Frequency hopping signal detection based on optimized generalized S transform and ResNet

The performance of traditional frequency hopping signal detection methods based on time frequency analysis is limited by the tradeoff of time-frequency resolution and spectrum leakage. Machine learning-based frequency hopping signal detection techniques have a high level of complexity. Therefore, this paper proposes a residual network and the optimized generalized S transform to detect frequency hopping signals. First, based on the time-frequency aggregation measure, the generalized S transform parameters $ \lambda $ and $ p $ are optimized using a multi-population genetic algorithm. Second, the optimized generalized S transform is used to determine a signal's time-frequency spectrum, which is then normalized to make this robust to noise power uncertainty. Finally, a residual network structure is designed which receives the time-frequency spectrum. To detect frequency hopping signals, the network automatically learns the time-frequency properties of signals and noise. Simulated findings indicate that the multi-population genetic algorithm not only increases optimization efficiency when compared to a regular genetic algorithm, but also has faster convergence and more stable optimization results. Compared with a hybrid convolutional network/recurrent neural network algorithm, the proposed technique is better at detection and has less computational and storage complexity.