AREA CLASSIFICATION FOR RF SIGNALS: A GRAPH NEURAL NETWORKS APPROACH

PhD Thesis Proposal Defence


Title: "AREA CLASSIFICATION FOR RF SIGNALS: A GRAPH NEURAL NETWORKS APPROACH"

by

Mr. Weipeng ZHUO


Abstract:

We study the problem of area classification for radio frequency (RF) signals,
where ambient RF signals, sampled by mobile Internet of Things (IoT) devices of
unknown location, are to be identified by the areas they are collected. The
solutions of the problem are applied in geofencing (for in-out area detection)
and floor identification (for indoor localization) tasks. We consider the
general but challenging scenario that the RF signals may exhibit significant
spatial and temporal heterogeneities. By modelling the heterogeneous RF signals
as a bipartite graph, we develop novel efficient graph neural network (GNN)
approaches for the tasks.

In geofencing, the ambient signals of a geofencing area is first collected for
offline training. In online process, the system would detect in real time
whether the signals, as collected and reported by a wearable attached to a
user, are outside the area or not. We propose GEM, a geofencing approach based
on network embedding for offline training and online detection. GEM efficiently
processes the RF signals using a novel GNN called BiSAGE, and captures the
multi-modal signal distribution within the area based on an enhanced
histogram-based classification model. Extensive experimental results show that
GEM outperforms state-of-the-art algorithms significantly (by up to 34% in
F-score).

For floor identification, we first consider the case where RF signals weakly
spill over across floors (e.g., office buildings). A large number of signals
are crowdsourced from users, and only a small number of them are labelled with
their floors. We propose GRAFICS, a graph embedding-based floor identification
system for the unlabelled signals. GRAFICS integrates a novel graph embedding
algorithm E-LINE to capture the signal relationship and to efficiently classify
the signals into different floors. It achieves higher than 96% in F-score with
much fewer labels per floor as compared with other advanced.

For floor identification with more substantial signal spillover (e.g., shopping
malls), we propose FIS-ONE, a novel floor identification system that requires
only one labelled signal record. FIS-ONE groups RF signals into different
clusters using our proposed RF-GNN, and then indexes the clusters with floor
numbers given the one labelled sample using a travelling salesman problem
formulation. Our extensive experiments demonstrates that FIS-ONE outperforms
other state-of-the-art algorithms substantially.


Date:                   Friday, 1 September 2023

Time:                   3:15pm - 5:15pm

Venue:                  Room 5510
                        lifts 25/26

Committee Members:      Prof. Gary Chan (Supervisor)
                        Prof. Andrew Horner (Chairperson)
                        Prof. Pedro Sander
                        Prof. Pan Hui (EMIA)


**** ALL are Welcome ****