Research On The Semi-supervised Learning Indoor Localization Algorithm Based On The Quality Estimation Of Unlabeled Data

With the development of mobile technology and the promotion of wireless communication network,people's demand for location based services is increasing,which provides a broad prospect for location based services and applications.Among them,indoor fingerprint-based localization technology based on Wireless Local Area Networks(WLAN)has gradually become a research hotspot because of its extensive deployment and strong applicability.However,in order to maintain high localization accuracy,it needs a lot of manpower and time to calibrate fingeiprint data.Therefore,there are still many challenges in practical application.In this thesis,we start from the research of indoor fingerprinting localization technology based on WLAN,and analyze the main problems of indoor localization.And we define the current demand for reducing the collection workload.Because that the collection of offline radio map costed a lot of manpower and material resources,and as time go by,the wireless signal will change greatly,so it is necessary to update or even rebuild the radio map.Therefore,reducing the collection of fingerprint is a vital problem to be solved urgently.To solve the problem,a semi-supervised learning indoor localization algorithm based on unlabeled data quality estimation is proposed.The main works are as follows:1)In order to reduce the collection amount of radio map,this thesis proposes a localization method CSELM,which combines composite neighboring order graph and semi-supervised extreme learning machine(SELM).The algorithm mainly uses the SELM to train the model for localization.In the construction of semi-supervised learning graph,because that traditional K-nearest neighbor graph ignores the symmetry of edges,our CSELM method combines the minimum-maximum neighboring order graph to improve the symmetry of graph.In addition,the traditional SELM method only uses the similarity between RSS data to calculate the edge weights,and does not make full use of the time stamp information and location information of labeled data.Compared with RSS data,the latter two kind of information helps recognizing the prior information of graph.Therefore,this thesis uses three kinds of information to construct composite graph to improve the accuracy of semi-supervised manifold alignment.The experimental results show that the localization accuracy of the proposed CSELM algorithm is higher than that of other methods under the condition of sparse data collection.2)Aiming at the problem that a large number of unlabeled data are easy to lead into noise,this thesis proposes a quality estimation method of unlabeled data based on Gaussian mixture model.Our algorithm named CSELM-QE uses the high quality unlabeled data after evaluation for CSELM algorithm,to improve the localization accuracy.Firstly,we cluster the unlabeled data and then delete the outlier data by outlier detection method.Secondly,we fit the data set in each cluster with Gaussian mixture distribution and get the mean value of each cluster.Finally,we determine the quality value of each data item according to the difference between the data and the mean value,so as to evaluate the quality of all unlabeled data.The experimental results show that the localization accuracy of the proposed method combined with quality estimation is higher than that of the method using all unlabeled data,and it outperforms other methods.