Research On Communication Signal Recognition With Deep Learning And Its Generalization

Communication signal recognition technology has important applications in fields such as spectrum sensing,spectrum monitoring and management,wireless network security,and electronic reconnaissance and countermeasures.Traditional communication signal recognition has hinged on feature engineering,which requires seasoned domain experts to design and select signal features meticulously.In recent years,communication signal recognition based on deep learning has employed deep neural networks with their powerful nonlinear representation to learn more comprehensive signal features directly from raw signal data,thus greatly reducing such dependence on domain knowledge and prior information.However,it is precisely because of its data-driven nature that communication signal recognition based on deep learning encounters many generalization issues when faced with some non-ideal data conditions.This dissertation focuses on two of these challenges.The first is how to avoid or mitigate performance degradation caused by overfitting when training under a limited sample of labeled signals.The second is how to detect signal classes that have not been seen during training.This dissertation delves into these two issues.The dependence on labeled data is alleviated with data augmentation and semi-supervised learning.The detection of unknown classes is achieved using open-set recognition.The research and main contributions are encapsulated as follows.1.In response to lessen such dependency on annotated data,this dissertation takes rotation and flipping as basic transformations,combined with stochastic permutation,to design a composite data augmentation scheme for communication signals.The implicit variance regularization effect of data augmentation is revealed through theoretical deduction.The intensity attribute of data augmentation is defined based on a support set of sample distribution,confirming that augmentation effects are positively related to augmentation intensity.Experimental results indicate that this composite data augmentation can significantly improve communication signal recognition performance under small sample conditions due to its considerable augmentation intensity.2.Given that conventional generative adversarial networks(GANs)struggle to model relatively complex signal data,this dissertation presents a framework named Radio GAN for signal data synthesis.By introducing prior information,such as signal models,Radio GAN unrolls its generator into a combination of multiple neural network modules designed using domain knowledge,significantly reducing modeling complexity.Experimental results demonstrate that,if sufficient signal priors are allowed,Radio GAN enables accurate modeling of signal data and can be used as data augmentation to improve communication signal recognition performance under small sample conditions to a level close to that of ideal sample conditions.3.Given that unlabeled signal data is relatively easy to obtain in real communication scenarios,this dissertation presents a semi-supervised communication signal recognition(SSCSR)algorithm to further improve generalization by consistency regularization that encourages a model to produce invariant predictions against random perturbations.The SSCSR algorithm adopts cross-entropy to measure predictive consistency but takes into account loss scaling and a symmetrically structured loss form,which can mitigate the adverse effects caused by the high-confidence erroneous predictions of unlabeled data.Experimental results demonstrate that SSCSR can achieve good recognition performance close to ideal sample conditions by providing a large number of unlabeled samples with only a small number of labeled samples(50 samples per class).4.In response to address unknown signal class detection,this dissertation presents a refined prototype learning(RPL)algorithm.The RPL algorithm builds on prototype regularization and then further reduces intra-class individual feature differences by data augmentation and enlarges inter-class distances between similar classes through label smoothing,which results in significantly improved open-set recognition performance.In addition,a statistical strategy is proposed for the adaptive configuration of detection thresholds for unknown classes.Experimental results demonstrate that RPL can effectively refuse to recognize potential unknown signals while maintaining good classification accuracy for known signal classes.