[Paper Review] Multi-Label Wireless Interference Identification with Convolutional Neural Networks

Sergej Grunau, Dimitri Block+1 more|arXiv (Cornell University)|Apr 12, 2018

Respiratory viral infections research11 citations

TL;DR

This paper proposes a multi-label convolutional neural network (CNN) for wireless interference identification (WII) in license-free 2.4 GHz ISM bands, classifying up to six interfering signals alongside a primary utilized signal within 10 MHz bandwidth and 12.8 µs snapshots. The approach achieves near-perfect accuracy (≈100%) for same-technology interference in narrowband IEEE 802.15.1 and 802.15.4, and at least 90% accuracy for cross-technology interference involving IEEE 802.11b/g, demonstrating robustness in realistic coexistence scenarios.

ABSTRACT

The steadily growing use of license-free frequency bands require reliable coexistence management and therefore proper wireless interference identification (WII). In this work, we propose a WII approach based upon a deep convolutional neural network (CNN) which classifies multiple IEEE 802.15.1, IEEE 802.11 b/g and IEEE 802.15.4 interfering signals in the presence of a utilized signal. The generated multi-label dataset contains frequency- and time-limited sensing snapshots with the bandwidth of 10 MHz and duration of 12.8 $μ$s, respectively. Each snapshot combines one utilized signal with up to multiple interfering signals. The approach shows promising results for same-technology interference with a classification accuracy of approximately 100 % for IEEE 802.15.1 and IEEE 802.15.4 signals. For IEEE 802.11 b/g signals the accuracy increases for cross-technology interference with at least 90 %.

Motivation & Objective

• Address the challenge of reliable coexistence management in crowded license-free 2.4 GHz ISM bands used by heterogeneous wireless technologies.
• Overcome limitations of prior single-label WII systems by enabling detection of multiple interfering signals simultaneously in the presence of a primary utilized signal.
• Develop a practical WII solution constrained to realistic sensing parameters: 10 MHz bandwidth and 12.8 µs snapshot duration, reflecting real-world hardware capabilities.
• Enable non-cooperative coexistence management for legacy industrial wireless communication systems lacking control channels.
• Generate a comprehensive multi-label dataset simulating real interference scenarios for training and evaluating the CNN-based WII system.

Proposed method

• Design a deep convolutional neural network (CNN) that performs end-to-end feature learning from in-phase and quadrature (IQ) samples of wireless signals.
• Train the CNN on a multi-label dataset containing 450,000 frequency- and time-limited snapshots, each combining one utilized signal with up to six interfering signals.
• Use a 10 MHz bandwidth and 12.8 µs duration per snapshot to simulate real-time, low-latency sensing constraints of industrial wireless systems.
• Construct the dataset by combining single-label snapshots from IEEE 802.15.1, IEEE 802.11b/g, and IEEE 802.15.4 technologies across 15 frequency channels each.
• Apply a signal-to-interference ratio (SIR) of 1 for all interfering signals to simulate realistic coexistence conditions with equal power levels.
• Implement multi-label classification where each snapshot is labeled with all active interfering technologies present, enabling simultaneous detection of multiple interference sources.

Experimental results

Research questions

• RQ1 Can a deep CNN effectively identify multiple interfering wireless signals in the presence of a primary utilized signal under strict sensing constraints?
• RQ2 How does the classification performance vary between same-technology and cross-technology interference scenarios?
• RQ3 To what extent does spectral overlap and signal bandwidth (narrowband vs. wideband) affect the accuracy of multi-label interference detection?
• RQ4 How does the number of interfering signals impact detection performance, especially for wideband IEEE 802.11b/g signals?
• RQ5 Does the frequency channel of the utilized signal significantly influence the detection accuracy of interfering signals?

Key findings

• The proposed CNN achieves approximately 100% classification accuracy for same-technology interference (STI) involving narrowband IEEE 802.15.1 and IEEE 802.15.4 signals due to minimal spectral overlap and full signal coverage within the 10 MHz sensing bandwidth.
• For same-technology interference involving wideband IEEE 802.11b/g signals, the classification accuracy drops to a minimum of 78%, primarily due to partial signal coverage and significant inter-signal spectral overlap.
• For cross-technology interference (CTI), the approach achieves at least 90% accuracy when the utilized signal is IEEE 802.11b/g, benefiting from the narrowband nature of interfering signals that reduces spectral confusion.
• When the utilized signal is IEEE 802.15.1 or IEEE 802.15.4, the accuracy slightly decreases to at least 95% due to partial spectral overlap with interfering IEEE 802.11b/g signals.
• The detection performance shows significant variation across frequency channels, indicating strong dependency on the center frequency of the utilized signal, particularly for IEEE 802.11b/g.
• The unexpected increase in true positive rate (TPR) for three IEEE 802.11b/g interferers may stem from combinatorial probability effects or model overfitting, suggesting a need for further validation in real environments.