Identifying LoRa parameters using convolutional neural networks

Анотація

This paper proposes an enhancement to the existing methodology of employing Convolutional Neural Networks (CNNs) for identifying parameters of the Long Range (LoRa) telecommunication standard. Building upon the approach of classifying signals through predefined combinations of modulation parameters, our method expands the detectable set by at least twofold through modifications to the signal processing. Employing the YOLO11 object detection model, our results demonstrate a notable improvement of Mean Average Precision (mAP50) up to 93.5%. This surpasses the previously established state-of-the-art result by 2.6%. Notably, our model reveals enhanced detection capabilities on small network sizes. It successfully identifies LoRa signals within spectrograms as small as 256×256, achieving an mAP50 of 87.1% without GPU acceleration, surpassing the 55.0% of the previous research. We validated our technique through an experiment conducted with a First-Person View (FPV) drone. With the drone control signal attenuated to the loss-of-connection level, our model still exhibited remarkable precision. This re-examines the training methodology, demonstrating that a model trained on artificial data can effectively discern modulation parameters from authentically captured signals. The outcomes of our research can be further integrated into a real-time radio intelligence module within discovery and defense systems.