【作者】 李玉榕；

【导师】 蒋静坪；

【作者基本信息】 浙江大学 ， 控制理论与控制工程， 2001， 博士

【摘要】 本文研究了多传感器信息融合与智能信息处理领域中的若干个问题。 首先研究多传感器融合在移动机器人导航中的应用。包括移动机器人的融合自定位问题：移动机器人利用光电编码器进行自定位，同时用扩展卡尔曼滤波器融合多个超声波传感器的测量值，采用回朔算法将融合值用于复位光电编码器，消除了光电编码器累积误差的影响，并能满足实时控制的要求：并提出一种基于Takagi-Sugeno模型的变结构模糊神经网络直接逆模型控制器，并应用于移动机器人的运动控制；利用模糊神经网络避障控制器融合CCD摄象机与超声波传感器探测到的环境信息，以实现机器人的安全避障。仿真实验说明了，提出的多传感器融合方法运用在移动机器人导航中的有效性。 将Dempster-Shafer理论与神经网络相结合，在时间域对多传感器的多次测量进行融合，以进行识别分类。该算法充分发挥了证据理论和神经网络各自的优点。工件识别系统实验表明该算法有效地提高了工件的识别率。 提出了分别基于Dempster-Shafer组合公式和两类模糊积分(Sugeno积分和Choquet积分)进行多个神经网络分类器组合的算法，这两种算法都考虑了每个分类器对不同类的识别能力的不同这一经验知识。将其应用于UCI数据库的分类和一个多传感器融合工件识别系统，结果表明了所提出算法的有效性。 粗糙集数据分析的主要优点在于它不要求任何关于被处理数据的先验或额外的知识，本文利用其对数据库进行分析计算， 自动获取数据库在各个层次上的规则集：在保证量化后的数据库具有最大一致性的前提下，利用遗传算法求取连续属性值的最优量化区间个数及各个区间分点值；同时将量化区间进行模糊化，将多层次清晰规则集转化为模糊规则集，利用模糊推理进行决策以提高鲁棒性。对UCI数据库的分类测试体现了该粗糙集数据分析算法的有效性。 提出了粗糙信息熵的概念，证明了粗糙集理论中知识不确定性与其所对应的粗糙信息熵之间的单调关系。并研究了与普通集合和模糊集合的不确定性相对应的粗糙信息熵，以及这两种集合的不确定性与其对应的粗糙信息熵之间的单调关系。利用粗糙信息熵的概念，能够从信息的角度充分反映不确定性的本质。并提出一种基于熵测量的属性简约算法，将其用于Iris数据库和Hsv数据库的属性简约，得到了理想的结果。更多还原

【Abstract】 Some aspects in the area of multisensor information fusion and intelligent information process are studied in this paper.The application of multiserisor fusion to the navigation of mobile robots is studied first. The fusion problem for self-localization is discussed. The mobile robot uses the photoelectric code-recorder to localize itself. At the same time, the Extended Kalman Filter is employed to fuse the multiple ultrasonic sensors, result of which is used to reset the photoelectric code recorder by the retroactive algorithm. The algorithm can not only eliminate the influence of the cumulative errors of the photoelectric code recorder, but also it can satisfy the requirement of the real-time control. A direct inverse model controller of fuzzy neural network with changeable structure based on Takagi-Sugeno inference is presented and it is used to the motion control of mobile robot. In order to avoid the obstacles successfully, detection results from CCD and ultrasonic sensors are fused by a fuzzy neural network, which acts as an avoidance controller. Simulation results show the validity of the proposed methods in the navigation of mobile robots.Multisensor information is fused in temporal field by combing Dempster-Shafer theory and neural networks in order to conduct recognition and classification tasks. The algorithm makes full use of advantages of both theories. The simulation shows that the method can effectively enhance the rate of the workpiece identification.Algorithms to combine the neural networks classifiers based on Dempster-Shafer theory and two kinds of fuzzy integral (Sugeno and Choquet integral ) respectively are proposed. The influences of the fact that every classifier has different classification ability for different class are all considered in these two kinds of algorithms. Several databases of UCI repository and a multiple sensor fusion system for workpiece identification are tested, showing the validity of these algorithms.The main advantage of rough sets data analysis is that it doesn’t require any prior or additional knowledge about the data, which is then used in this paper to analysis the database, acquiring automatically the hierarchical rule sets. In order to ensure maximum consistency of the quantiflcational data, the genetic algorithms is used to get the optimal number and points of division of quantification intervals. At the same time the quantification intervals is fuzzified and crisp rule sets are then transformed to fuzzy rule sets. Then the fuzzy inference is conducted to enhance the robustness. The validity of the proposed algorithm is proved through the test on some databases of the UCI repository.The concept of rough entropy is proposed. The monotony between the uncertainty of knowledge in the rough set theory and its corresponding rough entropy is proved. The rough entropy of the uncertainty of ordinary set and fuzzy set, and the monotonous relation between the uncertainty of these two kinds of set and their corresponding rough entropy, are discussed. Using the concept of rough entropy, the essence of the uncertainty can be fully reflected from the viewpoint of information. At last an entropy based attribute reduct algorithm is proposed. It is used to the attribute reduct of Iris and Hsv database, and the ideal results can be achieved.更多还原