Refined Autoencoder-Based CSI Hidden Feature Extraction for Indoor Spot Localization

Abstract

Wireless indoor localization technique has attracted wide attention recently. Fingerprint (FP) based method with received signal strength indicator (RSSI) is a popular approach due to easy implementation and robustness. Nowadays, fine-grained indoor spot localization resort to channel state information (CSI) owing to rich information property of CSI. However, due to a higher dimension of CSI compared to RSSI, CSI-based FP method requires higher storage and communication overhead, which is not suitable for most scenarios. In this paper, we propose a novel refined autoencoder-based CSI hidden feature extraction for indoor spot localization (RACHEL). Based on the concept of FP, we first introduce an autoencoder (AE) for the dimension reduction and feature discrimination. A low dimensional hidden feature of trained AE model is saved as FP database in the off-line stage. For indoor spot localization problems, users position is assumed to be close to one of the reference points. Therefore, CSI transforms to hidden feature space and users location is estimated by the nearest-neighbor algorithm in the on-line stage. Furthermore, to enhance the performance, instinctive AE is modified by considering corruption from time-varying environment and sensitivity between the hidden layer and input layer. Performance evaluations demonstrate that the proposed RACHEL can achieve 97.8% spot classification accuracy and yield a spaced savings of 94.3%.