Channel state information packet analysis and machine learning system for efficient wireless indoor localization

Abstract

Recently, many researchers have paid attention to a localization system using channel state information (CSI) obtained from radio WiFi communication on an IEEE 802.11 orthogonal frequency division multiplexing (OFDM) channel. Compared to the received signal strength index (RSSI) which contains only single-source information, CSI contains channel characteristics per-subcarrier that can improve localization accuracy. Nevertheless, due to radio noise, hardware offsets, and multi-path problems, the exact location of the transmitter device cannot be easily obtained using CSI. In this paper, by solving these problems, we intend to develop a practical localization system using machine learning that can be utilized in real building environments. We consider the CSI measurements received at multiple access points (APs) as multiview learning data so as to discriminate them to regression outputs, which are represented as normalized Cartesian coordinates.

At first, we introduce a method extracting useful characteristics of the CSI for localization. On a radio WiFi channel using an OFDM scheme, the complex-valued CSI phase is exploited as a form of normalized phasors consisting of in-phase and quadrature values. By finding the CSI phase difference between adjacent receiver antennas, we efficiently remove subcarrier-dependent hardware offsets. The extracted CSI data obtained from our construction method becomes an input for machine learning. We also take into account the data omission that occurs when a signal consisting only of non-line-of-sight (NLoS) paths cannot arrive at the receiver. In this case, our preprocessing using nonnegative matrix factorization (NMF) recovers the CSI of the original signal and completes the sparse CSI matrix that enables scalable and robust localization.

For application to grid topologies, we introduce a novel machine learning technique named CompFi using a partially connected neural network (PCNN). It consists of 3-layer partially and 1-layer fully connected neural networks (NNs) to make the best use of the CSI characteristics. CompFi is applied in a grid room environment with many structures and shows a localization accuracy of 1.66 m even under the slight movement of a transmitter device.

In addition, we consider the practical use of the localization system in corridor environments. The signals consisting only of NLoS paths are considered non-informative for localization since they have no LoS path information. Thus our variational inference-based machine learning techniques find informative CSI views within multiview CSI data. The proposed systems significantly outperform other existing machine learning-based systems, while still maintaining reliable localization accuracy even in a 30 % sparse network.

Based on our machine learning techniques, we design localization systems that operate in further complex building environments. The experiment space is divided into several sub-areas, and latent feature vectors are extracted by applying a separated NN to the CSI view corresponding to each sub-area. As non-informative latent features from multiple views are rejected, our view-selective deep learning (VSDL) and unsupervised view-selective deep learning (UVSDL) achieve localization accuracies of 1.28 m and 1.36 m in a complex building environment, which outperform the best-known accuracy in practical applications by 30 % and 25 % respectively. To the best of our knowledge, this is the first approach to apply variational inference and to construct an efficient and practical system for indoor localization. Furthermore, our work investigates a methodology for both supervised and unsupervised learning with multiview data where informative and non-informative views coexist.