Consumer electronics products are characterized by high unit prices, rapid iteration, and highly elastic demand, meaning embedded credit pricing directly impacts consumers' actual payment costs. AI-driven buy-now-pay-later and installment pricing may generate systematic interest-rate differences across consumer groups purchasing comparable devices, potentially affecting market accessibility and fairness. This study develops a multi-agent computational experiment to quantify fairness loss in AI-driven consumer credit pricing. The model integrates deep reinforcement learning pricing agents, heterogeneous consumer agents calibrated with regional Chinese consumer credit data, and a regulatory agent that implements missing-feature compensation and fairness-aware pricing constraints. Simulation results suggest that fairness loss may arise through two interacting channels. Feature sparsity among underrepresented consumers is associated with greater reliance on proxy-related variables and higher group-level interest-rate premiums. Shared market-state information among pricing agents is associated with pricing-strategy convergence, increased interest-rate dispersion, and reduced consumer surplus. Fairness loss is measured by the interest-rate Gini coefficient, relative group premiums, and consumer surplus changes. The regulated scenario reduces measured fairness loss but does not fully restore the competitive baseline. The findings of this study have direct implications for the consumer electronics market.
As pivotal consumer electronic devices, unmanned aerial vehicles (UAVs) have become indispensable for numerous daily tasks. Collision avoidance control, a profoundly significant UAV technology, is typically achieved by adding repulsive forces to the position loop controller in existing methods, which may restrict convergence and degrade performance. To address this problem, a dynamic trajectory generator (DTG) is constructed by introducing the artificial potential function into the distributed observer framework. The DTG can plan collision-free trajectories to guide UAVs to moving target fencing positions at an appointed-time. To improve control robustness and reduce computational load, an error-based adaptive signal is designed to compensate for unknown disturbances and the derivatives of the virtual control signals. Moreover, a quaternion-based controller is formulated to drive the UAVs to the generated trajectory with prescribed performance, which guarantees both control performance and safe operation for UAVs. Lastly, a simulation example is provided to demonstrate the availability and effectiveness of the proposed control method.
Breast cancer is the most common malignancy in women worldwide, and early-stage detection through accurate histopathological analysis is critical for effective treatment planning. AI-driven consumer health platforms in oncology are shaping new opportunities to bring automated diagnostic support in clinical workflows, directly benefiting women’s health outcomes. This paper presents a computational framework for invasive ductal carcinoma (IDC) detection from hematoxylin-and-eosin (H&E)-stained histopathology patches, designed to support intelligent women’s health monitoring systems under realistic clinical conditions. The proposed framework integrates ImageNet-pretrained convolutional neural network (CNN) backbones with a novel FastKAN-inspired classification head. This head performs a structured nonlinear feature expansion grounded in Kolmogorov–Arnold representations, enabling richer feature interactions within a compact and resource-efficient architecture. A primary methodological contribution is the enforcement of strict patient-level data partitioning to eliminate information leakage between training and evaluation splits, a critical but widely neglected requirement in patch-based histopathology pipelines. A proportion-based patient-level aggregation strategy is also proposed, translating patch-level predictions into clinically meaningful slide-level decisions. This reflects the clinical principle that malignancy assessment depends on the overall spatial extent of the tumor, rather than the appearance of any isolated image regions. Extensive experiments across four pretrained CNN backbones (VGG16, ResNet50, DenseNet121, and MobileNetV2) yield patch-level ROC–AUC values up to 0.9248 and patient-level ROC–AUC up to 0.8449 under a fully leakage-free protocol. Statistical validation via DeLong’s test and Holm-corrected Wilcoxon signed-rank tests confirms performance differences between backbones at the patch level. At the patient level, comparisons reveal the statistical challenges that arise when evaluation cohorts are small. Overall, this work contributes a principled, reproducible, and low-complexity AI pipeline for histopathology-based diagnostic support in women’s health, aligned with the goals of intelligent FemTech consumer systems.
Endoscopic imaging is essential for minimally invasive surgery and diagnosis and is widely used in modern medical-grade consumer devices such as surgical robots and advanced diagnostic systems. It provides direct visualization of internal organs and tissues, enabling early lesion detection and better patient outcomes. However, specular highlights from tissue surface reflections often degrade image quality and diagnostic accuracy, challenging both surgeons and automated analysis. Deep learning has improved highlight removal, but the complexity and variability of highlights in dynamic surgical scenes remain difficult. We propose SHRMamba, a Mamba-based architecture that unifies highlight localization and image inpainting in an end-to-end framework. Unlike conventional Mamba models with fixed scanning, SHRMamba introduces Dynamic Localization-Aware Scanning (DyLAS), which adaptively reorders scanning based on highlight distribution and defers processing of highlighted regions to preserve important features. SHRMamba also incorporates a Channel-Aware Interactive Attention (CAIA) module, which enables cross-channel information exchange via channel shuffling and uses linear attention to optimize the encoder–decoder feature propagation, prioritizing anatomically critical regions. On the public Hyper Kvasir dataset, SHRMamba outperforms state-of-the-art methods, achieving a PSNR of 34.89 and an SSIM of 94.85%. It also attains an SSIM of 81.41% on the external SCARED-2019 dataset, demonstrating strong generalization. These results show SHRMamba’s potential to improve endoscopic imaging performance in consumer medical devices, supporting more reliable, accurate, and user-friendly imaging solutions.
Balancing cost, emissions, energy efficiency, and autonomy remains a challenge in tropical microgrid planning. To address this, this paper sizes and evaluates solid oxide fuel cell, photovoltaic, and battery energy storage system (SOFC-PV-BESS) microgrids. Firstly, a mixed-integer linear programming model is formulated that co-optimizes capacities and dispatch while embedding load-dependent SOFC efficiency, ramp hysteresis, and two-stage waste-heat cascading through a gas turbine (GT) and combined cooling, heating, and power (CCHP). Then, because the least-cost design is not necessarily best on other objectives, the TERPE index is proposed to rank designs by total annual cost, emissions, renewable energy fraction, primary energy efficiency, and energy autonomy hours. Next, a two-stage sensitivity analysis is developed to identify the parameters that most affect the ranking while reducing computational cost. In the first stage, a sensitivity index identifies the most sensitive parameters; in the second, a polynomial chaos expansion (PCE) surrogate accelerates the search over the selected parameters. Finally, four case studies cover a 100-household community and an office building in Malaysia, both grid-connected and offgrid. The cascaded SOFC design is the least-cost option in most scenarios, and the load-dependent model prevents mis-sizing. TERPE ranks SOFC-GT-CCHP first in the commercial off-grid scenario, and diesel-PV-BESS first in the residential off-grid one, where renewable fraction and autonomy offset its higher cost. The sensitivity analysis shows that grid access determines which parameters dominate: battery aging off-grid, grid price or PV cost when grid-connected. The two-stage method combines high accuracy with low computational cost.
