Cushing's syndrome (CS) is a rare endocrine disorder characterized by chronic hypercortisolism, and facial image-based intelligent diagnosis has emerged as a promising non-invasive approach. However, existing diagnostic models suffer from two core bottlenecks: inefficient fusion of deep semantic features and clinical prior features, and insufficient multi-view facial feature disentanglement without CS-specific pathophysiological constraints. To address these limitations, we propose a novel Multi-View Facial Feature Disentanglement Network (MVFFD-Net) for high-precision automatic CS diagnosis. The network takes five standard facial views (frontal, bilateral 45 circle oblique, and bilateral 90 circle lateral views) as input, with three key innovations: a bidirectional cross-attention module for synergistic fusion of deep and clinical prior features, a multi-path disentanglement architecture with multi-objective loss for separating view-invariant and view-specific features, and a graph attention fusion framework for adaptive multi-view feature integration. Unlike state-of-the-art models that mainly rely on single-view representations or generic multi-view fusion, MVFFD-Net explicitly integrates clinically guided multimodal fusion, CS-oriented feature disentanglement, and pathology-aware graph-based multi-view aggregation. Experimental results demonstrate that MVFFD-Net achieves an F1-score of 98.78% for CS diagnosis, significantly outperforming representative state-of-the-art methods in the current experimental setting. In this study, the task is defined as subject-level AI-assisted diagnosis to distinguish individuals with CS from matched controls using five standard facial views. Accordingly, the proposed framework is intended to provide non-invasive auxiliary diagnostic support rather than replace standard endocrinological evaluation and biochemical confirmation. More broadly, this clinically informed multi-view learning framework shows promising potential as a non-invasive auxiliary screening tool, though external multi-center validation remains necessary prior to large-scale clinical application.
This paper presents the development of a digital twin of a heavy-duty electric truck powered by a hybrid energy system based on a hydrogen fuel cell and a battery pack. The objective of the model is to analyze different energy management strategies and to determine how the power demand of a real route can be shared between both energy sources, while keeping the fuel cell within safe operating limits to preserve its service life. The digital twin simulates vehicle dynamics, traction, and regenerative braking, and the main operational constraints of the fuel cell, including minimum and maximum power limits, power variation constraints, and startup conditions. The input data are obtained from real driving routes, using telemetry data recorded from conventional diesel trucks. These data provide realistic speed, acceleration, and slope profiles, which are used to reconstruct the power and energy demand of the vehicle and to simulate the behavior of an equivalent fuel cell-battery electric truck. Several rule-based energy management strategies are implemented and compared using indicators such as hydrogen consumption, minimum battery state of charge, unused generated energy, and fuel cell power variability. The results show that strategies limiting rapid fuel cell power changes provide a better balance between hydrogen efficiency, battery protection, and smoother fuel cell operation. The proposed digital twin offers a practical framework to evaluate hybrid energy management strategies under realistic operating conditions and to support future intelligent energy management developments.
Due to transportation disruptions after the earthquake, on-site surveys could not quickly assess house damage. Remote sensing can quickly acquire disaster-related information over large-scale affected areas and support rapid localization and morphological identification of damaged buildings from top-view observations. Nevertheless, on-site inspection by experienced engineers is still the most accurate method for judging real building structural damage, since UAV orthophotos only record surface features and cannot capture internal building damage. Field investigation data in our study served as ground-truth references for UAV image interpretation and labelling. Current remote sensing-based post-earthquake building assessment mostly adopts a binary classification system of collapsed and non-collapsed, which makes it difficult to distinguish between lightly damaged and intact buildings, and the imbalance between positive and negative sample sizes leads to low identification accuracy. This article created the orthophoto images from unmanned aerial vehicles in Shuanghe Town, Changning County, Sichuan Province and established a classification standard for building seismic damage based on visual interpretation. The earthquake-damaged buildings were ultimately classified into three damage levels: completely collapsed, partially damaged, and basically intact. Semantic segmentation can identify different objects in an image. This article compares the performance of different models, sample sizes, and loss functions. This study compares 50% overlapping cropping with non-overlapping cropping, and compares a Focal + Dice hybrid loss with cross-entropy, on UAV orthophoto patches of Shuanghe Town. Reported scores are mean metrics on the test block of this single mosaic. The results show that the MobileNetV3 model with Focal Loss and Dice Loss reported the highest mean mIoU, mPA, and F1 scores of 92.36%, 97.32%, and 96.77%, respectively, on the evaluated test block with 512 & times; 512 pixels and 50% overlapping sampling.
Random geometric imperfections strongly influence the buckling resistance of thin cylindrical shells. In random field imperfection modeling, Euclidean and geodesic distances may generate different spatial correlation structures on closed cylindrical shells. This study examines whether such distance model differences are transmitted to nonlinear stochastic buckling response. A paired stochastic finite element framework is developed, in which Euclidean-and geodesic-distance imperfection fields are generated from common nodal white noise inputs and rescaled to the same prescribed root-mean-square (RMS) amplitude. Eight normalized correlation lengths are considered, with 200 paired samples for each case, and the corresponding critical buckling loads are evaluated. The results show that the field-level discrepancy increases with correlation length, but its transmission to buckling response is not proportional. For the present benchmark, the Euclidean distance effect is very small at short correlation lengths but becomes increasingly evident as the correlation length approaches the shell radius scale, indicating a correlation scale-dependent trend rather than a universal acceptability threshold.
Hyperspectral image (HSI) classification requires models that can effectively capture long-range contextual dependencies while preserving fine-grained spectral-spatial variations under strict computational constraints. Recent transformer-based approaches, particularly Swin Transformers, have shown strong performance by leveraging localized self-attention; however, their reliance on generic attention mechanisms often overlooks frequency-sensitive information that is critical for discriminating spectrally similar materials. Moreover, existing frequency-aware designs typically introduce heavy parameterization or explicit spectral transforms, limiting their efficiency and practical deployment. In this paper, we propose FEAM-Swin, a lightweight frequency-aware Swin Transformer designed for efficient HSI classification. The proposed model introduces a novel Frequency-Enhanced Attention Modulator (FEAM), which captures local spectral-spatial frequency variations via gradient-based energy estimation and performs adaptive channel modulation with minimal computational overhead. Unlike conventional frequency modelling approaches, FEAM avoids explicit Fourier or wavelet transforms, enabling seamless integration into hierarchical transformer architectures. FEAM is embedded within Swin Transformer blocks to enhance discriminative feature learning while preserving the efficiency of window-based self-attention. Extensive experiments performed on four benchmark hyperspectral datasets widely used in the research community, including Indian Pines, University of Pavia, Salinas and KSC, demonstrate that FEAM-Swin consistently outperforms state-of-the-art CNN-and transformer-based methods. Specifically, FEAM-Swin achieves overall accuracies of 98.17%, 99.87%, 99.97%, and 98.41%, respectively, yielding consistent improvements over competing approaches while requiring fewer parameters and lower computational complexity. These results validate the effectiveness of lightweight frequency-aware modulation and establish FEAM-Swin as a practical and robust solution for real-world HSI classification.
