Search Results (117)

Physical sensor deployments in ground-based environmental monitoring networks are frequently constrained by high installation costs, hardware failures, and limited spatial coverage, resulting in incomplete observational datasets and degraded sensing capacity across monitoring stations. Data-driven virtual sensing offers a cost-effective alternative by estimating target environmental parameters through machine learning models trained on correlated sensor measurements, reducing dependency on dense physical infrastructure. This paper presents VS-DCFF, an applied virtual sensing framework for dual-target estimation of near-surface air temperature and relative humidity from ground-based sensor network data. VS-DCFF integrates: (i) a deterministic pipeline applying temporal encoding, rolling-window statistics, and mutual information-based feature selection to capture a linear trend/seasonal component and to select features predictive of the residual signal; (ii) a probabilistic pipeline employing a Gaussian copula model to generate statistically consistent synthetic residual samples preserving inter-variable dependencies; and (iii) an early feature-level fusion strategy feeding a copula-augmented XGBoost residual-boosting stage, whose output is combined with the linear trend component for the final prediction. Under a strict chronological evaluation protocol, VS-DCFF is benchmarked against persistence, linear and ensemble regression baselines, and a same-protocol re-implementation of the statistical core of the VSG-SGL framework, and achieves near-surface air temperature $\mathrm{RMSE}=0.7791{}^{\circ }$C, $R^2=0.9735$, and relative humidity $\mathrm{RMSE}=3.7747\%$, $R^2=0.9701$, outperforming all tested baselines. The framework is further validated through leave-one-station-out spatial generalization, robustness evaluation under simulated sensor faults and target-history loss, copula-variant and synthetic-data fidelity diagnostics, and a lightweight edge-deployment ablation. All findings, including cases where tested extensions such as spatial context features did not yield a robust improvement, are reported transparently. Results indicate that the proposed architecture provides a computationally efficient, extensively validated approach to dual-target environmental virtual sensing under realistic deployment conditions. Full article

The safe and stable operation of coal-fired generating units is of critical strategic importance for ensuring the reliable supply of power systems. However, the fault evolution of industrial thermal systems exhibits the characteristics of strong nonlinearity and a long incubation period, coupled with the extreme scarcity of key fault samples (Few-shot) in actual production, which severely limits the engineering application of traditional data-driven diagnostic methods. Existing deep learning models, which are highly dependent on massive and balanced labeled data, not only struggle to overcome the overfitting bottleneck in scenarios with scarce fault samples, but also frequently introduce severe label noise (Label Noise) by ignoring the physical incubation period of faults, resulting in the degradation of the model’s decision boundary. To address the above challenges, this paper proposes a novel fault diagnosis framework integrating multi-scale spatiotemporal feature engineering and the Tabular Prior-Data Fitted Network (TabPFN). Starting from the physical mechanism of the system, this paper develops a dynamic label cleaning strategy based on multivariate statistical deviation, which accurately defines the fault divergence point to eliminate the noise in the incubation period. The constructed multi-scale spatiotemporal feature engineering integrating first-order difference and sliding window statistics can effectively map the transient mutation and steady-state evolution trend of the system. The introduced pre-trained TabPFN model based on the Transformer architecture, relying on its Bayesian inference capability and in-context learning (In-Context Learning) mechanism, can realize parameter-tuning-free and efficient classification for scarce samples. Experiments based on high-fidelity dynamic simulation data from GE Steam Power show that under the strict setting of limiting the training set to only 2000 samples, the proposed method achieves a comprehensive diagnostic accuracy of up to 99.29% and an F1-score of 0.9929 for seven typical operating conditions. Multi-dimensional comparative experiments and ablation studies confirm that the proposed framework comprehensively outperforms six mainstream baseline models, including XGBoost and SVM, in terms of precision, recall, and anti-interference robustness, and also delivers outstanding performance when benchmarked against deep learning models. This provides a brand-new theoretical perspective and technical paradigm for equipment health management in the context of industrial big data. Full article

Return predictability on the Saudi Arabian Stock Exchange (Tadawul), the largest equity market in the Middle East, remains underexplored relative to its structural distinctiveness as an oil-linked, retail-dominated emerging market. We compare 11 predictive models spanning five linear (regularised) regressors, three tree-based ensembles, and three stacked hybrid architectures. This comparison quantifies the improvement that nonlinear and ensemble methods offer over linear benchmarks for daily return prediction in this setting and identifies which method delivers the best accuracy-versus-cost trade-off for practical deployment. Using 28,750 daily observations (January 2015–December 2025), we constructed a 40-feature technical signal space spanning six families and evaluated all 11 models under a strict chronological train–validate–test protocol with an 18-month sealed holdout. A Lasso–XGBoost stacked ensemble achieves the lowest test RMSE of 0.906 and an out-of-sample Information Coefficient of 0.133, outperforming linear benchmarks by 15–27% in forecast error. Translated into a long-short strategy subject to 0.6% round-trip transaction costs, the optimal model delivers a Sharpe ratio of 0.587, an 81.1% win rate and a maximum drawdown of $-5.53\%$ across 758 trades. Sensitivity analysis confirms robustness across hyperparameter grids and rolling estimation windows. Bollinger Band Width, cross-sectional stock identity and lagged MACD signals collectively dominate feature importance rankings. Full article

The rational design of high-performance carbon-filled polyimide (CF/PI) composites is challenged by complex interactions among processing parameters, composition, and interfacial effects. Here, we address thermo-mechanical co-optimization in CF/PI composites through two-stage processing and formulation design. An experimental dataset was employed to construct processing and composition datasets. A CatBoost model was developed to relate hot-pressing parameters to the tensile strength (TS) of pure PI, enabling inverse optimization of processing conditions, with quantitative experimental validation of the predicted processing windows (RMSE = 3.89 MPa). Within the optimized processing window, an XGBoost-based composition-property model was further established to perform high-throughput screening and multi-objective optimization of TS and thermal conductivity (TC). To quantitatively account for interfacial effects, the mass ratio of carbon-filled to sizing agent (CF/SA) was introduced as a composition-derived feature. Model interpretation reveals that CF and graphene positively contribute to TC but negatively affect TS. Notably, formulations with CF/SA > 2 tend to achieve a more favorable TS–TC balance. Experimental validation of model-selected composite formulations confirmed the predicted trends for both TS and TC. These findings suggest that interfacial regulation is a key lever for balancing mechanical strength and thermal transport in CF/PI composites. Full article

Deep excavations are high-risk geotechnical activities, and accurate prediction of diaphragm-wall settlement is important for construction monitoring and deformation control. This study investigates cumulative vertical settlement at 23 diaphragm-wall monitoring points from the deep excavation of Tianjin Goldin Finance 117 in Tianjin, China. Seven prediction models—a naïve persistence model, autoregressive integrated moving average (ARIMA), K-nearest neighbors (KNN), multilayer perceptron (MLP), gated recurrent unit (GRU), Transformer, and XGBoost—were evaluated using a unified five-fold rolling-origin expanding-window validation scheme. GRU achieved the best overall baseline performance, with a mean R² of 0.9172, a mean absolute error (MAE) of 0.1009 mm, a root mean square error (RMSE) of 0.1341 mm, and a mean absolute percentage error (MAPE) of 0.5875%. GRU was subsequently optimized using the crow search algorithm (CSA), the genetic algorithm (GA), and the whale optimization algorithm (WOA). GRU-WOA achieved the best numerical performance, with a mean R² of 0.9289 and an RMSE of 0.1215 mm. Relative anomaly levels were further identified from predicted settlement-change rates to characterize temporal concentration and spatial clustering of settlement-change activity. The proposed framework can support priority inspection and targeted monitoring, although the resulting anomaly levels represent project-relative statistical deviations rather than code-based engineering risk classes. Full article

Emerging industrial agglomeration drives green transformation through knowledge spillovers and economies of scale. However, its effects exhibit pronounced nonlinearity and heterogeneity, shaped by spatial externalities and development stages. This paper investigates 19 Chinese urban agglomerations over the period 2014 to 2023. Kernel density estimation based on enterprise-level Point-of-Interest (POI) data is used to characterize spatial agglomeration patterns across eight emerging sectors. A two-stage dynamic network super-efficiency SBM model decomposes Green Transition Efficiency (GTE) into resource utilization and pollution control sub-stages. An XGBoost-SHAP-GEO analytical framework, combined with partial dependence analysis, then identifies nonlinear driving mechanisms. The main findings are as follows: First, emerging industrial agglomeration intensifies and polarizes toward the eastern coast, whereas GTE displays a “high-west, low-east” pattern. This produces a significant spatial mismatch, rooted in the near-saturation of environmental carrying capacity in eastern regions, where congestion effects exceed knowledge spillover dividends. Second, geographic characteristics constitute the primary factor shaping GTE and operate through nonlinear interactions with industrial agglomeration and R&D investment. Notably, their moderation direction is reversible, suggesting that geographic endowments should be understood as “conditional assets” rather than fixed advantages. Third, nonlinear patterns across sectors are highly heterogeneous. The bio-industry is the only sector to achieve a J-shaped positive breakthrough. Information technology and new materials exhibit persistent inhibition, while related services display an extremely narrow threshold window with the deepest negative reversal. Thus, “moderate agglomeration” is a multidimensional concept that shifts dynamically with industry type and regional endowment. Fourth, driving mechanisms display stage-dependent evolution. The incubation stage relies on natural endowments and basic industrial pull, with the green bottleneck residing in resource utilization efficiency. The growth stage faces multiple tensions from coexisting positive and negative effects. The optimization stage shifts toward R&D innovation and industrial greening, marking a qualitative transformation from MAR externalities to Jacobs externalities. In addition, the non-significant linear coefficient in the 2SLS instrumental variable test is consistent with the inverted U-shaped nonlinear finding, further validating the necessity of a nonlinear analytical framework. These findings provide differentiated governance evidence for balancing industrial agglomeration with green sustainable development. Full article

Accurate air conditioning (AC) load prediction under extreme weather conditions is critical for power grid stability and energy management. While tree-based ensemble methods such as XGBoost, LightGBM, and Gradient Boosting have become the dominant paradigm in short-term load forecasting, their effectiveness for weather-sensitive AC load prediction—particularly during extreme weather events—remains insufficiently examined. This study presents a systematic comparison of four machine learning models—Support Vector Regression with RBF kernel (SVR-RBF), LightGBM, XGBoost, and Gradient Boosting—for daily AC load estimation conditioned on measured same-day meteorological and calendar features. Based on five years of processed load and weather data from a major city in Southwest China, we construct 31 features and evaluate model performance across four scenarios: normal days, weather-extreme days, high-load P85, and high-load P90. Hyperparameters are selected within the first four years by expanding-window validation, and the fifth year is held out for testing. Our results reveal that SVR-RBF achieves an overall $R^2$ of 0.9772, substantially outperforming LightGBM (0.9091), XGBoost (0.8972), and Gradient Boosting (0.9094); paired moving-block bootstrap intervals for the tree-minus-SVR MAE differences exclude zero. The advantage of SVR-RBF is most pronounced under extreme conditions: on weather-extreme days, SVR-RBF attains $R^2=0.8316$ versus $R^2=0.39$ for the best tree model. This pattern is consistent with the smooth U-shaped temperature–load relationship captured by the RBF kernel. Additional trend-sensitivity analysis shows that annual load growth and target-level extrapolation also explain a substantial part of the tree-model degradation. Furthermore, we quantify the performance limitation of the restricted feature set: while normal-day estimation achieves $R^2=0.984$, high-load P90 days reach $R^2=0.843$. These findings support SVR-RBF as a strong baseline for measured-weather conditional AC load estimation while emphasizing that temporal shift and training-domain coverage must be considered when interpreting model differences. For energy-system applications, the lower errors under high-load and weather-extreme conditions are relevant to peak-demand assessment, reserve planning, and demand-side management. Full article

Cross-chain bridges represent one of the most damaging attack surfaces in decentralized finance, with major exploits (e.g., Ronin, Wormhole, Nomad, Multichain) arising not from broken signature schemes but from failures in proof verification, replay protection, and signer-set management, gaps that conventional threshold-signature multisignature wallets do not address. This study presents an incident-aware multisignature architecture combining three on-chain predicates—block-height freshness windows, epoch-bound signer sets, and Merkle inclusion-proof verification—with a non-authoritative off-chain LightGBM classifier that generates SHAP-attributed risk explanations to support governance actions such as pausing, vetoing, or rotating signers, without directly blocking or approving execution. The framework was evaluated on a simulated benchmark of 78,600 Ethereum testnet transactions containing six injected anomaly classes (gas spikes, nonce jitter, malformed call data, stale intents, proof-delivery delays, and epoch-rotation replays). The LightGBM advisor achieved ROC-AUC 0.92 (95% CI [0.906, 0.926]) and F1 0.73 ([0.712, 0.749]), outperforming five baselines—logistic regression, Random Forest, XGBoost, isolation forest, and a rule-based detector—with the highest F1 (0.731) and PR-AUC (0.799), while the rule-based detector, which by construction covers only the anomaly classes addressed by the deterministic predicates, attained F1 0.282. Differences were statistically significant except for the LightGBM–XGBoost PR-AUC comparison. The deterministic layer itself is verified through 28 property-level contract tests covering all seven modeled attack objectives, with measured per-function gas costs (execute_Intent: 118,756 gas, of which 28,432 gas is Merkle-proof verification). Within this controlled setting, the results indicate that a machine learning advisor can extend anomaly-prioritization coverage beyond the scope of the deterministic predicates while leaving execution control fully deterministic. This work is presented as a controlled proof of concept: the reported metrics quantify recovery of scripted injection patterns, and validation against real-world exploit traces remains future work. Full article

Accurate joint forecasting of electricity, cooling, heating, and gas loads is essential to the coordinated operation of integrated energy systems. However, multivariate energy load sequences exhibit strong cross-carrier coupling, non-stationarity, and heterogeneous fluctuation characteristics, which limits the performance of conventional independent forecasting and fixed-model approaches. To address these challenges, this study proposes a joint load forecasting framework that integrates tabular Q-learning-assisted multivariate variational mode decomposition, sample-entropy-based reconstruction, and dynamic model selection. First, tabular Q-learning is employed to select the MVMD penalty factor and the four load sequences are synchronously decomposed to preserve the coupling relationships among components with common center frequencies. Second, sample entropy is used to reconstruct the decomposed modes into high-frequency, low-frequency, and residual subsequences, thereby reducing forecasting complexity while retaining relevant temporal features. Third, a dynamic model selection mechanism evaluates SVR, BiLSTM, XGBoost, and LightGBM and assigns an appropriate predictor to each reconstructed subsequence according to its forecasting performance. The framework is evaluated using daily electricity, cooling, heating, and gas load data collected from the Tempe Campus of Arizona State University from 2016 to 2020. A rolling input window of 56 days is used to forecast the subsequent seven days. Compared with the benchmark methods, the proposed framework achieved the best overall composite performance and competitive forecasting accuracy across the four load types. These results provide a potentially useful forecasting basis for operational decision-making in integrated energy systems. Full article

Reliable fault detection in nonlinear process systems requires both accurate classification and interpretable analysis of faults with weak or near-normal signatures. This study develops an explainable, data-driven, and temporally informed multiclass framework for the Tennessee Eastman Process (TEP), retaining all 21 operating conditions (20 fault types and the normal operating condition). Six sliding-window statistics were extracted from 52 process variables and classified using Extreme Gradient Boosting (XGBoost). Performance was evaluated at both sample and run levels through fault-specific window analysis, an a priori validation-driven hierarchical decomposition, SHapley Additive exPlanations (SHAP), a Relative Sensitivity Index (RSI) based on detection-delay sensitivity, and computational benchmarking. Aggregating sample-level predictions (macro F1 = 0.8823) into run-level decisions via majority voting improved performance substantially (macro F1 = 0.9515). Across five seeded repetitions, the mean run-level macro F1-score was $0.9522\pm 0.0035$ (95% CI: $\pm 0.0044$). Under the primary evaluation, 18 of 21 classes achieved F1 $\ge 0.97$. Fault 3 benefited strongly from extended temporal context, whereas Normal operation, Fault 9, and Fault 15 retained a structured but asymmetric confusion pattern dominated by Normal–Fault 9 errors. SHAP identified model-specific attribution patterns associated mainly with cooling-water-related variability features, while RSI indicated greater prediction-stream sensitivity for the historically difficult faults. Feature extraction and inference required approximately 6.2 ms on CPU and 35.3 ms on GPU, negligible relative to the 180 s sampling interval. These findings indicate that temporal context, run-level aggregation, and explainability can jointly support accurate data-driven fault diagnosis while revealing persistent fault-specific ambiguity. Full article

Continuous monitoring of Activities of daily living (ADLs) requires sensing systems that are privacy-preserving, low-power, and robust to environmental variation. Ambient sensing technologies provide an alternative to RGB video and wearable devices, but individual sensing modalities exhibit characteristic limitations. Sparse mmWave radar provides strong motion sensitivity but limited posture detail, low-resolution thermal sensing preserves posture-related spatial information, and smart plug telemetry captures only appliance-mediated behavioural interaction. To address these limitations, this paper proposes a layered multimodal ambient-sensing framework comprising a sparse-track 24-GHz FMCW radar, a $32\times 24$ low-resolution thermal sensor, and a Moko smart plug. It experimentally evaluates a radar–thermal HAR branch together with a separate smart plug appliance-recognition branch. The framework proposes three streams to enable continuous non-wearable monitoring while maintaining redundancy and reduced privacy exposure for intelligent-building and ambient assisted living environments. Radar and thermal streams are jointly evaluated on binary motion and four-class posture and activity recognition tasks collected across multiple environmental configurations using recording-grouped cross-validation, while the appliance stream is evaluated using per-plug telemetry from residential-grade appliances. The radar–thermal streams use a single-subject, fixed-placement dataset of binary-motion windows and four-class posture and motion windows collected across six furniture configurations. The separate intrusive load monitoring stream utilises smart plugs to classify appliances. Regarding binary motion recognition, radar $(F_{1,\mathrm{Transformer}}=0.882\pm 0.034)$ and thermal $(F_{1,\mathrm{XGBoost}}=0.870\pm 0.069)$ pipelines achieved similar macro F1 performance. On the four-class posture and activity recognition task, thermal features $(F_{1,\mathrm{thermal}}=0.775\pm 0.053)$ substantially outperformed radar $(F_{1,\mathrm{radar}}=0.609\pm 0.110)$. Weighted late fusion produced only modest descriptive gains. Separately, smart plug telemetry demonstrated strong appliance recognition performance using lightweight tree-based models suitable for constrained edge deployment. The results support a scoped redundancy argument. Sparse track-level radar carries gross motion, while low-resolution thermal sensing carries posture. The smart plug appliance monitoring extends the framework toward appliance-mediated instrumental activity of daily living (IADL) monitoring, with lightweight tree-based models achieving strong recognition performance under constrained edge deployment conditions. The findings support a layered multimodal sensing architecture for privacy-preserving ADL monitoring, where radar contributes motion-sensitive coverage, thermal sensing contributes posture-aware spatial context, and smart plug telemetry contributes appliance-level behavioural evidence within intelligent healthcare and ambient assisted living environments. Full article

Background: Type 2 diabetes (T2D) and insulin resistance (IR) are major global health challenges. Volatile organic compounds (VOCs) in exhaled breath offer a non-invasive window into metabolic dysregulation. This study aimed to predict HOMA-IR using machine learning (ML) by integrating biochemical markers and VOC profiles in a male cohort. Methods: This cross-sectional study included 1258 male participants from the Taiwan MJ cohort. Four ML algorithms (Elastic Net, MARS, Random Forest, and XGBoost) were trained to predict HOMA-IR. Model performance was evaluated using R², RMSE, and MAE. SHAP analysis was used to interpret feature contributions. Results: Ensemble tree-based approaches (Random Forest and XGBoost) demonstrated better predictive performance than Elastic Net and MARS. Random Forest achieved the highest predictive performance on the test set (R² = 0.323). SHAP analysis identified BMI (mean |SHAP| = 1.326) as the strongest predictor, followed by TG (1.005) and HDL-C (0.445). Notably, specific breath VOCs, including methanol and formic acid, ranked among the top 20 predictors, capturing distinct aspects of metabolic dysregulation orthogonal to standard blood tests. Conclusions: Integrating VOC profiles with clinical markers provides acceptable predictive performance for IR in men. While traditional metabolic markers dominate the prediction, specific VOCs capture distinct metabolic information, highlighting the potential of breath analysis as a complementary early screening tool. Full article

Natural fractures strongly influence fluid flow, hydraulic-fracturing performance, and production heterogeneity in tight oil reservoirs. Their identification from conventional logs remains challenging because image-log and core coverage is limited, fracture-related logging responses are non-unique, and discrete fracture interpretations are difficult to align with regularly sampled logging sequences. This study used conventional logging data and electrical image-log interpretations from 17 wells in the Xifeng Oilfield, Ordos Basin, together with core observations from selected intervals, to develop a fracture response mapping and triangular topology-optimized bidirectional long short-term memory model (FRM-BiLSTM-TTAO). After sliding-window construction and density-based undersampling, 1713 samples were retained and partitioned at the well level into 14 training wells and three independent test wells, yielding an approximate training-to-test sample ratio of 75:25. FRM extracts lithologic-background, local-abrupt-change, multiscale-fluctuation, and integrated fracture response features; BiLSTM captures bidirectional depth dependencies; and TTAO selects fracture response features and optimizes the network architecture and training parameters. On the test set, the model achieved a ROC-AUC of 0.9079, a recall of 0.8671, and an F1-score of 0.8464, outperforming CNN, MLP, ResNet1D, XGBoost, and the corresponding ablation models. The predicted high-probability intervals were generally consistent with image-log interpretations and core observations, indicating the feasibility of the proposed method for identifying fracture-prone intervals within the study area. Full article

Maintenance management of stationary combustion engines in the agricultural sector remains largely manual, increasing the risk of unplanned downtime. This study developed a machine learning-based predictive model to anticipate failures within a 60-day horizon, enabling the transition from reactive to proactive maintenance. Following the CRISP-DM (Cross-Industry Standard Process for Data Mining) framework, a sliding-window feature engineering pipeline was built from 2250 historical records spanning 59 engines. Four ensemble learners (Random Forest, LightGBM, XGBoost, and CatBoost) were then compared under two complementary protocols: a strict 60/40 chronological split simulating deployment, and a stratified leave-engine-group-out cross-validation withholding entire engines from training. Nonparametric testing (DeLong test and engine-level cluster bootstrap) showed that the four learners are statistically equivalent, whereas the feature engineering layer contributes a large, significant discrimination gain (ΔAUC $\approx +0.06$, $p\approx {10}^{-27}$) on engines unseen during training. Random Forest, selected as the final model, achieved an AUC of 0.90 with 84.2% recall under the deployment protocol and 0.96 with 90.7% recall under engine-grouped validation. Temporal extrapolation, rather than cross-engine generalization, appeared to be the primary challenge, indicating that rigorously engineered degradation features, more than the choice of ensemble algorithm, drive predictive performance in agricultural maintenance planning. Full article

In low-voltage distribution networks, load switching, induction motor start-up, photovoltaic output variations, and short-circuit faults may produce highly overlapping electrical characteristics, which can lead to maloperation or failure to operate in conventional protection. To address this problem, this paper proposes an adaptive protection method integrating physically guided and cost-sensitive learning. First, an incremental topology-constraint deviation and a voltage-current trajectory curvature are constructed based on the fault-superimposed network constraint and the variation characteristics of system equivalent impedance, enabling the discrimination of short-circuit faults from non-fault transient disturbances. Then, a cost-sensitive physically guided extreme gradient boosting (XGBoost) model is developed, in which a fault-current-increment-based weight is introduced into the objective function to enhance the learning capability for weak-fault samples. Furthermore, a temporal-consistency-based protection operation logic is designed using sliding-window confirmation and majority voting to suppress isolated abnormal predictions. Simulation and RTDS-based real-time validation results on a 0.4-kV low-voltage distribution network with distributed photovoltaic generation show that the proposed method improves weak-fault detection sensitivity and reduces maloperation under complex source–load disturbances. The method relies only on local measurements and has potential for deployment in low-voltage intelligent protection terminals. Full article