In recent years, research on knowledge graphs has exploded due to their capability of effective organization and representation for massive heterogeneous data. However, existing knowledge graphs are often incomplete and contain incorrect triples, which easily incurs a negative impact on the performance of downstream tasks. Besides, few works have been done on the construction of ship knowledge graph. For the former, the mainstream solution is link prediction, also known as knowledge graph completion. To this end, we propose a novel method of knowledge graph embeddings for link prediction, called QuatGAT, combining graph attention mechanism with quaternion embeddings. Specifically, multi-head attention mechanism is employed firstly to obtain entity embeddings by capturing features of both entities and relations of the neighborhood. Then, to represent relations more sufficiently, we use quaternion embeddings to explicitly represent relations as well as entities. Experimental results on benchmark datasets FB15k and FB15k-237 demonstrate the superiority of QuatGAT over existing state-of-the-art methods. Moreover, in terms of ship knowledge graph construction, we also build a multimodal ship knowledge graph named MSKG. Likewise, experimental results on this dataset verify the effectiveness of QuatGAT.

High-resolution range profile (HRRP) target recognition has garnered significant attention in radar automatic target recognition (RATR) research for its rich structural information and low computational costs. With the rapid advancements in deep learning, methods for HRRP target recognition that leverage deep neural networks have emerged as the dominant approaches. Nevertheless, these traditional closed-set recognition methods require labeled data for every class in training, while in reality, seen classes and unseen classes coexist. Therefore, it is necessary to explore methods that can identify both seen and unseen classes simultaneously. To this end, a semantic- and physical-guided generative network (SPGGN) was innovatively proposed for HRRP generalized zero-shot recognition; it combines a constructed knowledge graph with attribute vectors to comprehensively represent semantics and reconstructs strong scattering points to introduce physical constraints. Specifically, to boost the robustness, we reconstructed the strong scattering points from deep features of HRRPs, where class-aware contrastive learning in the middle layer effectively mitigates the influence of target-aspect variations. In the classification stage, discriminative features are produced through attention-based feature fusion to capture multi-faceted information, while the design of balancing loss abates the bias towards seen classes. Experiments on two measured aircraft HRRP datasets validated the superior recognition performance of our method.

In high-resolution range profile (HRRP) automatic target recognition (ATR), non-exemplar class-incremental learning (NECIL) is crucial for recognizing incremental classes while avoiding previously seen classes forgetting under strict storage constraints. Most existing NECIL methods train feature extractors on base classes and employ nearest class mean (NCM) classifiers during incremental phases. This setup faces two limitations: inductive bias toward base classes hinders generalization to incremental classes, and low-variance feature dimensions degrade the performance of NCM classifier. Thus, this paper proposes a compact NECIL framework, HGCI, integrating an HRRP local-scattering patching multi-layer perceptron (HPM) and a Gaussian-truncated Mahalanobis NCM classification module (GTM). HPM models general local scattering mechanisms instead of learning full sequences to mitigate inductive bias. However, patching can exacerbate the low-variance issue. To address this, GTM incorporates Gaussian variance adjustment (GVA) to elevate the variance floor while preserving Mahalanobis distances. In addition, a feature-truncated Mahalanobis NCM classifier (TMM) is introduced in GTM to enhance class separability through non-negative feature truncation. Experiments on measured and simulated HRRP datasets show that HGCI consistently outperforms leading NECIL methods.