Signal Modulation Recognition via Bias Adjustment-Based Class Incremental Learning

Automatic modulation classification (AMC) is crucial for electronic warfare, spectrum monitoring, and cognitive radios. Traditional methods, relying on maximum likelihood estimation and manual feature extraction, face challenges, such as complexity, inaccuracy, and limited adaptability. In this article, we present bias adjustment-based incremental learning (BA-CIL) for continuous classification of new modulation types. BA-CIL uses an enhanced colored constellation diagram, which is generated by the K-dimensional (K-D) tree algorithm to extract visual features from signals. An incremental learning ResNet classifier expands outputs and adds bias compensation layers. As old knowledge decreases and new knowledge improves, BA-CIL addresses catastrophic forgetting through methods, such as exemplar replay and distillation. Experiments with 24 modulation types show that BA-CIL retains old knowledge while adapting to new categories. At the signal-to-noise ratios (SNRs) of 20, 10, 5, and 0 dB, colored diagrams achieve average 7% higher accuracy than grid diagrams, which are 10%-20% better than normal diagrams. It outperforms the previous class incremental learning approaches, approaching the accuracy of full retraining. The proposed AMC system offers online adaptability, low storage and computational costs, and state-of-the-art accuracy.