Distance Metric Learning for Radio Fingerprinting Localization

摘要

Metric is not only a function to mesure the distance between data points, but also a main tool to evaluate the error of data analysis, so it is one of the important factors to be considered by expert systems and intelligent systems. A general metric function is difficult to adapt to all application scenarios. Thanks to the development of big data and machine learning technology, metric learning can be used to obtain an optimal metric for a specific application scenario. Considering the fingerprinting localization (FL) as an intelligent system for processing radio signals, this paper proposes two novel novel location fingerprint (LF) metric learning algorithms to improve the accuracy and adaptability of the system. The two proposed algorithms are named LMNN-LF and NCA-LF respectively, and they are based on two famous metric learning algorithms, LMNN and NCA. Considering the distribution characteristics of position fingerprints, we modified the original cost function in both schemes. In order to accommodate the low resolution of the fingerprint, an error radius is defined in our method. The distance relation of the LFs in high dimensional space is better described by the learned metric function, which improves the accuracy of the FL system compared with the general metric function or metric learning method. Moreover, the proposed methods can also effectively extract the features or reduce the dimension of LFs, improving the accuracy of other feature-based localization algorithms. Experiments on different data sets show that the metric obtained by the proposed method and several existing methods performs good in kNN localization, and performs good in LF feature extraction and dimensionality reduction.