DeepHome:一种基于深度学习的智能家居管控模型  

DeepHome:A Control Model of Smart Home Based on Deep Learning

作　　者：毛博[1] 徐恪[2,3] 金跃辉[1] 王晓亮[2,3] MAO Bo;XU Ke;JIN Yue-Hui;WANG Xiao-Liang(State Key Laboratory of Networking and Switching Technology,Beijing University of Posts and Telecommunications,Beijing 100876;Department of Computer Science and Technology,Tsinghua University,Beijing 100083;National Laboratory for Information Science and Technology,Tsinghua University,Beijing 100083)

机构地区：[1]北京邮电大学网络与交换技术国家重点实验室,北京100876 [2]清华大学计算机科学与技术系,北京100083 [3]清华大学信息科学与技术国家实验室,北京100083

出　　处：《计算机学报》2018年第12期2689-2701,共13页Chinese Journal of Computers

基　　金：国家自然科学基金(61170292;61472212);国家科技重大专项课题(2015ZX03003004);国家"九七三"重点基础研究发展规划项目基金(2012CB315803);国家"八六三"高技术研究发展计划项目基金(2013AA013302;2015AA015601);欧盟CROWN基金项目(FP7-PEOPLE-2013-IRSES-610524);清华信息科学与技术国家实验室(筹)学科交叉基金项目资助~~

摘　　要：作为物联网技术在日常生活领域的重要应用,智能家居产业近年来取得了快速发展.但是不同智能家居设备之间差异化的通信方式与割裂的功能,大大增加了用户管理与使用的复杂度.智能家居管控平台旨在整合异构网络环境下多种设备的数据监测与控制能力,为用户提供整体化家居服务.在面对多种设备协同、环境参数众多、用户需求难以确定等众多困难时,如何建立一种无感化、精确化、智能化的家居设备统一管控能力,是智能家居平台化过程中亟需解决的问题.为破解智能家居设备自动管控难题,文中基于深度学习方法提出DeepHome智能家居管控模型.DeepHome模型首先采用自编码网络构建设备模型,通过逐层无监督预训练,挖掘通用化设备特征;继而基于具体家居场景,综合多个独立设备模型构建多隐层学习网络,并使用家居环境数据进行模型整体训练.经过训练,DeepHome模型能够基于家居环境数据预测智能设备工作状态,并依照预测结果调整相应设备,实现对智能家居设备的自动化统一管控.由于智能家居平台尚处于起步阶段,现阶段仍缺乏能够有效描述家居环境整体的数据.文中设计调查问卷与数据收集网站收集不同用户的家居环境、设备数据与生活习惯,并基于所得数据构建智能家居环境仿真测试平台HomeTest,模拟生成批量家居环境数据,辅助模型训练.使用仿真数据进行10轮训练后,DeepHome模型对智能设备工作状态的预测准确率达到99.4%,较浅层神经网络模型提高了6.4%,较基于逻辑规则的控制方案提高了36.1%;与此同时,DeepHome在设备状态需要调整时的预测准确率达到74.1%,较浅层神经网络模型提高了4.7倍,较基于逻辑规则的控制的方案提高了13.2倍.在真实环境数据集下,DeepHome的对设备状态的预测准确率也达到了可以被用户接受的98.9%.实验表明,DeepHome模型能够充分发�As an important application of the Internet of things technology in the field of daily life, the smart home industry has made rapid development in recent years. However, the independent communications and the fragmented functions between different smart home devices increase the complexity of user management and use. Smart home control platform is designed to integrate monitoring data and control capabilities from lots of smart home devices and to provide integrated home services in heterogeneous network environment. To establish an automated, accurate and intelligent control model of smart home, we have to face the challenges of the coordination of multiple devices, the large number of environment parameters and the uncertainty of user requirements, which is a core problem in the process of smart home platformization. In order to solve the problem of automatic control of smart home devices, we propose DeepHome, a smart home control model based on deep learning algorithm in this paper. At first, the DeepHome model built device model based on autoencoder network model. Secondly, built the universal features of device model by unsupervised training layer by layer. Thirdly, integrated multi independent device model to build a multi hidden layer neural network and trained the complete model by stochastic gradient descent algorithm based on the dataset of home environment. After training, DeepHome model can predict the proper working status of smart devices and adjust the status by predictions, so as to control the smart home devices automatically and uniformly. Due to the smart home platform is still in its infancy, there is not enough effective data to describe the complete home scenarios currently. In this paper, we collected environment data, smart devices data and lifestyles from different users by questionnaires and a website we had established. Based on these data, we designed HomeTest, a simulation platform of smart home environment to generate more data of smart home to train and test the DeepHome model. Afte

关 键 词：智能家居 深度学习 机器学习 自编码网络 物联网 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]