Search Results (297)

Background/Objectives: Obstructivesleep apnea (OSA) is one of the most widespread forms of sleep disease, affecting over 936 million adults globally. The health consequences of obstructive sleep apnea (OSA) are well documented; however, it remains largely underdiagnosed because the current gold-standard diagnostic method, polysomnography (PSG), is often costly, time-consuming, and unavailable in many healthcare settings. To address these challenges, this study presents a novel explainable deep learning (DL) framework for automated multi-class OSA severity classification using single-lead electrocardiogram (ECG) signals. Methods: The proposed framework integrates a hybrid CNN–BiLSTM architecture with explainable artificial intelligence (XAI) techniques to generate clinically meaningful predictions and explanations across four OSA severity classes: Normal, Mild, Moderate, and Severe. The framework was evaluated using the publicly available PhysioNet Apnea-ECG dataset (70 recordings) together with an institutional ECG dataset (150 recordings), resulting in a combined cohort of 220 recordings. Results: The proposed framework achieved an overall classification accuracy of 94.7%, with sensitivity and specificity values of 92.3% and 96.1%, respectively. Furthermore, the proposed model consistently outperformed conventional machine learning algorithms, including Support Vector Machine (SVM), Random Forest, and XGBoost, by 5.5%, 4.2%, and 2.9%, respectively. To enhance transparency and clinical trust, SHAP (SHapley Additive exPlanations) was employed to identify the most influential physiological predictors driving model decisions. Heart rate variability features, particularly RMSSD and pNN50, emerged as the strongest indicators of OSA severity. Moreover, computational efficiency analysis revealed that the model required only 0.23 s to process a 60 s ECG epoch on a standard computing platform, supporting its suitability for real-time deployment. Conclusions: The findings demonstrate that explainable deep learning applied to ECG signals can provide accurate, interpretable, and computationally efficient assessment of OSA severity. The proposed framework may support OSA screening, clinical triage, and early intervention, particularly in resource-constrained healthcare environments. Full article

Agriculture is hypothesized by climate scientists as being the earliest contributor to the emergence of the Anthropocene—the human-induced geological epoch—due to related greenhouse gas emissions. The environmental effects of industrial agriculture since the eighteenth century support this hypothesis; however, when agriculture is understood in the context of family- and community-based farming in culturally and geographically specific places around the world, many of its regenerative approaches can be seen as land stewardship rather than as a cause of Anthropocenic climate change. Writing as a scholar-farmer, the author discusses the typically Christian religiosity found in the new agrarianism spoken through the works of Wendell Berry but finding its root in the Jeffersonian agrarian movement. The center of the discussion is on whether this Christianity-based religiosity is related to regenerative family farming in contemporary California. Walking with the reader through the Anthropogenic conditions of industrial agriculture from Europe to California and through the current state of religion and ecology, this article presents this case of family syntropic farming as a co-creative process with nature. It demonstrates that, in the multicultural and multi-religious context of California, it is challenging to adopt a single religious ecological knowledge from a single religious tradition. It concludes that advancing the farmers’ literacy of nature’s animacy is rather the foundation of the agrarian ecospirituality in the ethnically and religiously diverse California in the Anthropocene. Full article

This study proposes a sequential Hybrid GP-ANN architecture based on the concept of autonomous feature engineering for multi-output performance prediction of single-phase induction motors, serving as a high-fidelity surrogate model ready for digital twin integration. The required high-resolution motor dataset (88,375 samples) was generated using a parametric sweep approach with Ansys RMxprt. In the proposed architecture, Genetic Programming (GP) is not positioned as a final predictor but as an analytical filter that discovers hidden physical relationships in raw input data and converts them into super features, thereby structurally eliminating the polynomial explosion risk and scale sensitivity that arise when raw data are modelled directly. The nonlinear physical relationships discovered autonomously by GP are added to the network input matrix as mathematical vectors, breaking the black-box structure of standard artificial neural networks and transforming it into a physics-inspired grey-box model. The nonlinear terms discovered by GP are fed into the network after independent Z-score normalisation to preserve gradient stability. To comprehensively evaluate this framework, its predictive performance is benchmarked against industry-standard machine learning algorithms, including Random Forest (RF), Support Vector Regression (SVR), and Extreme Gradient Boosting (XGBoost). Results on variation-based unseen test data show that the hybrid model achieves competitive prediction accuracy relative to XGBoost—matching or exceeding it for Output Torque and Input Current, and remaining broadly comparable for Output Power, though XGBoost achieves substantially lower RMSE for Efficiency and Power Factor—while additionally offering transparent mathematical traceability, reducing error rates (RMSE) by 42.1% to 66.1% per parameter compared to the standard ANN. The model achieves an $R^2$ score of 0.9879 for the power factor parameter, where conventional approaches struggle. To rigorously validate the interpretability of this grey-box architecture, a global Permutation Feature Importance (PFI) analysis was conducted, showing that the GP-derived super features collectively account for 61.46% of the decision logic, outweighing the combined contribution of the raw inputs (38.54%). Furthermore, the autonomous feature engineering layer is found to reduce the learning burden on the ANN, allowing it to converge to a substantially lower error floor within the same fixed epoch budget. With an online inference time below 0.02 ms, the proposed architecture offers a robust methodological infrastructure for sustainable motor digital twins through a balanced trade-off between prediction accuracy and computational efficiency. Full article

Background/Objectives: Sleep disorders are heterogeneous conditions with diverse neural, muscular, and ocular manifestations, making polysomnography (PSG) the gold standard for accurate diagnosis. Artificial intelligence-based approaches, particularly deep learning (DL) models capable of integrating heterogeneous information, offer a promising solution for reliable decision-making in such clinical scenarios. However, most existing DL studies have focused on a single disorder, relied on limited datasets, or employed epoch-level labeling strategies that overlook the episodic nature of sleep pathophysiology, thereby limiting clinical applicability. To address these gaps, we propose a novel multimodal attention-enhanced fusion network (MAF-SleepNet) for automatic multi-class sleep disorder classification based on the International Classification of Sleep Disorders. Methods: MAF-SleepNet jointly processes electroencephalography (EEG), electrooculography (EOG), and leg electromyography (EMG) signals through modality-specific feature extraction and adaptive attention mechanisms, capturing both intra- and inter-modality dependencies. The model was evaluated on a combined dataset of 141 recordings from three public databases, including five PSG-requiring disorders and a healthy class. Results: Experimental results demonstrated that MAF-SleepNet achieved 86.07 ± 3.66% accuracy and 82.67 ± 4.46% macro-F1 under a strict subject-independent cross-validation, and 98.96 ± 0.66% accuracy and 98.93 ± 0.60% macro-F1 under subject-dependent cross-validation. Conclusions: These results demonstrate that the proposed approach provides a more reliable and clinically meaningful assessment compared to many existing studies that rely on subject-dependent evaluation or epoch-level labeling. The findings highlight the effectiveness of adaptive multimodal fusion for robust and clinically relevant sleep disorder classification. Future work should investigate the integration of respiratory and autonomic modalities and validation on larger multi-center cohorts. Full article

Accurate differentiation of Alzheimer’s disease (AD) and frontotemporal dementia (FTD) is clinically important because their management strategies differ. This study aimed to develop and evaluate a Dynamic Threshold Graph Convolutional Network (DT-GCN) and to systematically compare four EEG-based functional connectivity (FC) measures—Pearson correlation, phase-locking value (PLV), Granger causality, and copula analysis—for classifying AD, FTD, and healthy controls (HCs). In contrast to fixed graph binarization, DT-GCN updates the connectivity threshold at each training epoch according to training-fold loss. The model is trained under a multi-task objective combining classification, reconstruction, and contrastive losses, with the loss weights adjusted across three stages of training. Resting-state 19-channel EEG recordings from a public dataset (DS004504) comprising 36 AD, 23 FTD, and 29 HC participants were segmented into non-overlapping 8 s epochs. FC estimates were aggregated to obtain a subject-level adjacency matrix for each measure, with one-hot node identity and node degree as node features. Classification was performed at the subject level using stratified five-fold cross-validation with normalization and hyperparameter selection confined to training folds. Under the broadband setting, copula-based FC with DT-GCN yielded the highest observed three-class accuracy of $0.82\pm 0.07$ (macro-F1 $0.79\pm 0.07$), compared with $0.45\pm 0.06$ accuracy and $0.36\pm 0.06$ macro-F1 for the standard GCN baseline. However, the small single-center cohort ($N=88$, including 23 participants with FTD) and the resulting small test folds limit the precision of the performance estimates and preclude robust inferential comparisons among FC methods and classification scenarios. These exploratory, dataset-specific findings require independent external validation and should not be interpreted as generalizable clinical performance or validated clinical biomarkers. Full article

Vibration-based mechanical fault diagnosis has become a critical research area, mostly driven by the need to improve equipment reliability and reduce unplanned downtime in industrial settings. Time–Frequency Networks (TFNs) have shown strong potential here, combining interpretable time–frequency transformations with deep learning classifiers in a single framework. This work reproduces the original TFN model from the recent literature and evaluates its noise robustness under additive Gaussian noise (10 dB, 0 dB, $-5$ dB SNR) and impulsive noise at the same levels, across five architectures: Backbone CNN, Random CNN, TFN-Chirplet, TFN-Morlet, and a squeeze-and-excitation attention CNN baseline. The evaluation protocol corrects two methodological issues identified during peer review of an earlier version of this work—window-level data leakage between train and test splits, and selection of the best-performing training epoch rather than a fixed final-epoch result—both of which are shown to materially affect reported outcomes. Under the corrected protocol, TFN-Morlet remains the most noise-robust architecture, with only a 19.09% accuracy drop from clean to $-5$ dB AWGN, approximately 15.5 percentage points better than Backbone CNN under the same conditions; an architectural anomaly reported in the earlier version of this study, in which mild noise appeared to improve an unconstrained CNN’s accuracy, was not reproduced under the corrected protocol and is shown to be an artifact of the original methodological issues. Per-class analysis and multi-model confusion matrices further reveal that misclassifications under severe noise are dominated by confusion between the same defect severity at different fault locations, rather than between different severities at the same location as previously reported. These results indicate that time–frequency-aware convolutional kernels improve both classification accuracy and noise resistance under rigorous, leakage-free evaluation, and that this robustness is not replicated by a generic attention mechanism alone. Full article

Digital twins are increasingly adopted in geographic research as dynamic representations of environmental systems; however, their application to regulated river shorelines remains limited, particularly where bathymetric change, hydrological variability, and shoreline-state dynamics must be integrated within a single GIS-based framework. This study develops and validates a GIS-based spatiotemporal digital twin-oriented framework for the dam-affected shoreline downstream of the Włocławek Dam, Poland. The framework integrates four bathymetric surveys acquired in 2008–2011, water-level records, airborne laser scanning data, and three-dimensional hydrotechnical infrastructure within a unified geodatabase designed for dynamic shoreline-state reconstruction, multi-epoch analysis, and environmental monitoring. A key methodological element is the treatment of water level as a dynamic reference surface, enabling the automated delineation of inundation and exposure zones for observed and scenario-based hydrological conditions. The reconstructed bathymetric surfaces were organized as a multidimensional raster dataset with time as an explicit analytical dimension, supporting repeatable change detection, cross-sectional interpretation, and temporal trend analysis. To extend the framework beyond purely retrospective analysis, a near-real-time hydrological updating component was implemented through ingestion of operational water-level observations from the IMGW API into the geodatabase. Validation of the trend-based prediction for 2011 yielded R² = 0.967, RMSE = 0.44 m, MAE = 0.28 m, and bias = −0.06 m. The proposed framework provides a transferable geospatial basis for spatiotemporal modelling and monitoring of regulated river shoreline dynamics under changing hydrological conditions. Full article

High-temporal-resolution vertical total electron content (VTEC) inversion is important for ionospheric delay correction in maritime GNSS applications, but offshore single-station observations often suffer from limited satellite geometry, clustered ionospheric pierce points, and noise-sensitive least-squares (LSs) solutions. This study proposes a Virtual-Observation-Based Tikhonov Regularization (TVO) method for stabilizing ill-conditioned least-square VTEC inversion. TVO links the regularization factor to the condition number of the normal-equation matrix and selectively constrains higher-order spatial-gradient parameters while preserving background VTEC and receiver-bias terms. Experiments using the European mid-latitude station OBE4 and 17 surrounding stations on 1 July 2021 show that short epoch intervals and increased model complexity aggravate ill-conditioning, especially for the full quadratic model at 30 s. Compared with LS, TVO reduces the average RMS difference relative to the GIM-interpolated VTEC reference by 56.30% across the four VTEC models for the 17 stations. Maritime validation using South China Sea buoy data collected from 19 to 25 May 2025 further shows that TVO suppresses local discontinuities and amplitude anomalies, reducing the overall RMS difference relative to the GIM-interpolated VTEC reference from 26.07 TECU to 14.74 TECU. These results suggest that TVO can improve the numerical stability of maritime single-station VTEC inversion under constrained observation geometry. Full article

Background/Objectives: Automated cephalometric landmark detection using deep learning has the potential to streamline routine orthodontic diagnosis. However, the clinical relevance of artificial intelligence (AI) localisation accuracy depends on how detection errors propagate into derived angular measurements and skeletal classifications. We retrospectively evaluated 14 YOLO-based model configurations and quantified the agreement between AI-derived and expert-derived ANB-based skeletal classifications. Methods: Twelve working YOLO-based models (YOLOv5xu, YOLOv11 nano/small/medium/large variants) were trained on a single-centre dataset of 120 lateral cephalograms and evaluated on an independent test set of 11 cephalograms (stratified across skeletal Classes I, II, III). The four ANB-defining landmarks (Sella, Nasion, A-point, B-point) were the focus of the analysis. Each test cephalogram had been annotated by four orthodontists (44 measurements per image), yielding the expert reference. We assessed the effects of architecture, bounding-box size (40/100/150 px), training dataset scale (235–4255 images) and training epochs on localisation accuracy (mean radial error, MRE; Successful Detection Rate, SDR) and on the downstream ANB-based skeletal classification. Diagnostic concordance was quantified by classification agreement, Cohen’s $\kappa$ with bootstrap 95% confidence intervals (10,000 iterations), an exact one-sided binomial test for discordance, and Wilson exact CIs per class. Results: The best-performing model (Model 2; YOLOv11l, 40 × 40 px bounding box, 1175 training images) achieved an MRE of $3.10\pm 1.00$ mm and a SDR@4 mm of 87.2% for S, N, A, and B. ANB-based skeletal classification demonstrated 96.9% concordance with expert assessments (95% bootstrap CI: 93.8–99.2%; Cohen’s $\kappa$ = 0.946 [95% CI 0.89–0.99]; exact binomial test against a 90% concordance threshold $p=0.003$). Per-class concordance was Class I 95.8% (23/24), Class II 94.9% (56/59), and Class III 100% (47/47). Three of four discordant cases clustered near the Class I/II diagnostic threshold (expert ANB $\approx 4.5°$). Bounding-box size dominated localisation accuracy, with a 3.5-fold increase in MRE from 40 × 40 to 150 × 150 px configurations and SDR@4 mm collapsing from 82.8% to 0%. Conclusions: Within the constraints of a retrospective single-centre design with a small (n = 11) independent test set, YOLO-based AI landmark detection demonstrated promising diagnostic concordance with expert consensus for ANB-based skeletal classification. These findings warrant prospective, multi-centre external validation before clinical deployment and support a confidence-aware workflow in which AI predictions for borderline ANB values undergo mandatory clinician verification. Bounding-box calibration emerged as the single most impactful preprocessing decision. Full article

Breast ultrasound (BUS) is important for breast tumor detection, but speckle noise, low contrast, operator dependency, and limited medical datasets hinder robust deep learning. Although raw radiofrequency (RF) signals contain richer acoustic information than conventional B-mode images, multimodal fusion approaches often increase computational cost. To address these issues, this study proposes a curriculum transfer learning-based approach that sequentially exploits different ultrasound information representations during training. The proposed approach maintains a single detection model architecture rather than relying on complex multimodal input fusion. Phase, Envelope, and B-mode images generated from raw RF signals were defined as distinct input domains, and various training orders were evaluated. In addition, lightweight detection models based on YOLOv5, YOLOv8, YOLO11, and YOLO26 were compared to select the optimal model. A total of nine experimental settings, including single-modality training and curriculum learning conditions, were repeatedly evaluated using 100 random seeds. The experimental results showed that the proposed Phase-Envelope-B-mode (P-E-B) curriculum transfer learning strategy achieved the highest average mAP@50 among the evaluated training scenarios, with an approximately 2.08% relative improvement over single B-mode training under the fixed patient-level split and 100-seed repeated evaluation setting. The average convergence epoch was also lower than that of single B-mode training, indicating that the proposed strategy provided a favorable convergence profile while improving average validation performance. These results should be interpreted as proof-of-concept evidence obtained under a fixed patient-level split and 100-seed repeated evaluation setting, rather than as conclusive evidence of external clinical generalizability. Within this controlled small-data setting, the findings suggest that RF-derived representations may provide useful training-stage curriculum information for B-mode-based breast tumor detection while maintaining a single B-mode inference pathway. Full article

Transcutaneous spinal cord stimulation (tSCS) is a promising neuromodulation approach for motor recovery after spinal cord injury (SCI), yet clinical programming remains largely dependent on subjective parameter selection. This study evaluated high-density surface EMG (HD-sEMG)–derived spatial and temporal features as objective biomarkers for tSCS optimization in three adults with chronic cervical SCI. A 64-channel electrode array recorded stimulation-evoked responses across five cervical stimulation levels, four pulse widths, and graded amplitudes. Features describing activation magnitude, spatial distribution, cluster morphology, and temporal dynamics were extracted from epoch-based activation maps. Of the three enrolled participants, two demonstrated measurable stimulation-evoked responses and contributed to the paired-pulse analyses, whereas pulse-width analyses were limited to a single responsive muscle (left flexor carpi) in one participant. Paired-pulse analysis identified root mean square (RMS) as the most discriminative feature, revealing nonlinear, muscle- and level-specific dose–response relationships in which maximal suppression often occurred at intermediate rather than maximal amplitudes. Increasing pulse width expanded the spatial extent of recruitment (active area: p = 0.006; convex hull area: p = 0.004) without altering response timing. Polarity reversal analysis demonstrated stable innervation zone localization across stimulation levels and amplitudes. These findings establish a spatially resolved HD-sEMG framework that may support individualized tSCS parameter selection in SCI. Full article

Background/Objectives: Heterogeneous imaging protocols, a lack of labeled data, and domain shifts continue to make training deep learning models to analyze medical images a challenge. This study presents ETDACVO (Enhanced Tasmanian Devil Anti-Conservative Variable Optimization), a hybrid evolutionary optimization system designed to improve convergence stability and cross-domain robustness in brain tumor MRI classification. Methods: ETDACVO combines Tasmanian Devil Optimization (TDO), Anti-Conservative Variable Optimization (ACVO), and Exponentially Weighted Moving Average (EWMA) smoothing to stabilize evolutionary parameter updates. Unlike existing approaches that optimize augmentation policies or optimizer dynamics separately, ETDACVO simultaneously evolves both components within a single evolutionary loop. The framework was evaluated on four MRI datasets (Nickparvar, Mendeley, BRISC, and Figshare), comprising 28,151 images. In addition, a convergence-aware explainability mechanism, CA-EA-GradCAM, was developed by integrating gradient saliency, transformer attention, and evolutionary convergence confidence to generate confidence-sensitive tumor localization maps. Results: Experimental results demonstrated that ETDACVO achieved a 2.3–2.5% improvement in classification accuracy and converged 19–22 epochs faster than baseline optimizers. The statistical significance of these improvements was confirmed using paired statistical tests (p < 1 × 10⁻⁵). Cross-dataset transfer experiments further showed strong domain-shift resilience, with performance retention reaching 92.8%. The proposed CA-EA-GradCAM mechanism provided interpretable and confidence-aware tumor localization maps. Conclusions: ETDACVO provides a robust and computationally efficient optimization framework for deep-learning-based medical image analysis. By jointly optimizing augmentation strategies and optimizer dynamics, the framework enhances convergence stability, cross-domain robustness, and interpretability, making it a promising approach for reliable brain tumor MRI classification under heterogeneous imaging conditions. Full article

In this study, we propose SiStNet, a single-stage deep learning architecture for vehicle detection in autonomous driving scenarios. The proposed architecture is trained entirely from scratch on domain-specific data without relying on pretrained backbones and is evaluated against representative baseline detectors under identical training conditions. Experiments are conducted on the KITTI dataset under a consistent training and evaluation protocol. An ablation study conducted under a reduced training budget (20% of data, 30 epochs) revealed that multi-scale detection, data augmentation, and anchor-based prediction did not contribute positively to detection performance under the given training conditions. Based on these findings, the final SiStNet architecture was simplified by removing these three components and re-trained under the full training budget. The resulting model achieves a mean Average Precision (mAP) of $0.5033\pm 0.0072$ and a recall of $0.6935\pm 0.0214$, representing substantial improvements over the initially reported values ($0.3896$ and $0.439$, respectively). The inference speed of SiStNet is $24.41\pm 0.02$ FPS, which satisfies the real-time threshold of 20 FPS defined in this study. The model achieves lower mAP than baseline detectors that employ larger and deeper architectures; all models were trained from scratch under identical conditions, so the accuracy gap reflects architectural capacity differences rather than pretraining advantages. SiStNet is presented as a domain-specific, scratch-trained alternative that achieves competitive detection performance without reliance on large-scale pretraining, at the cost of lower mAP relative to deeper baselines. Full article

Wireless-power mobile edge computing, SWIPT-MEC, priority-aware WPT scheduling and swarm resource allocation already solve important parts of the energy-management problem. The novelty of WPT-JCCO is not any one of those elements; it is a single swarm-supervisory feasible set that couples decisions which the three adjacent method classes normally separate. Each epoch-level action jointly selects the robot to charge and one of three physically distinct WPT modalities: far-field radio-frequency, resonant near-field and directional lightwave transfer, together with the SWIPT split, local/edge task placement, CPU frequency, bandwidth and transmit power. Relative to SWIPT-MEC, the formulation adds discrete recipient–modality selection with pose, alignment, blockage and dwell-dependent feasibility. Relative to conventional WPT scheduling, charging is not a separate priority or routing stage but is solved jointly with computation and radio allocation. Relative to swarm resource-allocation methods, energy replenishment is endogenous and an individual minimum-battery constraint protects the weakest robot. A fourth coupling makes the centrally generated resource vector admissible only when the complete sense–compute–actuate age fits the one-second supervisory epoch; otherwise a previously feasible or local-safe action is applied. Nonlinear harvesting, partial offloading, priority scoring and augmented-Lagrangian primal–dual updates are treated as established techniques. This paper derives the continuous block updates, keeps the WPT variables binary through candidate screening, and declares convergence only when stationarity, feasibility, merit-change and binary-hold tests are jointly satisfied. Normalised primal steps are safeguarded by backtracking, dual and penalty updates are bounded, and a local tracking bound plus divergence monitor delimit real-time operation without claiming global mixed-integer optimality or closed-loop motion stability. Numerical evaluation over a 20-robot swarm and 30 Monte Carlo runs shows that WPT-JCCO reduces net energy depletion by 23.8% relative to communication–computation optimisation with static WPT and by 49.7% relative to local-only execution, while increasing task success from 93.5% to 97.3%. A released common-trace comparison shows normalised-cost reductions of 11.1%, 11.3% and 5.8% relative to two-stage WPT+CCO, fixed-SWIPT dynamic offloading and an offline Q-learning scheduler. Convergence and one-factor-at-a-time sensitivity studies further examine swarm size, task load, WPT budget, bandwidth, edge capacity, mobility and channel margin. The headline values remain scoped to the nominal independent-task case; mode-specific RF, near-field and lightwave operating envelopes, robust pose/CSI, WPT-safety and task-DAG extensions are formulated but not presented as hardware-validated results. Full article

Real-time precise point positioning (PPP) is increasingly supported by open satellite-broadcast state-space representation (SSR) services, yet standalone operation with a single service remains vulnerable to limited constellation support, correction outages, latency variations, and service-dependent modeling inconsistencies. In the Asia-Pacific region, MADOCA-PPP and PPP-B2b provide two publicly accessible and complementary SSR sources, but their consistent fusion before user-level PPP estimation remains insufficiently investigated. This paper proposes a correction-domain fusion framework that combines MADOCA-PPP and PPP-B2b orbit and clock corrections before PPP estimation, rather than merging final positioning solutions. Inter-service discrepancies and unknown cross-correlations are handled by a bias-state-aware structured covariance intersection strategy, in which the relative weighting is derived from the respective correction information (inverse variance), preserving statistical consistency and avoiding overconfident fusion. A unified multi-GNSS PPP scheme further supports signal-priority harmonization, broadcast-ephemeris adaptation, correction-age control, and GLONASS inter-frequency and differential code bias handling. Static-station per-epoch (pseudo-kinematic) and offshore kinematic experiments validate the framework. In the static-station test, fusion raised the mean number of valid satellites from 21.98 and 14.98 to 26.56 and improved the horizontal RMS to 0.033 m—better than either standalone service (0.037 m, 0.079 m)—confirming a genuine combination rather than source selection, while the 3D RMS (0.068 m) matched the best standalone service (0.066 m). In the offshore test, fusion achieved the best overall accuracy (0.232 m horizontal, 0.290 m 3D, versus 0.332 m and 0.313 m for the standalone services) and the most satellites (25.4). It also degraded most slowly with increasing elevation cut-off, outperforming both services about threefold at 40°. A normalized-innovation-squared check confirmed the fused covariance is consistent and not overconfident (median ≈ 1.1; within the 99% bound in 100% of epochs). Under single-service outages from 30 s to 600 s, fusion maintained 100.0% availability, confirming its advantage in redundancy, continuity, and resilience. Full article