applications have obtained rapid advances, which become to be an important research field. Among these methods, Neural Networks, Fuzzy Reasoning, Kalman Filter, Particle Filter and other data fusion methods have been a hot spot of research all along. In view of different actual application requiremrnts, how to combine these methods with each other is a new trend in present.In this paper, several kinds of data fusion methods are studied and they are applied to spacecraft reliability estimation. In the theoretical respect such as Neural Networks, Kalman Filter、Particle Filter and their combination methods, the following research work have been completed in this paper.Firstly, a new BP learning algorithm based on DFP method and the trust region method is presented in this paper in order to get the solutions for some learning problems of traditional BP Neural Network, such as the slow rate of convergence and poor stabilization. DFP method is a classical Quasi-Newton algorithm which is super-linear convergent in the optimization theory. This algorithm maintains the highly efficient searching and super-linear convergence. In the mean time, it needs less computation cost and is proper to deal with large residual problems. Furthermore, combining with trust region method, this algorithm possesses the global convergence property and stability which make it more applicable. The simulated experiments show that the new algorithm has the advantages of fast learning speed, little computation cost, high accuracy and properness for large residual problems.Secondly, aiming at the problem that too much time of the large scale samples clustering and the result of the clustering excessively relying on the experiential parameters, an efficient self-expanded clustering algorithm with parameter-adjustable based on density units(PASCDU)is proposed in this paper. The whole data space is divided into several equal density units, before each data point is mapped into the relevant density unit according to the data point charter. Then cluster extends around from large density unit, until average density under the low-limit or cluster extends to the cluster’s edge. During the clustering process, we can prevent clusters from excessively extending by setting low-limit and average-limit. If the ratio of effectual samples is under the designate value, then automatically adjusts this value, and then cluster again.Thirdly, although particle filter implements recursive Bayesian filter based on Monte Carlo simulation, and it shows the merits in dealing with the nonlinear and non-Gaussian models, particle degeneration is difficult to conquer. To combat particle degeneration, a new importance resampling algorithm, i.e. divisional resampling is proposed. The mainly idea of divisional resampling is to combine polynomial resampling method and layered one, divide stochastic set of data into several sets which can insure that each of the sets has stochastic array of data and all the sets have an array according to sort ascending. Compared to the other general resampling algorithms, the method proposed improves the average performance of particle filter.Finally, aiming at the problem of the learning algorithm of neural network based on Kalman Filter being good at dealing with samples with great noise, being not good at dealing with batch manipulation and with low learning accuracy, this paper develops improved BP NN learning algorithm based on Kalman Filter. The main idea of improved BP learning algorithm based on Kalman Filter lies in: First, updates the estimation state parameters by other learning algorithm in the time updating process, and then modify the formula of Kalman Gain by these results. In this way, some new time update formulas and measure update formulas come up. This new method solves the problems of dimension disaster and large computation. And it enhances the robust of Neural Network and improves the learning ability by batch learning.The above mehods are proved to be efficient and work well by the simulation experimental results.更多还原