Improving Recognition of Sub-GHz LPWANs: A Deep Learning Approach With the UPC-LPWAN-1 Dataset

Abstract

Deep neural networks (DNNs) have emerged as an effective technique for modulation/system recognition but rely heavily on representative datasets. This paper introduces the “UPC-LPWAN-1” dataset, a comprehensive collection of 40 Sub-GHz LPWAN transmission modes acquired using real hardware. Publicly available to the scientific community, this dataset includes raw and pre-processed samples across different Signal-to-Noise Ratios (SNRs) and features multi-carrier modulations, which are typically underrepresented in existing datasets. The variability in studies using different neural network architectures and small, unrepresentative datasets complicates research comparisons. To address this, this paper compares seven proposed architectures using UPC-LPWAN-1, providing a standardized evaluation. To further enhance accuracy, we propose four new convolutional neural network (CNN) architectures adapted to four forms of signal representation. Our results demonstrate that while some existing models perform well under high SNR conditions, their performance degrades significantly in low SNR environments. The proposed spectrogram-based CNN consistently outperforms other models, achieving a classification accuracy of 99.71% at SNR = 0 dB, above 90% at SNR =−10 dB, and above 70% at SNR =−15 dB, while still being able to differentiate between systems.