Feature Extraction Based on Manifold Learning for Radio Fingerprint

Wireless Local Area Network (WLAN) fingerprinting has been extensively studied for indoor localization due to the pervasive facilities. Conventional fingerprint database is composed of a set of raw Received Signal Strength (RSS) which is not processed to features. Even though it provides adequate results in some cases, but for large-scale environment, it brings the storage problem and computational complexity due to the high dimensionality. To address these problems, this paper presents a feature extraction algorithm using a manifold learning called T-distributed Stochastic Neighbor Embedding (TSNE) which extracts these non-linear fingerprint features and reduces the dimensionality simultaneously at offline stage. Then to increase positioning accuracy, out-of-sample extension method is proposed to process the online record to achieve the same dimensionality as the reduced offline database. Furthermore, when facing the major bottleneck of dimensionality reduction (DR) technologies that determining the proper value of dimensionality, we utilize intrinsic dimensionality estimation method to obtain the best dimensionality previously. Experiments are conducted in an actual indoor large-scale environment, and the results demonstrate our approach performs perfectly which reduces the original dimensionality 168 to 10 and achieves better position accuracy simultaneously.