Enhancing Automatic Modulation Recognition for IoT Applications Using Transformers

TL;DR

The paper adapts Transformer architectures for automatic modulation recognition in IoT contexts, proposing four tokenization-based variants (TransDirect, TransDirect-Overlapping, TransIQ, TransIQ-Complex) to balance recognition accuracy and model size. Through extensive experiments on RadioML2016.10b and CSPB.ML.2018+, TransIQ variants achieve state-of-the-art-like accuracy with far fewer parameters than conventional baselines, and they demonstrate practical latency and throughput suitable for edge devices. The study highlights the importance of tokenization strategy and convolutional feature augmentation for effective AMR in resource-constrained environments, with results showing accuracy around 65% on both datasets. These findings support deploying Transformer-based AMR on IoT edge nodes to improve spectrum sensing, interference localization, and cognitive radio capabilities while preserving power and compute budgets.

Abstract

Automatic modulation recognition (AMR) is vital for accurately identifying modulation types within incoming signals, a critical task for optimizing operations within edge devices in IoT ecosystems. This paper presents an innovative approach that leverages Transformer networks, initially designed for natural language processing, to address the challenges of efficient AMR. Our transformer network architecture is designed with the mindset of real-time edge computing on IoT devices. Four tokenization techniques are proposed and explored for creating proper embeddings of RF signals, specifically focusing on overcoming the limitations related to the model size often encountered in IoT scenarios. Extensive experiments reveal that our proposed method outperformed advanced deep learning techniques, achieving the highest recognition accuracy. Notably, our model achieves an accuracy of 65.75 on the RML2016 and 65.80 on the CSPB.ML.2018+ dataset.

Figures (6)

• Figure 1: The overall architecture of our proposed transformer-based model for the AMR task includes three main components: a Tokenization module that converts the signal into tokens, a Transformer-encoder module, which captures information and extracts relevant features through self-attention mechanism, and a Classifier module for the final classification step.
• Figure 2: TransDirect architecture. In the Tokenization module of this architecture, IQ samples segments into shorter sequences referred to as tokens, each having a size of $l$.
• Figure 3: Tokenization module of TransDirect-Overlapping architecture, which divides IQ samples into tokens, each with a length $l$, where each token overlaps the preceding one by $l/2$.
• Figure 4: TransIQ architecture. In the Tokenization module of this architecture, the input signal segments into tokens. Each token then undergoes one-dimensional convolutional before being processed by the Transformer-encoder module.
• Figure 5: Modulation recognition accuracy comparison between two variant of TransIQ and other baseline models on CSPB.ML.2018+ dataset with a change in SNR, where SNR ranged from -19 dB to +20 dB.