Differentiable architecture search-based automatic modulation recognition for multi-carrier signals

Considering the lack of a general multi-carrier signal dataset in urban multipath channels, and the challenge of recognizing the modulation types of distorted signals at low signal-to-noise ratio (SNR), a differentiable architecture search-based (DARTS) automatic modulation recognition algorithm for multi-carrier signals was proposed. Firstly, the received signal datasets of commonly used multi-carrier signals, i.e., orthogonal frequency division multiplexing, filter bank multi-carrier, and orthogonal time frequency space, were generated over typical urban multipath channels. The time-frequency images, which were insensitive to modulation parameters, were selected as feature vectors to train the neural network. Secondly, DARTS was employed to automatically search the optimal network architecture. Finally, a joint attention mechanism was introduced into the feature learning process. This mechanism spatially transforming distorted signal features to mitigate the impact of multipath interference, while also calculating and sorting the information weights for each channel of the feature maps to improve the recognition performance of the relevant feature map channels. Simulation results demonstrate that the proposed algorithm improves accuracy in urban multipath channels, especially at low SNR, while simultaneously providing better robustness to modulation parameter variations and small-sample scenarios.