In modern communication systems, the electromagnetic environment has become increasingly complex and highly dynamic. This dynamism is reflected in two key aspects: firstly, the emergence of new communication protocols and signal devices has led to a continuous growth of novel modulation types, demanding that existing systems be capable of identifying new modulation schemes in dynamic settings; secondly, signals are subject to time-varying channels, multi-path effects, and noise interference during transmission, causing feature drift even within the same modulation type and significantly degrading the recognition accuracy and robustness of models. Although traditional deep learning methods perform well in static scenarios, they require retraining with both historical and new data when encountering new signal types, resulting in substantial computational overhead and time delays, which fail to meet the urgent need for real-time adaptation in modern cognitive radio systems. Therefore, developing an incremental recognition mechanism that can continuously and efficiently learn new signal types while resisting feature drift has become a critical technical challenge in the field of signal processing and a core driver for the development of next-generation intelligent communication systems. This study addresses this important issue by proposing a lightweight and adaptive incremental learning paradigm, providing theoretical and methodological support for intelligent signal recognition in open environments.

To achieve comprehensive adaptability in dynamic environments, a feature and semantics-driven incremental modulation recognition method was proposed. The core of this approach involved the construction and dynamic maintenance of an evolutionary semantic space with a robust geometric structure. It specifically comprised the following three key components:

Robust Multi-Dimensional Feature Extraction Network: To counteract feature drift, multi-dimensional physical features such as instantaneous amplitude, phase, and frequency were first constructed. A parallel temporal convolutional network was then designed to process these features in a decoupled manner, preventing premature mixing of information from different physical dimensions and enabling the network to learn more intrinsic and interference-resistant temporal patterns.

Semantic Graph-Based Incremental Classification Strategy: To address the class imbalance caused by sample replay in incremental learning, the conventional Softmax classifier was abandoned. Instead, a "modulation semantic graph" composed of dynamically evolving class center vectors was constructed, and a nearest neighbor strategy was adopted for classification. This method does not rely on gradient-optimized classifiers but rather on geometric distances in the feature space, effectively mitigating classifier bias induced by data imbalance.

Multi-Objective Joint Loss Function: The optimization of the semantic space was driven by a joint loss function combining three objectives. Hierarchical knowledge distillation loss. A frozen previous model guided the current model, ensuring that the deep feature outputs of the current model for old-class samples remained consistent with those of the teacher model. This deep constraint effectively suppressed drastic weight updates triggered by new tasks, thereby alleviating catastrophic forgetting. Semantic space distance loss. Metric learning was used to directly optimize class centers, pulling samples of the same class toward their correct centers while pushing apart samples from different classes, enhancing feature discriminability. Laplacian spectrum optimization loss. A manifold regularization term based on spectral graph theory was innovatively introduced. By minimizing the sum of the first k eigenvalues of the Laplacian matrix, this loss globally enhanced the separation among different class clusters in the semantic space, actively promoting a compact intra-class and separable inter-class geometric structure.

To validate the effectiveness of the proposed method, an incremental learning scenario was designed, consisting of seven consecutive tasks on a public dataset containing 17 modulation types. Experimental results demonstrate that our method achieves significant improvements in both performance and efficiency:

Experimental Result 1: After completing all seven incremental tasks, the proposed method achieved an average recognition accuracy of 84.46%, significantly outperforming mainstream incremental learning methods LwF (71.28%) and iCaRL (80.56%). These results indicate that our method not only effectively learns new categories but also substantially mitigates catastrophic forgetting, with a performance improvement exceeding 10%. Notably, when learning amplitude-dominated signals such as 4ASK, both LwF and iCaRL suffered from severe performance degradation due to feature space collapse, whereas our method maintained stable recognition, demonstrating superior robustness.

Experimental Result 2: Complexity analysis shows that our model contains only 116,000 trainable parameters and requires 996 MFLOPs for inference, which is considerably lower than LwF (192,000 parameters, ≈1.6 GFLOPs) and iCaRL (207,000 parameters, >1.7 GFLOPs). This confirms the dual advantages of our approach in model compactness and computational efficiency, making it more suitable for resource-constrained edge devices and applications demanding high real-time performance.

This study addresses the challenge of incrementally recognizing new modulation schemes in dynamic environments by proposing a semantic space-based incremental modulation recognition model. It effectively tackles the issue of insufficient recognition accuracy caused by feature drift and imbalanced sample distribution across different tasks. Experimental results show that the proposed method achieves an average recognition accuracy of 84.46% across multiple incremental tasks, outperforming mainstream incremental learning approaches by over 10%, which demonstrates its significant effectiveness. However, certain limitations remain. In high-dimensional and complex semantic spaces, the current classifier is prone to confusion in regions with dense category distributions, limiting its accuracy in identifying fine-grained modulation types. Future research will focus on constructing nonlinear decision boundaries and modeling more refined probabilistic distributions of category features to effectively enhance fine-grained discriminative capability.