Wireless Interference Recognition With Non-Exemplar Class Incremental Learning

When considering the challenge introduced by dynamic wireless environment, there is an urgent demand for robust wireless interference recognition (WIR) techniques that can continuously adapt to emerging interference types. While deep learning (DL) based WIR methods perform well in static environment, they suffer from catastrophic forgetting when incrementally learning new classes without access to old data. To tackle this challenge of class incremental learning (CIL), this letter proposes a non-exemplar method that employs a Gaussian mixture model (GMM) based prototype augmentation in combination with a composite time-frequency self-supervised learning (CTFSSL), termed GPA-TFSS. Specifically, the GMM based prototype augmentation could represent old class distributions and generate diverse prototypes through probabilistic sampling, while the CTFSSL strategy incorporates rotation, segment-wise shifting, and cyclic overlay augmentations to improve the feature generalization. Simulations on a 28-class interference dataset under Rician fading show that GPA-TFSS consistently outperforms existing non-exemplar methods and even surpasses a typical exemplar-based baseline, achieving a strong balance between the model’s plasticity and stability across different incremental conditions.