Search Results (1,653)

Super-large aggregate asphalt mixtures (SLAM-50) are expected to improve rutting resistance while reducing asphalt demand, but the quantitative relationships between volumetric parameters and key performance indicators remain unclear, limiting performance-oriented mixture design. This study experimentally evaluated the volumetric properties and key performance indices of SLAM-50 across four aggregate gradations and seven asphalt contents. The measured performance indices included compressive strength, splitting strength, compressive resilient modulus, fracture energy, and dynamic stability. With increasing asphalt content, the air voids (VV) decreased, the voids in mineral aggregate (VMA) decreased initially and then increased, and the voids filled with asphalt (VFA) increased monotonically. All performance indices exhibited a non-monotonic trend, increasing first and then decreasing, with a balanced overall performance at 3.0–3.2% asphalt content. Linear regression models showed limited predictive capability (R² = 0.5944–0.8845). To address this gap, a Backpropagation (BP) neural network was developed using asphalt content, volumetric parameters, mixture density, and gradation type as inputs, and the measured performance indices as outputs. This framework enables simultaneous multi-output prediction and captures the nonlinear, coupled relationships among variables. The model achieved R² > 0.91 for both training and testing datasets, demonstrating its ability to accurately predict SLAM-50 performance. These findings provide a practical, data-driven basis for performance-oriented mixture design and optimization, addressing the current scientific gap and offering guidance for engineering practice. Full article

Drug discovery is still an expensive and time-consuming process where finding the right drug associations is important for therapeutic development. In this paper, a new system is proposed for drug design called PharmaGraphFormer (PGF). It consists of five stages: (i) Data acquisition and preprocessing (DAP), (ii) Feature extraction and feature fusion (FEF), (iii) Molecular representation (MR), (iv) Multi-task prediction, and (v) Explainable artificial intelligence (XAI). This study employs a hybrid graph neural network (GNN)-transformer architecture that combines structural and sequence-based representations. Through DAP, several processes are executed, including the imputation or removal of missing values, outlier rejection, and class balancing. Next, through FEF1, features are extracted to represent the input data efficiently. Initially, compound-protein features are generated to document the interactions and relationships between chemical compounds and their corresponding target proteins. Secondly, drug characterizations are computed to encapsulate the physical, chemical, and structural attributes of each drug. After that, MR is performed using a graph-based molecule representation. Then, a novel model integrating GNNs and graph transformers, termed GNN-T, is proposed. Initially, GNNs represent the most promising deep learning models adept at processing non-Euclidean data. The Graph Transformer layer enhances atom representations by consolidating the representations of adjacent atoms through an attention mechanism. Finally, XAI is applied to explain the internal mechanisms of AI systems, rendering them comprehensible and interpretable. Across five independent runs, the proposed model achieved an accuracy of $0.963\pm 0.002$, a precision of $0.971\pm 0.002$, a recall of $0.958\pm 0.003$, an F1-score of $0.964\pm 0.002$, and a ROC-AUC of $0.993\pm 0.001$. These results demonstrate an outstanding performance when compared with all other models and emphasize that the proposed model is reliable in solving the problems of prioritizing compounds in line with the latest developments in AI-powered virtual screening and drug–target interaction modeling. Full article

Pathological gait classification plays an important role in objective motor function assessment, early clinical screening, and rehabilitation evaluation. However, traditional clinical gait assessment methods are highly dependent on expert experience and may fail to detect subtle gait abnormalities. Moreover, existing inertial measurement unit (IMU)-based gait recognition methods often rely on single-sensor configurations or single-scale temporal models, limiting their ability to capture complex pathological gait patterns. In this study, a convolutional neural network–bidirectional long short-term memory–temporal convolutional network (CNN-BiLSTM-TCN) multi-branch feature fusion framework was proposed for pathological gait classification using a publicly available clinical multi-inertial measurement unit dataset containing 260 subjects. The proposed model employs three parallel branches to extract local instantaneous motion variations, continuous temporal dynamics, and relatively broader temporal dependencies within the 2 s input window, respectively, followed by feature-level fusion and end-to-end joint optimization. Experimental results show that the proposed model achieves a test accuracy of 0.9818 and an F1-score of 0.9700, outperforming conventional machine learning methods, single-branch models, voting-based fusion methods, and other temporal models, including Support Vector Machine (SVM), Temporal Convolutional Network (TCN), and Convolutional Neural Network-long short-term memory (CNN-LSTM). Five repeated experiments with stratified random splits demonstrate minimal performance variation, indicating good robustness and stability. The proposed framework provides a potential approach for pathological gait screening and quantitative rehabilitation assessment. Full article

Antenna selection is investigated as a means of reducing radio-frequency (RF) chain power in massive multiple-input multiple-output (MIMO) base stations under realistic channel state information (CSI) impairments. The study is motivated by the mismatch between conventional selection objectives and multi-cell operation with estimation error, pilot contamination, spatial correlation and inter-cell interference. APCS-Boost-R is introduced as the primary contribution. An interference-whitened D-optimal seed is combined with projected rank-one exchanges and a calibrated surrogate that incorporates a user-side interference-plus-noise report and a closed-form estimation-error correction. APCS-Boost-RG is retained as an optional graph neural network (GNN) refinement in which residual exchanges are ranked after the algebraic solution has been formed, while feasibility and non-degradation of the calibrated surrogate are verified deterministically. In a three-cell urban macro configuration derived from Third Generation Partnership Project (3GPP) TR 38.901 with 64 antennas, 16 active RF chains and eight users per cell, APCS-Boost-R achieves 19.364 bit/s/Hz over 200 paired realizations. Improvements of 2.58 percent over APCS-Boost, 6.76 percent over greedy search and 10.16 percent over a genetic algorithm are obtained. APCS-Boost-RG adds 0.019 bit/s/Hz but is treated as an optional refinement because it requires a second-stage neighborhood evaluation and offline model maintenance. In the archived common timing record, APCS-Boost-R requires 20.376 ms per three-cell realization, compared with 12.728 ms for APCS-Boost, 57.775 ms for norm-initialized greedy search and 41.302 ms for the genetic algorithm, while APCS-Boost-RG requires 24.0 ms versus 20.4 ms for APCS-Boost-R in the separate archived learned-stage record. Separate reconstructions on the documented reproducibility host require $55.3\pm 14.5$ ms for APCS-Boost-R and $592.2\pm 181.9$ ms for a complete APCS-Boost-RG rebuild. Additional paired examinations confirm robustness across stronger search budgets, report imperfections, regularized precoding, coordination, near-field sensitivity, hardware perturbations, and configurations ranging from 32 to 128 antennas and one to seven cells. Full article

Computational fluid dynamics methods range from Reynolds-Averaged Navier–Stokes (RANS) simulations to Direct Numerical Simulations (DNSs). RANS offers low computational cost at the expense of accuracy, while DNS provides high accuracy but at a prohibitive cost. The aim of this study is to evaluate whether multi-fidelity neural networks can learn a corrective mapping from RANS to DNS for a small canonical dataset and to determine how training-set composition and model architecture influence generalisation across geometries. In this study, multi-fidelity neural networks for data fusion between low-fidelity RANS and high-fidelity DNS data were applied to turbulent flow (Re = 5600) over parameterised periodic hills, defined by a geometry parameter characterising the steepness ratio. The inputs to the models were the coordinates and the corresponding RANS velocity components, and the outputs were the DNS velocity components, with data from both fidelities mapped onto the same mesh. Both a single-branch and a two-branch architecture were considered. Generalisability was assessed within a small dataset of five periodic hills defined by different values of the geometry parameter $\alpha$ (0.5, 0.8, 1.0, 1.2, 1.5). Both model architectures were trained on data from different combinations of the geometry parameter to evaluate interpolation and extrapolation capabilities. Both networks successfully corrected RANS flow fields for unseen geometries in interpolation regimes. When interpolating, the single-branch architecture achieved more than a 69% reduction in error, while the two-branch architecture achieved more than a 60% reduction, with both improving key flow features such as recirculation zones and jet structures. A key finding is that the single-branch architecture consistently outperformed the two-branch formulation, particularly in low-data regimes. The results show that multi-fidelity neural networks can improve RANS predictions using small datasets and simple inputs, provided that the training set spans the relevant geometric space. As the model does not require the geometry parameter as an explicit input, it is applicable to geometries lacking straightforward parameterisation. The demonstrated advantage of the single-branch architecture highlights the importance of architectural simplicity when training data is limited. Full article

In the TJ-II stellarator, a finite net toroidal plasma current can modify the rotational transform profile and shift the radial position of low-order rational surfaces. Because these surfaces can influence edge gradients, turbulence and confinement in low-shear configurations, estimating the current waveform is relevant for the preparation and interpretation of scenarios programmed with the ohmic heating (OH) coil. This work compares two data-driven estimators using 87 selected OH-programmed discharges from the fixed magnetic configuration 100_44_64, divided into 50 training, 20 validation and 17 test cases. The first estimator uses MultiGene Genetic Programming (MGGP) to obtain explicit equations. The second is a nonlinear autoregressive recurrent neural network with exogenous inputs (NARX-RNN), implemented with long short-term memory (LSTM) branches. The symbolic equations and their settings were selected using the training and validation discharges, while the test set was used only for final evaluation. On the original 17-discharge test partition, the general OH-normalized equation achieved an overall root-mean-square error (RMSE) of 0.4898 kA, whereas the compact equation achieved 0.5700 kA with substantially lower expression complexity. After routine signal quality control, two test records were excluded, and the final comparison used 15 discharges. On this set, the mean discharge-wise RMSE was 0.350 kA for the NARX-RNN, 0.515 kA for the general symbolic equation and 0.576 kA for the compact equation. The NARX-RNN also gave a median RMSE of 0.230 kA. Whether this current error level is sufficient for positioning a particular rational surface depends on the local rotational transform response and magnetic shear, so no universal current error threshold for edge control is assigned here. The symbolic models were less accurate on average but provided explicit relations involving OH amplitude, previous current information, electron cyclotron resonance heating, fuelling and wall-conditioning variables. Full article

The BeiDou-3 Satellite Navigation System (BDS-3) serves as a critical national space-based information infrastructure. However, the signals received at user terminals are characterized by power levels that lie significantly below the noise floor, rendering them highly vulnerable to both intentional suppressive jamming and unintentional interference in complex electromagnetic environments. Traditional anti-jamming techniques predominantly rely on blind suppression strategies. While effective against simple jamming scenarios, these methods suffer from inherent limitations, including low resource utilization and poor adaptability to complex, time-varying interference. To achieve precise suppression and real-time detection of interference, it is necessary to simultaneously perform classification and parameter estimation of the jamming signal. To this end, this paper proposes a multi-task dual-branch convolutional neural network (CNN) learning framework that encompasses four typical types of suppressive jamming: continuous wave interference (CWI), wideband noise interference (WNI), pulse interference (PI), and chirp interference (CI). Taking time-frequency spectrograms of signals as input, the framework achieves accurate identification of interference types and simultaneous estimation of key physical parameters. Simulation and experimental results demonstrate that the proposed method maintains stability and feasibility under low-INR conditions. Full article

Accurate fault location in low-observability active distribution networks is hindered by uncertain inter-node relationships, underutilized early transients, and insufficient global context. To address these challenges, this paper proposes an adaptive Mamba-driven spatiotemporal graph neural network (AM-STGNN). It provides a unified task-driven spatiotemporal representation framework that progressively integrates fault-propagation modeling, global dependency modeling, transient-state learning, and topology-aware discriminative enhancement. Specifically, a prior-guided adaptive implicit topology is first learned to characterize task-dependent electrical coupling relationships among sparse observation nodes. Based on the resulting topology, topology-conditioned multi-order feature propagation and a dual-axis linear-attention module based on the spatiotemporal graph transformer (STGformer) are employed to capture local and global spatiotemporal dependencies. The resulting global spatiotemporal representation is subsequently processed by a Mamba selective state-space encoder to model input-dependent temporal evolution and emphasize informative fault transients. Finally, element-wise gated fusion, topology-aware differential output, and a margin constraint are employed to integrate the STGformer and Mamba representations and enhance the separability of adjacent faulted line sections with similar response characteristics. Extensive experiments demonstrate the effectiveness of AM-STGNN, while additional evaluations confirm its applicability to larger-scale networks, strongly phase-unbalanced conditions, and field-measured operating backgrounds. Robustness tests under individual and multi-level joint disturbances further demonstrate the practical relevance of the proposed architecture. Compared with the baseline models, AM-STGNN achieves consistent improvements in the macro-averaged F1 score (Macro-F1), exact accuracy, and one-hop accuracy under the clean IEEE 123-node condition. More importantly, it maintains clear performance advantages under identical mild, moderate, and severe joint disturbances, demonstrating improved robustness and practical relevance under simulated non-ideal operating conditions. Full article

Background: Elite soccer performance emerges from the interplay of cognitive, emotional, psychophysiological, and tactical processes that operate in real time during matches. Advances in wearable sensors and artificial intelligence (AI) now allow continuous monitoring of physiological and psychological states. They also allow modeling of how these states relate to tactical and physical performance. Existing reviews have examined machine learning in soccer, heart rate variability (HRV) monitoring, and psychological determinants of performance separately. No scoping review has mapped the intersection of AI analytics, wearable psychophysiological monitoring, and psychological performance constructs as one integrated decision-support framework in elite soccer. Aim: The aim of this study was to map the available evidence on the integration of AI and machine learning with psychophysiological monitoring for performance modeling in elite soccer, to identify the psychological constructs already used as model inputs, to describe the wearable technologies and AI methods applied, and to set out the translational challenges and evidence gaps that need priority attention. Methods: The review followed the PRISMA extension for Scoping Reviews (PRISMA-ScR) and the updated Joanna Briggs Institute (JBI) methodology. The protocol was registered on the Open Science Framework (OSF). Six databases (PubMed/MEDLINE, Scopus, Web of Science, SPORTDiscus, IEEE Xplore, and PsycINFO) were searched from January 2000 to March 2026 using the Population–Concept–Context (PCC) framework. Two reviewers independently screened titles, abstracts, and full texts (Cohen’s kappa = 0.82). Results: Thirty-six sources met the eligibility criteria after screening of 3104 records. AI and machine learning have been applied widely to predict physical and tactical performance in soccer, yet they rarely include psychological constructs. Reported models (decision trees, gradient boosting, and artificial neural networks) reach high accuracy for physical outcomes in internal validation, for example, above 66% for injury risk. Multi-modal models that add physiological and psychological inputs report stronger prediction. These figures come mostly from internal validation, and external validation and overfitting controls are seldom reported, so they should be read as optimistic upper bounds. Psychological and psychophysiological inputs remain under-represented. Explainable AI (XAI) methods, in particular Shapley Addictive exPlanations (SHAP) values, are appearing, but validation with domain experts is scarce. HRV has been reviewed as a psychophysiological marker in soccer, yet its use within AI decision-support tools for real-time psychological readiness has not been mapped. Three translational challenges stand out: the ecological validity gap between laboratory cognitive tests and match-embedded psychophysiology; the interpretability problem of opaque AI in high-stakes decisions; and the data fragmentation problem created by disconnected physical, tactical, and psychological data streams. Conclusions: Integrating AI with wearable psychophysiological monitoring offers a credible route toward integrated performance modeling in elite soccer. Closing this gap calls for multi-modal frameworks that combine psychological constructs, physiological markers, and tactical data within explainable AI. Research priorities include ecologically valid psychophysiological assessment protocols, position-specific psychological profiling, and practitioner-validated tools that turn AI outputs into usable coaching recommendations. Full article

Globally, respiratory diseases such as asthma, chronic obstructive pulmonary disease (COPD) and pneumonia affect populations significantly, requiring early and accurate diagnosis for effective clinical management. Manual auscultation and expert interpretation are the common shortcomings in conventional diagnostic approaches, as they lead to time-consuming and inconsistent analysis. To address these limitations, an artificial intelligence-driven framework for respiratory disease classification using multi-spectral feature extraction and deep learning architectures is proposed to classify four different respiratory conditions: Asthma, COPD, Pneumonia and Healthy. The dataset is collected from Kaggle’s respiratory sound database and the COUGHVID V3 database, which together contain 322 Asthma signals, 746 COPD signals, 323 Pneumonia signals and 174 Healthy signals. Subsequently, the features are extracted using four different feature extraction techniques—Constant Q Transform (CQT), a Gammatone spectrogram, Mel-Frequency Cepstral Coefficients (MFCC) and Perceptual Linear Prediction (PLP)—and these extracted spectral representations are provided as inputs to various deep learning models such as a Deep Convolutional Neural Network (Deep CNN), a Temporal Attention Network (TAN) and an Autoencoder for automated feature learning and disease classification. The proposed framework is evaluated using several performance metrics, and the experimental results clearly indicate that the performance of the proposed classification framework strongly depends on the selection of spectral feature extraction techniques and deep learning models. Among all the evaluated combinations, it is evident that the Autoencoder model integrated with CQT features exhibited the best classification performance, with an accuracy of 98.72%, precision of 98.74%, recall of 98.72%, Matthews correlation coefficient (MCC) of 98.11%, Cohen’s kappa value of 98.10% and the least log loss of 0.025. The proposed artificial intelligence (AI)-enabled respiratory disease classification framework has demonstrated the ability to produce a reliable computer-aided diagnostic system which is suitable for smart healthcare applications and automated pulmonary disease screening. Full article

Connected vehicles and vehicle-to-everything (V2X) communication create new opportunities for estimating urban origin–destination (OD) demand from continuously collected mobility data. However, in realistic deployment conditions, only a fraction of vehicles may be connected, making OD reconstruction a highly underdetermined problem under low penetration rates. This paper proposes a supervised deep-learning framework that reconstructs full OD matrices from synthetic connected-vehicle data in a simulated Manhattan network from New York City. Vehicle movement information is aggregated into intra-zonal and adjacent-zone traffic counts using K-means traffic analysis zones. These partial connected-vehicle observations, represented by the zonal movement matrix, outgoing and incoming zonal-movement summaries, diagonal movement counts, and penetration-rate features, form a compact input to a conditional Multi-Layer Perceptron (MLP) that predicts the complete OD matrix of all vehicles. The training objective separates OD spatial shape from total traffic volume and adds losses on marginals, diagonal elements, and log-space reconstruction to embed basic flow-conservation properties. A single conditional MLP is trained across multiple connected-vehicle penetration-rate scenarios by appending the penetration rate $\rho$ and $\log \left(\rho \right)$ to the input representation. The model is evaluated over ten random connected-vehicle sampling seeds. Results show that the proposed estimator remains stable down to 20% penetration, with test sMAPE increasing only from $20.70\pm 0.00\%$ at full penetration to $21.71\pm 0.38\%$ at 20% penetration. Marginal and total-flow errors increase more gradually as penetration decreases, while clear degradation appears below approximately 2–1% penetration. Baseline and ablation comparisons further show that penetration-rate conditioning and the conservation-aware loss are essential for improving OD reconstruction and total-flow consistency. Within the evaluated simulated Manhattan scenarios, these findings suggest the potential of conditional neural estimators for OD reconstruction under limited connected-vehicle penetration. Validation across longer periods, additional demand regimes, and real-world data is required before the results can be generalized to broader urban traffic conditions. Full article

Against the backdrop of global climate change and energy structure transition, district energy systems have garnered significant attention for their efficiency and sustainability. Accurate load forecasting is crucial for enhancing the operational efficiency of district cooling systems. However, as typical dynamic time-varying systems, building complexes within district energy stations exhibit load characteristics influenced by multi-scale features. Furthermore, traditional load forecasting models employ single-scale analysis without specifically modeling these multi-scale characteristics, resulting in insufficient generalization capabilities of data-driven models under dynamic, time-varying scenarios. This paper proposes a multi-step forecasting model structure based on baseline-morphology decomposition to address the coupling of multi-scale characteristics. By decomposing load into baseline and morphological components, separate prediction models—a backpropagation neural network (BP) and a K-means clustering-decision tree (DT) classification prediction model—are constructed, overcoming the challenge of capturing multi-scale features in traditional methods. The results show that, during the four-month test period from August to November 2024, the proposed model achieves MAPE values ranging from 8.91% to 12.57% under the peak and transitional cooling conditions represented in the dataset. Compared to direct structure, recursive structure, and multi-input multi-output (MIMO) structure, it reduces errors by 1.49% to 19.62% while achieving remarkable advantages in training efficiency. Full article

This study presents a methodology for predicting the geometric features of single tracks of 316L stainless steel produced by Powder-Bed Arc Additive Manufacturing (PBAAM) from four independent process parameters using a multi-modal Physics-Informed Neural Network (PINN). PBAAM shares the same powder-deposition and layering scheme as Laser Powder Bed Fusion (LPBF) but uses a low-current micro-TIG arc rather than a laser as the heat source. A multi-task PINN architecture was developed that simultaneously predicts five geometric features measured from two imaging modalities (top-view and side-view arc), namely the arc core diameter (D_q), the arc cone angle (α_c), the heat-affected zone width (w_HAZ), the track core width (d_core) and the areal equivalent track width (w_iz), from four input parameters (arc current, traverse speed, work angle and working distance). The model was assessed on a full-factorial training matrix of 36 experiments and on four pure speed extrapolation experiments above the training range. A composite quality score filter classified 23 of the training experiments as stable and 13 as unstable. On the pure validation set, the mean absolute percentage error (MAPE) was 4.25% (95% confidence interval 0.91–8.49) for the arc core diameter, 6.29% (5.07–7.59) for the arc cone angle, 8.06% (6.08–9.82) for the heat-affected zone width, and 17.02% (10.77–21.62) for the track core width. Classical regression baselines attain comparable aggregate errors on this narrowly distributed validation set; the distinguishing property of the proposed model is the joint, physically ordered prediction of all five outputs. The Ayrton voltage sub-module of the model converged to U(I) = 11.33 + 97.13/I without any direct voltage measurement, purely through the physics loss term; this function is consistent with the order of magnitude expected from the physics of low-current TIG arcs. The results indicate that physics-informed learning can be applied to the PBAAM process parameter space under small-sample conditions. This capability is demonstrated for 316L stainless steel, for the micro-TIG electrode configuration and the process window investigated here, for single tracks rather than multi-layer builds, and against a validation set of four experiments varying in a single direction. Full article

Three-dimensional (3D) imaging systems, including depth cameras, LiDAR sensors, and multi-view scanning pipelines, often produce point clouds with noisy normals, outliers, sparse sampling, and non-uniform density, which can degrade downstream mesh reconstruction. Poisson surface reconstruction is lightweight and training-free, but its global implicit formulation is sensitive to unreliably oriented samples and fixed density-trimming thresholds. This paper presents RG-PSR, a reliability-guided enhancement framework for Poisson-family surface reconstruction from degraded 3D-imaging point clouds. RG-PSR estimates a deterministic per-point reliability score from local density regularity, spacing variation, and normal consistency, and propagates this score through conservative point filtering, reliability-guided normal refinement, adaptive density-reliability trimming, and structure-aware postprocessing. The main pipeline requires no manual labels, neural network training, or ground-truth meshes at inference time. Experiments on three groups of object meshes under five deterministic degradation types show that RG-PSR improves Poisson-family reconstruction under degraded inputs. Compared with fixed density-trimmed Poisson reconstruction, RG-PSR reduces the overall Chamfer-$L_1$ from 0.0218 to 0.0172, improves $F_{0.01}$ from 0.6618 to 0.6836, and reduces ${\mathrm{Artifact}}_{0.02}$ from 0.3090 to 0.2632. In the broader classical comparison, local triangulation methods achieve stronger point-wise accuracy, while RG-PSR yields the fewest connected components and the highest largest-component ratio. These results position RG-PSR as a practical reliability layer for coherent Poisson-family reconstruction rather than a universal replacement for all surface-reconstruction methods. Full article

Occupancy detection is fundamental to the operational intelligence of smart buildings, driving critical functions in energy management, HVAC automation, and physical security. While modern Deep Learning (DL) models have achieved high accuracy in parsing complex environmental sensor data, they remain highly vulnerable to adversarial examples, imperceptibly perturbed inputs designed to deceive neural networks. These vulnerabilities pose severe real-world risks, ranging from energy sabotage, in which systems heat empty rooms, to critical security breaches in which intruders go undetected. To address this security gap, we propose ADS-Guard, a novel Adversarial Detection and Sanitization (ADS) framework rooted in sequence-to-sequence autoencoder purification. Unlike standard denoising techniques, ADS-Guard incorporates a latent consistency regularization mechanism that encourages alignment between clean and adversarial representations in the latent feature space. We evaluated ADS-Guard using a comprehensive experimental pipeline comprising five distinct DL architectures (LSTM, GRU, 1D-CNN, MLP, and Transformer) across three diverse datasets: (1) The UCI Occupancy dataset (20,699 samples) for standard binary detection; (2) Building59 dataset (7200 samples) for three-class occupancy-level classification (Low, Medium, High); and (3) Room Occupancy dataset (10,129 samples), representing a highly imbalanced binary occupancy-detection task. We evaluate ADS-Guard against both Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD) attacks across diverse occupancy datasets and model architectures. We further assess the framework under adaptive white-box attacks and compare its performance with FGSM-based and PGD-based adversarial training baselines. Our results demonstrate that adversarial attacks can substantially degrade occupancy-detection performance across datasets and model architectures. ADS-Guard consistently improves robustness relative to undefended models against both FGSM and PGD attacks, recovering a substantial portion of the lost performance in binary occupancy tasks and providing meaningful gains in the more challenging multi-class setting. Furthermore, ADS-Guard remains effective under stronger adaptive threat models while providing a practical retraining-free defense that can be integrated with existing occupancy-detection systems without modifying downstream classifiers. Full article