ABSTRACT The pulse sequence data of radar is an important type of passive detection data, which contains rich pulse modulation features and type features of radar. Due to the limitations of detection methods and noncooperative environments, the obtained pulse sequence data often exhibits incomplete characteristics, such as short segment sequences, high pulse loss rates and high false pulse rates, which pose great processing difficulties for various recognition methods. Due to the low data dimension of pulse sequences, the effective data dimension is further reduced at large mask rates, making it difficult to learn the features of pre training modes using mask reconstruction strategies and making it difficult for the MAE‐like model to converge. To this end, we introduce a class of generalised autoencoder (GAE) structures with Ulam stability preservation capability to improve the representation ability of MAE on low dimensional data, whilst making model training more stable. Furthermore, we combine stability contrastive learning strategies to improve the performance of pre training models. The experimental results show that the GAE can significantly improve the performance of the pretraining model on downstream tasks, such as domain adaptation, fine‐tuning and linear probing.
Ren et al. (Thu,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: