Cross-domain fault diagnosis with inaccessible source data is challenging in industrial systems because privacy and communication constraints prevent centralized access to source signals. In multi-source-free domain adaptation (MSFDA), repeated feature extraction and prediction across multiple source-model branches increase computational workload as the number of source models grows. Existing methods face difficulties in preserving fine-grained fault characteristics, estimating sample-dependent source reliability, and controlling pseudo-label noise during self-training. To address these issues, this paper proposes a Multi-Source-Free Reliability-Aware Progressive Adaptation Framework (MSF-RAP) for rolling bearing fault diagnosis. A Fine-grained Embedding Projector (FEP) preserves discriminative patterns through patch-level feature representation, while an Adaptive Matrix-guided Dynamic Fusion (AMDF) estimates source reliability for target samples using prediction uncertainty and structural matching. A Confidence-Guided Progressive Pseudo-labeling (CGPP) strategy improves target supervision through consistency checks, neighborhood consensus, and threshold adjustment. Experiments on eight cross-condition tasks across two bearing datasets show that MSF-RAP achieves the highest average accuracy of 97.01
IoT-based healthcare monitoring systems generate diverse patient data with varying urgency and processing requirements. Patient mobility in hospital and home environments further complicates task scheduling, as the connections between patients and processing nodes may change dynamically. Existing approaches often fail to jointly address healthcare task prioritization, patient mobility, and resource-aware scheduling while minimizing latency and energy consumption. To address these challenges, we propose the Meta-Heuristic Adaptive Hybrid Scheduler for Healthcare (MHASH), a mobility-aware scheduling framework for IoMT Cloud-Fog environments. MHASH classifies healthcare tasks according to their urgency and employs a hierarchical allocation strategy that prioritizes emergency tasks for fog-level execution, while allocating non-emergency tasks to fog or cloud resources according to fog resource availability. The proposed framework integrates HHO and SA with problem-specific mechanisms, including an intelligent capacity-aware boundary checking mechanism, a customized neighbor-generation strategy, and a mobility-aware penalty incorporated into the latency component of the fitness function to preserve data locality during patient mobility. Compared with MobMBAR and HHO, MHASH achieves average reductions of up to 19.5
Accurate traffic prediction is essential for proactive Virtual Network Function orchestration in Network Function Virtualization environments. However, centrally collecting edge traffic histories and repeatedly transmitting full prediction models incur substantial communication overhead. This paper proposes a communication-efficient edge-cloud prediction framework integrating a dense Mixture-of-Experts architecture with LoRA-enhanced PatchTST. Each edge server trains a RevIN_PatchTST model using its local traffic history. The backbone parameters are transferred only once, while subsequent collaborative rounds transmit only low-rank updates. At the cloud, a dense Softmax gate combines all experts using time-aligned aggregate traffic. Individual edge traffic series are not stored separately at the cloud. Experimental results show that the proposed framework achieves competitive forecasting accuracy and downstream NFV performance while reducing the repeated-round uplink payload by 90.66
Modern deep learning-based neural rendering models enable high-fidelity three-dimensional (3D) visual synthesis, yet their intense computational complexity creates substantial bottlenecks during real-time inference. Layer fusion a post-training optimization technique reduces runtime latency by merging structural redundancies without requiring costly retraining or architecture redesigns. This study presents a mathematically grounded, quality-gated framework for post-training layer fusion to optimize neural rendering workloads. The framework decouples a one-time, offline similarity-analysis stage evaluating four complementary metrics (Kullback–Leibler divergence, Wasserstein distance, cosine similarity, and centred kernel alignment) from the subsequent inference-time acceleration it enables via freezing or mean-based merging. A Lipschitz-continuity analysis provides theoretical motivation by establishing how weight and representation-space similarities constrain downstream output deviations, while a quality-gated verification loop accepts candidate fusions only if PSNR, SSIM, and LPIPS remain within defined quality budgets relative to the unfused baseline. Evaluated on an NVIDIA T4 GPU across the eight scenes of the NeRF Synthetic dataset, the proposed framework is applied specifically to a fully connected neural radiance field (NeRF) and a feature grid-based convolutional neural radiance derivative field (CeRF) variant. The method achieves mean inference-time reductions of up to 56
The rapid increase in the number of motor vehicles has led to a rise in the risk of traffic accidents. To improve the accuracy and real-time performance of autonomous driving object detection, we propose a Real-Time Detection Transformer (RT-DETR) network based on a Context-Aware Reparameterized (CAR) module, dubbed CAR-RTDETR. Specific improvements include: designing a lightweight Feature Extraction Network (FEN) based on the CAR module to replace the ResNet backbone in RT-DETR; improving the hybrid encoder of RT-DETR; and replacing the GIoU loss with the Inner-EIoU loss. The proposed model is validated on the VisDrone2019, KITTI, and SODA10M datasets. Compared with existing improved DETR-series models, CAR-RTDETR reduces parameters and computational complexity by at least 41.6
