Objective: Photovoltaic (PV) fire accidents caused by arc faults and hotspot effects pose growing safety risks globally. Conventional Class A foam fails in PV fire scenarios owing to poor adhesion on smooth inclined glass surfaces and wind dispersion susceptibility, leaving modules vulnerable to sustained heat exposure and thermal-stress-induced fractures. In this study, we develop a novel gel foam (NGF) with superior adhesion, wind resistance, and thermal protection performance specifically tailored for PV fire suppression. Methods: The NGF was formulated from sodium dodecyl sulfate (SDS), konjac glucomannan (KGM), and sodium sulfate (Na2SO4). A three-factor, three-level orthogonal experiment was used to optimize their concentrations using the expansion ratio (E), the drainage half-life (T1/2), and a composite foam comprehensive index (FCI) as evaluation metrics, yielding three candidate formulations. Foam retention was assessed on a 45° inclined glass platform at wind speeds of 0, 6, and 25 m/s. Thermal insulation performance was quantified under a far-infrared radiation system at 15 kW/m2, with thermocouples at the foam surface and at 10 and 20 mm depths; total heat absorption was calculated via radiative heat transfer modeling. Finally, two sets of full-scale thermal protection experiments were performed using a 35 cm-diameter gasoline pool fire beneath a standard-size PV module (1 640 mm × 990 mm), with visible and infrared cameras recording surface damage and temperature distribution in real time. Results: Among the three factors, KGM concentration had the greatest influence on E and T1/2, followed by Na2SO4 and SDS concentrations. Increasing KGM and Na2SO4 concentrations enhanced solution viscosity and promoted three-dimensional gel network formation through hydrogen bonding and polar interactions, thereby increasing T1/2 while decreasing E. The optimal formulation (0.5 wt% SDS, 0.8 wt% KGM, and 1.0 mol/L Na2SO4) achieved the highest FCI of 3 528.7, with E = 5.2 and T1/2 = 904.8 min. In retention tests on a 45° inclined glass surface, this formulation maintained retention rates of 82.04%, 80.39%, and 82.13% at wind speeds of 0, 6, and 25 m/s, respectively. Meanwhile, conventional Class A foam retained only 15.27% at 0 m/s and collapsed entirely at 6 and 25 m/s. These results demonstrate that NGF provides considerably more stable surface adhesion under static and high-wind conditions. In thermal insulation experiments, compared with conventional Class A foam, the foam collapse time of the optimal formulation was extended by 243.2% and total heat absorption capacity increased by 250.5%. In full-scale fire tests, NGF-coated PV modules exhibited substantially more uniform surface temperature distributions than uncoated modules. During early combustion, the foam absorbed heat through liquid drainage, stabilizing module temperature; as burning continued, the foam drained downward and hardened. Post-experiment inspection confirmed that NGF-protected modules sustained substantially less physical damage. Conclusions: The developed NGF, leveraging the synergistic salting-out gelation between KGM and Na2SO4, forms a robust three-dimensional network that delivers superior adhesion, wind resistance, and thermal stability on inclined smooth PV glass surfaces. Compared with conventional Class A foam, the optimal formulation offers markedly higher retention under wind exposure, a 243.2% longer thermal insulation duration, and a 250.5% improvement in heat absorption capacity. Full-scale fire experiments validate its effectiveness in suppressing temperature rise and preventing thermally induced module fracture under realistic fire conditions. This study provides a scientific foundation and technical reference for developing next-generation fire suppression and thermal protection materials tailored to PV fire scenarios.
Objective: Industrial Internet of Things (IIoT) has emerged as a critical infrastructure for safety monitoring in high-risk industrial environments such as petrochemical plants and power systems. Real-time identification of safety risks, including abnormal behaviors and hazardous events, is essential for ensuring operational reliability and preventing accidents. Federated learning (FL), a distributed machine learning paradigm, enables collaborative model training across multiple edge devices while preserving data privacy. However, its application in IIoT-based safety risk identification still faces considerable challenges, including unreliable edge nodes, heterogeneous computing and communication capabilities, and high communication overhead due to frequent parameter exchanges. These issues lead to degraded model accuracy, reduced system robustness, and increased latency. To address these limitations, this paper proposes a dynamic clustering framework for FL (DCF-FL) to improve the reliability and efficiency of safety risk identification in IIoT environments. Methods: The proposed DCF-FL framework integrates three key components: a reputation evaluation mechanism, dynamic cluster head selection, and joint optimization of device selection and resource allocation. First, a comprehensive reputation evaluation mechanism based on subjective logic is designed to assess the trustworthiness of edge devices. This mechanism incorporates multiple factors—such as data contribution, local model accuracy, and channel conditions—to dynamically select reliable devices as cluster heads, thereby enhancing training robustness. Second, the FL process is organized into dynamically formed clusters, where the cluster head coordinates local training and model aggregation. A dynamic replacement strategy for cluster heads is introduced to mitigate single-point failures and adapt to varying network conditions. Third, the joint optimization problem of device participation and resource allocation is formulated as a mixed-integer nonlinear programming problem, which considers communication and computation costs while maintaining model accuracy. To efficiently solve this complex problem, a deep reinforcement learning (DRL)-based decoupling approach is adopted, where device selection and resource allocation are handled in separate stages. The DRL agent iteratively learns optimal decision policies through interaction with the environment, improving resource utilization and reducing the overall system cost. Results: Extensive simulations are conducted using public datasets, including BoWFire and UCF-Crime, along with a custom dataset for industrial safety scenarios. The experimental results demonstrate that DCF-FL substantially outperforms the FedAvg and FedProx baseline methods. In a typical scenario with 15 devices, DCF-FL's total communication and computation cost is reduced by 76.9% and 61.3%, respectively, compared with the baseline algorithms. Moreover, the DRL-based optimization achieves near-optimal performance, with only a 4.6% gap compared with the theoretical optimum obtained by traditional branch-and-cut methods, while maintaining much lower computational complexity. In terms of model performance, DCF-FL exhibits faster convergence and lower test loss, especially in non-independent and identically distributed data settings. The accuracy of risk identification remains consistently higher than that of the baseline methods across most training rounds, demonstrating the effectiveness of dynamic clustering and reputation-based node selection. Additionally, the proposed reputation mechanism effectively distinguishes unreliable devices, preventing their negative impact on model training and improving system stability. Conclusions: This paper presents DCF-FL, a novel FL framework tailored for industrial safety risk identification in IIoT environments. By integrating subjective logic-based reputation evaluation, adaptive cluster head selection, and DRL-driven resource optimization, DCF-FL effectively addresses the challenges of node unreliability and high resource consumption in traditional FL systems. The results indicate that DCF-FL achieves a favorable balance between model accuracy and system efficiency, making it suitable for resource-constrained and heterogeneous industrial scenarios. Future work will focus on further improving the stability of model accuracy under highly dynamic conditions and extending the framework to more complex real-world applications.
Objective: Toxic gas leakage accidents in hazardous chemical industries can rapidly threaten human life safety, property, and the surrounding environment. The emergency response following such accidents depends on timely knowledge of the source term, including the source location and source strength. However, direct access to the leakage point is often difficult because the accident scene may involve high toxicity, poor visibility, and rapidly changing dispersion conditions. Against this backdrop, source-term inversion based on sparse monitoring data provides an important computational approach for supporting emergency decision-making. This study aims to develop a stable source-term inversion framework for toxic gas leakage accidents under sparse monitoring conditions and to evaluate how meteorological parameters, monitoring noise, monitoring-point density, and data augmentation affect the inversion results. Methods: A Monte Carlo–differential evolution (DE) optimization method, referred to as the MDO method, was constructed by combining Monte Carlo preconstraints, DE global search, and local optimization using the limited-memory Broyden–Fletcher–Goldfarb–Shanno algorithm with bound constraints (L-BFGS-B). The Gaussian plume model was used as the forward dispersion model to calculate toxic gas mass concentrations at monitoring points. First, Monte Carlo sampling was used to generate candidate source-term parameters within predefined search ranges and to identify a feasible search region according to the mismatch between simulated and monitored concentrations. This step reduced the search space for subsequent optimization. Second, DE was used to globally search for the leakage source coordinates, thereby reducing the dependence of inversion results on a single initial guess. Third, L-BFGS-B was used to locally refine the source strength under bound constraints. To examine the role of sparse monitoring data, interpolation-based data augmentation was treated as an optional preprocessing step rather than a mandatory component. The method was tested using four controlled source-term scenarios with six monitoring points. Additional numerical simulations were conducted to examine atmospheric stability class, wind direction deviation, wind speed perturbation, monitoring noise level, and the number of monitoring points. Baseline comparisons with full-domain DE, Monte Carlo-only estimation, and local multi-start optimization were also performed. Results: Under the baseline-controlled conditions, the proposed MDO framework completed the joint inversion of source location and source strength within a short computation time. Meteorological sensitivity analysis showed that atmospheric stability mismatch strongly affected the location result. When the stability class used in inversion was consistent with the forward simulation condition, namely class C, the mean location error was 0.174 1 m. When the stability class was B, D, or E, the mean location errors increased to 9.564 0, 12.315 4, and 11.423 4 m, respectively. Wind direction deviation also increased the location error. When the deviation changed from 0° to ±10° and ±15°, the error generally increased, indicating that wind direction accuracy is important for source-term inversion. In contrast, small wind speed perturbations around the baseline wind speed had a weaker influence on location error in the present setting. Robustness analysis showed that when the monitoring noise level increased from 0% to 15%, the mean location error increased from 2.125 9 to 2.417 0 m, and the relative error of source strength increased from 8.451 7% to 10.917 5%. The degradation was gradual. Increasing the number of monitoring points improved the inversion accuracy: when the number of monitoring points increased from 4 to 10, the mean location error decreased from 3.018 8 to 1.163 3 m. The data augmentation comparison showed that interpolation did not always improve inversion accuracy. In the default scenario, the mean location error obtained using the original monitoring points was 3.620 9 m, whereas the error after interpolation-based augmentation was 6.596 3 m. The baseline comparison further showed that the MDO method produced results of the same order of magnitude as those of the comparison methods; however, it did not show a distinct advantage over all alternatives across every metric. Conclusions: The MDO method provides a feasible multi-stage framework for source-term inversion of toxic gas leakage accidents under controlled numerical conditions and sparse monitoring data. The results indicate that atmospheric stability and wind direction accuracy are key meteorological factors affecting inversion precision, whereas higher monitoring-point density helps improve source localization. Data augmentation should be used cautiously because interpolated points are not independent measurements and may introduce additional uncertainty under unfavorable monitoring-point distributions or interpolation errors. The present study is still based on the Gaussian plume model and controlled numerical scenarios. Complex terrain, building interference, unsteady wind fields, and real accident monitoring errors were not fully considered. Future work should incorporate higher-fidelity dispersion models and field monitoring data to further evaluate the engineering applicability of the proposed method.
Objective: Oil spills in the ocean resulting from accidents can cause significant water pollution across large areas. When a dispersed oil film on the water surface comes into contact with an ignition source, it may ignite and lead to a severe marine fire. Currently, research on the microscopic mechanisms through which oil properties and substrate environments influence the dynamic diffusion behavior of oil is relatively limited. Methods: In this study, a multiphase flow model was developed using the ANSYS FLUENT software to investigate the dynamic evolution and diffusion mechanisms of oil on the substrate under various influencing factors. Using numerical simulation, diffusion kinetics of oil films on substrates with different properties were studied under various environmental conditions, focusing mainly on three variables: (1) substrate salinity (deionized water, 20 g/kg, 30 g/kg, and 40 g/kg), (2) water content of the oil (crude oil, 20%, 40%, and 60%), and (3) substrate temperature (20 ℃, 30 ℃, and 40 ℃). This study mainly emphasizes the effects and regulatory roles of these three influencing factors on the dynamic diffusion behavior of oil films. Results: The results indicate a direct correlation between the salinity of the substrate and the diffusion rate of the oil film. Increased salinity elevates the density and surface tension of the substrate, promoting the diffusion of the oil film. The water content of oil affects its physical properties, such as density, viscosity, and surface tension. Higher water content leads to higher viscosity, which inhibits the diffusive movement of the oil on the substrate and ultimately reduces the diameter of the diffused oil film. Analysis of the oil film diffusion rate contour map revealed that during the initial diffusion stage, gravity acts as the primary driving force, causing rapid diffusion in a radial direction. Over time, surface tension dominates and slows down the diffusion process at the oil film edges. With an increase in substrate temperature, the viscosity and surface tension of the oil decrease, thus promoting the diffusion of the oil film. The ratio of the diameter to the thickness of the oil film also affects the diffusion trend. At higher temperatures, this ratio varies significantly, indicating vigorous movement of the oil film in the initial stages. Formulas for predicting the diameter of the diffused oil film under varying conditions of salinity, water content, and temperature were derived based on power-law functions and showed good agreement with simulation results. Conclusions: Oil spills in marine environments spread at a faster rate with an increase in salinity. The water content in oil films also affects the diffusion rate: oil films with high water content spread at a slower rate compared with those with low water content. Furthermore, water temperature has a significant effect on oil spilling; high water temperature makes the diffusion behavior of the oil film more intense. Numerical simulation characterizes the dynamic diffusion features of oil under diverse environmental conditions, laying a theoretical foundation for the prediction of slick diffusion of oil during spill incidents.
Objective: Heat strain and skin burn injuries severely affect the health of emergency rescue personnel and reduce their rescue efficiency in disaster environments. Predicting these injuries is thus a crucial aspect of early warning approaches. Furthermore, the heat transfer coefficient, influenced by factors such as wind velocity and wind direction, is a critical input for simulations of heat transfer, heat strain, and skin burn injuries. To improve the health and safety of emergency rescue personnel, a comprehensive understanding of the effects of wind velocity and wind direction on the heat transfer coefficient across different body segments is essential. Methods: First, a three-dimensional body scanning technique was applied to obtain a numerical thermal manikin, based on a 20-zone thermal manikin named "Newton" (including the head, face, chest, stomach, shoulders, back, thighs, calves, feet, forearms, hands, hips, and upper arms). Subsequently, a numerical climate chamber measuring 5 m × 2.7 m × 3 m was developed, with the numerical thermal manikin placed at the center. Second, the angle between the numerical manikin and the inlet was adjusted to set five wind directions (0°, 45°, 90°, 135°, and 180°), and eight wind velocities (0.2, 1.0, 2.0, 4.0, 8.0, 12.0, 16.0, and 20.0 m/s) were used in the numerical simulation. Third, heat transfer and airflow were simulated using software based on the finite volume method, and the whole-body and local convective heat transfer coefficients were calculated under various wind directions and wind velocities. Finally, the simulation performance for heat transfer and the convective heat transfer coefficient was validated against experimental measurements and simulations in the literature. Results: The results indicated that dry heat transfer increased with wind velocity, with total heat transfer increasing by 17.74 times when wind velocity rose from 0.2 to 20.0 m/s. Moreover, the corresponding ratio of convection to dry heat transfer increased by 42.6%. By contrast, wind direction had a negligible effect on whole-body convective heat transfer but showed large differences among body segments. Specifically, the convective heat transfer coefficient at the torso was considerably higher than that at the four limbs (upper arms, forearms, thighs, and calves). At a wind direction of 0° and an wind velocity of 16.0 m/s, compared with the right limb, the convective heat transfer coefficient at the left upper arm, left forearm, left thigh, and left lower leg decreased by 21.1, 25.1, 9.1, and 10.8 W/(m2·℃), respectively. Conclusions: The convective heat transfer varied greatly across body segments, with the hands exhibiting the highest convective heat transfer coefficient in all cases. Specifically, the difference in convective heat transfer coefficient among body segments was amplified by wind velocity. The simulated convective heat transfer coefficients under different wind velocities and wind directions can serve as inputs for human thermoregulation models, heat strain evaluations, and skin burn injury assessments. Subsequently, these outputs can provide fundamental knowledge for developing early safety warning systems, high-performance personal protective clothing, and decision-making tools for rescue personnel during rescue operations.
Objective: This study systematically identifies the key causal factors of unmanned aerial vehicle (UAV) disturbance accidents and elucidates the directed transmission paths and interaction intensities among multilevel risk factors across the "regulation–operation–equipment–environment" dimensions. Unlike previous studies that predominantly constructed undirected networks based on co-occurrence relationships, this research proposes an integrated analytical framework capable of capturing both the causal directionality and the structural fragility of the risk propagation network. Methods: A total of 1,089 UAV disturbance accident reports were collected from the UK Air Accidents Investigation Branch and the US Federal Aviation Administration for the period 2015–2025. First, the KeyBERT model (using the all-MiniLM-L12-v2 embedding) was applied to extract keywords from the accident narratives, followed by hierarchical clustering to derive six thematic clusters. Second, the Systems Theoretic Accident Model and Processes (STAMP) and its associated System Theoretic Process Analysis (STPA) were employed to map the extracted keywords onto a five-level hierarchical control structure encompassing policy/regulation, operation/risk management, pilot decision-making, equipment/software, and airspace/environment. Third, all risk factors were standardized into 38 items across five dimensions: flight behavior (FB), accident occurrence (AO), accident consequence (AC), equipment and environment (EE), and management and supervision. The FP-Growth algorithm (minimum support = 0.015, minimum confidence = 0.25) was then used to mine strong association rules, yielding 3,359 rules. Based on these rules, a directed weighted causal network was constructed, with nodes representing risk factors and edge weights corresponding to confidence values. Network topology analysis (including degree, strength, clustering coefficient, and betweenness centrality), community detection (via modularity optimization), and robustness analysis (random attacks via Monte Carlo simulation and targeted attacks based on DomiRank centrality) were conducted to identify critical nodes, vulnerable paths, and overall structural fragility. Results: The resulting directed weighted causal network comprised 38 nodes and 86 edges, with an average degree of 4.72 and an average strength of 4.12. Five communities were identified (modularity = 0.482). Nodes with high degree and strength included AO4 (geographical environment occlusion), AO5 (loss of control/abnormal flight attitude), and AO7 (navigation system failure). Nodes exhibiting high betweenness centrality—AO3 (proximity to manned aircraft, 179.75), EE3 (geofencing failure, 155.5), and AO5 (123.9)—served as critical "bottlenecks" in risk propagation. Notable strong association rules (confidence > 0.8) included FB1 (beyond-visual-line-of-sight [BVLOS] operation) → AO1 (communication link interruption) (0.855), AO3 → AC1 (emergency avoidance) (1.000), EE2 (strong wind) → AO5 (0.891), and FB2 (regulatory avoidance) → AO3 (0.833). Robustness analysis revealed that the network is relatively resilient to random node or edge failures, with normalized reachability remaining above 0.85 even after 30% node removal. In contrast, targeted attacks based on DomiRank centrality demonstrated high structural fragility: removal of the top 20% of DomiRank nodes (such as AO7 navigation system failure and FB2 regulatory avoidance) reduced reachability to below 0.1. DomiRank-based edge attacks (AUC = 0.642) proved more effective than edge-betweenness attacks (AUC = 0.657). The most complex risk propagation community consisted of AO7, FB5, EE9, EE7, FB4, and AC2. Ultimately, eight key causal factors were identified: regulatory avoidance behavior, missing procedures, BVLOS operation, communication link interruption, proximity to manned aircraft, geographical environment occlusion, navigation system failure, and geofencing breakthrough. Conclusions: This study successfully integrates STAMP, association rule mining, and complex network theory to construct a directed weighted causal network of UAV disturbance accidents. The findings demonstrate that human operational behaviors and equipment/environmental factors dominate the critical causal chains, with certain hub nodes (e.g., navigation system failure and regulatory avoidance) exerting disproportionate control over risk propagation. Drawing on the eight key causal factors, the study offers hierarchical defense recommendations spanning technical equipment enhancements (redundant navigation/geofencing systems and real-time communication monitoring), operational management (restrictions on BVLOS operations, pilot qualification standards, and penalty mechanisms), regulatory enforcement (electronic fence infrastructure and big-data surveillance of regulatory avoidance), and institutional improvements (airworthiness certification and standardized accident data collection). These recommendations provide actionable guidance for airspace managers, airport operators, and UAV regulators. Future research will focus on dynamic cascading failure modeling and the development of risk propagation control strategies in low-altitude airspace.
Objective: Electrical fires occur frequently in complex environments, where environmental factors such as ambient temperature and wind speed contribute to fire occurrence and spread by changing the behavior of series arc faults. This study investigates the patterns of low-voltage AC series arc faults under the combined effects of ambient temperature and wind speed, with a focus on temperature field evolution, spatial heat transfer, current and voltage responses, and arc energy. Methods: A two-dimensional axisymmetric magnetohydrodynamic model is developed in COMSOL Multiphysics to simulate the thermal-fluid-electromagnetic coupling behavior of an AC series arc fault. This model integrates magnetic and electric fields, heat transfer, laminar flow, and an external circuit. The simulation domain consists of a copper electrode, a graphite electrode, a 3 mm arc gap, an ignition heat source, and the surrounding air. The circuit comprises a 5 Ω resistor and a 220 V/50 Hz AC source. Arc formation is initiated by setting the ignition heat source to 12,000 K. A two-dimensional orthogonal combined design is adopted, including three ambient temperatures (288.15, 298.15, and 308.15 K) and three wind speeds (0, 2, and 4 m·s−1). Average temperature, temperature integral, the root mean square (RMS) of current and voltage, and arc energy are selected as evaluation indicators. Observation points, arranged from the arc center to the outer region, are used to quantify spatial variations in the thermal response. Results: The results show that the arc temperature field expands and contracts with the power-frequency cycle, and this behavior is closely associated with voltage variations. In the absence of wind, the temperature field exhibits a spindle-like shape, with heat accumulating around the arc. At a wind speed of 2 m·s−1, the temperature field shifts in the direction of airflow, and the high-temperature region moves toward the graphite electrode. The effects of wind speed on the temperature field demonstrate clear spatial dependence. In the arc core region, temperature slightly increases with rising wind speed due to thermal contraction concentrating energy near the arc column. In the peripheral region, however, convective cooling dominates. At 4 m·s−1, the average temperature in the outer region decreases by more than 12.06%, and the temperature integral at wind-cooling-dominated observation points decreases by up to 42.27%. The current RMS remains stable between 34.4 and 34.7 A. The voltage RMS is more sensitive to wind speed than ambient temperature, decreasing by approximately 2.55 V when wind speed increases to 4 m·s−1. The arc energy remains stable at 2 m·s−1 and decreases slightly at 4 m·s−1. Ambient temperature has a limited effect on arc energy, exerting only a weak influence under strong wind conditions. Conclusions: Wind speed is the primary environmental factor controlling the temperature field and electrical response of low-voltage AC series arc faults, whereas ambient temperature has a limited effect. Wind enhances heat concentration in the arc core while increasing cooling in the outer region. The current conduction channel remains stable, but voltage and arc energy respond more distinctly to airflow disturbance. These findings provide a theoretical basis for early warning and prevention systems for electrical fires in complex environments.
Objective: The Guangdong–Hong Kong–Macao Greater Bay Area is geographically situated in a region frequently impacted by severe tropical cyclones, which pose persistent threats to infrastructure resilience and public safety. Among critical facilities, airport terminals—especially those located along coastal zones with high typhoon landfall probability—face significant challenges regarding structural integrity under extreme wind events. Wind uplift forces during typhoons can cause catastrophic damage to roofing systems, cladding elements, and even primary structural components if not adequately mitigated. Studies have highlighted the necessity of proactive monitoring mechanisms that capture real-time structural behavior under extreme wind loads. Nonetheless, integrated solutions tailored specifically for large-span airport terminals remain limited. This study addresses this gap by proposing a comprehensive structural safety monitoring and early warning system explicitly designed for coastal airport environments, with a focus on improving preparedness for wind disasters, enhancing structural performance, and minimizing operational downtime during extreme weather events. Methods: The study is grounded in the engineering context of Zhuhai Airport Terminal 2, a representative large-span spatial steel structure exposed to frequent typhoons. A multi-layered structural health monitoring framework was developed, comprising more than 250 sensor units across 14 distinct categories, strategically deployed to capture critical environmental and structural parameters. The sensing system includes three-dimensional anemometers for wind speed and direction measurements, differential pressure sensors for surface wind pressure, strain gauges and fiber-optic sensors for stress–strain monitoring, and accelerometers for dynamic response characterization. An advanced data acquisition and processing platform integrates safety assessment algorithms with intelligent alarm logic, enabling continuous evaluation of structural conditions. Notably, the system employs enhanced fiber-optic intelligent sensing reinforcements directly bonded to continuously welded roof panels, providing high-resolution strain measurements while improving resistance to wind uplift. Multiple sensor types are co-located at identical monitoring points to enable multi-source heterogeneous data fusion, thereby enhancing redundancy, accuracy, and robustness in data interpretation. Results: The proposed system was empirically validated during Typhoon Wipha in 2025, which made landfall near the Pearl River estuary with sustained winds exceeding design-level thresholds. The monitoring system successfully recorded detailed time histories of wind velocity, gust factors, and localized pressure peaks across different roof zones. The data indicate that maximum wind pressures occurred at roof corners and leading edges, consistent with aerodynamic theory, although their magnitudes slightly exceeded initial design assumptions. Strain measurements further show that the proposed fiber-optic reinforcement effectively reduced peak tensile stresses in roof panels, confirming its dual function in structural health monitoring and mechanical enhancement. Multi-source heterogeneous data fusion enabled precise correlation between wind field characteristics and structural dynamic responses, allowing the identification of potential overstress conditions before they reached critical levels. Within seconds of their detection, anomalous trends triggered intelligent alarms, demonstrating the system's responsiveness and reliability. Comparative analysis against conventional single-source monitoring approaches showed a marked improvement in detection accuracy and situational awareness. Conclusions: This study demonstrates that an integrated, sensor-rich monitoring and early warning system can significantly improve the resilience of coastal airport terminals to typhoon-induced wind hazards. By combining advanced sensing technologies, intelligent data fusion, and targeted structural reinforcement, the proposed framework not only delivers real-time safety assurance but also generates valuable datasets for design code and maintenance strategy refinement. The approach is readily adaptable to other large-span spatial structures in hurricane-prone regions, offering broad applicability in civil infrastructure protection. Future work may explore integration with predictive modeling tools, machine learning-based anomaly detection, and automated mitigation measures to further strengthen disaster preparedness.
Objective: Accurate identification of ship maneuvering motion models is essential for motion prediction, heading control, trajectory tracking, and intelligent navigation. However, ship maneuvering responses are affected by hydrodynamic nonlinearities, rudder action, environmental disturbances, and measurement noise. Traditional response models, such as the Nomoto model, are physically interpretable, but their simplified structure limits their ability to describe complex free-running maneuvering data. Purely data-driven models, such as long short-term memory (LSTM) networks, can learn nonlinear time-series mappings but may exhibit limited generalization performance when the validation sequence differs from the training data. To improve physical interpretability and prediction accuracy, this paper proposes a physics-informed LSTM (PI-LSTM) method for system identification of ship maneuvering response models. Methods: A second-order nonlinear Nomoto response model was selected as the physical prior. The continuous Nomoto equation was discretized using finite-difference approximation, and the resulting discrete coefficients were directly identified as model parameters. A parametric physical branch was constructed to describe the ship's main yaw dynamics, whereas an LSTM network was introduced as a nonparametric residual branch to compensate for unmodeled dynamics, environmental disturbances, and measurement errors that cannot be explicitly represented by the simplified Nomoto model. The PI-LSTM prediction was expressed as the sum of the Nomoto physical prediction and the LSTM-based residual correction. The loss function consisted of a data-fitting term, a physical branch term, and a residual regularization term. An adaptive weighting strategy based on the exponential moving average was adopted to dynamically balance these terms during training. Free-running zigzag test data for the Esso Osaka scaled ship model under three conditions (30°/30°, 20°/20°, and 15°/15°) were used for training and validation. Given the temporal continuity of the data, each dataset was divided chronologically, with the first 70% used for training and the remaining 30% for validation. The standard LSTM, the identified Nomoto physical model, and the proposed PI-LSTM model were compared using the coefficient of determination (R2), root mean square error (RMSE), and symmetric mean absolute percentage error (SMAPE). Results: The identified discrete Nomoto parameters remained stable in magnitude across the zigzag conditions and followed physically reasonable trends. Parameters related to historical yaw-rate states were consistent across the three conditions, indicating that the main yaw dynamics were reliably identified. The nonlinear yaw-rate coefficient and the rudder-input-related coefficient varied with rudder angle amplitude, reflecting changes in maneuvering excitation intensity. The time-domain results showed that all three models could fit the training segment well. However, in the validation segment, the standard LSTM showed amplitude attenuation or trend deviation in some conditions, especially the 20°/20° and 15°/15° tests. The Nomoto model remained more stable owing to its physical structure, but its simplified form limited prediction accuracy. In comparison, the proposed PI-LSTM model better reproduced the yaw-rate peaks, sign changes, and heading-angle trends in both segments. Quantitatively, PI-LSTM generally achieved higher R2 values and lower RMSE and SMAPE than the standard LSTM and the Nomoto model alone, especially in the validation segment. Conclusions: The proposed PI-LSTM method combines the physical interpretability of the second-order nonlinear Nomoto model with the nonlinear compensation ability of LSTM. It can simultaneously identify discrete maneuvering parameters and model nonparametric residual dynamics. Experiments on the Esso Osaka zigzag data show that the method improves the prediction accuracy and generalization performance of ship maneuvering response modeling, providing an effective approach for motion prediction, heading control, and intelligent navigation assistance.
Objective: Large-scale oil pool fires release intense thermal radiation, posing significant threats to personnel, equipment, and adjacent facilities. In actual accidents, such fires are often influenced by unsteady environmental wind conditions rather than ideal calm or steady wind. Wind fluctuations alter flame inclination, smoke diffusion, air entrainment, and the spatial distribution of thermal radiation intensity, thereby increasing uncertainty in personnel risk assessment. Existing research has primarily focused on no-wind or steady-wind conditions, leaving the coupled effects of wind speed and turbulence intensity insufficiently understood. This study investigates the thermal radiation and personnel risk associated with large-scale oil pool fires under unsteady environmental wind conditions based on numerical simulations. Methods: A square-shaped diesel oil pool fire model of 50 m2 was developed using Fire Dynamics Simulator. Unsteady environmental wind fields were generated by combining the Kaimal spectrum function with the harmonic superposition method, and the resulting wind velocity time histories were incorporated into the numerical simulations. Three wind speeds (5 m/s, 10 m/s, and 15 m/s) and five turbulence intensities (0%, 5%, 10%, 20%, and 30%) were utilized to generate 15 distinct cases. The 0% turbulence intensity case served as the steady-wind reference. Heat flux gauges were positioned at z=1.5 m and on the vertical section to capture the spatial distribution and temporal variation of thermal radiation intensity. The model was validated against full-scale controlled wind tunnel oil pool fire data for wind speeds of 5 m/s and 10 m/s. Representative gauge values at similar distances from the oil pool center were compared with experimental measurements to assess the reliability of the numerical simulations. Following validation, the effects of wind speed, turbulence intensity, and measuring distance on thermal radiation intensity were analyzed. Cumulative probability was introduced to quantify the stochastic fluctuations of thermal radiation intensity, and a logistic model was used to fit the relationship between thermal radiation intensity and cumulative probability. Finally, the thermal radiation intensity threshold model was combined with the simulated distribution to evaluate personnel safety distances under various wind conditions. Results: The numerical simulations yielded the following results: 1) Compared with full-scale controlled wind tunnel experimental data, the relative errors in the far-field region, critical for personnel risk assessment, ranged from 0.43% to 32.49%, with an average of 16.43%. Notably, 16 of 20 cases had errors within 25%. 2) The validated model showed that unsteady environmental wind increased the complexity of thermal radiation intensity variation. At low turbulence intensity, fluctuations were primarily controlled by flame entrainment and flame turbulence, whereas at high turbulence intensity, ambient wind turbulence was the dominant factor. 3) The relationship between thermal radiation intensity and cumulative probability followed a logistic model. For a given thermal radiation intensity, the cumulative probability decreased with increasing turbulence intensity. 4) Personnel safety distance was jointly influenced by wind speed and turbulence intensity. The downstream safety distance increased with wind speed and decreased with turbulence intensity, whereas the lateral safety distance exhibited the opposite trend. 5) The peak thermal radiation intensity generally occurred within 30 s after ignition, indicating that the early stages of fire development represent the critical period for personnel exposure. Conclusions: By incorporating unsteady environmental wind into the numerical simulation of large-scale oil pool fires, this study clarified the coupled effects of wind speed and turbulence intensity on thermal radiation intensity, cumulative probability, and personnel safety distance. The logistic model proved effective in characterizing the probabilistic distribution of thermal radiation intensity. The results show that assessments based solely on steady-wind assumptions may underestimate lateral and upwind risks, thus providing a more realistic foundation for hazard-zone determination and emergency evacuation planning.
Objective: As a critical environmental perception sensor, millimeter-wave radar offers notable advantages such as privacy protection and immunity to lighting conditions, making it well-suited to complex environments. However, accurately detecting static targets, such as furniture and walls, remains a substantial challenge because static objects lack relative motion, causing their Doppler shifts to approach zero so that they are easily filtered out as noise. Furthermore, traditional radar scene reconstruction methods, such as radar Simultaneous Localization and Mapping (SLAM), rely heavily on static point cloud accumulation, which provides only rough geometric outlines and fails to extract semantic categories or fine-grained three-dimensional (3D) bounding boxes. Existing vision-based scene reconstruction methods depend largely on RGB-D data and remain constrained by privacy and lighting issues. Although recent studies have attempted to use human–object interaction for scene reasoning, applying these vision-driven methods directly to millimeter-wave radar is difficult. Radar point clouds are inherently sparse and accompanied by high-frequency jitter, and directly using traditional absolute global coordinate features introduces cascading errors. Therefore, exploring how to effectively use the dynamic human posture information captured by millimeter-wave radar to accurately infer static indoor scene layouts is critically important. Methods: To address the problems of sparse point clouds and unstable absolute pose features, this paper proposed a novel millimeter-wave radar-driven indoor scene reconstruction method that indirectly infers static scene layouts from human motion patterns. The proposed model consisted of three core modules: spatiotemporal feature extraction, a scene-aware voting mechanism, and Gaussian mixture decoding. Initially, human activity point clouds were captured, and a 3D skeletal posture sequence was extracted. In the feature extraction stage, a centroid displacement encoding scheme was designed. Using the relative displacement trajectory of the human centroid rather than absolute joint coordinates, the model captures the macroscopic dynamic features of human–object interactions while effectively filtering out local high-frequency noise. Subsequently, a spatial graph attention mechanism and one-dimensional temporal convolutions were encapsulated into a stacked spatiotemporal residual module to extract deep interactive representations. In the scene-aware voting phase, the human centroid served as a seed position, and a learnable offset function calculated the center votes for potential interactive objects; these votes were then clustered and weighted into stable voting clusters. Finally, considering the inherent uncertainty of predicting scenes from single-frame postures, a hybrid prediction module was introduced. It used a Gaussian mixture distribution to model the 3D bounding box parameters (center, size, and orientation), generating diverse and plausible scene hypotheses that are jointly optimized by classification and Huber regression losses. Results: Extensive experiments were conducted on a self-constructed real-world millimeter-wave radar dataset encompassing 6 indoor spatial layouts, 12 interacting object categories, and 20,000 human posture sequences. The quantitative results demonstrated that the proposed method achieved an overall mean average precision (mAP) of 50.81% in 3D scene reconstruction. This performance surpassed that of mainstream visual scene reconstruction models, achieving a 4.50 percentage point improvement over the best baseline. Notably, for highly challenging categories such as "sofa," the mAP reached 78.17%. Comprehensive ablation studies confirmed the necessity of each component: introducing the centroid displacement encoding raised the mAP from 8.66% to 33.89%, while integrating spatiotemporal encoding and the hybrid prediction module further improved the accuracy to the final 50.81%. Furthermore, cross-room generalization evaluations using minimal matching distance and total mutual diversity metrics under different data-splitting strategies (S1 and S2) showed that the proposed multimodal decoding effectively balances prediction accuracy and scene generation diversity. Qualitative visualization results further indicated that the generated 3D bounding boxes were highly consistent with real environments in terms of structural configuration and spatial rationality, with no severe unnatural penetrations. Conclusions: The proposed scene reconstruction method based on centroid displacement encoding reduces dependence on absolute posture coordinates and overcomes the instability caused by the sparsity and local jitter of millimeter-wave radar point clouds. By capturing macroscopic motion trends and employing a hybrid multimodal decoding mechanism, the method significantly improves both reconstruction accuracy and prediction stability for multiple typical human–object interaction targets. In addition, it demonstrates strong cross-scene generalization capabilities. Although the system relies on the accuracy of frontend human posture extraction and may be affected by severe multipath effects in complex metallic environments, it establishes an effective framework and offers a novel perspective for intelligent indoor environmental perception using millimeter-wave radar. Future work will focus on multisensor fusion and robust feature extraction under severe multipath conditions to further enhance the system's engineering applicability.
Objective: Criminal cases constitute one of the primary threats to social security. Serious violent crimes, such as arson, pose a substantial risk to public safety. Moreover, the highly destructive nature of arson scenes and the vulnerability of evidence to damage or destruction significantly complicate fire investigations, making arson cases particularly difficult to solve, especially for investigators with limited experience. Faced with increasingly complex and evolving criminal patterns, traditional investigative approaches, primarily dependent on accumulated experience and manual screening, are no longer sufficient to meet the growing demand for rapid and accurate resolution of arson cases. Consequently, the development of a universal investigative reasoning model for arson cases has become a critical priority requiring urgent attention. Therefore, this study aims to construct a generalized investigative model to guide arson case investigations and support intelligent and standardized collection of evidence. Methods: The study first analyzed the key elements of arson cases by examining both the conditions under which the criminal act occurred and its resulting consequences. The conditions under which a criminal act occurs were defined as conditional elements; the criminal act itself was defined as the behavioral element, and the resulting consequences were defined as outcome elements. Subsequently, an element correlation model was constructed to analyze the relationships among conditional, behavioral, and outcome elements in arson cases. The element correlation model further decomposed physical evidence into two components: item-trace elements and their associated relationships. Then, a Bayesian network inference model was employed to establish an investigative framework for arson cases. Results: The proposed model was validated through analyses of representative case studies. First, a simulated case was used to illustrate the application procedure and the effectiveness of the model. Through this case study, the process of constructing Bayesian network nodes and their interconnections for arson investigations was examined. Bayesian probability calculations were performed to estimate the likelihood that each conditional element hypothesis was true. Furthermore, the practical case of the "Xiamen Bus Arson Incident" was used to demonstrate the model's effectiveness in addressing complex real-world cases. Through a phased analysis of available investigative data and the recommendations generated by the model, combined with comparisons against actual case conditions, the application process of the model was demonstrated. The result showed a high degree of accuracy in constructing the incident and supporting investigative reasoning. Conclusions: The results demonstrate that the proposed investigative model can effectively reconstruct incident scenarios and generate accurate investigative hypotheses. The model can assist criminal investigators in handling arson cases, improve investigative efficiency, and support the development of intelligence-led investigative systems. Notably, the investigative reasoning model proposed in this study focuses primarily on directly quantifiable crime scene elements and behavioral causal chains and does not yet incorporate more complex criminal motivations into the reasoning framework. Although the model primarily addresses reasoning processes during the investigation phase of arson cases, it provides limited consideration of the evidentiary standards associated with evidence collection. In practical applications, the model primarily provides analytical support during the preliminary stages of an investigation, facilitating the intelligent and standardized collection of evidence rather than replacing the expertise and judgment of professional criminal investigators.
Objective: As China's urbanization continues to accelerate, cities are expanding in size, populations are becoming increasingly concentrated, and the demand for energy is growing. Consequently, the construction of underground natural gas pipeline networks continues to advance, with steadily expanding coverage and increasing pipeline density. However, safety incidents have become more frequent during the construction and operation of gas pipeline networks, resulting in serious casualties, property damage, and the disruption of urban production and daily life activities. Urban gas pipelines traverse the underground space of cities, making it difficult to avoid intersections with or co-location alongside other municipal pipelines, such as those for water supply and drainage, district heating, and electricity. This is primarily due to limited underground space and a lack of unified planning and coordination among various municipal construction projects, which often leads to mutual interference between different pipeline systems during construction. This complex underground pipeline layout exposes gas pipeline networks to numerous safety risks. Methods: A model for gas accumulation in cable ducts with adjacent cable runs is established using Fluent, a powerful computational fluid dynamics software package with a rich set of physical models that enables high-precision numerical simulations of complex fluid flow and mass transfer processes. Model development considered various factors, including the actual structure of the cable duct, location and intensity of gas leaks, and ventilation conditions within the duct. Through reasonable simplifications and assumptions, a mathematical model was constructed that accurately simulated the real-world conditions. Subsequently, numerical simulations of gas accumulation within the cable duct were performed, and different operating conditions were included to systematically investigate the effects of varying the slope and ventilation rate on the patterns of gas accumulation within the duct. Results: The results indicate that the slope of a cable duct substantially affects the methane diffusion path, accumulation locations, and concentration field distribution. In the absence of a slope, methane is distributed uniformly around the leak point within 100 s. When a slope is present, methane diffusion is deflected by the slope, and the accumulation effect increases significantly as the slope steepens, making it likely for high-concentration accumulation zones to form beneath the supports at the bottom of the cable duct. Mechanical ventilation effectively reduces methane accumulation concentrations within cable ducts and alters the direction of methane diffusion; its suppression effect is influenced by the synergistic interaction between ventilation frequency and duct slope. Increasing the ventilation frequency improves the overall air exchange efficiency of the duct, causing the accumulation zone to shift from the top (under nonventilated conditions) to the bottom near the exhaust outlet. Conclusions: This study provides multidimensional technical support for the prevention and control of gas safety accidents in areas with overlapping utility lines, including real-time gas leak monitoring, intelligent hazard detection, and emergency response technologies, thereby effectively enhancing the precision and efficiency of gas safety management. In addition, it provides a comprehensive and in-depth scientific basis for the safe design of cable ducts, the layout of monitoring points, and the optimization of risk management strategies in engineering applications, thereby facilitating the development of safer and more reliable urban underground pipeline networks to ensure the stable operation of urban infrastructure and reduce risks to people's lives and property.
Objective: As a primary cause of road traffic injuries, fatigued driving requires efficient and accurate detection to improve traffic safety. Traditional single-signal approaches face limitations in capturing fatigue states, including high data collection intrusiveness, complex data structures, difficulty in real-time prediction, and low accuracy. Integrating information from different modalities has emerged as a new direction for development. Methods: This study developed a multimodal fatigue detection model for drivers by using electrocardiogram signals, vehicle trajectory data, and driver facial video data collected during long-term real-vehicle driving experiments. For ECG signal processing, an improved two-step adaptive filtering method was adopted for denoising, followed by time-domain and frequency-domain analyses to extract the driver's ECG feature set. For vehicle trajectory data, the quartile method was first used to remove outliers, backward filling was applied to fill in missing values, and two-dimensional discrete wavelet analysis was then employed for data denoising. For facial image data, the LabelMe annotation tool was used to construct a personalized training dataset by manually annotating each driver's face with at least 300 images per subject. The YOLOv8 deep learning model was then fine-tuned and trained on this dataset, and the optimal model weights were saved. Next, the optimized model was used to automatically annotate the remaining unlabeled images, and the Hopenet algorithm was applied to extract head pose angles from the cropped facial regions. The model incorporated modules for data processing, feature extraction, feature selection, and fatigue prediction, utilizing a self-attention mechanism to capture long-term dependencies and generate predictive outputs. Results: The model achieved a maximum accuracy of 97.89% in predicting fatigue state categories, with overall recall and F1 scores exceeding 80%, demonstrating strong predictive accuracy. The detection model utilizing data from all three modalities served as the control group, while six experimental groups were formed using a single modality or a combination of two modalities. The experiments revealed that the fatigued driving detection model employing data from all three modalities achieved optimal performance across various metrics. Conclusions: This study demonstrates that the proposed model successfully integrates information from different modalities, exhibits high accuracy and adaptability, and enhances the assurance of driving safety. Specifically, this model outperforms all comparison models in terms of accuracy, precision, recall, and F1 score, achieving the best performance in fatigued driving detection tasks, and maintains stable detection performance even under complex data structures, diverse sources, large time spans, average-quality facial images, and individual driver differences. The model achieves a maximum accuracy of 97.89% in identifying fatigue states, indicating that it correctly learns and recognizes fatigue patterns. Furthermore, models using a single modality or a combination of two modalities yield lower evaluation metrics than the three-modality model. Moreover, two-modality models consistently outperform single-modality variants, confirming that ECG signals, trajectory data, and facial images contribute positively to accurate fatigued driving detection.
Objective: Traditional longitudinal ventilation systems for tunnel fires typically adopt a fixed critical velocity for control design, which fails to adapt to the dynamic evolution of smoke characteristics during fire growth. To solve this problem, this study conducts a large-scale experimental investigation on the precise dynamic ventilation control of tunnel fire smoke. A closed-loop longitudinal ventilation control system based on the proportional–integral–derivative (PID) algorithm is proposed. Taking real-time temperature signals from detectors as control inputs and dynamic fan frequency regulation as control outputs, the system realizes the adaptive regulation of tunnel fire smoke. Based on a 1:5 geometric scaled large-scale tunnel test platform, systematic tests are performed to explore the suppression effect of smoke back-layering and the maintenance mechanism of smoke stratification during smoke migration. Methods: The test tunnel has an internal dimension of 260 m (length) × 2 m (width) × 2 m (height), and a 99% methanol pool fire is deployed as the fire source. Multipoint thermocouple arrays along the longitudinal direction, air velocity monitoring points, and a real-time fuel mass acquisition system are arranged to synchronously collect key parameters, including the temperature field distribution, longitudinal ventilation velocity, and fire heat release rate. A series of comparative experiments are carried out involving PID parameter tuning, system repeatability verification, and control point position variation tests. The control performance of the proposed system under diverse working conditions is clarified, and the influence of control point layout on smoke confinement efficiency is quantitatively analyzed. Results: The test results indicate that the optimized PID control system can dynamically adjust the longitudinal ventilation velocity in real time according to the temperature deviation at the detection point. Under various working conditions, the system effectively suppresses upstream smoke back-layering and maintains the stable stratification of downstream hot smoke layers. Repeatability tests demonstrate highly consistent temperature distribution characteristics and control effects, verifying the excellent robustness and repeatability of the proposed system. The control point position significantly affects the smoke control performance. When the control point is arranged near the tunnel ceiling and close to the fire source, the system exhibits faster response speed and higher control accuracy. In contrast, vertical downward offset of the control point aggravates the upstream migration trend of ceiling smoke and weakens the smoke front confinement capability of the system. When the control point is arranged axially farther from the fire source, the system can still confine smoke downstream of the target position; however, the decreased average ventilation velocity and heat exhaust capacity lead to higher temperature above the fire source and obvious control response delay. Conclusions: This study verifies the feasibility and effectiveness of the PID-based intelligent smoke dynamic control method for tunnel fires through large-scale model experiments. The research findings provide solid experimental support and technical references for the improvement of tunnel fire ventilation control theory and the engineering promotion of intelligent tunnel smoke control systems.
Objective: In disaster emergencies, edge cloud service resource pre-caching faces two major challenges. First, post-disaster service demands exhibit strong burstiness and temporal dynamics. Insufficient pre-cached resources may cause request queuing and prolonged response latency, whereas excessive pre-caching wastes computing resources and incurs additional operating costs. Second, because disaster events are rare, historical data are insufficient for newly affected regions, new service types, and newly deployed edge nodes. Conventional prediction and reinforcement learning methods therefore struggle to obtain effective pre-caching policies rapidly under few-sample conditions. To address these challenges, this paper proposes an edge cloud service resource pre-caching method based on meta-reinforcement learning (Meta-RL). Methods: Because user behavior data from real disaster scenarios are scarce, this paper analyzes service traffic in routine operational settings and trains the agent through interactions with daily service environments. The goal is to learn transferable workload evolution patterns and adapt quickly to emergency service modes. User request traffic is analyzed across Internet data centers (IDCs) and service types. The results show that traffic variations among IDCs exhibit correlations at monthly and intraday scales, indicating transferable common structures across regional workloads. Meanwhile, differences in request baselines, peak intensities, and fluctuation ranges reveal environmental heterogeneity. Across service types, different services share temporal trends but differ in request scale, peak duration, dispersion degree, and tail-load distribution. These observations indicate that daily service scenarios provide transferable experience, whereas data distribution differences hinder independent reinforcement learning training. Therefore, a Meta-RL framework is constructed to learn transferable policy initialization parameters through multitask training. Results: In the proposed framework, the edge cloud service resource pre-caching process is formulated as a Markov decision process. The agent determines the number of resources to pre-cache in the next time slot based on the current resource queue length and request waiting queue length. The reward function jointly constrains redundant cached resources and waiting requests, guiding the agent to balance service responsiveness and resource efficiency. To evaluate this trade-off, a comprehensive pooling indicator is designed from request hit capability and resource waste rate. The hit rate measures the ability of pre-cached resources to immediately satisfy user requests, whereas the waste rate characterizes the proportion of cached resources that remain unused. For policy optimization, proximal policy optimization (PPO) is adopted as the basic framework, and a meta proximal policy optimization (METAPPO) algorithm is developed. METAPPO performs task-specific adaptation through inner-loop updates and optimizes shared initial policy parameters via outer-loop updates, improving adaptability to new IDCs, service types, and emergency tasks. Conclusions: Simulation experiments compare METAPPO with PPO and other baseline methods across edge cloud service environments. The results show that METAPPO accelerates convergence across multiple scenarios. After a scenario change, METAPPO still exhibits favorable convergence performance, demonstrating that the meta-learning mechanism improves policy adaptation efficiency. Although some conventional methods may achieve competitive steady-state performance after sufficient training, they usually require more historical samples and longer training processes. Conversely, METAPPO leverages cross-task experience from daily service scenarios and rapidly generates effective resource pre-caching policies under few-sample conditions, supporting fast resource allocation adaptation in multiregion, multiservice, and newly emerging emergency edge cloud scenarios.
Objective: Thermal runaway propagation within lithium-ion battery modules endangers electric vehicles and energy storage systems. The state of charge (SOC) and charge/discharge rates are key operational parameters that considerably influence battery thermal stability. However, the combined electrothermal effects of these parameters on the propagation behavior of thermal runaway, especially under dynamic charging or discharging conditions, remain insufficiently understood. This study aimed to quantify how different SOC levels and charge/discharge rates affect the propagation characteristics of thermal runaway in ternary 18650 battery modules and to reveal the underlying electrothermal coupling mechanisms. Methods: Commercial ternary 18650 lithium-ion battery cells (nominal capacity: 2.6 Ah) were assembled into modules comprising four cells connected in series with a fixed spacing. Thermal runaway was initiated by locally overheating the first cell using a heating film. The experiments were conducted under a controlled ambient temperature of 25±2 ℃. Four SOC levels (25%, 50%, 75%, and 100%) were tested under static conditions (no current). For the charge/discharge rate tests, modules at 100% SOC were subjected to galvanostatic charging or discharging at 1 C and 2 C during thermal runaway propagation, with a no-current condition serving as the baseline. Temperature evolution was recorded using K-type thermocouples attached to the center of the surface of each cell and to the busbars. The voltage of each cell was monitored simultaneously. The module mass before and after each experiment was measured using a high-precision balance. Propagation time was defined as the interval between the onset of the temperature increase in the triggering cell and that in the adjacent cell. The trigger temperature and maximum cell temperature were extracted from the temperature curves. Each experimental condition was repeated three times to ensure reproducibility. Results: The results showed that charging accelerated thermal runaway propagation, whereas discharging delayed it. At charging rates of 1 C and 2 C, the average propagation time decreased by 7% (from 45.2 s to 42.0 s) and 12% (to 35.3 s), respectively, compared with the no-current condition. Concurrently, the trigger temperature of the adjacent cell decreased by 5.8 ℃ (1 C) and 12.4 ℃ (2 C), whereas the maximum temperature of that cell increased by 12.2 ℃ and 24.5 ℃, respectively. In contrast, at discharging rates of 1 C and 2 C, the propagation time increased by 17.3% (to 53.0 s) and 19.6% (to 54.1 s), respectively, compared with the no-current condition. The trigger temperature increased by 9.1 ℃ (1 C) and 10.6 ℃ (2 C), whereas the maximum temperature decreased by 15.4 ℃ and 18.9 ℃, respectively. The mass loss measurements indicated that higher SOC levels and charging rates aggravated electrolyte venting and internal material ejection. At 100% SOC, the accelerating effect of charging was most pronounced, whereas at a low SOC (25%), thermal runaway propagation was considerably suppressed regardless of the current condition. Conclusions: The observed differences originated from the chain-reaction propagation characteristics of thermal runaway and the opposing feedback polarities under electrothermal coupling mechanisms. During charging, Joule heating generated by the high current, combined with increased reversible capacity, enhanced internal side reactions (e.g., solid electrolyte interphase decomposition and anode–electrolyte reactions), forming a positive feedback loop that accelerated heat accumulation and thermal runaway propagation. During discharging, the capacity decreased as the cell delivered energy, reducing the available lithium inventory and mitigating the intensity of subsequent side reactions, thereby creating a negative feedback effect that slowed propagation. Higher charging rates (2 C vs. 1 C) amplified the Joule heating effect, further accelerating propagation. Higher discharging rates increased capacity reduction, but the propagation delay saturated above 1 C because of competing ohmic heating. These findings provide quantitative insights for designing safer battery thermal management systems and suggest that controlled discharging or low-SOC operation can serve as emergency mitigation strategies for delaying thermal runaway propagation. Future work should explore the influence of module configuration and cooling conditions.
Objective: Curved aramid fiber-reinforced composites are important load-bearing and energy-dissipating components in protective helmets and other shell-type protective structures. Unlike flat laminates, curved shells resist transverse impact through the coupled actions of local indentation, bending, membrane stretching, and geometric flattening. Therefore, impact indices obtained from flat panels cannot fully characterize the load transfer and damage evolution of a locally spherical shell. Specifically, the coupled effects of bending radius (R), ply count (L), and impact energy (E) on the transition from localized failure to global deformation remain insufficiently understood. This study aimed to clarify how these parameters govern the low-velocity impact response of curved aramid composites, establish a correspondence between response curve signatures and physical damage, and provide an experimental basis for the structural design of curved personal protective equipment. Methods: Curved aramid/epoxy laminates were fabricated from 1500D aramid prepreg on custom split stainless-steel spherical molds by vacuum bag/autoclave molding. The 200 mm × 200 mm prepreg sheets were laid onto the molds ply by ply, locally softened during forming to reduce wrinkling and trapped air, vacuum-sealed, and consolidated at a pressure ramp of 0.09 MPa/min to a maximum pressure of 3 MPa. A full test matrix was constructed using bending radii (R) of 80, 120, and 160 mm, ply counts (L) of 6, 8, and 10, and nominal E values of 12, 17, and 22 J. Drop-weight impact tests were performed according to ASTM D7136/D7136M using a 5.5 kg impactor fitted with a 16 mm diameter hemispherical tup; the curved specimens were restrained at four corners. Force–displacement, force–time, and energy–time histories were recorded. Maximum dynamic displacement (Umax), peak impact force (Fmax), time to peak force, and energy absorption ratio were extracted and compared across the parameter combinations. Post-impact damage morphologies were then correlated with the stages identified from the response curves. Results: Two reproducible response modes were identified. Mode Ⅰ, which occurred mainly in specimens with a larger radius or fewer plies, exhibited an initial slow increase and slight decrease, followed by a rapid increase to a relatively high peak and an abrupt post-peak drop. Its deformation was initially dominated by local indentation and matrix cracking; subsequent yarn stretching and breakage induced concentrated delamination and sudden loss of load-carrying capacity. Mode Ⅱ, favored by a smaller radius or a larger ply count, showed a gradual increase, a stable plateau, and a slow unloading stage. Wide-area bending, yarn rotation and stretching, interfacial sliding, and crack deflection across multiple interfaces enabled a greater proportion of the specimen to participate in load bearing, producing distributed and progressive damage. At fixed values of E and L, increasing R reduced Umax but increased Fmax. For the six-ply specimens, increasing R from 80 to 120 mm reduced Umax by 18.07%–21.77% across the three impact energies, whereas increasing R from 80 to 160 mm reduced Umax by 33.26%–34.98%. Under the 12 J, six-ply condition, the time to Fmax decreased from 33.82 ms at R = 80 mm to 23.76 ms at R = 120 mm and 17.66 ms at R = 160 mm. Thus, the flatter specimens developed a higher peak force more rapidly and with less global displacement but were more susceptible to abrupt local failure. The sensitivity to E also depended on the response mode. Relative to 12 J, Umax of the R80-L8 specimens increased by 37.57% at 17 J and by 70.38% at 22 J, while the corresponding increases for R80-L10 were 18.80% and 65.07%. When E was increased from 12 to 22 J, Fmax increased by 13.25%–64.03% for Mode Ⅰ configurations but remained within 743.79–776.59 N for Mode Ⅱ R80-L8 specimens, indicating a comparatively stable force response. The dissipated energy fraction generally increased with E, and at 22 J, it was approximately 90% or higher for most specimens. Increasing L, however, tended to reduce this fraction, indicating that a stiffer laminate retained a greater recoverable elastic energy share. For example, at R = 80 mm and E = 12 J, the energy absorption ratio decreased from 99.0% for six plies to 82.5% for eight plies, with a concurrent transition to Mode Ⅱ. The damage observations confirmed that a high dissipated energy ratio should be interpreted together with the spatial concentration and severity of irreversible damage. Conclusions: Bending radius and ply count jointly determine whether a curved aramid laminate resists low-velocity impact through localized high-load bearing or through global, progressive deformation. A larger radius reduces maximum displacement and increases peak force, but it also aggravates contact zone deterioration and promotes abrupt failure. A smaller radius facilitates load redistribution and global bending, although it allows a larger overall deflection. Additional plies improve cooperative load sharing and stabilize progressive damage, while the associated increase in stiffness can lower the dissipated energy fraction. Consequently, curved protective structures should not be optimized solely by peak force, displacement, or absorbed energy. Curvature and laminate thickness must be coordinated to balance force transmission, allowable deformation, and damage localization. The proposed two-mode classification connects measurable response curves with damage mechanisms and provides a practical framework for optimizing aramid composite helmet shells and related curved protective components.
Objective: Nickel–cobalt–manganese oxide (NCM) lithium-ion batteries are widely used in electric vehicles because of their high energy density. However, external heating, mechanical damage, internal short circuits, or other abnormal conditions can trigger thermal runaway, resulting in rapid temperature rise, fire, explosion, and the release of hazardous gases such as hydrogen (H2) and carbon monoxide (CO). Inside a closed transport container, heat and gases may accumulate rapidly, increasing transportation risks. However, conventional monitoring methods, which mainly rely on battery voltage, current, or contact temperature measurements, are difficult to directly apply to transported batteries. This study therefore investigates the thermal runaway characteristics of NCM cells and modules in a closed transport container and develops a long short-term memory (LSTM)-based early-warning method using multisource, noncontact sensor data. Methods: A thermal runaway experimental platform in a transport container was established, and NCM single cells and battery modules at 50% state of charge were selected as the experimental objects. Heating-triggered thermal runaway experiments were conducted under comparable conditions, and the valve-opening time, thermal runaway time, temperature variation, and characteristic gas release of the two battery forms were analyzed. Noncontact monitoring was achieved using electrochemical H2 and CO sensors and an infrared temperature sensor installed inside the container. The gas sensors detected characteristic gases generated by battery decomposition reactions, whereas the infrared sensor monitored battery temperature changes. The collected infrared temperature, H2, and CO time-series data were input simultaneously into an LSTM neural network, which extracted temporal correlations and synergistic changes among the multisource sensor signals. Based on these outputs, a risk probability criterion was proposed to determine whether the transported battery was in a thermal runaway risk state; the model output was used for binary risk-state identification rather than for classifying multiple risk levels or different thermal runaway stages. Results: The experimental results showed clear differences between the thermal runaway behaviors of the NCM cell and module. The module’s valve-opening and thermal runaway times were 770 and 896 s, respectively, which were earlier than the corresponding times of 890 and 1 037 s for the single cell. Owing to thermal interactions within the module, its thermal runaway onset and maximum temperatures were 98.0 ℃ and 672.8 ℃, respectively, which were lower than the corresponding values of 139.8 ℃ and 734.5 ℃ for the single cell. These results indicated that thermal interactions among cells substantially changed the module’s thermal runaway evolution, accelerating triggering and causing the module to enter a hazardous state earlier than the single cell. Therefore, NCM modules presented more prominent thermal runaway risks and imposed higher requirements on early-warning technologies. Infrared temperature and electrochemical gas sensors showed good detection capability during the early development of thermal runaway, with infrared temperature reflecting the external thermal response of the batteries and H2 and CO signals reflecting the internal gas-generating reactions. Together, these combined signals provided complementary information on battery temperature and gas release and were suitable for monitoring NCM cells and modules under transportation conditions. By integrating temporal changes in infrared temperature, H2, and CO signals and analyzing the synergistic variations in temperature gradients and characteristic gas concentrations, the LSTM-based multisource early-warning model accurately identified the thermal runaway risk state of NCM cells or modules. The model achieved an average warning lead time of 120 s, providing additional time for vehicle stopping, cargo isolation, cooling, fire suppression, and personnel evacuation. Conclusions: Thermal interactions within an NCM module substantially accelerate thermal runaway development in a closed transport container. Compared with a single cell, the module exhibits earlier valve opening and thermal runaway, as well as lower thermal runaway onset and maximum temperatures. Noncontact infrared temperature and electrochemical gas sensors can effectively monitor the thermal and gas-release characteristics of NCM batteries during transportation. The proposed LSTM-based method, which integrates multisource sensor data and determines thermal runaway risk using a risk probability criterion, enables accurate early warning for NCM cells and modules. The results provide a reference for intelligent monitoring and emergency response during the road transportation of NCM lithium-ion batteries.
