Fire Risk Assessment of Lithium-Ion Power Battery Shipping Containers in Maritime Transportation Scenarios

Abstract

As the demand for maritime transportation of power battery shipping containers grows rapidly, the incidence of fire accidents has increased in tandem. However, most studies focus on analyzing fire causes through the thermal runaway mechanism; few analyze fire risk across the full maritime transportation process from a safety science perspective. To fill this gap, based on the thermal runaway mechanism of lithium-ion batteries, this study couples the loading characteristics of shipping containers with maritime operating conditions and employs the Fault Tree (FT) model, Bayesian Network (BN) model, and Attack–Defense Game Theory for investigation. The results are as follows: Starting from three core factors—battery thermal runaway mechanism, scenario characteristics of shipping container maritime transportation, and failure of initial emergency response—and combining the FT model, it qualitatively identified and systematically sorted accident-causing factors. Via the FT-BN conversion criteria and expert assessment results, the fire probability of po’wer battery shipping containers on the target route was calculated to be 35%. According to Attack–Defense Game Theory, two key risk evolution pathways were identified with occurrence probabilities of 3.77% and 4.35%, respectively. Meanwhile, their action mechanisms were elaborated on, and the targeted preventive measures were proposed. This study provides theoretical support and methodological reference for the systematic assessment of fire risks associated with power battery shipping containers in maritime scenarios.

1. Introduction

The total global shipment volume of lithium-ion batteries reached 1545.1 GWh in 2024, increasing by 28.5% year-on-year [1]. In the cross-regional transportation of lithium-ion power batteries, maritime transportation is one of the core transport modes, featuring significant advantages such as large shipping capacity, low cost, low energy consumption, and low pollution [2,3,4,5]. However, lithium-ion battery-related accidents have occurred frequently in recent years [6,7], including the 2020 lithium-ion battery shipping container accident involving COSCO Shipping [8], the 2021 electric ship accident of MS Brim, the 2023 lithium-ion battery fire accident aboard MV Genius Star XI, and the 2024 lithium-ion battery container fire accident at the Port of Montreal, Canada. These accidents not only caused significant casualties and property losses but also attracted widespread attention from all sectors of society. Therefore, it is necessary to conduct fire accident risk analysis for lithium-ion power battery shipping containers in maritime transportation scenarios.

In the field of lithium-ion battery fire research, scholars have conducted in-depth investigations on core factors associated with shipping containers via maritime transportation, including state of charge (SOC) [9,10,11,12], storage configuration [8,13,14,15], fire-extinguishing agents [16,17,18], and confined spaces [19,20,21,22]. For instance, Yan et al. [9] investigated the thermal runaway and combustion behaviors of lithium-ion batteries under different SOCs in a pool fire environment, and further analyzed the thermal safety performance of lithium-ion batteries in hybrid ships. Liu et al. [8], on the other hand, studied the influence of storage configuration on the self-heating and spontaneous combustion behaviors of batteries during transportation. They proposed that reserving gaps between battery layers and implementing effective cooling measures can significantly increase the critical temperature for self-heating and spontaneous combustion. Cao et al. [16] examined the suppression effect of liquid nitrogen on the thermal runaway of lithium-ion batteries under varying conditions of heating power, injection direction, SOC, and injection timing. The results demonstrated that liquid nitrogen exhibits excellent suppression performance, capable of rapidly reducing the surface temperature of batteries.

Beyond these investigations, some researchers have also carried out fire risk analyses of lithium-ion batteries in maritime transportation scenarios [23,24,25]. Fault Tree (FT) and Bayesian Network (BN) models are the two most commonly used methods in risk analysis [23,26]. The FT enables qualitative analysis and logical organization of accident-causing factors [27], while the BN facilitates the determination of influence weights of different factors via quantitative calculations [28]. Zhang et al. [23] focused on fire accidents involving lithium-ion battery energy storage systems under maritime transportation conditions. By applying FT-BN models, they pointed out that high humidity and the absence of real-time alarm functionality in monitoring devices are the most critical factors affecting fire incidents. Jiang et al. [24] conducted a risk assessment on fires related to the transportation of lithium-ion electric vehicles via ro-ro ships (roll-on/roll-off ships), also via FT-BN models. They found that the highest risk concentrates in the fire ignition stage, and accordingly put forward improved recommendations tailored to this scenario. Furthermore, the core of Attack–Defense Game Theory aligns closely with risk analysis. A central tenet of this theory is that the weakest link in a system represents the critical point most susceptible to attacks, which facilitates the theory’s in-depth integration with risk analysis [29,30]. For instance, by integrating Attack–Defense Game Theory, Li et al. [31] systematically evaluated the potential risk pertaining to the transportation of hazardous materials amid terrorist attack threats.

In summary, while existing studies have yielded fruitful outcomes, two aspects remain to be further refined: first, most studies prioritize exploring the thermal runaway characteristics of lithium-ion batteries under different influencing factors, and have not yet conducted fire risk analysis for maritime transportation scenarios from the perspective of the entire process (encompassing production, transportation, and loading–unloading), integrated with accident evolution mechanisms; second, a limited number of relevant risk analyses mostly focus on assessing maritime transportation risks of lithium-ion battery energy storage systems and electric vehicles, yet specialized research on lithium-ion power battery shipping containers remains relatively scarce.

Therefore, this study focused on fire accidents involving lithium-ion power battery shipping containers in maritime transportation scenarios, relied on the thermal runaway mechanism of lithium-ion batteries as the theoretical basis, and conducted a systematic analysis by integrating the unique environmental conditions of shipping containers in maritime transportation scenarios. Employing the Fault Tree-Bayesian Network (FT-BN) model, this study performed qualitative analysis and quantitative assessment of the identified hazard-causing factors, and concurrently clarified the key risk evolution paths by incorporating Attack–Defense Game Theory. The research findings not only provide a systematic risk assessment basis for fire accidents during the maritime transportation of lithium-ion power battery shipping containers, but their quantitative results and key risk evolution paths also offer targeted recommendations for accident prevention. This, in turn, enhances the safety of lithium-ion power battery maritime transportation and ensures the safety and efficiency of shipping.

2. Methods

To conduct a more comprehensive and precise study of lithium-ion battery shipping container fire accidents in the maritime transportation context from a safety perspective, this study incorporated the Fault Tree (FT) model, Bayesian Network (BN) model, and Attack–Defense Game Theory to conduct qualitative analysis and quantitative evaluation of such accidents and ascertain key risk pathways. The specific research approach was illustrated in Figure 1: First, the FT model was employed by incorporating the thermal runaway mechanism of lithium-ion batteries and the unique operational conditions of the maritime transportation context to qualitatively identify the causal factors of the accidents. Second, the FT-to-BN conversion approach was applied to transform the results of FT analysis into a BN model and perform quantitative calculations, in turn enabling the achievement of quantitative evaluation of the causal factors. Finally, by integrating the Attack–Defense Game Theory with the aforementioned quantitative analysis results, the key risk pathways were ascertained. The models and methods utilized in this study were elaborated upon in detail in the following sections.

A supplementary note was warranted here: why was it still necessary to introduce the Attack–Defense Game Theory for risk analysis subsequent to the application of FT-BN? Specifically, in conventional risk analysis scenarios, while sensitive analysis of BN could identify root nodes exerting the most significant influence on the probability of the top event. However, this approach only clarified critical risk points and their evolutionary trajectories, lacking the capability to conduct a quantitative assessment of the occurrence probabilities of these paths.

The core tenet of Attack–Defense Game Theory was that the node most susceptible to attacks in a system constituted its weakest link, a characteristic highly aligned with the essence of key risk paths. It was worth emphasizing that BN sensitivity analysis relied primarily on posterior probabilities when identifying critical root nodes. In contrast, following the integration of Attack–Defense Game Theory, the identification of key evolutionary paths was grounded in the prior probabilities of individual root nodes. While both represented extended applications of the BN computational framework, the latter could delineate the occurrence probability of each risk path by incorporating the dynamic evolutionary process of accident chains. Based on this attribute, by comparing probability discrepancies across accident chains at different levels, the risk chain most likely to lead to the top event was identified, ultimately achieving dual precise characterization of key risk evolution paths: namely, “what they are” (qualitative) and “how probable their occurrence is” (quantitative).

The following section provides a detailed overview of the methodologies employed.

(1)

Fault Tree model

The FT was a system reliability and risk analysis method grounded in deductive reasoning. By establishing logical causal relationships, it retroactively traced all potential contributing events from the “top event”—the incident that could lead to system failure. This model was widely employed in risk early warning and mitigation across diverse industries. Its core components fell into four categories: top events, intermediate events, basic events, and logic gates.

A top event represented a system failure incident that demanded prioritized prevention. An intermediate event was a transitional occurrence, triggered collectively by multiple subordinate events through logic gates. A basic event refers to an underlying contributing event that could not be further decomposed. Logic gates primarily consisted of two core logical relationships: “AND gates” and “OR gates.” Specifically, an “AND gate” signified that the output event was only triggered if all input events occurred; an “OR gate” indicated that the output event was triggered when any input event occurred.

The primary purpose of applying the FT was to visually illustrate the inherent logical relationships between various contributing factors using graphical representations. This, in turn, enabled qualitative analysis of fire incidents involving lithium-ion power battery shipping containers in maritime transportation scenarios.

(2)

Bayesian Network model

In 1988, Judea Pearl proposed the Bayesian Network model, building on a synthesis of prior research. As a probabilistic graphical model, its core consisted of a directed acyclic graph, which was theoretically based on the integration of probability theory and graph theory. It characterized uncertainties within a system through probability distributions. Not only did this model exhibit robust probabilistic reasoning capabilities, but it also clearly revealed causal relationships between variables; meanwhile, it could more accurately calculate system failure probability and the posterior probabilities of basic variables by utilizing the belief propagation algorithm and evidence update mechanism. Moreover, the construction of BN was often based on FT, and the specific conversion schemes for this process would be elaborated in subsequent applications.

The core advantage of applying the Bayesian Network model to accident analysis lies in its capacity to advance from qualitative analysis to the stage of quantitative analysis. Specifically, by evaluating the prior probabilities of respective basic events and employing quantitative calculation methods, the overall occurrence probability of fire accidents was derived. Meanwhile, the contribution degree of each basic event to accident occurrence was quantified—a step that in turn provided a quantitative basis for in-depth analysis of accident mechanisms and implementation of targeted risk management and control.

In the context of the Bayesian network model, the core formula was that for conditional probability, as follows:

P(A|B) = P(B|A) × P(A)/P(B)

(1)

where P(A|B) denoted the posterior probability; P(B|A) represented the likelihood; P(A) stood for the prior probability; P(B) indicated the evidence probability.

In prior analyses of safety engineering scenarios, P(B) was typically not provided directly. Instead, it required derivation via the law of total probability, which followed:

P(B) = P(B|A) × P(A) + P(B|¬A) × P(¬A)

(2)

where ¬A denoted the complementary event of A; P(¬A) represented the probability of A not occurring; P(B|¬A) indicated the conditional probability of B occurring given that A had not occurred.

When Equation (2) was substituted into Equation (1), Equation (1) could be rearranged to:

P(A|B) = [P(B|A) × P(A)]/[P(B|A) × P(A) + P(B|¬A) × P(¬A)]

(3)

(3)

Attack–Defense Game Theory

The Attack–Defense philosophy served as a core concept in strategic and tactical domains. Its essence lay in opposing parties adopting defensive, attack, or hybrid strategies and measures to safeguard their own interests, diminish the opponent’s capabilities, and ultimately achieve their respective strategic objectives. Within Attack–Defense Game Theory, the most critical element was the Nash equilibrium: it mandated that both parties anticipate and incorporate the opponent’s potential strategies into their decision-making processes, thereby formulating their respective optimal strategies. This “prediction-response” game logic had been adapted to risk analysis scenarios; by applying Attack–Defense Game Theory and integrating it with the Bayesian network model, the key risk evolution pathways for fire accidents could be delineated.

The specific implementation steps were as follows:

(a)

The initial assessment probabilities (i.e., prior probabilities) of root nodes at the same hierarchical level were determined.

(b)

A specific root node was assumed to have occurred, with its occurrence probability set to 100%. Using this as the analytical starting point, the conditional probabilities of the corresponding next level were clarified.

(c)

A layer-by-layer iteration method was employed to construct a complete causal network for the causal branches of each root node.

(d)

The prior probability of each root node was cumulatively multiplied stepwise with the posterior probabilities of its subsequent levels, yielding the overall occurrence probability of the corresponding induced terminal event.

(e)

All overall occurrence probabilities were compared, and the root node with the highest probability value was selected; the extended pathway of the causal branch corresponding to this node was identified as the key risk evolution pathway.

In the identification and quantitative calculation of critical risk evolution paths, the application of Attack–Defense Game Theory relied on two core elements: the first being the prior probabilities of each root node in Bayesian network models, and the second being the evolution chains of accident-causing factors, which were constructed based on fault tree models.

Integrating these two elements, with the prior probabilities of each root node as the calculation basis, and by coupling the dynamic evolution characteristics of accident chains with the quantitative calculation methods of Bayesian networks, the accurate calculation of the occurrence probability of accident chains (triggered by each root node leading to the top event) was enabled. Among the results, the path with the highest occurrence probability corresponded to the most vulnerable nodes of the system from the perspective of Attack–Defense Game Theory, while within the risk analysis framework, this path was clearly defined as the critical risk evolution path.

3. Results

Based on the fundamental theory of fault tree models, a fault tree model for fire accidents involving lithium-ion power battery shipping containers during maritime transportation was constructed, as shown in Figure 2 and

Table 1. The specific meanings of each symbol in the fault tree.

Symbol	Meaning	Symbol	Meaning	Symbol	Meaning
T	Fire accidents of lithium-ion power battery shipping containers in maritime transportation	A1	Thermal runaway in power battery shipping containers causing open flames	A2	Failure to handle the initial stage of thermal runaway, leading to further fire expansion
B1	Abnormality in the quality and status of the power battery itself	B2	Effect of internal and external inducing factors on the container	B3	Failure of the container protection system in waterway transportation
B4	Early fire detection and warning failure	B5	Failure of the container and ship fire extinguishing system	B6	Lag in emergency rescue response for waterway transportation
C1	Substandard manufacturing quality of power batteries	C2	Abnormality in the pre-delivery storage/pretreatment status of the power battery	C3	Violation of the pre-transportation loading status of the power battery
C4	Improper operation in container loading and unloading	C5	External mechanical impact during ship navigation	C6	Loss of control of environmental parameters inside the container
C7	Electrical connection failure of the power battery pack	C8	Impaired structural integrity of the container	C9	Failure of special protective accessories for the container in the waterway
C10	Fault of the fire detection device built in the container	C11	Inadequate on-site personnel monitoring and inspection	C12	Failure of the container-built-in fire extinguishing system
C13	Ship deck fire-fighting equipment fails to provide effective coverage	C14	Failure of emergency organizational coordination within the vessel	C15	Delay in the linkage of shore-based rescue forces
C16	Restriction of rescue operations by the waterway environment	D1	Deficiencies in power battery design	D2	Counterfeit power batteries
D3	Deficiency in quality control during the production process	D4	Excessive temperature and humidity in the power battery storage environment	D5	Overcharge/discharge during power battery storage
D6	Power battery storage period exceeds shelf life	D7	Mixed loading of power battery cells	D8	Loose fixing of the battery pack
D9	Failure to remove batteries with apparent damage during loading	D10	Personnel have not received special training on power batteries	D11	Over-limit operation of the shore bridge/gantry crane
D12	Container collision occurs during stacking	D13	The ship encounters severe sea conditions, inducing severe rolling	D14	Collision between the ship and the dock during berthing/departing
D15	Loss of stability and collapse of container stacks	D16	Failure of the container temperature control system	D17	Blockage of the container ventilation device
D18	Excessive humidity inside the container caused by water accumulation	D19	Insulation layer breakage and short-circuit of the battery pack connecting wires	D20	Poor contact of the external power supply lines connected to the battery pack
D21	Failure of the battery management system	D22	Substandard structural strength of the container at the factory	D23	Uninspected and unrepaired dents/cracks on the container body before transportation
D24	Container wall perforation due to seawater corrosion	D25	Aging and cracking of the waterproof sealing strip of the container	D26	Blockage/damage of the waterproof and air-permeable valve
D27	Failure to install anti-seawater immersion pads at the bottom of the container	D28	Wrong selection of fire detectors	D29	False alarm/missed alarm caused by non-periodic calibration of the detector
D30	Fracture of the detection signal transmission line due to vibration	D31	No scheduled inspection system formulated	D32	Omission in inspection by patrol personnel
D33	Inspection interruption in low-visibility environments	D34	The fire extinguishing device expired	D35	Improper selection of the fire extinguishing medium
D36	Insufficient pressure of the ship’s fire hydrant	D37	Aging and leakage of the fire hose/interface	D38	The fire extinguishing range cannot cover the upper containers
D39	Internal communication system failure of the vessel	D40	Insufficiency of crew emergency drills	D41	Fire occurrence area far from the coastline
D42	Wind waves/tides hinder rescue vessels’ approach	D43	The container fire spreads and causes rescue channel blockages	E1	Inferior thermal stability of battery cathode materials
E2	Short service life of SEI film design	E3	Insufficient compressive strength of the battery casing	E4	Absence of an effective quality management system certification
E5	Failure to implement the latest power battery production standards	E6	Loosening of container stacking and fixing fixtures	E7	Failure of anti-skid measures on the ship’s cargo hold deck
E8	The stacking height exceeds the ship’s stability requirements

. Due to space constraints, Figure 2 only employed symbols to denote the accident’s causing factors, while the specific meanings of each symbol were detailed in Table 1.

It should be noted that the accident-causing factors of the fault tree model concerning lithium-ion power battery fires in the maritime transportation scenario, illustrated in Figure 2, originated from two aspects. First, they were explored based on the direct and indirect causes extracted from existing accident cases. Second, they were expanded and developed by integrating scenario-specific characteristics of maritime transportation, including high-salt and high-humidity conditions, ship perturbations, the thermal runaway mechanism of lithium-ion batteries, container protection system failures, and early-stage emergency response failures. Thus, the analysis of accident-causing factors was not only grounded in existing accident cases but also aligned with the unique environment of the maritime shipping scenario.

This model took “fire accidents of lithium-ion power battery shipping containers in maritime transportation” as the top event. The first-level intermediate events were defined as “thermal runaway in power battery shipping containers causing open flames” and “failure to handle the initial stage of thermal runaway leading to further fire expansion,” with a logical AND gate connecting the two.

3.1. Factors Resulting from the Battery Quality and Maritime Transportation

For such fire accidents, the occurrence of thermal runaway in power battery shipping containers and subsequent ignition of open flames was the core prerequisite. This prerequisite mainly stemmed from the synergistic effect between the quality defects of the power batteries themselves and the external inducing factors unique to maritime transportation—the former laid inherent risks for thermal runaway, while the latter provided external triggering conditions for the exposure and development of such risks.

(1)

Sources of quality risks in batteries themselves

The quality risks of lithium-ion batteries themselves mainly originated from the confluence of three stages: production, storage, and loading, as shown in Figure 3A.

Production stage: First, design defects, such as poor thermal stability of cathode materials and short lifespans of the Solid Electrolyte Interface (SEI) film, directly reduced the thermal runaway threshold of batteries. Second, management oversights, such as the lack of effective quality management system certification and failure to follow the latest production standards, led to unqualified batteries entering the transportation link.

Storage stage: Excessive temperature and humidity accelerated the degradation process of batteries; overcharging/overdischarging triggered lithium dendrite precipitation; and issues such as storage periods exceeding the shelf life further exacerbated the residual risks inherited from the production stage.

Loading stage: Mixed loading of new and old batteries, loose fixing devices, and neglected inspection of batteries with external damage not only created risks for battery structural damage in a vibrational environment but also aggravated the overall quality risks.

(2)

Impact of maritime transportation on power battery shipping containers

Compared with transportation modes such as railway and aviation, the long-term high-salt and high-humidity environment, ship propulsion-induced vibrations, and large-scale loading/unloading operations of maritime transportation exerted continuous impacts on power battery shipping containers, as displayed in Figure 3B.

First, ship roll, pitch, and heave generated dynamic impacts on shipping container stacks: if securing ropes loosened, deck anti-slip measures failed, or stacking height exceeded stability design limits, stack displacement and collapse were likely to occur. This damaged battery casings, electrode sheets, and insulation layers of connecting wires, inducing short circuits and heat generation; even without collapse, long-term low-frequency vibrations accelerated internal reactions in defective batteries, shortening the time to thermal runaway.

Second, long transportation cycles led to the formation of a microenvironment characterized by “heat accumulation and moisture retention” in sealed shipping containers: ventilation devices clogged by salt accumulation impaired battery heat dissipation, thereby creating a high-temperature environment—a critical condition for electrolyte vaporization and electrode material decomposition. Water accumulation in the hold exacerbated high environmental humidity, accelerating corrosion of battery casings, reducing the breakdown strength of insulation layers, and ultimately inducing leakage and overheating issues.

Finally, port loading/unloading centered on “high-efficiency turnover,” with risks concentrating on two aspects: on the one hand, workers lacked specialized training and had insufficient awareness of operational taboos such as “no tipping” and “avoid impact,” resulting in excessive impact on shipping containers during operations; on the other hand, if equipment such as quay cranes and gantry cranes exceeded limits during emergency lifting operations, the excessive impact acceleration damaged the internal structure of batteries. In particular, batteries with pre-existing external defects were more prone to internal short circuits.

(3)

Failure of shipping container protection systems

Failure of shipping container protection systems constituted the critical transmission path from “external inducements to inherent defects,” manifested specifically as impaired structural integrity and failure of dedicated protective accessories—both directly related to the maritime transportation environment, as shown in Figure 4A.

Structural issues stemmed from initial defects or subsequent erosion: factory-out strength failed to meet standards, such as thin plates and weak welds, or unrepaired dents and cracks before transportation, all of which resulted in an inherent insufficient corrosion resistance. In the marine environment, long-term salt spray erosion rendered the shipping container prone to perforation, with deck layers additionally subjected to the impact of seawater splashes. Seawater intrusion not only reacted with the electrolyte to release flammable gases but also directly induced short circuits; even without perforation, corrosion weakened the shipping container’s impact resistance, causing it to deform during ship vibrations and thereby squeezing the batteries.

Failure of dedicated protective accessories was mostly attributed to inadequate maintenance: aged and cracked sealing strips not replaced in a timely manner allowed rainwater infiltration, exacerbating moisture retention; breathing valves clogged by salt lost their function of allowing ventilation while preventing water ingress, accelerating battery degradation; some ports failed to install bottom anti-seawater immersion pads in accordance with specifications, such that when water accumulated in the cargo hold, batteries were directly immersed in seawater, potentially triggering large-scale short circuits or thermal runaway.

3.2. Factors Resulting from Pre-Fire Emergency Response Measures

In the early stage of thermal runaway of lithium-ion power batteries, effective detection and scientific handling could have effectively prevented the occurrence of fire accidents. Therefore, the failure of emergency response measures—particularly the failure in early detection and warning—was one of the key triggers for fires in lithium-ion power battery shipping containers in maritime transportation scenarios. In terms of composition, such failure specifically encompassed three core links: failure of early detection and warning, failure of fire suppression systems, and delays in emergency rescue. Their failure mechanisms were not only closely related to the unique characteristics of lithium battery fires but also deeply coupled with the emergency response limitations inherent in maritime transportation scenarios.

(1)

Failure of early detection and warning

The early stage of lithium battery thermal runaway exhibited characteristics of “flameless smoking and local temperature rise,” requiring coordinated prevention and control by specialized detection equipment and manual inspections. However, in practical applications, there were two types of blind spots, as displayed in Figure 4B:

First, limitations of detection equipment—some shipping containers were only equipped with ordinary smoke detectors, which could neither identify early flameless smoke (such as CO and electrolyte decomposition gases) nor capture weak local temperature rise signals, resulting in basic detection blind spots. If detectors were not regularly calibrated, their sensitivity continuously decreased, further impairing the ability to capture early abnormalities. Additionally, ship vibrations could cause wire breakages in detection equipment, directly interrupting the warning link.

Second, oversights in manual inspections—24-h shift operations on ships easily led to fatigue among inspection personnel. Particularly in low-visibility environments (e.g., nighttime or foggy weather), it was difficult to detect subtle signs such as abnormal temperature rises on shipping container surfaces or smoke from gaps through visual inspection or preliminary touch. Moreover, enclosed cargo hold spaces and narrow passages rendered some high-level stacking areas inaccessible, ultimately forming inspection blind spots and delaying the detection of early thermal runaway signals.

(2)

Failure of fire suppression systems

Failure of fire suppression systems stemmed from insufficient medium compatibility and inadequate equipment reliability, as shown in Figure 4C:

First, mismatched or ineffective fire-extinguishing agents—some ships and shipping containers still employed conventional water-based agents (e.g., plain water, water mist) and dry powder: water-based agents reacted with reactive lithium compounds in thermally runaway batteries to release hydrogen, exacerbating the risk of explosion; dry powder, though capable of temporarily suppressing flames, failed to cool the core area of batteries, making re-ignition highly likely. Even when equipped with gas agents suitable for lithium battery fires (e.g., heptafluoropropane, perfluorohexanone), their fire-extinguishing capability was lost if they exceeded their expiration date or leaked.

Second, abnormal operation of fire suppression equipment—ship vibrations easily caused jamming of activation buttons and loosening of pipeline joints in fire-extinguishing devices, thereby preventing normal activation or causing leakage of fire-extinguishing agents. Some ships had issues such as insufficient pressure in fire hydrants and aging/leakage of fire hoses, meaning that even if the system was activated, it failed to deliver an effective fire-extinguishing dose. Additionally, if the designed range and coverage of fire monitors could not meet the needs of high-level shipping container stacking, fires in upper layers could not be effectively extinguished.

(3)

Delays in emergency rescue

Maritime transportation, characterized by remoteness from shore-based support and strong constraints from sea conditions, made rescue delays a unique risk, distinguishing it from other transportation modes, as illustrated in Figure 4D.

Externally: If a fire occurred in the open sea (over 200 nautical miles from the coastline), the arrival time of shore-based rescue forces (fireboats, helicopters) would far exceed the “golden 10-min” emergency response window for lithium battery fires. Sea conditions with winds of force 6 or higher hindered rescue vessels from approaching. Moreover, weak satellite communication signals in the open sea easily caused alarm delays, further prolonging emergency response times.

Internally: If crew members had not received specialized drills for lithium battery fires, operational errors (e.g., misuse of conventional fire-extinguishing agents) were likely to impair response efforts. Internal communication failures interrupted the transmission of commands, making it difficult for rescue teams to coordinate. Fire spread blocked cargo hold passages, preventing rescuers from reaching the fire site, ultimately causing fires in individual shipping containers to spread to entire stacks.

3.3. Conversion of FT to BN

The construction of the Bayesian Network model required the fault tree model as a prerequisite, and their conversion followed the following criteria: basic events, intermediate events, and the top event in the fault tree corresponded to root nodes, intermediate nodes, and leaf nodes in the Bayesian Network, respectively; input events of logic gates corresponded to parent nodes in the Bayesian Network, and output events corresponded to child nodes; for repeated events, only one corresponding node was required to be established to avoid overestimation of event occurrence probabilities due to neglect of conditional dependencies between events. Meanwhile, the occurrence probabilities of basic events in the fault tree corresponded to the prior probabilities of root nodes in the Bayesian Network, and the conditional probabilities of each node were determined based on the logical relationships of the fault tree. Figure 5A visually illustrated the logical relationships of OR gates and AND gates, as well as the conversion process of repeated events into the Bayesian Network structure, while Figure 5B showed the final conversion result, i.e., the Bayesian Network model constructed for fire accidents involving power battery shipping containers during maritime transportation.

The Bayesian Network model shown in Figure 5B contained five hierarchies, 26 intermediate nodes, and 49 root nodes. Due to space constraints,

Table 2. Correspondence between the coding in the Bayesian network model and the coding in the fault tree.

Bayesian Code	Fault Tree Code	Bayesian Code	Fault Tree Code	Bayesian Code	Fault Tree Code
IE1	A1	IE2	A2	IE3	B1
IE4	B2	IE5	B3	IE6	B4
IE7	B5	IE8	B6	IE9	C1
IE10	C2	IE11	C3	IE12	C4
IE13	C5	IE14	C6	IE15	C7
IE16	C8	IE17	C9	IE18	C10
IE19	C11	IE20	C12	IE21	C13
IE22	C14	IE23	C16	IE24	D1
IE25	D3	IE26	D15
X1	C15	X2	D2	X3	D4
X4	D5	X5	D6	X6	D7
X7	D8	X8	D9	X9	D10
X10	D11	X11	D12	X12	D13
X13	D14	X14	D16	X15	D17
X16	D18	X17	D19	X18	D20
X19	D21	X20	D22	X21	D23
X22	D24	X23	D25	X24	D26
X25	D27	X26	D28	X27	D29
X28	D30	X29	D31	X30	D32
X31	D33	X32	D34	X33	D35
X34	D36	X35	D37	X36	D38
X37	D39	X38	D40	X39	D41
X40	D42	X41	D43	X42	E1
X43	E2	X44	E3	X45	E4
X46	E5	X47	E6	X48	E7
X49	E8

only presents the mapping of corresponding numbers between each node in the Bayesian Network model and those in the fault tree model, and their specific meanings are not reiterated here.

4. Discussion

4.1. Quantitative Assessment and Analysis

This study employed Netica for quantitative calculations of the Bayesian network. Its core advantages included the following: a graphical interface that simplified modeling; efficient learning algorithms that reduced manual parameter settings; support for flexible reasoning and graphic–text output; the capability to conduct sensitivity analysis to aid decision-making; and visualization tools that enhanced result readability. In addition, the prior probabilities of root nodes and conditional probabilities of child nodes in the Bayesian network were obtained through historical data statistics and expert scoring via the Delphi method. Some prior probabilities could be determined based on public data or literature, but due to the limitations of historical data and the difficulty in directly obtaining the occurrence probabilities of combined events under specific scenarios, experts were invited to evaluate the conditional or prior probabilities of some nodes to clarify their quantitative composition.

A 20-member expert panel was established for this assessment, with members from diverse backgrounds: five university researchers, five regulatory representatives, five shipping managers, and five frontline operators. This composition enabled a comprehensive and systematic evaluation across four dimensions: academic, regulatory, managerial, and operational.

Prior to the evaluation, the factors to be assessed were fully presented to the experts for feedback. Following consensus-building, scoring was administered. Through rigorous deliberation, the expert panel confirmed that all evaluated accident-causing factors were objectively existing safety hazards in practical work, thereby fully validating the rationality of these accident-causing factors.

Experts used anonymous, independent scoring and, after three to four rounds of feedback, reached consensus; the average value of node probabilities was then calculated. Furthermore, expert evaluations were prone to subjective differences due to personal experience-based judgments. Therefore, weights needed to be assigned to the evaluation results. Participants in the evaluation were categorized into four levels based on their experience: “extremely rich,” “relatively rich,” “rich,” and “average,” with initial weights of 4, 3, 2, and 1 in sequence. Shipping managers and frontline operators, who were familiar with on-site conditions of maritime transportation of lithium-ion power battery shipping containers, were classified into the “extremely rich” level; regulatory personnel and university researchers, who had less on-site exposure but sufficient relevant experience, were classified into the “relatively rich” level. The initial weights underwent normalization, and the specific weight values are shown in

Table 3. Proportional weight of different types of evaluation experts.

Expert Type	Weight Proportion
Regulatory Representatives	0.214
Shipping Managers	0.286
Frontline Operators	0.286
University Researchers	0.214

.

Furthermore, this study employed Cronbach’s alpha method to assess the consistency of expert scoring results, with the corresponding calculation formula presented in Equation (4). Only when the computed Cronbach’s alpha coefficient exceeded 0.8 could the consistency of experts’ opinions be confirmed.

$$\alpha ={\displaystyle \frac{K({\sigma }^{2}X-{\sum }_{i=1}^K{\sigma }^{2}Y)}{{\sigma }^{2}X(K-1)}}$$

(4)

where $\alpha$ denotes the Cronbach’s alpha coefficient; ${\sigma }^{2}X$ represents the variance of scoring results across all root nodes; ${\sigma }^{2}Y$ stands for the variance of scoring results for a specific root node; $K$ is the number of root nodes.

To summarize, to ensure the reliability of prior probabilities for each root node in the Bayesian network model, the study strengthened this aspect through three key measures: firstly, it recruited experts with interdisciplinary backgrounds to participate in the evaluation; secondly, it calibrated the weights of the scoring results provided by these experts; and thirdly, it utilized the Cronbach’s alpha method for verification to guarantee the consistency of the final scoring outcomes.

Figure 6 presents the final probability assessment results. For the full-chain process of power battery shipping containers on a specific shipping route—encompassing product manufacturing, loading, transportation, and unloading—the expert panel conducted assessments and combined them with quantitative calculations from the Bayesian network model, ultimately determining that the occurrence probability of fire accidents in maritime transportation of power battery shipping containers on this route was 35%.

Further analysis of Figure 6 revealed that the probability of thermal runaway occurring in power battery shipping containers and causing open flames was 61.8%, while the probability of failure of early emergency measures leading to fire expansion was 49.7%. This indicated that thermal runaway of lithium-ion batteries was the core risk factor inducing such fire accidents; meanwhile, effective early emergency response measures could also exert a critical impact on fire control effectiveness.

The Bayesian network model not only estimated the occurrence probability of the final accident on the basis of the prior probabilities of individual root nodes but also exhibited another core function: integrating the posterior probabilities output by the model with sensitivity analysis methods to pinpoint the root nodes that exerted the most significant impacts on the top event (fire accident). In the Fault tree model established in this study, the occurrence of fire accidents necessitated the combined action of IE1 and IE2 as a prerequisite. Consequently, sensitivity analysis was employed for IE1 and IE2 as the initial analytical tiers.

Figure 7 depicts the sensitivity analysis results for the first-level components of IE1 and IE2. For IE1, the sensitivity coefficient ranking of its parent nodes was IE4 (53.6) > IE3 (48.4) > IE5 (32.3); in the case of IE2, the sensitivity ranking of its parent nodes was IE6 (43.9) > IE7 (39.2) > IE8 (37.3). These findings demonstrated that IE4 constituted the key influencing factor for IE1, whereas IE6 represented the core sensitive indicator for IE2.

Guided by this analytical framework, the second-level sensitivity analysis traced hierarchically starting from IE4 and IE6 until the bottommost root nodes were localized. Given space limitations, Figure 8 directly illustrates the sensitivity analysis results of the ultimate root nodes.

As illustrated in Figure 8, the ultimately highly sensitive root nodes were concentrated in two branches, namely IE14 and IE19. Within the IE14 branch, X15 demonstrated the highest sensitivity with a coefficient of 31.5, followed by X16 (23.0), whereas X14 exhibited the lowest sensitivity at 18.2. In the IE19 branch, the sensitivity coefficient of X30 (42.5) was significantly higher than that of other nodes, followed sequentially by X31 (22.0), while X29 had the lowest sensitivity at merely 0.71.

Synthesizing the aforementioned analysis, among the two influence pathways centered on IE1 and IE2, the root nodes exerting the most prominent effects on the fire accident were X15 and X30, respectively. Notably, while sensitivity analysis could identify key influencing nodes, it failed to quantitatively characterize the probability of these critical risks evolving into accidents, nor could it accurately compare the probability discrepancies between different risk pathways. To remedy this limitation, the present study incorporated Attack–Defense Game Theory to achieve the qualitative determination of key risk evolution pathways and the quantitative calculation of their occurrence probabilities, with relevant details elaborated in the subsequent section.

4.2. Identification of Key Risk Evolution Paths

Based on the key risk identification method, which was established by integrating Attack–Defense Game Theory in Section 2, and with the quantitative calculation results of the Bayesian network incorporated, the key risk evolution paths for shipping container fire accidents involving lithium-ion power batteries in maritime transportation scenarios were finally identified. The visualized results of these paths were presented in

Table 4. Summary of causal chains for key risk evolution paths in fire accidents of power battery shipping containers during maritime transportation.

Root Node	Evolution Path	Occurrence Probability (%)
X15	X15-IE14-IE4-IE1	3.77
X30	X30-IE19-IE6-IE2	4.35

and Figure 9. For the intermediate nodes IE1 and IE2, the logical relationship between them was an AND gate structure; hence, they were, respectively, identified as the terminals of the key risk evolution paths.

First, in the scenario of “Thermal runaway occurs inside the power battery shipping container and triggers open flame”, the key risk evolution path was: X15 (Blockage of the shipping container ventilation device) → IE14 (Loss of control of environmental parameters inside the shipping container) → IE4 (Effect of internal and external inducing factors on the shipping container) → IE1 (Thermal runaway occurs inside the power battery shipping container and triggers open flame), and the occurrence probability of this path was 3.77%. The calculation process of this probability was presented in Equation (5).

$$P_{X15}=0.052\times 0.891\times 0.903\times 0.902\approx 0.0377=3.77\%$$

(5)

In the scenario of “Failure of initial handling of thermal runaway results in fire expansion”, the key risk evolution path was: X30 (Omission in inspection by patrol personnel) → IE19 (Inadequate on-site personnel monitoring and inspection) → IE6 (Early fire detection and warning failure) → IE2 (Failure of initial handling of thermal runaway results in fire expansion), and the occurrence probability of this path was 4.35%. The calculation process of this probability was presented in Equation (6).

$$P_{X30}=0.060\times 0.891\times 0.904\times 0.899\approx 0.0435=4.35\%$$

(6)

Upon clarifying the evolution path of key risks, it was imperative to further elucidate the intrinsic mechanisms underpinning their ascendance as dominant risks from a systemic perspective. Taking root node X15 as a case in point, blockage of ventilation devices in shipping containers represented the core factor triggering thermal runaway and open flames in power batteries, whose action mechanism could be elaborated along three dimensions: environmental regulation, battery response, and system coupling.

Ventilation systems constituted the core unit for thermal environment regulation in shipping containers. In the marine transportation setting, salt crystallization under a high-salt-fog environment, cargo packaging debris, and sundries splashed by sea waves were prone to gradual adhesion to the surface of ventilation grilles, leading to progressive blockage. Notably, this type of blockage was not an abrupt failure, but accumulated slowly over a shipping cycle of several days to weeks; its concealed nature rendered it undetectable in routine inspections, ultimately emerging as the initial trigger for risk evolution.

Blockage of ventilation devices directly precipitated imbalances in environmental parameters within shipping containers, and the thermal stability of power batteries exhibited extreme sensitivity to temperature variations. The dual effects of elevated temperature and humidity manifested as follows: on the one hand, accelerating the heterogeneous nucleation and growth of lithium dendrites on electrode surfaces; on the other hand, exacerbating electrochemical corrosion and structural degradation of electrode active materials, thereby triggering a significant reduction in the thermal runaway temperature thresholds of batteries. This positive feedback loop—environmental degradation → battery performance decay → accelerated heat generation rate → further environmental degradation—underpinned the core mechanism by which risks propagated rapidly from the environmental domain to the battery.

The dynamic nature of marine transportation further exacerbated the risk propagation effect: notably, hull sway imparted uneven stress distribution within the battery pack, elevating the likelihood of micro-deformation of electrode sheets and separator breakage. The superimposition of multiple risk factors ultimately precipitated thermal runaway and open flames in the batteries.

The risk evolution mechanism for root node X30 could be dissected from the following dimensions. First, in marine container stacking scenarios, container stacking density reached 4–6 layers; the ventilation devices and monitoring points of some containers were obscured by upper units, rendering inspection dependent on personnel climbing or auxiliary tools and thereby significantly heightening labor intensity. Second, most shipping enterprises implemented a routine inspection regime of once every 4 h, with a single inspection only providing 15–20 min of effective coverage. The tension between high-intensity tasks and narrow time windows induced selective inattention among inspectors due to physical and visual fatigue, resulting in a higher likelihood of risk oversight. Most critically, existing inspection criteria exhibited notable limitations: they focused solely on verifying the external integrity of equipment and failed to incorporate latent risks—such as ventilation grille blockage severity and abnormal battery surface temperatures—into core inspection metrics.

Upon the direct conversion of root node X30 to IE19 via risk propagation, the quasi-risk that could have been mitigated through timely ventilation grille cleaning persisted latently. Stationary monitoring devices available at the time only detected significant abnormal signals, while oversight in manual inspections disrupted the synergistic mechanism of manual inspection and equipment monitoring, thereby creating a risk monitoring blind zone. More critically, IE19 further catalyzed IE6 (early-warning failure risk): electromagnetic interference during ship navigation readily induced high false-alarm rates in monitoring equipment, and the blind spots in manual inspections precluded operation and maintenance personnel from promptly verifying the validity of early warnings. This persistent failure to validate warnings accumulated over time, triggering early-warning fatigue and ultimately resulting in the disregard of early-warning signals. Under the superposition of multiple risk factors, the risk ultimately propagated to IE2, thus finalizing the entire risk evolution chain.

Based on the intrinsic mechanisms underlying the evolutionary pathways of the above-identified key risks, targeted optimization recommendations were proposed to improve the fire risk resilience of lithium-ion power battery container transportation in shipping enterprises.

Centered on the key risk evolution chain corresponding to root node X15, specific prevention and control measures were divided into the following three aspects.

First, the intelligent upgrading of container ventilation systems and optimization of redundancy design were facilitated, with the adoption of active cleaning-redundant ventilation dual-guarantee technology. When the ventilation resistance exceeded the preset rated threshold, the cleaning procedure was automatically triggered to remove filter blockages. Meanwhile, standby air vents were added on both sides of the container to ensure the continuous and stable operation of the ventilation system.

Second, a tripartite integrated multi-dimensional parameter monitoring network for “temperature-gas-humidity” was established, and high-precision sensing units were integrated, enabling real-time and accurate perception of key environmental parameters throughout the entire battery transportation process as well as rapid early warning response to abnormal signals.

Third, a new generation of intrinsically safe transport enclosures was developed, and integrated transport enclosures with thermal insulation, explosion-proof, and pressure relief functions were promoted to enhance the active prevention and control of fire risks and emergency disposal capabilities at the structural design level.

For the key risk evolution chain corresponding to root node X30, a systematic prevention and control system was mandated to be established from three dimensions: inspection, early warning, and emergency response, with specific measures as follows.

First, an air-ground coordinated intelligent inspection system of “drone-robot” was established, incorporating sensing modules including high-definition visual recognition and infrared thermography temperature measurement. For inspection of blind areas formed by high-stacked containers, three-dimensional patrolling was conducted via on-site drones.

Second, the intelligent fire detection and early warning system was upgraded, and a neural network algorithm was incorporated to perform fusion analysis on multi-source monitoring data (i.e., temperature, gas concentration, and smoke concentration) to mitigate the false positive rate and false negative rate of individual sensors.

Third, an integrated emergency response system of “training-equipment-process” was established. Specialized training on the fire characteristics and disposal essentials of power batteries was conducted, which strengthened the crew’s emergency operation skills and team collaborative disposal capabilities, thereby improving the emergency response efficiency for sudden fires.

5. Conclusions

This study focused on fire accidents involving lithium-ion power battery shipping containers during maritime transportation. It integrated the Fault Tree Model, Bayesian Network Model, and Attack–Defense Game Theory, systematically sorted out accident-causing factors, established a qualitative-quantitative comprehensive assessment system, and identified key risk evolution paths. The main research findings were as follows:

(1)

A systematic analysis was performed on the accident-causing factors of maritime fire accidents involving lithium-ion power battery shipping containers from three dimensions: the battery quality, maritime transportation characteristics, and pre-fire emergency measures. The logical correlations among the various factors were qualitatively elucidated via a fault tree model. Based on the transformation criteria concerning FT-BN, a Bayesian network model consisting of five hierarchical levels, 26 intermediate nodes, and 49 root nodes was established.

(2)

According to the Bayesian network model, the probability of fire accidents involving lithium-ion power battery shipping containers on the target route was estimated to be 35%. Via posterior probability inference of the Bayesian network, two root nodes exerting the most prominent influence on the top event were identified, namely X15 (Blockage of the container ventilation device) and X30 (Omission in inspection by patrol personnel).

(3)

Attack–defense game theory was employed to qualitatively identify the critical risk evolution pathways, namely X15-IE14-IE4-IE1 and X30-IE19-IE6-IE2, while their occurrence probabilities were quantitatively estimated to be 3.77% and 4.35%, respectively. From a systematic perspective, the intrinsic mechanisms were revealed: failure of container ventilation facilities triggered battery thermal runaway, and mission failure in personnel inspection resulted in the failure of pre-fire emergency response. Targeted optimization suggestions for risk prevention and control were put forward, targeting these mechanisms.

Furthermore, future research can be expanded to different route scenarios, involving more experts in the field to conduct multiple rounds of verification to systematically test the generalizability of the accident-causing factors. Additionally, based on the findings of this study, a risk calculation application tool can be developed to support risk assessment throughout the entire lifecycle of lithium-ion power battery transportation, thereby significantly enhancing the technical capabilities of shipping companies to resist related fire risks.