【作者】 李慧；

【导师】 张庆范；

【作者基本信息】 山东大学 ， 电力电子与电力传动， 2011， 博士

【摘要】 随着生活水平的提高人们对居住环境的要求也越来越高,如何最大化地满足人们的舒适性要求,减少能源损耗和其它潜在的浪费是智能居住环境研究的目标。智能居住环境是指将各种内嵌的智能设备和住宅设备通过一定范围的内部网络连接起来,能够通过这种网络提供各种服务,并与外部世界相连接,同时保持这些设备与住宅的协调,并能根据用户的需求主动做出响应的环境。其具有泛在性、透明性和智能化的特点。智能居住环境作为信息社会体系的边缘结构,既与人们的日常生活密切相关,又与其作为信息社会有机单元所具有的功能和特性分不开。因此,既要在宏观理论指导下建立完善合理的开放式体系构架,又要从微观角度考虑用户的具体需求,使智能居住环境系统更具个性化和智能性。随着计算机、微电子、通信、智能控制等技术的发展,最近十几年智能居住环境的理论与应用技术的研究受到广泛关注。但目前的研究在宏观上缺乏完善的系统构建的理论体系,在微观上没有考虑时间的历史性、系统的自学习和自适应能力。由各大公司开展的项目重点是如何实现家庭和设备之间的通信,大多数的家庭自动化技术的方案,通常只采用简单的自动化,而很少采用人工智能,也很少强调对用户行为的学习和自适应。本文致力于智能居住环境理论体系的深入研究,建立具有开放性的、基于Multi-Agent的系统体系结构。根据智能居住环境的特点,研究智能居住环境的推理、学习、控制策略,提高整个系统的决策水平。在满足用户舒适性的前提下,最大限度地节能。本文的主要研究内容和创新成果如下：1.提出了一种基于ZigBee无线传感器网络的智能居住环境MAS(Multi-Agent System)体系结构,其中ZigBee无线传感器网络的协调器对应MAS中的管理Agent,路由器对应功能Agent,终端设备对应现场Agent。同时建立了现场Agent、功能Agent、管理Agent的内部结构模型,为智能居住环境MAS中各Agent之间的协作、冲突消解、组织联盟提供了基础构架。2.用户的行为识别是实现透明的、泛在的、智能的居住环境关键技术之一。提出了一种新型One Pass神经网络,该神经网络可根据智能居住环境中低水平传感器(例如,位移传感器、压力传感器等)的信息识别用户的高水平行为(例如,睡眠、学习、听音乐等)。该神经网络的结构包括三层：输入层、中间层、输出层。输入层为传感器状态层；输出层为用户行为层；中间层为隐含层,其维数与输出层相同。在该神经网络上增加时间信息可用于识别用户的非正常行为。One Pass权系数学习方法由于其学习算法简单,不需要迭代,占用内存少,可便于嵌入式微处理器实现在线学习。实验结果表明该算法是透明的、简单的、有效的。3.分别提出了基于用户喜好学习的智能模糊Agent和基于Hebb规则的输入输出动态关联算法。并将其结合到一起,形成一种新型动态关联智能模糊Agent,实现对用户喜好的学习,并主动控制居住环境的设备。在智能居住环境中,不同用户的喜好是不同的,同一用户的喜好也会随着时间的推移而发生变化,这就要求智能Agent应具有演化功能。提出的新型智能模糊Agent包括5个阶段,①采集输入输出数据对；②隶属度函数学习；③萃取模糊规则；④Agent控制；⑤在线自适应学习。依据采集到的传感器和执行器的信息,学习模糊Agent的隶属度函数和模糊规则。当用户的喜好发生变化时,系统会快速地学习隶属度函数和最优化模糊规则。此外,在智能居住环境中,嵌入式Agent通过网络联系到一起,具有智能推理、规划和学习的能力,而大量的互相连接的嵌入式Agent必将导致系统通信和计算负荷的增大,降低了系统的执行效率。在Hebb神经网络的基础上,提出了一种新型智能居住环境嵌入式Agent动态结构关联算法。当系统产生一个事件,将重新计算传感器Agent和设备Agent之间的关联权值。根据得到的关联权度矩阵,将嵌入式Agent动态划分为多个组,从而将模糊规则库划分为多个模糊规则子库。提出的上述方法可有效地降低Agent的模糊规则数,提高Agent的学习效率并减少嵌入式Agent之间的网络通信。4.居住环境是一类异常复杂和难以控制的高维非线性系统,用常规的建模方法很难建立系统的数学模型。针对智能居住环境的特点,充分利用系统长期运行积累的传感器状态及设备操作的历史数据,提出了一种基于聚类的超闭球CMAC神经网络算法,用于智能居住环境预测控制动态系统非线性建模。通过输入数据的模糊聚类确定神经网络节点数和节点位置,并采用输入输出数据模糊推理优化算法确定神经网络初始权值。与超闭球CMAC神经网络算法比较,该算法可有效地降低神经网络节点数,提高学习精度。5.静态的热环境易造成人体热适应能力降低,对健康不利。动态的热环境与自然环境相似更有利于用户的健康。提出了一种基于用户热舒适区学习的智能动态热舒适控制系统。为了满足不同用户对热舒适的需要,提出了基于PMV指标的个人热舒适区模糊学习算法,通过用户热舒适喜好的学习在线修改个人热舒适区。在计算实验的基础上提出了动态热舒适控制策略,使热舒适度在舒适区和节能区周期性交替变化。实验表明,该方法既满足用户的热舒适性需求,对用户的健康有利,与静态热舒适控制相比节能效果明显。总之,本论文对智能居住环境的学习和控制策略进行了深入研究,研究成果可为智能居住环境的系统平台构建提供技术支持,为智能居住环境的用户行为识别、自适应、长期的学习机制、系统的非线性建模、动态热舒适控制等提供理论依据。更多还原

【Abstract】 With the improving of living level, the people have higher demands for their inhabited environments. How to maximally meet their comfortable need and reduce energy consumption and other potential waste are the goals of HE (intelligent inhabited environments). HE are enclosed living spaces equipped with embedded intelligent equipments that connect with residential equipments through a range of internal network, which can provide a variety of services by connecting with external world and actively respond to the environments according to the needs of the users, while maintaining these equipments coordinating with residences. HE can be characterized by its ubiquity, transparency and intelligence. As the information edge structure of social system, the HE closely relate to the people’s daily lives and they are inseparable with the functions and features as the information society units. Therefore, not only the perfect and reasonable open architecture should be established under the guidance of macro theory, but also the user’s specific needs should be considered from the micro point, which make the system more personal and intelligent. With the development of computer, microelectronic, communication and intelligent control, the theory and application technology for HE were researched widely in recent 10 years. But the current researches lacked of perfect theoretical system in macro and did not consider historic time, the self-learning and adaptive capacity of system in micro. The study projects of the major companies focused on how to achieve communication between homes and equipments. Most of the home automation technology solutions usually only adopted simple automation, and rarely used artificial intelligence, and little emphasized the learning and adaptive of user behaviors.This paper is devoted to further study the theoretical system of HE and built an open Multi-Agent System (MAS) architecture. According to the characteristics of HE, the reasoning, learning and control strategies are studied to improve the decision-making of system, which can meet the needs of users’comfort and maximize the energy efficiency. The main research work and innovative fruits are as follows.1. A MAS architecture developed for HE based on ZigBee wireless sensor network is presented. The management agent in the management layer corresponds to the coordinator of ZigBee network. The function agents in the function layer correspond to the routers of ZigBee network. The bus agents in the bus layer correspond to the end-devices of ZigBee network. Moreover, the structure models of bus agent, function agent, and management-agent are built to provide the infrastructure for the cooperation, conflict elimination, and organizational alliance of the agents.2. The user’s activities recognition is one of the key technologies to realize ubiquitous, transparent and intelligent inhabited environments. A novel One-Pass neural network system is presented which is able to recognize different high level activities (such as "sleeping", "learning", "listening music", et al) based on simple sensors(such as "move sensor", "pressure sensor", et al) for HE. The neural networks architecture includes three layers which are input layer, middle layer and output layer. The input layer is sensor states layer. The output layer is user behaviors layer. The middle layer is hidden layer and its dimension is same as the output layer. The neural network system adding temporal informations is able to recognize abnormal behaviors. Due to the fact that the One-Pass learning method of weight ratios has the characters of simplicity, no iteration, and lower memory, the embedded computers can be trained in an online mode. Experiment results show that this method is transparent, simple and effective.3. The intelligent fuzzy agent based on user preference learning and the input-output dynamic associated algorithm based on Hebb neural network are proposed, respectively. On the basis, a novel dynamic associated intelligent fuzzy agent is build that associates the dynamic associated algorithm with the intelligent fuzzy agent. It can realize the learning of user preferences, and then actively control the devices of inhabited environment. The different user’s preferences and needs are different In HE, and the same user’s preferences will change over time, which desire that the intelligent agents should have evolvement function. The intelligent fuzzy agent proposed includes five phases:①Capturing input-output data pairs.②Membership function learning.③Extracting fuzzy rules.④Agent control.⑤Adaptive learning algorithm online. Initially the system’s fuzzy rules are extracted from the collected information of sensors and actuators in HE, after that the system rapidly optimizes the fuzzy rules when the user’s preferences change. Moreover, the interconnected embedded agents by network have the capabilities of intelligent reasoning, planning and learning in HE. However, the multitude of interconnected embedded agents can result in major load in network communication and calculation, which decreases the system’s execution rate. A novel dynamic structure association algorithm for embedded agents in HE is proposed based on Hebb neural network. The association weight values between sensor agents and device agents will be calculated and updated if an event occurs. The embedded agents can be dynamicly divided into multiple groups according to association weight matrix, and then the fuzzy rules base can be divided into multiple sub-fuzzy rules bases. The methods proposed can reduce the number of fuzzy rules, improve the learning rates of agents, and decrease the network communication among embedded agents.4. The inhabited environment is a kind of very complex and control difficultly high-dimensional nonlinear system. It is hard to build the system’s dynamical mathematical model with the conventional methods. According to the characteristics of HE, fully using the accumulated historical data of sensor status and equipment operations during the system’s long-term running, an improved hyperball CMAC neural network algorithm based on clustering is proposed for HE nonlinear dynamic modeling of predicted control. A fuzzy clustering algorithm is adopted to determine the node number and node locations by clustering the input data. A fuzzy inference optimization algorithm is proposed to determine the initial weight values of neural network based on input-output data. Compared with the hyperball CMAC, the improved algorithm can effectively reduce the neural network nodes and improve the learning accuracy.5. The static thermal environment is unfavorable to the human’s health as it can reduce the ability of human’s heat adaptation. The dynamic thermal environment is favorable to the human’s health as it is similar to the natural environment. A dynamical thermal comfortable control system is presented for the inhabited environment based on the learning of human’s thermal comfort zone. The fuzzy learning algorithm of personal thermal comfort zone is proposed based on predicted mean vote (PMV) index, which can modify the personal thermal comfort zone on line with the learning of human’s thermal preference to meet the needs of different humans. The dynamical thermal comfort control strategy is proposed with computational experiments, which make the thermal comfort zone and energy saving zone change periodically. The experiment results demonstrate that this method can meet the human’s thermal comfort need and reduce the energy consumption, whilst it is favorable to the human’s health.In sum, the learning and control strategies for HE are studied deeply in this paper. The research results can be used to provide technology support of system platform building for HE. Moreover, the research results can be used to provide the theoretical basis for user’s behavior identification, user’s adaptive, long-term learning mechanism, system nonlinear modeling, dynamic thermal comfort control, et al.更多还原