Adaptive Construction Of Semi-supervised Fingerprint Database In Indoor WLAN

The indoor Wireless Local Area Networks(WLAN) fingerprinting localization technology positions using the widely deployed WLAN network currently and smart mobile device, no additional hardware overhead, and it has a free working frequency band and high positioning accuracy, so the technology is wide used in indoor environment. However, the manpower and time cost is highly required in the offline phase when fingerprint database is constructed in the algorithm, so the reduction of database construction costs is of great practical significance in the condition of positioning precision is ensured. We gathered unlabeled fingerprints samples through users’ cooperation in this thesis, and then designed a semi-supervised fingerprint database construction scheme, the scheme greatly reduces the workload of fingerprint acquisition in the off-line phase, and ensure the high positioning accuracy. The main research contents of this thesis are as follows:Firstly, the number of location labels in fingerprint samples collected by manpower is less, to solve this problem, we design a scheme of adaptive construction of label database and the scheme can be used when the feedbacks’ locations are stochastic distributed. On the basis of analysing the indoor WLAN signal propagation characteristics, we applied radial basis function interpolation on received signal strength estimation of non-feedbacks aimed at WLAN signal scattering properties. At the same time, we constructed some label databases in different interpolation algorithms, and compared their positioning performance though simulation experiment, to verify the validity of presented algorithm in this thesis.Secondly, because of the large number and the wide distribution of unlabeled fingerprint samples collected by users, we studied a method of preprocessing unlabeled samples. Namely mining the hidden features of unlabeled samples using Isomap dimension reduction and Clara clustering algorithm, and narrowing their regional scope. The method not only provides a convenient for the classifier training behind, but also improves accuracy of location label.Finally, after the adaptive construction of label database, a design scheme of constructing semi-supervised fingerprint database is presented based on Co-training algorithm. The algorithm extracts sufficient redundant views and trains the classifier in the continuously updated tag database, and mark the location label on unlabeled samples, ultimately the semi-supervised fingerprint database is constructed. The results of experiments show that the positioning performance of semi-supervised fingerprint database in this thesis is better, the cumulative error probability within 3m is increased 15.47% compared with the feedback database.