Construction Of Semi-supervised Indoor WLAN Location Fingerprint Database Based On Manifold Learning

With the rapid development of short-range radio technology and popularization of Wireless Local Area Network(WLAN),the users become to enjoy the convenient and fast information interaction function from WLAN,the demand for the efficient and accurate indoor location services is also growing.Since the WLAN based indoor localization technique only needs software upgrade and update to realize localization function through software calculation,it is widely used in indoor environment.However,due to the fluctuation of localization signal in the complicated indoor environment,to guarantee localization accuracy,the conventional WLAN fingerprint based localization algorithm need to spend a large amount of time and labor cost in offline phase to construct location fingerprint database.Therefore,how to preserve localization performance as well as reduce the construction cost of location fingerprint database is a significant problem.In response to this problem,based on a large batch of the unlabeled fingerprint samples collected by users,this thesis proposes a multi-information manifold learning based semisupervised location fingerprint database construction scheme,which guarantees high localization accuracy as well as greatly reduces the database construction cost in off-line phase.The main research contents of this thesis are summarized as follows.1.According to the number of required labeled fingerprint,this thesis discusses four different database construction approaches used in indoor WLAN fingerprint based localization system.At the same time,this thesis analyzes the theoretical foundation and typical approaches of manifold learning,and then gives the manifold learning based semisupervised indoor WLAN location fingerprint database construction approach.2.This thesis modifies the conventional manifold learning algorithm to fit WLAN fingerprint based localization system.First of all,aiming at the problem that the unsupervised graph construction approach cannot describe the information of labeled data,a new graph construction approach based on labeled data is proposed to construct highdimensional manifold,and then the corresponding theoretical proof is given.Second,to further fit the actual WLAN environment,different distance measurement approaches are proposed to measure the distance of adjacent points with respect to different feature spaces with the purpose that the manifold can better depict high-dimensional data feature and consequently improve the usability of location fingerprint database.3.This thesis utilizes hypothesis testing theory to derive the minimum number of samples required at each reference point to effectively depict signal feature,and meanwhile integrates the cubic spline interpolation algorithm to realize the initial database construction.Then,by using the multi-information manifold learning algorithm,the physical location information of labeled data with the timestamp information of unlabeled data are integrated to a large amount of unlabeled data and consequently construct the semi-supervised location fingerprint database.At the same time,extensive experiments are conducted to verify the effectiveness and reliability of the designed location fingerprint database construction approach.The confidence probability of positioning error within 3 meters of the system is 80.78%,which is 27.48% higher than that of the traditional RADAR system.