Neural Network Based Radio Fingerprint Similarity Measure

The radio signal Received Signal Strength Indicator (RSSI) based localization method is one of the most often adopted indoors localization methods, due to the fact that it can be used by any handhold devices on the market without any modification. This paper will present a new deep learning inspired model to predict locational distance/similarity between two points based on their RSSI measurement. This model uses carefully designed input features, neural network architecture, as well as purposely crafted pre-training to combine the features and statistics from multiple hand crafted RSSI to locational similarity models (reference models). The reference models include a RSSI difference based model (euclidean distance), number of visible signals by both measurements (Jaccard distance) and the rank difference of commonly visible signals in the comparing measurements (Spearman's footrule). Our evaluation shows the three reference models have very distinct strengths and weaknesses: the euclidean distance based model generates the most detailed prediction and has best estimation results when the two measurement points are close to each other; the Jaccard distance based model can only provide a very coarse estimation, however, it can distinct points that are far away; the Spearman's footrule based solution has overall good but coarse estimation, and its estimations represent the relative distance very well, especially in the middle range. The proposed method, as we expected, combines the best features from all three reference models and generates the best locational distance estimation.