Behavioral modeling is becoming increasingly important in antenna development, particularly in accelerating simulation-based design processes such as optimization. However, the construction of accurate data-driven models remains a challenging task, typically limited to electrical responses such as reflection coefficients or integral performance metrics (e.g., gain). In contrast, the behavioral modeling of radiation patterns—although crucial from a design perspective—has been relatively scarce in the existing literature. This research introduces a new surrogate modeling framework for broadband antennas that leverages customized deep learning techniques. The proposed approach enables accurate and comprehensive modeling of both electrical responses (e.g., reflection coefficient, gain) and far-field radiation patterns across a broad range of frequencies, elevation/azimuth angles, and geometric parameters. As a result, the developed surrogates are highly versatile and applicable for addressing a wide spectrum of design and optimization tasks. For accurate representation of radiation patterns, a hybrid deep learning architecture is developed that integrates multi-branch recurrent processing with cross-attention–based feature fusion. Six independent RNN branches (each combining FC, LSTM, and GRU layers) extract complementary representations of the geometric input space, which are subsequently fused through a cascade of attention modules. The fused embedding is decoded using a transposed-convolutional refinement network to reconstruct the full radiation pattern. Extensive validation studies, performed using two ultra-wideband antenna designs, confirm the excellent predictive accuracy and robustness of the proposed surrogates. The results also demonstrate their advantages over multiple benchmark methods, including conventional regression models and various deep learning schemes. Furthermore, the practical utility of the framework is demonstrated through real-world design tasks, including impedance matching enhancement, size reduction, omnidirectional radiation, and gain stability. Overall, the presented methodology represents a significant advancement toward comprehensive multi-response antenna modeling and substantially extends the capabilities of behavioral metamodeling in contemporary antenna design practice.
Accurate remaining useful life (RUL) prediction is essential for intelligent maintenance in complex industrial systems. However, multi-source sensor sequences are often affected by noise interference, weak early degradation patterns, long-range dependency degradation, and insufficient cross-channel interaction. To address these challenges, this paper proposes BEP-Net, an incremental learning-enhanced multi-source framework for industrial RUL prediction. BEP-Net first employs a bidirectional Mamba module to extract noise-resistant local degradation features from forward and backward temporal contexts. A cross-sensor attention-enhanced PatchTST module is then introduced to capture long-term degradation trends and coupling relationships among heterogeneous sensor channels. An Improved Black Widow Optimization algorithm is further used to determine adaptive fusion weights between local and global feature representations. To accommodate newly arriving monitoring data without retraining the entire model, BEP-Net also incorporates a lightweight incremental update strategy that selectively adjusts task-relevant parameters while retaining previously learned degradation knowledge. This strategy freezes the main feature extraction backbone and introduces historical knowledge constraints to mitigate catastrophic forgetting, which lowers computational overhead for model iteration in practical industrial deployment. Experiments on the NASA CMAPSS and XJTU-SY datasets show that BEP-Net achieves the best prediction performance among 11 representative baselines. Compared with the strongest baseline, it reduces the average RMSE by 11.5% on CMAPSS and by 15.0% on XJTU-SY. These results demonstrate that BEP-Net provides an accurate, efficient, and extensible solution for multi-source industrial RUL prediction under evolving operating environments.
Medical image segmentation with limited annotation depends strongly on the transferability of representations learned from unlabeled data. Masked image modeling is promising for this setting, yet conventional random masking ignores the highly non-uniform structural informativeness of medical images and may devote substantial pretraining capacity to homogeneous or weakly informative regions. We propose a structure-aware pixelwise masking framework with encoder-centric pretraining for annotation-efficient medical image segmentation. A reinforcement-learning-based policy generates image-dependent pixelwise masks under a fixed masking budget, while a composite reward combines reconstruction behavior, normalized gradient-based structural guidance, and mask-ratio regularization. The sampled mask is applied directly in image space before patch embedding, preserving sub-patch spatial variation while retaining a Transformer encoder. To strengthen downstream transfer, the encoder incorporates local structural enhancement and multi-level feature fusion, whereas a lightweight reconstruction decoder limits decoder-side compensation and places greater representational responsibility on the encoder. Experiments on BraTS and five Medical Segmentation Decathlon tasks under strict case-level splits and a unified 2D protocol show that the proposed method achieves the highest DSC and IoU among the evaluated baselines, reaching 0.8777/0.7860 on BraTS and average values of 0.7610/0.6932 on MSD. Controlled patch-versus-pixel comparisons, pairwise interaction analyses, structural-cue sensitivity experiments, and joint WT/TC/ET segmentation further support the proposed design. Case-level paired testing on BraTS also confirms statistically significant DSC improvements over all baselines (). These results demonstrate improved representation transfer for annotation-efficient medical image segmentation.
Anomaly detection in dynamic graphs has received significant attention in recent years, as it uncovers potential anomalies in nodes by aggregating temporal neighbor information. Existing graph anomaly detection methods typically rely on valid node attributes for message aggregation. However, node attributes are often unavailable, unreliable, or restricted by privacy constraints in real-world dynamic graphs. In this paper, we focus on unattributed dynamic graph anomaly detection, where raw node attributes are unavailable and node representations must be inferred from temporal topology. Moreover, anomaly nodes often attempt to reduce their suspiciousness by connecting to many normal nodes. The anomaly nodes will receive excessive representation information from normal nodes during message aggregation, which causes the anomaly feature coverage problem. In dynamic unattributed graphs, the lack of node attributes and dynamic message aggregation can more severe dynamic anomaly feature coverage problems. In this paper, we propose Tracing Structural EvolutIon: Representation Reconstruction for Unattributed Dynamic Graph Anomaly Detection (SIGAD), which can address the lack of node attributes and dynamic anomaly feature coverage problems on the unattributed dynamic graphs. Specifically, SIGAD captures dynamic structural information and assigns similar representations to nodes that exhibit similar structural roles. These representations encode the temporal evolution of node’s structural roles, which enriches their expressive capabilities. The experimental results on two real-labeled dynamic graph datasets and two synthetic-anomaly dynamic graph datasets demonstrate that SIGAD consistently outperforms competitive baselines under the same node-unattributed input protocol.
Session-based recommender systems aim to capture the latent preferences of the anonymous user and predict the next item based on the historical interactions within the current session. Most of existing works rely on attention-based modules to capture item dependencies derived from graph learning or sequence learning. However, these works suffer from the effectiveness and efficiency dilemma regarding the design of the overall framework and the modules, which hinders more precise recommendations on the larger scale datasets. To address these challenges, we propose Conan, an asymmetric contrastive framework that uses effective session-level augmentation and generates positive and negative samples with just a single forward propagation. Conan enhances the feature learning capacity of the graph learning module by leveraging the meaningful contrastive signals derived from sequence learning. For the specific module design, we introduce the Mamba block and the Fourier Kolmogorov-Arnold network to further enhance both effectiveness and efficiency. Extensive experiments conducted on four real-world datasets (Diginetica, Retailrocket, Dressipi, and KuaiRand) demonstrate the superiority of our approach over state-of-the-art methods, achieving average gains of 2.32% to 2.70% in Recall and 2.74% to 3.14% on two ranking metrics (MRR and NDCG). The ablation studies suggest the effectiveness of the overall framework and main components. The efficiency experiments demonstrate that our Conan achieves a well-balanced performance in terms of both time and space complexity.
