A method for adversarial sample modelling and dynamic decision boundary adaptive optimization in transient signal open-set recognition

Transient signals produced by radar systems operating under complex conditions inherently encapsulate both target dynamic features and indicators of equipment anomalies, making accurate classification essential for effective radar status monitoring. Traditional label-based recognition techniques encounter significant limitations due to the challenges of few-shot learning and the presence of open-set interference from previously unseen categories. To address these issues, this study proposes a meta-learning-based dual-phase adaptive framework that combines adversarial sample generation with dynamic threshold decision-making. In the first phase, adversarial samples are produced using a conditional generative adversarial network with an attention mechanism (CGAN-AM), enabling the construction of a cross-task augmented training set that aligns with actual sample distributions. In the second phase, task relevance is evaluated for few-shot query sets using inter-class covariance matrices. By integrating contrastive loss, the framework dynamically adjusts decision boundaries, allowing for accurate classification of known classes while actively rejecting unknown categories. Experimental validation on a dual-channel MIMO RF front-end transceiver dataset confirms that the integration of the meta-learning framework with CGAN-AM substantially improves few-shot classification performance. Furthermore, the implementation of a class-adaptive threshold strategy based on inter-class correlation analysis dynamically optimizes parameters, significantly reducing false alarm rates in cross-class conditions and effectively addressing open-set threshold adaptation. This framework offers a robust and scalable approach for radar status monitoring in complex operational environments.