Search Results (464)

Navigation lock manifolds are key components of filling-and-emptying systems, and port-flow distribution affects chamber flow stability and filling efficiency. Under unsteady filling conditions, port-flow distribution is governed by discharge variation and manifold geometry, making rapid prediction and engineering-constrained screening challenging. This study develops a surrogate-assisted prediction and Pareto-screening framework for a large-scale navigation lock manifold. Three-dimensional computational fluid dynamics (CFD) simulations were used to examine unsteady port-flow evolution. The peak-flow condition was selected as a representative control condition, and the flow non-uniformity coefficient α and system resistance coefficient ξ were used as performance indicators. Based on 243 parametric CFD samples and 144 independent external test samples, artificial neural network (ANN), Gaussian process regression (GPR), and support vector regression (SVR) models were evaluated. ANN performed best, with independent-test R² values of 0.9999 and 0.9928 for α and ξ. Feature-attribution analysis identified port width, culvert height, and port number as dominant variables. Pareto screening within a predefined engineering design space identified representative candidates with CFD verification errors below 1.1%. The TOPSIS-based candidate reduced ξ by 32.2% while maintaining α nearly unchanged. Full article

In safety-critical industrial environments such as nuclear power plants (NPPs), early detection of pipeline leakage is essential for maintaining operational safety. However, leakage events are rare, abnormal samples are difficult to collect, and obtaining sufficient condition-specific normal data is also challenging. To address these limitations, this paper proposes SiameseGAD, a generative few-shot anomaly detection framework for pipeline leakage detection in the secondary systems of NPPs. The proposed method formulates leakage detection as a few-shot normality-modeling problem rather than as a problem of directly learning anomaly patterns. The Siamese network learns similarity relationships among normal samples and constructs a normal feature manifold, while anomaly scores are computed based on the distance from the estimated normal distribution. To improve normal distribution estimation under limited data, a denoising diffusion probabilistic model (DDPM) is used to generate in-distribution normal variants to augment the support-set. The main contribution of SiameseGAD lies in combining metric-learning-based few-shot normality modeling with normal-to-normal generative augmentation, enabling anomaly detection using only a few normal samples without relying on real anomaly data or synthetic anomaly generation. In the three evaluated target classes, SiameseGAD achieved an average AUROC of 93.59% and an average accuracy of 95.26%. These results indicate the potential of SiameseGAD for few-shot anomaly detection using only normal support samples, without requiring real or synthetically generated anomaly samples during inference. Full article

Finger vein recognition (FVR) has significant potential in biometrics due to its high accuracy and intrinsic liveness detection capabilities. However, the increasingly stringent privacy regulations have presented severe data security challenges for traditional centralized training. While federated learning (FL) mitigates these privacy concerns through a decentralized training paradigm, conventional FL algorithms that seek a single global model experience significant performance degradation on non-independent and identically distributed (Non-IID) data in real-world cross-institutional deployments. This degradation stems primarily from a dual-heterogeneity issue that involves domain shift caused by hardware discrepancies across acquisition devices, and label skew resulting from nonoverlapping user identities. To address this dual-heterogeneity challenge, we propose a personalized federated learning framework driven by hierarchical parameter decoupling and subspace metric. First, we designed a hierarchical parameter decoupling architecture. Macroscopically, the architecture retains the classifier locally to isolate label heterogeneity; microscopically, it introduces an additive parameter decomposition that decouples the feature extractor on a global full-rank basis (to capture domain-invariant semantics, namely, the shared physiological vein topologies) and a local low-rank adapter (that accommodates device-specific characteristics, such as hardware-induced noise and illumination discrepancies). Furthermore, we propose a subspace similarity matching strategy based on principal angles on the Grassmann manifold. By exploiting the geometric properties of low-rank projection matrices, this strategy accurately quantifies the underlying distribution discrepancies among clients to guide personalized weighted aggregation. Extensive experiments on six public finger vein datasets demonstrate that the proposed framework significantly improves the overall recognition performance and mitigates performance degradation caused by data heterogeneity. Full article

Few-shot remote sensing scene classification aims to accurately recognize unseen scene categories using only a scarce number of labeled samples, which has emerged as a research hotspot in the field of remote sensing image interpretation. However, remote sensing images intrinsically suffer from large intra-class variations, high inter-class similarities, and complex background interferences. Traditional few-shot learning methods typically perform feature metric learning in Euclidean space, making it difficult to capture the non-Euclidean geometric distribution characteristics of remote sensing features, and they often neglect the spatial structural information embedded in feature maps. To address these issues, this paper proposes a novel few-shot remote sensing scene classification method, termed ZSFS-RGBN, which integrates a Zigzag Scanning Feature Sequence with a Riemannian Geometric Barycenter Network. Specifically, ResNet12 is first employed as the backbone to extract deep convolutional feature maps from both the support and query sets. Second, a Zigzag scanning strategy is introduced to reorganize the two-dimensional feature maps into one-dimensional feature sequences, thereby effectively preserving the spatial locality and structural continuity of the features. Third, an autoregressive moving average (ARMA) model is constructed to characterize the spatial dependencies of the feature sequences, and its state parameters are mapped onto a symmetric positive definite (SPD) matrix manifold, enabling the deep semantic representations of remote sensing scenes in a non-Euclidean geometric space. Finally, a Riemannian geometric barycenter network is designed to learn the Riemannian barycenter of each category on the SPD manifold, where a joint loss function is introduced to simultaneously optimize intra-class compactness and inter-class separability. Comprehensive experiments are conducted on three public remote sensing scene datasets: NWPU-RESISC45, UC Merced Land-Use, and WHU-RS19. Experimental results demonstrate that the proposed method consistently outperforms several representative state-of-the-art approaches under both 5-way 1-shot and 5-way 5-shot settings. Furthermore, ablation studies verify the effectiveness of each component within the proposed framework. Full article

Occlusion introduces severe structural asymmetry into pedestrian representations by corrupting body topology, breaking cross-scale semantic continuity, and destabilizing identity geometry. Rather than treating occluded person re-identification (ReID) as a local visibility completion problem, this work reformulates it as a topology-preserved symmetry restoration problem: recovering symmetric identity structure from asymmetrically corrupted observations. Under this view, we present the Topology-Stable Person Re-identification Network (TSP-Net), a unified visual framework with three coordinated components: structural restoration, cross-scale symmetry alignment, and prototype-stabilized identity geometry. Specifically, Topology-Guided Occlusion and Visibility Modeling (TOVM) serves as the structural restoration component, and is realized by a closed loop of the Topology-Aware Occlusion Simulator (TOS) and the Topology-Aware Visibility Estimation (TVE) branch; Semantic-Anchored Cross-Scale Fusion (SACF) performs symmetry-consistent semantic recovery across hierarchical features; and the Prototype-Stabilized Supervision Loss (PSS Loss) regularizes identity embeddings toward topology-consistent manifold centers through momentum-updated prototypes. Experimental results on both occluded and holistic benchmarks show that TSP-Net is effective for learning occlusion-robust person representations. These findings suggest that restoring topology-preserved symmetry is a promising route for robust person re-identification under structural corruption. Full article

Residual stress in brittle films on compliant substrates is routinely inferred from fragmentation experiments by combining an elastic stress-transfer model with a fracture strength criterion. This inversion is inherently coupled because the observed crack spacing depends jointly on the residual stress and the film fracture strength. Conventional closed-form estimators typically rely on a single feature, such as the cracking onset strain, and prescribe the fracture strength a priori, often at its bulk value. This practice discards most of the information encoded in the full crack-spacing evolution. It also obscures two sources of uncertainty: the intrinsic variability of thin-film fracture strength and the limited sensitivity of any single observable to individual parameters. Here, we recast the inversion as a Bayesian physics-informed neural network (B-PINN) in which the entire measured curve of the mean crack spacing versus applied strain is likely to occur. Stochastic gradient Langevin dynamics then sample the joint posterior of residual stress and fracture strength. A central finding is that crack-spacing data alone constrain only the difference between fracture strength and residual stress, confining the posterior to a one-dimensional manifold in parameter space and leaving each quantity individually unresolved. A single substrate curvature measurement, which, through the Stoney relation, depends on the residual stress but not on the fracture strength, provides the missing orthogonal constraint and collapses the posterior to a tight, well-resolved region. We further derive an identifiability condition under which buckle-wavelength observations serve as a third independent channel for recovering interface adhesion energy, and provide a synthetic proof-of-concept of this three-channel extension on DLC/Si and Mo/Si datasets; an experimental validation of the adhesion channel is identified as the natural next step but lies beyond the present scope. Requiring only standard fragmentation measurements and a single non-destructive curvature scan, the framework converts a point-estimate procedure into a posterior-quantified inverse method that makes explicit what can, and cannot, be learned from thin-film mechanics experiments. Full article

Single-cell RNA sequencing data are characterized by high dimensionality, sparsity, and strong nonlinearity, hindering conventional single-view clustering methods from capturing linear and nonlinear feature subspaces simultaneously. Features from distinct dimensionality reduction approaches are inherently complementary: PCA (Principal Component Analysis) preserves global linear structures, UMAP (Uniform Manifold Approximation and Projection) maintains topology and local neighborhoods, and PHATE (Potential of Heat-diffusion for Affinity-based Trajectory Embedding) depicts gradual transitions in cell differentiation. To fuse these complementary sources, we adopt an inter-view correlation maximization paradigm. Canonical Correlation Analysis (CCA) integrates two views by maximizing projection correlation but is limited to pairwise scenarios. We extend it to Generalized Canonical Correlation Analysis (GCCA) for multi-view alignment and introduce a deep autoencoder to construct the DeepGCCA (Deep Generalized Canonical Correlation Analysis) framework. This method generates three views via PCA, UMAP, and PHATE, extracts nonlinear latent features with the autoencoder, projects multi-view representations into a unified subspace under weighted GCCA constraints, and performs K-means clustering. Experiments on the two simulated and three real single-cell datasets evaluated in this study show that DeepGCCA demonstrates competitive performance against all single-view baselines and performs favorably compared to several widely adopted methods. Moreover, downstream marker gene analysis supports the biological interpretability of the resulting clusters within these datasets. Within the scope of this benchmark, DeepGCCA provides a valuable reference for high-precision clustering of single-cell transcriptomic data, offering practical insights into multi-view integration and biological interpretability. Full article

Enterprise financial risk prediction is a fundamental task in financial risk management, yet its performance is often hindered by severe class imbalance, cross-enterprise heterogeneity, and the limited availability of labeled risky samples. These challenges are particularly pronounced in few-shot settings, where conventional machine learning and deep classification models tend to suffer from unstable representation learning, feature collapse, and weak decision boundaries. To address this issue, this study proposes a hierarchical metric learning framework for few-shot enterprise financial risk prediction on imbalanced tabular data. The framework integrates a state-space feature embedding network, an Adaptive Spectral Decomposition and Multi-Scale State Embedding module, and a Hierarchical Metric Manifold Alignment mechanism to enhance risk-sensitive representation learning, preserve geometric consistency across embedding levels, and improve prototype-based discrimination in the metric space. Experiments are conducted on three public datasets, namely American Bankruptcy, Corporate Financial Risk Assessment, and Enterprise Financial Network, under a unified 2-way 20-shot setting. The proposed method consistently achieves the best overall performance across Precision, Recall, Accuracy, F1-score, and AUC, with AUC values of 0.9526, 0.9687, and 0.9716 on the three datasets, respectively. Ablation studies and visual analyses further show that the proposed framework improves intra-class compactness, inter-class separability, and classification robustness under highly imbalanced conditions. These findings indicate that the proposed method provides an effective and robust machine learning solution for enterprise financial risk prediction and early warning in data-scarce financial scenarios. Full article

Batch effects arising from variations in hardware, acquisition protocols, and reconstruction parameters present a critical challenge in radiomics, limiting the generalizability of models across multicentre studies. Existing harmonization methods, such as ComBat, CovBat, z-score normalization, and Generative Adversarial Networks, exhibit significant limitations when applied to high-dimensional radiomic data. ComBat assumes a linear feature space and tends to leave residual center-specific information recoverable by downstream classifiers. This paper introduces RiTex (Riemannian Texture Harmonization), a framework that solves a generalized eigenvalue problem between class-aware biological scatter and Ledoit–Wolf-regularized per-batch covariances, with the SPD-manifold Fréchet mean used as a principled averaging step. We evaluate RiTex on the 50-dataset radMLBench benchmark and on a new four-center head-and-neck benchmark with known center labels ($n$ = 380 patients, k = 4 centers from TCIA: HGJ, MDACC, Maastro, QIN). On radMLBench, RiTex reduces the batch auto-detection AUC in 48/50 (96%) datasets, 42/50 (84%) reductions remain significant after Benjamini–Hochberg correction; the mean Batch AUC reduction is ΔBatch = −0.365 (95% bootstrap CI [−0.418, −0.312]), with no significant degradation in biological AUC (mean ΔBio = +0.018, 95% CI [−0.011, +0.047]). On the H&N benchmark with real center labels, RiTex reduces the Batch AUC from 0.74 to 0.59, while ComBat and CovBat leave it at ≈0.98. A component-wise ablation shows that the dominant source of empirical performance is the GEVD step, together with Ledoit–Wolf shrinkage. The SPD Fréchet mean acts as a theoretical scaffold with a negligible empirical contribution (ΔBatch AUC = −0.014 vs. arithmetic mean). Full article

Non-stationary signal detection is challenging when discriminative information is not reflected in global energy, mean spectra, or a single covariance statistic, but is instead embedded in the organization of local time–frequency structures. This paper proposes an information-geometric detector defined on local symmetric positive definite (SPD) structure fields. Time–frequency patches are transformed into a spatially distributed field of second-order tensors to characterize local directional organization and anisotropy. Under a locally isotropic Riemannian Gaussian approximation on the SPD manifold, the local distance-difference evidence is monotonically related to an approximate log-likelihood ratio, providing an information-geometric interpretation without implying strict Neyman–Pearson optimality. Instead of forming a single global statistic or stacking patch-level features, the proposed method constructs a spatially distributed field of structured SPD objects and derives local distance-difference evidence, which is subsequently aggregated into a sample-level detection statistic. Experiments under a controlled SPD structure-field locality benchmark show that performance gains are primarily driven by the proposed SPD structure-field representation, with the Riemannian metric providing only secondary refinement. Full article

Continuous monitoring of lower-limb kinematics in natural environments is essential for gait analysis and rehabilitation but remains challenging due to the limitations of optical systems and the inaccuracy of sparse inertial sensor methods. To address this, we propose a high-precision, minimalist wearable system utilizing only three inertial measurement units placed on the pelvis and shanks. In the data preprocessing stage, engineering modifications are made based on the traditional gradient descent algorithm to implement adaptive channel adjustment on the acceleration and magnetic data of a single IMU, aiming to alleviate the impact of motion acceleration and external magnetic interference on the temporal feature manifold. Subsequently, a pure Transformer neural network is utilized to capture long-range temporal dependencies, reconstructing full lower-limb kinematics without relying on rigid biomechanical assumptions. The model was optimized and deployed on an STM32N647 microcontroller to achieve real-time edge inference with a low latency of approximately 17 ms. Experimental results demonstrate that the proposed method achieves a mean absolute error of 2.41° for level walking, significantly outperforming traditional constrained Kalman filter approaches. Furthermore, it maintains high tracking robustness during complex nonlinear movements such as squatting and lunging. In conclusion, this edge-computing-enabled framework provides an accurate, comfortable, and real-time solution for unconstrained human motion capture in daily scenarios. Full article

Emotion recognition through electroencephalogram (EEG) signals is crucial for brain–computer interfaces (BCIs), yet existing methods often struggle with heterogeneous feature fusion and capturing long-range temporal dependencies. To address these challenges, we propose MAF-TransNet, a novel unified spatiotemporal framework. Specifically, parallel Fully Connected Neural Network (FCNN) modules first non-linearly align heterogeneous differential entropy (DE) and power spectral density (PSD) features. Subsequently, an Adaptive Channel-wise Feature Encoder (ACFE) recalibrates spatial–spectral responses to highlight emotion-relevant cortical activations. Finally, a Transformer encoder dynamically models the global temporal evolution of emotional states. Evaluated on the SEED-IV and DEAP datasets, MAF-TransNet achieves superior subject-dependent (SD) accuracies of 88.80% and 96.58%, respectively, alongside robust subject-independent (SI) performance. Furthermore, Granger causality analysis reveals distinct emotion-dependent prefrontal asymmetry, while t-SNE visualizations confirm the formation of a highly discriminative, linearly separable feature manifold. Ultimately, MAF-TransNet effectively unifies local spatial–spectral extraction with global temporal modeling, providing an accurate and robust approach, while offering preliminary insights into the spatiotemporal dynamics of emotion for future affective BCI applications. Full article

Micro-expression recognition (MER) remains a formidable challenge in affective computing due to the subtle, localized, and fleeting nature of facial muscle actuations. Conventional spatial-temporal networks are easily overwhelmed by static facial topologies, leading to feature representations that are heavily biased toward identity-specific noise. To address this, we propose the Motion-Anchored Cross-Modal Fusion Network (MACFN), a novel dual-stream ViT architecture that explicitly decouples and synergizes spatial appearance and optical flow dynamics. Specifically, we introduce a motion-anchored spatial attention module, which translates latent motion features into a sparse spatial probability mask. It acts as an enhancement gate, forcing the texture stream to bypass static backgrounds and attend to genuine ME-related regions. Furthermore, we design a cross-modal bilinear fusion module to capture the second-order interactions across modalities, mapping the coupled features into a discriminative semantic manifold. Extensive experiments conducted on the CASME II, SAMM, and SMIC databases under the rigorous leave-one-subject-out composite database evaluation protocol demonstrate that MACFN is effective and achieves competitive performance compared to several recent methods. Full article

Predictive modeling of thermal aging in solid propellants is challenging because HTPB-based propellants are highly filled particle-reinforced polymer systems with sparse experimental data, nonlinear parameter coupling, and partially unclear aging mechanisms. This study proposes a Physics-Informed Manifold Neural Operator (PIMANO) framework for multi-parameter prediction of polymer aging in HTPB solid propellants. Accelerated thermal aging, stress relaxation, and swelling experiments were used to obtain aging temperature, aging time, crosslinking density, and viscoelastic Prony-series parameters. A continuous aging-state field was first reconstructed over the temperature–time domain by radial basis function interpolation. Crosslinking density was then introduced as a physically interpretable bridge-state variable linking aging conditions with viscoelastic responses. Among three candidate kinetic models, the modified Arrhenius–Avrami model gave the best fitting performance for crosslinking-density evolution, with R² = 0.988 and MRE = 0.0199. By combining local multi-scale neighborhood features, manifold latent representations, and DeepONet-based operator learning, PIMANO established a unified mapping from aging conditions to multi-parameter viscoelastic responses while incorporating bridge-state consistency, parameter non-negativity, and evolution-direction constraints. Under the RBF-augmented validation setting, PIMANO-ae achieved RMSE = 0.7847, MAE = 0.3366, R² = 0.9995, and MRE = 0.0027. Compared with the traditional model, RMSE, MAE, and MRE were reduced by 94.93%, 96.47%, and 96.85%, respectively. Temperature leave-one-out validation further yielded average R² values of 0.9469–0.9647 and MRE values of 4.98–6.21% at unseen aging temperatures. These results demonstrate that PIMANO provides an accurate, stable, and physically interpretable framework for multi-parameter aging prediction and life-assessment modeling of polymer-based energetic materials. Full article

Bipolar plates are critical components of proton exchange membrane fuel cells (PEMFCs), strongly influencing thermal management, mechanical stability, and overall system efficiency. In this study, an integrated framework combining multiphysics simulation, artificial intelligence (AI), and data augmentation techniques was developed for PEMFC bipolar plate optimization. A coupled thermal–structural finite element model was established in COMSOL Multiphysics to evaluate temperature distribution, thermal stress, and structural deformation under varying operating conditions. A total of 80 parametric design cases were generated by varying six key parameters: hole radius, plate thickness, heating power, manifold pressure, plate number, and heat transfer coefficient. The dataset was expanded using SMOTE, GAN, and LLM-based augmentation techniques and used to train ANN, LR, RF, XGBoost, and SVR models. Model performance was evaluated using 5-fold cross-validation with MAE, RMSE, and LogCosh metrics. The results showed that ensemble tree-based methods, particularly RF and XGBoost, achieved the highest prediction accuracy and computational efficiency. XGBoost produced the best temperature prediction performance for the SMOTE-based dataset (RMSE = 3.668), while RF achieved the lowest stress prediction error (RMSE = 0.0490). GAN-augmented datasets provided stable and reliable predictions, whereas LLM-generated datasets resulted in higher prediction errors and lower physical consistency. Feature importance analysis revealed that plate thickness dominates displacement prediction (≈0.72 importance), manifold pressure governs stress behavior (≈0.999), and heating power is the primary factor affecting temperature prediction. The proposed AI-assisted surrogate modeling framework enables rapid and accurate thermo-mechanical prediction while significantly reducing computational cost compared to conventional multiphysics simulations. The findings demonstrate that integrating physics-based simulations with data-driven approaches provides an efficient strategy for the optimization of next-generation PEM fuel cell bipolar plates. Full article