DeepTxFinder: Multiple Transmitter Localization by Deep Learning in Crowdsourced Spectrum Sensing

Abstract

As the radio spectrum has become the bottleneck resource with increasing volume of mobile data and ultra-dense network deployments, it is crucial to use spectrum more flexibly in time, space, and frequency dimensions. However, higher efficiency in spectrum usage facilitated by flexible spectrum allocation comes with a cost, namely the increased complexity of spectrum monitoring and management. Identifying the transmitters is at the interest of particularly spectrum enforcement authorities to ensure that spectrum is used as intended by the legitimate users of the spectrum. For a scalable, efficient, and highly-accurate operation, we propose a crowd-sensing based solution where sensing devices report their measured receive power levels to a central entity which later fuses the collected information for localizing an unknown number of transmitters. Our solution, referred to as DeepTxFinder, leverages deep learning to handle many sources of uncertainty in the operation environment: namely number of transmitters, their transmission power levels, and channel conditions (shadowing). Using deep-learning, DeepTxFinder distinguishes itself from the prior state-of-the art which requires knowledge of the number and transmission power of transmitters or require the transmitters to be well separated in space by tens to hundreds of meters making them ill-suited for application in expected ultra-dense deployment of smallcells. Moreover, we propose a tiling-based approach to increase the scalability of our proposal by reducing the computational complexity. Our simulation studies show that DeepTxFinder can provide a high detection accuracy even only by collecting data from a very small number of sensors. More specifically, with 1 %-2 % sensor density DeepTxFinder can estimate the number of transmitters and their locations with high probability which proves that sparse sensing is feasible.

Toward Integrated Large-Scale Environmental Monitoring Using WSN/UAV/Crowdsensing: A Review of Applications, Signal Processing, and Future Perspectives

“…To further increase the source identification accuracy, the maximum likelihood ratio test can be re-peated multiple times as a sequential probability ratio test (SPRT) [239]. Recently, some preliminary work investigated the adoption of deep learning approaches to cope with the presence of very few crowdsensing nodes, especially when operating in harsh propagation environments [240].…”

An RFML Ecosystem: Considerations for the Application of Deep Learning to Spectrum Situational Awareness

“…Positioning and localization play a crucial role in both military and commercial communications. For example, as the quantity of consumer-focused wireless devices continue to grow, positioning and localization become increasingly useful in emergency and safety applications, such as search and rescue operations [45], [46].…”

DeepMTL Pro: Deep Learning Based MultipleTransmitter Localization and Power Estimation

“…A large pixel subarea could certainly lead to high localization errors, due to very coarse granularity. We can address this by using a "tiling" technique, wherein we divide the given area into tiles, represent each tile by 100×100 size image and use our localization techniques, and do some post-processing to handle crosstiling issues (e.g., [6] uses overlapping tiles and employs a voting scheme inside the overlapping tile area).…”

“…A straightforward representation that represents the TXs with locations is to just use an array of (x, y) elements where each (x, y) element is the location of a transmitter, as in [6]. However, this simple representation is less conducive to efficient model learning, as the representation moves away from spatial representation (by representing locations as positions in the image) to direct representation of locations by coordinate values.…”

“…Prior Works. The MTL problem has been recently addressed in a few prior works, among which SPLOT [4], MAP * [5], and DeepTxFinder [6] are the most prominent. SPLOT essentially decomposes the MTL problem to multiple single-transmitter localization problems based on the sensors with the highest power readings in 1: Multiple transmitter localization using a distributed set of sensors.…”