HybNet: A Hybrid Deep Learning -- Matched Filter Approach for IoT Signal Detection

TL;DR

IoT networks employing random-access schemes suffer from interference that degrades LoRa symbol detection. The paper introduces HybNet, a hybrid detector that switches between a conjugate-matched filter path and a CNN-based detector based on the detected interference level, preserving AWGN performance while gaining robustness under interference. It systematically evaluates three data modalities (I/Q, time-frequency spectrogram, and magnitude spectrum) through CNNs, and demonstrates that the spectrum-based DL detector offers strong performance with lower complexity; HybNet further improves robustness by adaptively selecting the detector path, achieving superior BER across interference regimes. The results suggest a practical, low-complexity approach for reliable LoRa symbol detection in dense IoT deployments and random-access networks.

Abstract

Random access schemes are widely used in IoT wireless access networks to accommodate simplicity and power consumption constraints. As a result, the interference arising from overlapping IoT transmissions is a significant issue in such networks. Traditional signal detection methods are based on the well-established matched filter using the complex conjugate of the signal, which is proven as the optimal filter under additive white Gaussian noise. However, with the colored interference arising from the overlapping IoT transmissions, deep learning approaches are being considered as a better alternative. In this paper, we present a hybrid framework, HybNet, that switches between deep learning and match filter pathways based on the detected interference level. This helps the detector work in a broader range of conditions, optimally leveraging the matched filter and deep learning robustness. We compare the performance of several possible data modalities and detection architectures concerning the interference-to-noise ratio, demonstrating that the proposed HybNet surpasses the complex conjugate matched filter performance under interference-limited scenarios.

Figures (12)

• Figure 1: Dataset generation process for training each CNN.
• Figure 2: Illustration of the proposed HybNet architecture switching between a deep-learning branch and a matched filter-branch based on the interference-to-noise ratio.
• Figure 3: Illustration of how the energy detector switching works.
• Figure 4: The plot of the overall accuracy for the energy detector against a varying threshold value.
• Figure 5: The plot of the overall accuracy against a varying INR value to compare the performance of a threshold energy detector and the Selector CNN.