Research On Incremental Learning Algorithm For Modulation Type Recognition

Modulation type recognition is an important technology in the fields of communication reconnaissance,cognitive radio,and spectrum supervision,and has been deeply studied.The breakthrough development of deep learning has made modulation type recognition using deep learning a research hotspot.Compared to traditional modulation classification methods,modulation recognition based on deep learning has a strong classification performance.However,because the modulation classification process based on deep learning is often static,trained classification models cannot maintain the recognition ability of old modulation type when learning new modulation type,which can lead to catastrophic forgetting problems.Therefore,the continuous learning ability of classification models has been paid attention.Incremental learning algorithms for modulation type recognition are mainly studied in this paper.To solve the problem of catastrophic forgetting of old modulation type of classification models during learning new modulation type,an incremental learning algorithm of modulation type recognition based on vector space separation(ILVSS)is proposed.The normalization and similarity loss functions are introduced to increase the distance between the classification layer weight vectors corresponding to new and old tasks in vector space to minimize the impact on old tasks when learning new tasks.And the algorithm does not need semantic information for task discrimination and old modulation type data.The incremental learning process of the ILVSS algorithm is simulated on two datasets of computer-simulated radio signals and collected radio signals.The experimental results show that the ILVSS algorithm is superior to the fine-tuning method,feature extraction method,and learning without forgetting(LWF)method,and can alleviate the catastrophic forgetting of classification models.Different task sequences have a certain impact on the algorithm.To further alleviate the catastrophic forgetting problem,an incremental learning algorithm of modulation type recognition based on sample recall(ILSR)is proposed.The algorithm consists of a memory and recalls module,a discriminator module,a classification network,and an incremental process.The memory and recall module implements accurate memory and recall of samples,introducing a code embedding layer to store sample labels and class label information.The discriminator module is used to convert the data distribution into a normal distribution,helping the memory network better remember samples.The classification network realizes the recognition of sample modulation type.The incremental process combines recall samples with the dataset of the new task to form a new dataset for training the classification network.To reduce the storage space of the ILSR algorithm,the ILSR algorithm that only stores the sample data information with partial signal-to-noise ratio is called the incremental learning algorithm of modulation type recognition based on sample recall-memory partial(ILSR-MP).To improve the performance of the ILSR-MP method,the ILSR algorithm for data augmentation of recall samples is called the incremental learning algorithm of modulation type recognition based on sample recall-partial memory and augment(ILSR-MPA).Some appropriate dataset class labels and sample label information are saved,and the recall module can accurately recall the original sample.The incremental learning process of the ILSR-MP method and ILSR-MPA method is simulated and analyzed on two datasets of computer-simulated radio signals and collected radio signals.The experimental results show that the ILSR-MPA method of performance is close to or even better than that of the joint training method.The performance of the ILSR-MP method is slightly worse than that of the ILSR-MPA method,but better than that of fine-tuning method,feature extraction method,and incremental classifier and representation learning(i Ca RL)method.The proposed ILSR method effectively alleviates catastrophic forgetting during incremental learning,and can effectively solve the problem of catastrophic forgetting.