Lithium-ion batteries pose fire risks due to the potential for thermal runaway, yet the influence of abuse method on the resulting hazards remains insufficiently understood for large-format lithium iron phosphate cells. This study experimentally investigated hazards associated with thermal runaway of twenty 305 Ah prismatic lithium iron phosphate cells subjected to overheating and overcharging with pilot ignition to study the fire scenario. Data collected included heat release rate, total heat release, mass loss, CO and CO2 concentrations, smoke production, heat flux, flame behavior, and post-mortem analysis. The results showed that the progression toward thermal runaway depended strongly on the abuse method. Safety vent opening occurred after approximately 46 min during overheating and 103 min during overcharging, at average surface temperatures of 129 °C and 44 °C, respectively, and was associated with different release mechanisms and combustion behavior. Despite these differences, several characteristics of the resulting thermal runaway events were comparable under the investigated conditions, with peak surface temperatures of approximately 500 °C, total mass losses of 21%, total heat release of approximately 7–8 MJ and comparable post-mortem damage. Hazard characteristics were found to depend strongly on the failure stage. Venting under overheating abuse produced the highest peak heat release (284 kW) and smoke production rates (1.3 m2 s−1) , whereas overcharging was characterized by liquid electrolyte release. Furthermore, quantities commonly reported in battery fire studies, including effective heat of combustion, smoke production varied between abuse methods despite the use of the same cell chemistry. These findings highlight the importance of considering abuse method and failure stage when interpreting battery-fire experiments and developing engineering design fire scenarios.
The minimum ignition energy (MIE) is a parameter used to assess electrostatic hazards in industrial dust dispersions. This study evaluates the application of the FEST algorithm, a procedure originally developed for the characterisation of energetic materials, to lycopodium dust in a Hartmann tube. By constructing sensitivity curves based on the log-normal cumulative distribution, the FEST method, which applies the delivered energy concept according to ASTM E2019, is compared with the EN 13821 standard evaluation using the nominal energy concept, and testing according to ASTM E2019, using the delivered energy concept, is also discussed. Although the higher number of trials required for the FEST method can limit its routine industrial use, the approach complements standard testing by providing a probabilistic context to discrete energy values, making risk assessments more reliable. Due to the large number of trials required to achieve acceptable accuracy, the FEST method is recommended for use at the dust concentration at which the standard method achieved the lowest ignition energy.
Chemical accidents, including natural hazard-triggered technological (Natech) events, can cause toxic gas releases, posing significant inhalation exposure risks to on-site workers and emergency responders. Respirator performance depends not only on contaminant removal by the canister, but also on canister flow resistance, in-mask airflow, and breathing demand. To evaluate these coupled effects, this study developed a sequentially coupled CFD framework integrating a manikin head model, a breathing zone, a respirator, an equivalent porous-medium canister domain, and periodic breathing boundaries representing different activity intensities. Breathing-zone peak concentration, cycle-averaged breathing-zone concentration, exposure reduction efficiency, and canister pressure drop were used to assess the effects of equivalent canister porosity parameters and breathing intensity on contaminant transport and protective performance. Under the current equivalent-parameter settings, increasing the equivalent porosity parameter reduced both the peak and cycle-averaged breathing-zone concentrations, with the reduction from 0.40 to 0.882 in the equivalent porosity parameter reaching approximately 12.6% and 14.1%, respectively, indicating a nonlinear response. In contrast, increasing breathing intensity more strongly amplified breathing-zone exposure and markedly increased canister pressure drop. Overall, equivalent canister porosity parameters and breathing intensity jointly govern the exposure-resistance trade-off of respirators under accidental toxic gas scenarios. The results provide a quantitative basis for respirator performance comparison and parameter optimization in emergency toxic-gas protection, including Natech-related process-safety scenarios.
Hydrogen refueling station (HRS) is a critical infrastructure to promote the energy transition, ensuring HRS operation safety is essential since highly flammable hydrogen release may be inevitably triggered by disruptive events such as accidents or natural disasters. To prevent escalation of release accidents (e.g., fire and explosion), it is indispensable to enhance the resilience of HRS while limited attention has been paid on quantitative resilience of HRS. This study thus developed a quantitative resilience assessment methodology for HRS based on dynamic Bayesian network (DBN). Resilience is quantified as the probability of the HRS maintaining the normal operation state under disruption and recovering from a low functional state to the normal operation state. The performance of HRS is defined by four functionality states: disruption, absorption, adaptation, and restoration. A Markov chain model consisting of four functional states was developed, and the transition probabilities between states were determined by the mean time between failures and mean time to repair. By mapping the temporal processes of the four functionality states into the system’s functionality analysis, the Markov chain model is transformed into a resilience assessment model based on DBN. Furthermore, sensitivity analysis is conducted to identify critical factors influencing resilience. By hydrogen leakage scenario, the DBN-based methodology quantifies HRS resilience (stabilizing at 0.991) and identifies redundancy and robustness as most critical, with failure reducing resilience to 0.607. Priority upgrades include backup equipment, embrittlement-resistant materials, and optimized ventilation and leak detection. The framework adapts to other scenarios via fault tree reconfiguration.
Moisture is known to reduce the explosion hazards of organic combustible dusts via heat absorption, fuel dilution, and increased inter-particle cohesion. Its enhancing effects on certain light metal dust explosions are also recognized, attributed to hydrogen release or oxide film disruption. In industrial scenarios where dust explosion hazards coexist with humid environments, moisture can also affect the inert substances for their explosion mitigation performance. For example, rock dust applied to suppress coal explosions in mines may cake severely after prolonged humidity exposure; metal dust wetted by water may require an additional suppression level. To address these issues, this paper explores two contrasting strategies: rendering inert dust hydrophobic to resist moisture-induced cohesion, or employing a superhydrophilic material to agglomerate combustible particles.Results show that modified sodium bicarbonate (SBC) and magnesium hydroxide (MH) achieve water contact angles >110° and >120°, respectively, while a 20 wt% admixture of hydrophobic additive based on tetra-needle-like zinc oxide whiskers (ZnOw) also proved sufficient to impact hydrophobicity. The suppression effectiveness of the hydrophobic modified inert dusts on coal explosions is maintained after moisture exposure, whereas the raw materials exhibit a significant decline in mitigation performance. For the moisture-utilization strategy, aluminum-magnesium (Al-Mg) alloy dust premixed with sodium polyacrylate (PAAS) forms strong cakes upon moisture absorption; notably, it achieves complete suppression for Al-Mg explosions at 40% addition, whereas its dry form fails even at 90%. Further tests with coal, wood, and polyethylene dusts confirm the effectiveness and broad applicability of PAAS in humid environments. Overall, these results demonstrate the feasibility of two contrasting strategies for dust explosion mitigation and provide a proof-of-concept basis for their further development under humid conditions.
Although pure ethanolamines are hazardous substances widely used in chemical processes (e.g., gas sweetening and pretreatment agents), a dedicated assessment of their fire and degradation properties is missing. Within this scope, an experimental analysis was carried out by using a cone calorimeter to quantify the burning characteristics of the ethanolamines most widely used: monoethanolamine (MEA), diethanolamine (DEA), triethanolamine (TEA), and methyl diethanolamine (MDEA). The effects of an external heat flux on ignition temperature, maximum mass burning rate, mass loss rate, heat release rate, and exhaust gas composition of liquid samples were observed. The collected data were employed for the development and quantification of an apparent kinetic rate. The resulting rates were validated against empirical data from the literature, showing excellent agreement. Besides, the radical reactions most relevant for the activation of the investigated amines were identified. The experimental overall reaction rates were compared with theoretically based kinetics of the identified reactions to identify the rate-determining step in the process. The results indicate that the hydrogen abstraction from the nitrogen site by HO2 can be considered as the rate-determining step for the MEA and DEA cases, and the same agent is ruling the activation of TEA from a carbon site, highlighting the role of low-temperature chemistry in the ignitability of the ammines. These observations are useful for the accurate modelling and design of chemical processes involving ethanolamines, for the analysis of prevention and mitigation, and for the evaluation of pool fire likelihood and consequence assessment.
Dust explosions pose a hazard in various branches of the process industry, including facilities that produce, handle, and transport silicon and silicon alloys. This paper describes an experimental study of silicon dust explosions in pipes of varying diameter, representative of accident scenarios involving dust lifting and flame propagation in dust extraction systems. The experimental program comprised tests with silicon and ferro-silicon-magnesium in experimental configurations that comprised a 32-L vessel connected to one or several pipes. After distributing a uniform layer of dust in the pipes, corresponding to a given nominal dust concentration, flame propagation supported by dust lifting was initiated by dispersing and igniting a dust cloud in the vessel. Piezoelectric transducers measured the pressure development in the vessel and in specific positions along the pipes, and a video camera captured the flames emitted from the open end of the pipe. The results from the present study include pressure recordings obtained for dust explosions in a configuration with the 32-L vessel connected to four pipes: a 245 mm (inner diameter) pipe, a 157 mm pipe, and finally two 62 mm pipes. The maximum values of the recorded overpressures obtained with fine silicon dust were in the range 50 to 100 bar. Pressures of this magnitude were not observed for the same material in other configurations, nor for a less reactive dust in the same configuration. Because explosive dust clouds may be present inside process equipment during normal operation, and accumulated dust deposits can support flame propagation and flame acceleration in ducts and pipes, it is important for the design of process facilities, as well as systems for explosion protection, to understand the phenomena involved and be aware of the potential hazard.
Dense gases released at ground level without directed momentum, from a pool for example, typically disperse downwind while remaining adjacent to the ground. The HEGADAS model, a component of the HGSYSTEM suite of gas dispersion models, is designed to deal with this situation. Buoyant gases released at ground level, such as from pools of liquid hydrogen or liquid ammonia may lift off from the ground. HEGADAS, by design, cannot deal with this situation. The AEROPLUME model, another component of the HGSYSTEM suite of gas dispersion models, can deal with grounded gas clouds and can deal with lift off and plume rise. However, AEROPLUME was devised to be initialized as a jet and cannot be initialized as a grounded plume so a situation with a grounded gas cloud can only occur at some downwind location as part of the solution. If AEROPLUME was able to start from a grounded plume, then it would naturally capture the rise of the plume but some means would be required to supply initial conditions to the code. In this work updated versions of both HEGADAS and AEROPLUME have been created which allow greater flexibility. The new version of AEROPLUME allows initialization as a grounded plume. This is done starting at output from a calculation over the pool using the new version of HEGADAS. The resulting methodology is tested by comparison to wind tunnel and field experiments.
Hot work risk assessment in chemical industrial parks is limited by insufficient accident labels, subjective weighting, and weak real-time deployment. This study proposes an intelligent framework integrating Criteria Importance Through Intercriteria Correlation (CRITIC), VIseKriterijumska Optimizacija I Kompromisno Resenje (VIKOR), and Extreme Gradient Boosting (XGBoost). Guided by the accident causation 2-4 Model, a 16-indicator system covering management factors, inherent hazards, human factors, and on-site operational risks was developed. Based on 460 hot work records from 21 enterprises, CRITIC was used to calculate objective indicator weights, and VIKOR generated the comprehensive risk score Q as a continuous surrogate label for model training. Because the VIKOR-based surrogate label inherently penalizes degradation of any single safety barrier, the XGBoost surrogate not only achieved high predictive accuracy (R2 = 0.9815, MAE = 0.0246, RMSE = 0.0325) but also preserved the short-board sensitivity that dominates real hot work accidents, outperforming support vector regression and multiple linear regression. SHAP analysis further verified that the model’s internal logic was consistent with the original CRITIC-VIKOR framework (Spearman correlation = 0.8941). Compromise coefficient sensitivity analysis, indicator-score perturbation tests, data-partition robustness analysis, and single-sample timing tests confirmed that the proposed framework was robust to parameter settings, input-score uncertainty, and sample partitioning, while achieving millisecond-level prediction for rapid pre-approval risk assessment.
Traditional accident analysis methods struggle to process large volumes of unstructured accident texts and fail to fully excavate latent association patterns and co-occurrence relationships among risk factors. To address this limitation, a novel chemical accident risk analysis framework is proposed by integrating text mining and the GNN-Apriori association rule mining algorithm. First, TextRank, BM25, and BERT are combined via multi-strategy fusion to extract candidate keywords, which elevates keyword quality and enriches semantic feature representations. Second, a keyword co-occurrence network is built and Graph Convolutional Neural Networks (GNNs) are adopted for deep feature learning to mine implicit structural and semantic correlations between risk keywords. Subsequently, the GNN-Apriori association rule mining algorithm is built by fusing graph neural network embeddings into the classic Apriori algorithm. The GNN-Apriori framework overcomes the drawbacks of frequency-only association mining by incorporating structural topology extracted from keyword co-occurrence networks. It prioritizes statistically meaningful association rules that correspond to central nodes within the accident factor network, offering a new analytical perspective for exploratory accident investigation and safety governance. Finally, taking the “degree of improvement” as the evaluation metric, a correlation network model for accident risk factors is constructed based on mined association rules. This model supports visual presentation and in-depth analysis of inter-factor correlations. Experimental results verify that the proposed approach can efficiently capture structurally critical risk factors and their frequent co-occurrence combinations, enabling more scientific accident prevention and safety management practices.
Hydrogen is rapidly emerging as a cornerstone of the global clean energy transition, yet its unique physicochemical properties—wide flammability range, low ignition energy, high diffusivity, and invisible flame—pose substantial safety challenges across the entire value chain. Simultaneously, artificial intelligence (AI) and machine learning (ML) techniques may augment established safeguards through data-driven leak detection, real-time risk assessment, predictive maintenance, and physics-informed simulation. Following a structured, semi-systematic methodology, this review analyses the role of AI in hydrogen safety by examining applications in six key domains: (i) leak detection and sensor intelligence, (ii) dispersion modelling via ML–CFD surrogates, (iii) explosion and consequence prediction, (iv) probabilistic risk assessment, (v) fuel cell diagnostics and prognostics, and (vi) digital twin frameworks. The paper further evaluates physics-informed learning, reinforcement learning, natural language processing, and explainable AI from a process-safety perspective. Throughout, we distinguish demonstrated hydrogen-specific evidence from methods transferred from adjacent domains, and appraise reported performance critically, noting that high headline accuracies seldom address false-negative rates, dataset imbalance, uncertainty, or out-of-distribution behaviour. Key challenges including data scarcity, model interpretability, real-time deployment, and regulatory integration are discussed, and a research roadmap is proposed. Language-model applications are treated as expert-supervised knowledge assistance rather than substitutes for formal safety studies or regulatory decisions.
Process industry plants are crucial for society, as they are the key producers of chemistry, energy etc. during industrial processing. For this reason, loss prevention in industrial plants is of great importance, as loss of production facilities interrupts the continuing delivery of products. Beyond that, the owner of the plant will obviously face logistical, competitive, and economic challenges.Many industrial processing facilities include or handle hydrogen during some processes, (ammonia etc.). Hydrogen is highly combustible and has a wide range of flammable limits (4% −74.4%). This fact imposes a certain risk on the facilities, including the surrounding construction (walls/slabs). A severe hydrogen fire in a process industry plant could result in loss of structural load capacity and collapse of the entire building. Therefore, chemical and plant process safety must be a high priority for society.One of the challenges to be investigated more thoroughly is the thermal behaviour of the high-strength concrete structures used in industry plants, tunnels etc. The spalling behaviour of high-strength concretes is still not entirely predictable and limited knowledge is available for hydrogen impinging jet flames. This paper reports experimental results on the spalling behaviour of three concrete slabs at different compressive strengths under exposure of impinging propane gas flames and more severe high-pressure hydrogen jet flames. The slabs are tested under a compressive load. The results are discussed in terms of the spalling-reducing effect of polypropylene (PP) fibres.This study presents an experimental study on the thermal response and spalling behavior of concrete slabs exposed to impinging propane and high-pressure hydrogen jet flames.
Thermal runaway (TR) in lithium-ion cells produces high-velocity jets, flames, and hot particles that interact with nearby structures. This study transfers a CFD framework previously developed and validated for NMC TR configurations to NCA-based 18650 literature experiments, namely a cell venting toward a horizontal aluminum plate and an equivalent open-space jet. Internal heat generation is represented by an abuse-chemistry and internal-short model that triggers a two-stage vent with SOC-based composition. External combustion is modeled using a compressible Reynolds-averaged Navier-Stokes (RANS) approach with the k-ω shear-stress transport (SST), finite-rate/eddy-dissipation chemistry for H2/CO/CH4/C2H4/C2H6, P1 radiation, and a two-way-coupled discrete phase for solids including particle-wall sensible-heat deposition. The model is evaluated against under-plate thermocouple records and image-derived flame extension for the plate-obstacle case, and against image-derived flame length for the open-space jet. Results show that the model reproduces the main under-plate temperature trends and timing in the plate-obstacle case, while the same flame-extension metric captures the mixing-controlled flame envelope in both confined and open-space configurations. Using a single, input-driven framework across sealed, open-end, and top-confined configurations supports its potential use as a design-support tool for vent routing, stand-off, and shielding, provided future work refines time-resolved vent composition and wall-deposit evolution for realistic module-scale applications and safety assessments in transport, stationary storage, aviation, and maritime systems.
The accumulation of wood dust in dust collection ducts of wood processing enterprises represents a critical safety concern related to dust explosion hazards, and accurate identification of hazardous accumulation states is essential for effective monitoring and risk mitigation. Existing assessment approaches mainly rely on physical mechanisms or single-feature analysis, which are limited by insufficient feature weighting strategies and weak integration between feature interpretability and feature selection. This study proposes an interpretable image feature-driven framework integrating feature extraction, multi-model SHAP-based feature selection, subjective–objective weight fusion, and threshold determination for hazardous dust accumulation identification. A total of 3448 high-resolution images of dust deposition inside dust collection ducts were collected. Based on a dust thickness threshold of 1 mm, 30-dimensional visual features were extracted from the collected images. Subsequently, the ReliefF algorithm combined with multi-model SHAP analysis was employed to identify key features. Specifically, SHAP values derived from four machine learning models, including Random Forest (RF), Support Vector Machine (SVM), Extreme Gradient Boosting (XGBoost), and Gaussian Naive Bayes (GNB), were integrated with ReliefF results to determine 12 critical visual features. Furthermore, subjective weights calculated using the G1 method and objective weights obtained using the entropy weighting method were integrated, and the particle swarm optimization (PSO) algorithm was applied to optimize the fusion coefficient for constructing the final feature weights. Finally, the integrated feature weights were used to establish the risk score-based identification frameworkRR, and the RF model was employed to determine the optimal discrimination threshold (R∗ = 0.3561 R^* = 0.3561 R∗ = 0.3561) for hazardous dust accumulation identification. Experimental results demonstrate that skewness, fractal dimension, contour area, and grayscale standard deviation (gray_std) are the most important features for distinguishing hazardous dust accumulation states. The proposed framework enables effective identification of critical visual features, reduces feature redundancy, and provides an interpretable technical approach for image-based hazardous dust accumulation identification under limited sample conditions.
Industrial fires often involve flammable materials, high-pressure equipment, and complex spatial layouts, posing significant risks of fire escalation, explosion, and toxic release. Effective and timely suppression is therefore critical to preventing major accident scenarios and mitigating consequential losses. Although robotic firefighting systems offer a safer alternative to manual intervention in hazardous environments, their performance is fundamentally constrained by the accuracy and consistency of visual perception. Existing approaches typically treat fire detection, flame segmentation, and water-jet landing-point localization as independent tasks. This separation limits the consistency of the visual feedback available for downstream spray-direction adjustment To address this gap, this study proposes FireJet, a coupled visual perception system that integrates these three tasks in a unified video-processing architecture. FireJet uses an edge–cloud design in which YOLO-Fire performs lightweight fire detection on the robot, while first-frame-prompted SAM2 segmentation and water-jet landing-point localization run in the cloud. The coupled outputs provide spatial feedback that can be used by a downstream spray controller Experiments on DetectiumFire and the evaluated industrial-like scenarios showed consistent performance across the three perception tasks. These results support FireJet as an engineering-oriented perception module, while closed-loop targeting and suppression efficiency remain to be validated. Our project homepage is https://widemountfirejet.github.io/.
