Search Results (688)

To address the coexistence of random measurement delays and intermittent data loss in multi-UAV cooperative navigation under weak communication scenarios, an adaptive cooperative localization method based on online diagnosis is proposed. A timestamp-driven diagnosis mechanism first identifies the physical reachability and temporal validity of cooperative measurements and classifies the channel state as normal, delayed, or lost. The estimator then adaptively switches between timestamp-diagnosed multi-step augmented filtering for delayed packets and Gray Wolf Optimizer (GWO)-optimized Gated Recurrent Unit (GRU) virtual measurement reconstruction for complete dropouts. In a four-UAV hybrid-anomaly simulation with continuous random delays and a 10 s communication blackout, the proposed method reduces the mean east and north RMSE by 69.1% and 75.2%, respectively, and lowers the mean horizontal-channel RMSE from 2.200 m to 0.608 m. Ablation experiments isolate the contribution of each module (diagnosis and multi-step delayed filtering and GWO-GRU virtual reconstruction), and 20 paired Monte Carlo runs with paired significance tests confirm the improvement. Relative to the strongest delay-aware baseline, the full-trajectory mean accuracy is comparable, and the main advantage of the proposed method is its robustness during complete communication outages and fast post-outage recovery. The simulation scenarios are driven by flight data collected with the DJI Matrice 350 RTK platform; the communication anomalies are included in the simulation, and all evaluations are performed offline. These results indicate that the proposed diagnosis-driven scheduling strategy can improve cooperative localization robustness when delay and data loss occur simultaneously. Full article

Synthetic aperture radar (SAR) object detection remains challenging, as target evidence is often sparse, discontinuous, and heavily influenced by speckle noise, sidelobes, shadows, and clutter-like background scattering. Existing dense detectors usually process SAR images as ordinary grayscale images, which causes feature modulation to rely on unrestricted saliency responses that may simultaneously enhance true targets and bright background scatterers. To address this issue, we propose a SAR-inspired Scattering-Center Field (SCF) modulation framework for multi-scale dense object detection. The SCF module is designed as a tensor-preserving feature adapter that decomposes feature modulation into three internal latent fields: a center-like evidence field, a response-amplitude field, and an orientation-anisotropy field. These fields are inferred from intermediate features through lightweight local, strip-convolution, and contextual branches, and then they are subsequently fused into an identity-initialized residual spatial gate. The proposed module requires no scattering-center annotations, segmentation masks, auxiliary field supervision, phase history, polarimetric data, or additional post-processing. Consequently, it can be seamlessly integrated into standard dense detection pipelines without altering labels or prediction heads. Experiments are conducted on four publicly available SAR detection datasets, namely SSDD, SAR-AIRcraft-1.0, SAR-Ship, and MSAR-1.0. The proposed detector achieves mAP50/mAP50-95 scores of 0.954/0.682 on SSDD, 0.930/0.661 on SAR-AIRcraft-1.0, 0.971/0.676 on SAR-Ship, and 0.688/0.482 on MSAR-1.0. These results indicate that SCF provides a lightweight and detector-compatible modulation mechanism for improving SAR object detection under sparse target responses and cluttered imaging conditions. Full article

Accurate vibration prediction is critical for condition monitoring and predictive maintenance of hydropower units, yet it remains challenging due to strong noise interference, severe non-stationarity, and multi-scale coupling characteristics of raw vibration signals. This paper proposes a novel hybrid prediction framework that integrates cooperative denoising, multi-scale signal decomposition, and deep learning-based sequential modeling to achieve high-precision long-term vibration forecasting. First, a two-stage cooperative denoising stategy combining wavelet threshold denoising (WTD) and singular spectrum analysis (SSA) is designed to suppress high-frequency noise while effectively preserving the global trend and critical transient features. Then the denoised signal is decomposed into a set of physically interpretable intrinsic mode functions (IMFs) and a residual component via complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), which alleviates mode mixing and improves decomposition completeness. Subsequently, each IMF component is independently predicted using a gated recurrent unit (GRU) network optimized by the Adam algorithm with adaptive learning rate decay, enabling efficient capture of nonlinear temporal dependencies. The proposed framework is validated using 3.5-year real-world vibration data from an lower guide bearing of a pumped-storage hydropower unit. Experimental results demonstrate that the model achieves MAE = 0.3831, RMSE = 0.6964, MAPE = 0.2832%, and $R^2=0.9911$, compared to 0.6666 for the conventional CEEMDAN-GRU model, a 32.45 percentage point increase and a 54% reduction in unexplained variance. Ablation studies and comparative analyses verify the superiority and statistical significance of the cooperative denoising mechanism and the overall hybrid architecture. This work provides a reliable, interpretable, and deployable tool for the condition monitoring and predictive maintenance of hydropower units, supporting proactive operation and reducing unplanned downtime in clean energy systems. Full article

Rapid cellular assessment of cell-cycle perturbation can complement conventional developmental toxicology in zebrafish, but rapid DNA-content workflows developed for mammalian tissues require adaptation for whole embryos. Here, we applied an embryo-adapted version of rapid DNA-content flow cytometry analysis within a 96-h cisplatin exposure model in zebrafish (Danio rerio) and integrated this cellular readout with morphometric, cardiac, and neurobehavioral analysis. Dechorionated embryos were exposed to 10–400 μM cisplatin and evaluated for body length, eye surface area, heart rate, locomotor activity, thigmotaxis, touch-evoked and vibrational startle responses, and whole-embryo cell-cycle distribution. The adapted workflow combines immediate mechanical dissociation, filtration, brief propidium iodide/RNase staining, singlet gating, and direct flow cytometric acquisition. Importantly, each flow cytometry sample consisted of a single embryo, enabling DNA-content analysis at the individual-embryo level without pooling. Cisplatin produced time- and concentration-dependent growth effects and high-dose behavioral alterations, whereas heart rate was comparatively preserved. Flow cytometry detected the clearest cell-cycle redistribution at 400 μM, with a significant reduction in G0/G1 and an increase in G2/M. These findings demonstrate the practical applicability of a combined rapid embryo-adapted DNA-content workflow as a complementary cellular endpoint in a zebrafish developmental-toxicity platform and support further methodological application across experimental settings. Full article

Under the vehicle as a service (VaaS) paradigm, intelligent connected vehicles continuously generate large-scale, computation-intensive perception data processing tasks. However, limited onboard computing resources and power supply prevent vehicles from handling these tasks efficiently. Vehicular edge computing (VEC) extends available computing resources through vehicle–infrastructure collaboration. Nevertheless, continuous vehicle mobility causes intermittent communication links between vehicles and roadside units, posing new challenges for VEC task scheduling. Therefore, this paper proposes a reinforcement learning-based VEC task scheduling approach that integrates a fuzzy reward mechanism with multi-agent proximal policy optimization (FRMPPO). First, a system architecture is developed by integrating the directed acyclic graph task model, dynamic communication model, and computation model. The scheduling problem is formulated as a partially observable Markov decision process, with the objective of minimizing task completion latency and vehicle energy consumption. Second, a fuzzy reward mechanism is designed to guide the training of multi-agent proximal policy optimization. It takes edge node load pressure and communication state as inputs to adaptively combine local immediate rewards and the global reward, eventually guiding the agents toward a globally optimized cooperative policy. Finally, real-time scheduling decisions under communication intermittency are enabled through a centralized training and decentralized execution framework and gated recurrent unit-based state encoding. Simulation results demonstrate that FRMPPO effectively solves the VEC task scheduling problem, achieving significantly superior performance over existing algorithms in terms of both task completion latency and vehicle energy consumption. The proposed method thereby satisfies the real-time processing demands of perception tasks in VaaS scenarios. Full article

Human pose estimation in factory surveillance is challenged by scale variation, limb occlusion, complex backgrounds, and spatial detail loss during upsampling. This paper proposes MFA-Pose, an improved YOLO11s-Pose framework that enhances contextual representation and cross-scale feature reconstruction. The Multi-Scale Context Fusion (MSCF) module preserves directional positional information and models multi-range cross-channel dependencies using parallel one-dimensional convolutions with kernel sizes of 3, 5, and 7. The complete C2MSCF module contains approximately 0.790 M learnable parameters, compared with approximately 0.991 M parameters in the original C2PSA module. The Multi-Receptive-Field Adaptive Gated Upsampling (MRAG) module predicts sampling offsets through standard and dilated convolution branches, suppresses unreliable offsets using gating mechanisms, and combines dynamic resampling with a stable bilinear interpolation reference. The two MRAG modules used in MFA-Pose contain approximately 0.457 M and 0.130 M learnable parameters, respectively, corresponding to approximately 0.587 M parameters in total. On COCO 2017, MFA-Pose achieves 89.0% ${\mathrm{AP}}_{50}^{\mathrm{pose}}$ and 61.7% ${\mathrm{AP}}_{50:95}^{\mathrm{pose}}$, outperforming YOLO11s-Pose by 2.7 and 1.6 percentage points, respectively, with only 0.40 M additional parameters. Quantitative evaluation on the annotated industrial surveillance test set further shows that MFA-Pose achieves 87.2% Precision, 77.0% Recall, 81.8% F1-score, 88.7% ${\mathrm{AP}}_{50}^{\mathrm{pose}}$, and 61.1% ${\mathrm{AP}}_{50:95}^{\mathrm{pose}}$, outperforming the YOLO11s-Pose baseline across all evaluated accuracy metrics. Qualitative comparisons further show more coherent pose predictions under self-occlusion, non-standard working postures, and cluttered equipment backgrounds. Overall, MFA-Pose provides a favorable balance between pose estimation accuracy and model complexity for industrial surveillance applications. Full article

Accurate semantic segmentation of large-scale outdoor LiDAR point clouds remains a challenging endeavor, primarily due to ambiguous class transitions at object interfaces, non-uniform sampling density across the surveyed area, and shared geometric signatures among distinct object categories. This paper proposes GFE-Net (Geometry-Enhanced Feature Extraction Network), a hierarchical encoder–decoder architecture that systematically improves per-point feature characterization through three complementary design contributions: First, to mitigate the shortcomings of conventional fixed-neighborhood queries in regions of variable point density, a Structure-Guided Neighborhood Adaptation (SGNA) module is devised. At its core lies a morphology-driven contextual gating (MCG) unit that synthesizes neighbor-wise calibration weights from hierarchical shape descriptors fused with elevation difference statistics, allowing the network to preferentially amplify morphologically congruent neighbors while dampening spurious or cross-boundary contributions. Second, to strengthen semantic discrimination beyond what spatial locality alone affords, a Local–Global Interactive Enhancement (LGIE) module is presented. The LGIE module simultaneously distills precise local structure through Euclidean-space neighborhood graphs and captures scene-wide co-activation patterns via compact bilinear factorization of the latent feature space, merging both streams through a residual refinement mechanism that markedly improves inter-class separability. Third, to enforce label consistency at object interfaces without relying on post-processing heuristics, a Neighborhood Prediction Consistency (NPC) loss is introduced. Built upon a Gaussian distance-decay weighting kernel, the NPC loss assigns progressively stronger penalties to label mismatches between a query point and its geometrically proximate neighbors, thereby promoting spatially coherent predictions and attenuating boundary noise. GFE-Net is rigorously benchmarked on two widely adopted large-scale datasets—S3DIS and SensatUrban—yielding OA/mIoU of 89.6%/73.1% and 93.3%/61.1%, respectively. These results demonstrate competitive performance under the reported protocols. Detailed ablation studies and computational profiling further substantiate the efficacy of each individual component. Full article

Earthquake rescue priorities are often determined from incomplete evidence that is partly shared across information channels. Under such conditions, an apparently plausible mean score may conceal substantial directional conflict. This study investigates whether that conflict can be carried forward into the action boundaries themselves. We develop the Conflict-Driven Action Boundary Generation Model (CABGM), which links composite-conflict diagnosis to a constructed association proxy, normalized action propensities, and feedback-sensitive three-way decision boundaries. The structural coefficients are screened against prespecified sign, normalization, boundedness, and boundary-feasibility constraints and are treated as theory-constrained operating settings rather than estimates fitted to rescue outcomes. For the retrospective empirical application, we reconstructed ten named settlement-scale units from public records using a prespecified five-level documentary coding protocol, a fixed source hierarchy, and a fixed evidence cutoff before applying the frozen model specification. We also implemented a scalar-score adaptation of adaptive-consensus logic and a reliability-informed Dirichlet comparator for channel-weight uncertainty. Both comparators are transparent adaptations to the present $10\times 4$ score matrix rather than exact reproductions of the original linguistic or event-network models. In the public-record audit, all five methods yielded identical rank-concordance statistics: a Spearman correlation of 0.912, Kendall’s tau-b of 0.839, a p-value of 0.0016 from an exact two-sided permutation test, and 100% top-four overlap. The fixed-threshold conflict gate, opinion-distance update, adaptive-consensus adaptation, and CABGM each produced an acceptance/deferment/rejection split of 6/4/0, whereas the Bayesian weight-uncertainty comparator produced a 4/6/0 split. CABGM offers a new methodological option for emergency group decision-making by integrating composite-conflict diagnosis, a constructed association proxy, normalized action propensities, and feedback-sensitive three-way decision boundaries within a unified framework. Compared with fixed-boundary and consensus-contraction approaches, the model makes the transmission of diagnosed conflict into boundary adjustment explicit and traceable, while distinguishing boundary adaptation from score smoothing. It therefore provides a transparent mechanism for preserving unresolved or threshold-adjacent alternatives for further verification before definitive action is taken. Full article

Wind direction time series exhibit angular periodic discontinuity, multi-scale non-stationary fluctuations and abrupt wind shifts, which hinder the precision of short-term forecasting for wind turbine yaw control. In this paper, a dual-branch forecasting framework based on Variational Mode Decomposition (VMD) and Transformer is developed to address the above drawbacks, with a hysteresis gating and zoning residual compensation module embedded for targeted error correction. First, sine–cosine encoding is adopted to eliminate the numerical discontinuity between 0° and 360° for wind direction angular data, and valid meteorological input features are screened to discard redundant covariates. Second, the sine–cosine-encoded wind direction sequence is decomposed into multiple band-limited intrinsic mode functions (IMFs) via VMD, extracting frequency-specific features that reduce non-stationarity and facilitate subsequent Transformer modeling. The standard Transformer encoder serves as the normal branch to capture long-range temporal dependencies across the whole time series, while a lightweight multilayer perceptron (MLP) constitutes the compensation branch to learn prediction deviations between baseline predictions and ground-truth values. The hysteresis gating unit activates residual compensation based on historical prediction errors and angular variation, without requiring access to the current ground-truth value, and compensation intensity is adaptively adjusted via the zoning strategy; relevant coefficients are optimized by random search. Verified on a real wind farm dataset consisting of 10,421 15 min sampling points, the proposed model achieves the lowest MAE of 9.64° among six benchmark models. For the improved genuine mutation samples (angle change > 70°), the model achieves a mean improvement of 6.02°. Ablation experiments verify that VMD preprocessing, the MLP compensation branch, and the hysteresis gating mechanism play indispensable roles in forecasting performance. The proposed framework can support accurate yaw control of wind turbines, and the decomposition–compensation workflow can also be generalized to other periodic non-stationary forecasting tasks. Full article

The effective resolution of non-deterministic polynomial time hard (NP-hard) combinatorial optimization problems requires a delicate balance between global exploration and local exploitation. While Quantum-Inspired Algorithms (QIAs) leverage principles of superposition to explore vast search spaces, they often lack the fine-grained exploitation capabilities of classical heuristics. To address this limitation, we propose the Quantum-Memetic Hybrid Algorithm (QMHA), a component-based framework that synergistically integrates qubit-based global search with adaptive classical refinement. The QMHA architecture explicitly coordinates five distinct algorithmic components: (1) quantum rotation gates for exploration, (2) a problem-aware memetic operator for immediate solution refinement, (3) an adaptive learning rate schedule, (4) periodic local search, and (5) a stagnation-based population reset for diversity management. We rigorously evaluate the framework against nine established metaheuristics, including Genetic Algorithms (GA), Differential Evolution (DE), Particle Swarm Optimization (PSO), Simulated Annealing (SA), Ant Colony Optimization (ACO), MAX-MIN Ant System (MMAS), Memetic Algorithms (MA), Quantum Evolutionary Algorithm (QEA), and Harmony Search (HS), across a comprehensive benchmark suite comprising six NP-hard problem families: constrained combinatorial (Knapsack), graph-based (Max-Cut), permutation-based (TSP), constraint satisfaction (Graph Coloring), bin optimization (Bin Packing), and scheduling (Flow Shop Scheduling), as well as real-world machine learning (Feature Selection) problems and the continuous Congress on Evolutionary Computation (CEC) 2022 benchmark. Statistical analysis using Friedman tests and Nemenyi post hoc comparisons confirms that QMHA achieves a statistically significant performance advantage ($p<0.004$) and superior average rank (1.5) compared to component baselines and state-of-the-art competitors. Comprehensive analyses include computational complexity profiling, parameter sensitivity mapping, scalability testing up to $D=2000$, noise robustness evaluation, variable correlation degradation analysis, a six-component ablation study, exploration–exploitation dynamics tracking, integration mechanism comparison across five architectures, and a multi-objective extension feasibility study. The proposed framework offers a robust, verified approach to hybrid optimization without relying on biological metaphors. Full article

In early-stage projects, incomplete geological data can affect the selection of retaining structures for deep excavations. This paper proposes an adaptive two-layer Case-Based Reasoning (CBR) framework to address this issue. First, a case database was constructed; feature interaction terms were incorporated into the similarity measure to capture the nonlinear coupling among geological parameters, and a Gradient Boosted Decision Tree (GBDT) was employed to achieve an objective, data-driven allocation of feature weights. The first layer implements a gating mechanism based on logical conjunction and adaptive tolerance thresholds, which automatically switches between the two inference paths when information is missing, preventing invalid hard matches; the second layer applies K-means++ clustering and local inductive reasoning to mitigate the biases caused by data sparsity. Experiments demonstrate that, under conditions of parameter incompleteness and noise interference, the method’s Top-3 recommendation accuracy significantly outperforms traditional models and machine learning baseline models and exhibits strong resistance to interference, providing a solid methodological foundation for retaining structure selection in complex data scenarios. Full article

When a lithium-ion cell’s series resistance is underestimated, the pre-update terminal-voltage innovation contains the first-order term $e_k\approx -I_k{R}_{0,k}\delta b$. This term breaks conditional sign symmetry and creates an odd response under current reversal. We test that mechanism before using it as an activation rule. The operational null is a near-zero conditional innovation centre with weak innovation–current coupling; declared falsifiers are comparable coupling under the nominal model, the wrong correlation sign under positive resistance error, failure of charge/discharge polarity reversal, or negative-control activation approaching ohmic-mismatch activation. A persistence-confirmed gate combines normalised-innovation-squared exceedances, innovation–current compatibility, a positive local resistance correction, and five consecutive qualifying windows. Sixty settings were ranked on 10 calibration seeds and frozen before disjoint holdouts. From 1.0× to 2.0× R₀, the sign-imbalance index increased from 0.0040 to 0.0786, and |$corr\left(e,I\right)$| increased from 0.0658 to 0.7777. The 30-seed static holdout produced 0/30 nominal activations, 27/30 detections at 1.5×, and 30/30 detections at 2.0–3.0×. A disjoint linear-drift audit yielded 0/30 pre-ramp activations and 30/30 detections at a median 1.71× multiplier, reducing late-drift SOC RMSE from 3.129% to 0.895%. A signed-current audit confirmed the predicted polarity reversal. An estimator-unseen audit gave 30/30 ohmic detections but retained 2/30 current-linked non-ohmic and 1/30 current-bias triggers, so the rule is not a unique fault classifier. NASA and LG records remain diagnostics of fixed versus always-on adaptation; they do not validate the gate or independent absolute SOC. The contribution is a falsifiable residual-symmetry mechanism with an explicit evidence boundary, rather than a post hoc symmetry label or hardware claim. Full article

Electric control valves (ECVs) are core control components in process industries such as petrochemicals and power generation, and their operational reliability directly affects system safety and energy efficiency. However, frequent changes in the working conditions of ECVs cause vibration signals to exhibit strong nonlinearity and non-stationarity, which leads to the loss of high-frequency transient features, difficulty in extracting weak faults, and cross-condition domain shifts. These issues severely limit the generalization ability of existing fault diagnosis methods. To address this, this study proposes a collaborative fault diagnosis framework that combines a miniaturized high-frequency data acquisition system with a multi-scale attention-conditioned domain adversarial network (MS-ACDAN). First, a miniaturized high-speed data acquisition system is developed based on a field-programmable gate array (FPGA) to enable lossless acquisition of high-frequency transient signals. Subsequently, the original vibration signals are decomposed, filtered, and reconstructed using Adaptive Noise-Complete Empirical Mode Decomposition (CEEMDAN) and the Comprehensive Sensitivity Index (CSI) to generate feature-enhanced signals with high signal-to-noise ratios. Next, a feature extractor combining a one-dimensional convolutional neural network (1D-CNN) with a channel attention mechanism is constructed to automatically focus on key fault frequency band features while suppressing redundant information. Finally, a Conditional Adversarial Network (CDAN) is introduced for transfer learning. By establishing a conditional dependency between class prediction and feature representation. This approach achieves domain alignment while preserving the discriminative features of the data, thereby overcoming the limitation of traditional domain adaptation methods that ignore category information. The experimental results show that the proposed fault diagnosis framework demonstrates high recognition performance in various transfer tasks. Furthermore, even under extreme industrial noise conditions of 0 dB, the framework exhibits good diagnostic robustness. This research provides a theoretical basis and technical solution for addressing the fault diagnosis of critical control equipment under complex and variable working conditions. Full article

Near-surface ozone (O₃) pollution is a growing environmental concern, particularly in the Beijing–Tianjin–Hebei (BTH) region, one of China’s most densely populated megacity clusters experiencing increasingly severe O₃ episodes. Existing data-driven forecasting models systematically underestimate high-concentration events and offer limited lead times. To reveal the meteorological drivers of extreme O₃ episodes, we conducted composite anomaly analysis over 2019–2023 and identified the dominant meteorological mechanism as a coupled pattern of mid-tropospheric anticyclonic circulation with high temperature, low humidity, and deep subsidence inversion, which suppresses vertical diffusion while southerly advection drives rapid near-surface O₃ accumulation. Motivated by meteorological diagnostics, we proposed ARC-Net, a Transformer-encoder-based Adaptive Residual Correction Network that ingests numerical weather prediction data from the European Centre for Medium-Range Weather Forecasts (ECMWF) and air quality observations to produce hourly O₃ forecasts up to 240 h (10 days) ahead. The model features a dual-branch regression-classification architecture enhancing feature discrimination at high concentrations and an Adaptive Residual Correction module that dynamically calibrates outputs through a triple-gating mechanism conditioned on pollution-level priors. In independent forecast tests for the year 2023 across 13 cities in the BTH region, ARC-Net achieved R² = 0.879 and a root mean square error (RMSE) of 17.03 μg/m³ at 0–24 h, retaining R² = 0.749 and RMSE = 24.57 μg/m³ at 0–240 h. For extreme episodes (maximum daily 8 h average ozone (MDA8_O₃) ≥ 215 μg/m³), the Critical Success Index improved by 63.9% over the baseline, and RMSE decreased by 33.15% within the 215–265 μg/m³ range in a representative case. These results indicate that meteorology-guided predictors combined with adaptive residual correction can partially alleviate high-O₃ underestimation and provide practically useful medium-range warning skill. Full article

As an important part in the mining crushing and grinding circuit, the high-pressure grinding roll (HPGR) is essential for energy efficiency, cost reduction, quality improvement and efficiency gains; therefore, it plays a key role in sustaining stable production and intelligent control of mining operations. Its instantaneous throughput and processing capacity directly influence the efficiency of the comminution system, the compatibility of production scheduling, and overall energy consumption. Therefore, these metrics serve as important indicators for intelligent optimization and stable operation. However, it is a difficult task to predict the throughput of HPGRs by traditional purely data-driven models. The process is characterized by strong nonlinearity, significant time-lag effects, and limited physical consistency. To address these challenges, this work develops a tailored throughput prediction framework for HPGRs by combining Gated Recurrent Units with Physics-Informed Neural Networks (GRU-PINN), which integrates an HPGR-specific volumetric throughput mechanism as a dedicated physical constraint. This approach first exploits the GRU network’s “reset” and “update” gates to selectively filter historical operating information while adaptively updating the current process features. This allows the model to better capture complex long-term temporal dependencies in the production data. Additionally, based on volumetric analysis and the principles of bed comminution, the HPGR throughput formula is incorporated into the loss function as a physical constraint. Together, these components establish a joint optimization framework that combines data-driven learning with physical constraints to correct prediction biases generated by the neural network during abrupt changes in operating conditions. The proposed GRU-PINN model was validated using real operational data collected from an industrial mining site. The results show that the proposed framework significantly improves the physical consistency of HPGR throughput predictions and, at the same time, effectively corrects prediction biases, which are commonly observed in conventional purely data-driven models under complex operating conditions. It also exhibits improved robustness and lower inference latency. Compared with other benchmark models, the proposed method displays superior overall performance across key evaluation metrics, including root mean square error (RMSE), mean absolute error (MAE), prediction accuracy, and inference speed. These findings confirm that the proposed method greatly enhances the physical interpretability of the prediction model without compromising accuracy. Consequently, it lays a solid foundation for process parameter optimization and intelligent control of HPGR system. Full article