Analysis on Causative Factors and Evolution Paths of Blast Furnace Gas Leak Accident

Abstract

Although the interest in metallurgical accident investigation of blast furnace gas (BFG) leakage has increased to explore the engineering failures, more effort is needed to address the individual and organizational causative factors to clear and determine the weak links for improving safety management and accident prevention to achieve green metallurgical manufacturing. This study aims to examine the causative factors and evolution paths of BFG leakage by introducing a combined method, the 24 model and Bayesian network (BN), based on 50 cases of fire, explosion and suffocation accidents caused by BFG leakage. A BN model of BFG leakage was established based on the identification of 25 causative factors by the 24 model. Results showed that eight nodes, including A1 (unsafe operation), A2 (unsafe behavior), A4 (unsafe condition), B1 (valve failure), B2 (improper gas safety operation), X4 (use of BFG violates regulations), X5 (water gas is not cut off before shutdown reduction) and X6 (incomplete steam purging), were more sensitive than others, and the posterior probability of nodes A1, A2, A3 (unsafe command), A4, B1, B2, B4 (improper emergency behavior), B5 (unsafe behaviors on BFG site) increased compared to prior probability. Three main accident causal chains were obtained which indicate that control the unsafe operations (A1) related to gas (B2) and valve (B1) are suggested to be improved. Another important factor is A4 (unsafe condition), which is related to intrinsic safety conditions. Considering the results, the key points of 3E strategy about BFG leakage prevention are suggested. This study provides useful insights to understand the organizational and individual factors and their relative influence in BFG leakage accidents, which will support BFG leakage prevention and safety management.

1. Introduction

Metallurgical gas is an important energy and by-product in steel manufacturing [1,2]. According to the production process, it is usually divided into blast furnace gas (BFG), Linz–Donawitz gas (LDG), coke oven gas (COG), etc. Due to different gas-making materials and gas production and recovery processes, the proportions of each component of metallurgical gas are different. BFG is from blast furnace ironmaking, which is the core part of steel plants [3]. BFG is typically composed of CO (28–33%), CO₂ (6–12%), H₂ (1–4%), N₂ (55–60%) and a small amount of sulfur dioxide (SO₂). Its explosion limit range is from 30.8% to 89.5%. As the largest stream of waste energy in any steelworks, the energy consumption of BGF accounts for about 70% of the total energy consumption [2,4]. Though the BFG’s fuel value is “weak” at about 10% that of natural gas, BFG has many valuable in-plant uses. Meanwhile, due to its toxicity, flammability and explosiveness, once BFG leakage occurs, it may cause fire, explosion or asphyxia accidents. Therefore, BFG leakage accident prevention is the guarantee of energy saving, emission reduction and efficiency improvement in steel manufacturing [1,5].

China is the world’s largest steel producer. Although a short-term decline in the global steel industry was inevitable due to the COVID-19 epidemic, the crude steel output of China remained 1.053 and 1.003 billion tons in 2020 and 2021, accounting for more than 50% of the world’s total [6]. Large production capacity is accompanied by more BFG leakage accidents in total. According to the statistics by the authors, from 2001 to 2021, there were more than 50 fires, explosions or asphyxias caused by BFG leakage in China, resulting in 121 deaths, 321 injuries and significant loss of property. Considering the frequent occurrence of BFG leakage accidents in China, there is an urgent need to collect enough cases to study the generated characteristics of BFG leakage accidents.

Previous research mostly focused on engineering technology to prevent BFG leakage accidents. Due to the few fault data and large fluctuation in data distribution of the blast furnace ironmaking process, they usually consider fault diagnosis rather than BFG leakage alone. Liu et al. [7] proposed a cost-conscious least squares support vector machine (LSSVM) model to realize the rapid fault diagnosis of blast furnaces. Stefan et al. [3] introduced the probability density function to describe the modeling errors in both time scales and space scales based on 21 points of a temperature measuring device in the blast furnace during different time periods. Dali et al. [4] proposed a deep weighted joint distribution adaption network to align the probability distributions of the dataset more comprehensively, which can obtain good results in diagnosing the abnormal conditions of the BF ironmaking process. Some scholars focus on the detection and prediction technology of gas leakage risk, which includes robust data-driven models to study dispersion of vapor clouds [8], using an infrared camera and Faster R-CNN technique to achieve real-time leak detection [9], using a physics-guided deep learning probability model to achieve real-time gas release forecasting [10] and combining a naive Bayes classifier (NBC) and probabilistic neural network (PNN) to predicate the diffusion of gas leakage [11]. These fault diagnosis models and gas leakage detection and prediction technology have important superiority in improving the intrinsic safety of blast furnace to prevent BFG leakage, but they cannot explain the causative factors of BFG leakage, as they do not consider the organizational and individual problems in blast furnace operation and safety management [12].

There is strong evidence that 80–90% of accidents could be attributed in some way to human factors [13,14,15]. Lallemand [16] reported an ergonomic intervention in a blast furnace plant by setting up working groups to solve health and safety issues. Results showed that the 80 employees who participated in the working group had safer attitudes and behaviors than the other 150 employees after four months. Sundararaj et al. [17] adopted a mock drill exercise strategy to improve the risk prevention and management in blast furnace operation. Verma et al. [18] point out the importance of studying the significant amount of safety-related data in steel enterprises. They analyzed incident investigation reports of an integrated steel plant in India, found that causes of incidents differ depending on the activities performed in a department and found that fire/explosion and process-related incidents are more common in coke-making and blast furnace departments.

Although these safety-related incident report analyses and BF operation strategies revealed the risk factors as weak links of safety management in the operation of blast furnaces, they cannot quantify the causative factors’ effect and their evolution path to the consequence. In order to quantitatively explore the characteristics of BFG leakage accidents, it is necessary to understand the organizational and personal factors comprehensively, and find out the key points and accident causation chains. The choice of appropriate accident causation model and the quantifiable analysis of causative factors are two important considerations for establishing an accurate BFG leakage accident analysis model to reveal the causative factors and evolution paths.

The Accident Causation “2–4” Model (24 model for short) is based on Heinrich’s, Bird’s and Reason’s accident causation models as well as organizational behavior theory, which considers that accidents are related to two important factors: organizational and individual [19,20]. Compared to HFACS, the Trajectory Intersecting Model and other popular accident causation models, it has advantages in both the clear categories and definitions of accident causes [21], and more consideration of “software” factors such as individual and organizational factors than “hardware” factors such as technical factors for improving safety management. Therefore, using the 24 model can better find the causative factors of BFG leakages more comprehensively.

Meanwhile, as a qualitative model, the 24 model has a disadvantage in the quantifiable analysis of causative factors and their influence on the consequences of accidents. The Bayesian network (BN) is an important method to deal with quantitative risk assessment and can be used to deal with the influence of causative factors and accident causation chains [22,23,24]. For instance, Li at al. [25] built a cause model of coal mine gas explosion accidents based on a Bayesian network (BN), and found that the main factors were gas accumulation and fire source explosion. Liu et al. [26] proposed a novel method to quantify risk coupling of subsea blowout accidents based on a dynamic Bayesian network (DBN) and NK model, and the risk coupling mechanism of subsea blowout was revealed from the perspective of interactions among different types of risk factors. Thus, it is possible to combine the 24 model and BN to quantitatively analyze the causative factors and evolution paths of BFG leakage accidents, which will provide more powerful and multi-dimensional insights into the generated characteristics of BFG leakage than applying a general statistical analysis.

In this study, the 24 model and BN method are first combined to quantitatively reveal the causative factors and evolution paths of BFG leakage accidents. Specifically, based on the 50 BFG leakage cases from 2001 to 2021 in China, the causative factors and their prior probabilities in the BN are more factual than subjective expert assignment. This analysis could help us understand the question of, “which are the key causative factors and causative chains of BFG leakage accidents?”. The results will provide useful insights for solving the organizational and individual problems in safety management of blast furnace operation, which can provide a basis for BFG leakage accident prevention and management.

2. Materials

Through sorting out the cases of BFG leakage accidents published by the local emergency management departments and safety management web pages in China, a total of 50 cases with relatively detailed investigation reports from the years of 2001 to 2021 were used as the materials. Each report has a qualitative description about the accident time, process and direct causes and indirect causes. Some of these cases contain more detailed information about the legal treatment of the responsible persons for the accident. The 50 cases include 42 asphyxia cases accounting for 84%, 7 fires accounting for 14% and 3 explosions at 6%, which matches the types of accident caused by BFG leakage. The brief description of accidents is shown in Appendix A.

3. Methodology

The workflow of the proposed approach consists of five key steps, including identifying causative factors based on the 24 model, establishing a BN model using the causative factors as network nodes, analyzing the posterior probability of each factor, quantifying the sensitivity of each factor and analyzing the accident cause chain of BFG leakage accidents.

3.1. Causative Factor Identification

3.1.1. Structure of 24 Model

The 24 model is a general accident analysis method. It was first proposed by Professor Fu Gui at China University of Mining & Technology (Beijing), Beijing, China in 2005. In the last 17 years, the model has had four visions. The latest version of the 24 model is shown in Figure 1.

Figure 1. Latest version of Accident Causation 24 Model adapted from [17,18].

Figure 1. Latest version of Accident Causation 24 Model adapted from [17,18].

The model splits the accident causes into internal causes and external causes. The internal causes are the causes within the organization, while the external causes are social environment and government regulatory factors, which cannot be controlled by the organization itself. However, the influence of external factors can be transformed into the state of internal factors. Therefore, the accident cause analysis needs to focus on the internal causes.

As shown in Figure 1, the internal causes are divided into an organizational level and individual level consisting of four phases. In the individual level, the unsafe acts of people and unsafe conditions of objects (machineries, equipment, tools, etc.) are the immediate causes in Phase I, while safety knowledge, safety awareness, safety habits, physiological status and psychological status are the indirect causes in Phase II. In the organizational level, the defects of safety management systems are the radical causes in Phase III. The weakness of safety culture is the root cause in Phase IV.

3.1.2. Identify the Causative Factors of BFG Leakage

By using 24 model, the causative factors of BFG leakage can be comprehensively identified. In the 24 model, indirect causes are the causes that lead to direct causes and, from the perspective of building a Bayesian network, it is only necessary to analyze the direct causes and the intermediate events they may cause. Thus, only direct causes need to be considered when constructing a Bayesian network. This section describes the identification process of immediate causes of BFG leakage as an example.

For instance, following the guide of the 24 model on the two types of immediate causes, unsafe acts refer to actions that cause an accident or have an important impact on the occurrence of the accident, which can be divided into unsafe operation, unsafe behavior and unsafe command according to their functions contributing to the organization management. Unsafe conditions refer to the unsafe state of objects (machineries, equipment, tools, etc.) causing the accident. First, based on the accident investigation reports of the 50 cases, 25 kinds of direct causative factors of BFG leakage accidents were identified and counted. Then, according to the characteristics of these causative factors, they were sorted into unsafe acts and unsafe conditions. Finally, 17 direct causative factors were identified as unsafe acts, including 8 kinds of unsafe operations, 7 kinds of unsafe behaviors and 2 kinds of unsafe commands. Eight direct causative factors were identified as unsafe conditions, including 4 kinds of inadequate intrinsic safety precautions and 4 kinds of equipment problem. The detailed factors are shown in

Table 1. Causative factors as immediate causes of BFG leakage accident.

Classification	Sub-Classification	Direct Causative Factors	Frequency	Proportion
Unsafe operations	Improper valve operation	Valve not closed in place	7	14%
		Violation of operation of valve	8	16%
		Violation of operation of blind flange valve	4	8%
	Improper gas operation	Use of BFG violates regulations	1	2%
		Water gas is not cut off before shutdown reduction	1	2%
		Incomplete steam purging	1	2%
	Improper BF body operation	Improper maintenance and inspection of furnace body	3	6%
		Unauthorized entry into the tank for operation	3	6%
Unsafe behaviors	Unsafe behaviors on BFG site	No safety condition confirmation	7	14%
		Unclear safety technical disclosure on BFG leakage prevention	5	10%
	Improper emergency behavior on BFG leakage	Going into hazardous areas without gas detection	6	12%
		Failure to take decisive shutdown measures in case of overtemperature	1	2%
		Choosing escape routes blindly	1	2%
Unsafe commands	Dispatching personnel without operation qualification		4	8%
	Not dispatching the site supervisor		4	8%
Unsafe conditions	Inadequate intrinsic safety precautions	Lack of relevant alarm and cut-off devices	9	18%
		Lack of safety deterrent measure	2	4%
		Failure of personal protective equipment	1	2%
		Safety device is not repaired in time	1	2%
		Working environment does not meet the safety requirements	1	2%
	Equipment problem	Equipment itself has defects at the start	7	14%
		Long-term corrosion of BFG equipment and pipeline	5	10%
		Unreasonable equipment construction and installation	2	4%

.

Taking the first accident in Appendix A as an example, on 4 December 2021, when the No.4 blast furnace of Chaoyang Mingxin Casting Co., Ltd., located in Liaoning Province, China was running, the worker opened the valve without authorization, which caused BFG leakage and led to poisoning of in-plant workers. The accident investigation report showed that the direct cause of this accident was that the worker who opened the valve without authorization during BF operation. When the direct cause was analyzed by the 24 model, it was classified as “violation of operation of valve” under “improper valve operation” and further under “unsafe operations”. According to this analysis method, the direct causes of each accident were analyzed and classified statistically, and the frequency of each cause leading to BFG leakage in Table 1 was obtained.

3.2. BN Model Building

3.2.1. Bayesian Network

The Bayesian network, also known as a belief network, is a mathematical model based on probability reasoning. It is widely used to deal with complex systems and intuitively express the causal relationship between variables in the form of graphics and has a powerful probability inference function. It uses the known variable information to calculate the probability information of location variables based on the Bayesian formula and graph theory. Similar to a neural network, it is also a directed acyclic graph (DAG) composed of node circles and directed edges. The node represents a variable, the edge represents the relationship between the variables and the node stores the conditional probability distribution of this node equivalent to its parent node. The basic Bayesian network is shown in Figure 2.

Its mathematical theory is based on the Bayesian formula as shown in Formula (1), where $P\left(A|B\right)$ is the conditional probability, $P\left(B|A\right)$ is the posterior probability and $P\left(B\right)$ is the prior probability.

$$P\left(A|B\right)=\frac{P\left(B|A\right)P\left(A\right)}{P\left(B\right)}$$

(1)

The reason why the Bayesian network is very practical in solving problems is that in some cases, the conditional probability is easier to obtain than the posterior probability. When a Bayesian network is applied to the study of BFG leakage accidents, the prior probability of the causative factors leading to the accidents can be obtained by collecting a large number of accident cases, which will avoid subjectivity compared to the way to obtain the prior probability based on expert judgment. Additionally, under the condition that a certain factor occurs, the probability of BFG leakage is often difficult to determine. At this time, Bayesian calculation and reasoning analysis can be used to solve the problem.

3.2.2. Transfer Causative Factors to Directed Acyclic Graph in Bayesian Network

The parent node T is set as the accident caused by BFG leakage. The root nodes are the 25 direct causative factors obtained from 24 model analysis. The intermediate nodes are the upper classification of the root nodes, including unsafe operation, unsafe behavior, unsafe command, unsafe condition. The node information is shown in

Table 2. Names and symbols of nodes in BN of BFG leakage accidents.

Symbol	Node	Symbol	Node
T	BFG leakage accident	X8	Unauthorized entry into the tank for operation
A1	Unsafe operation	X9	Failure to take effective protective measures
A2	Unsafe behavior	X10	Failure to notify people to evacuate
A3	Unsafe command	X11	Choosing escape routes blindly
A4	Unsafe condition	X12	Failure to take decisive shutdown measures in case of overtemperature
B1	Valve failure	X13	Going into hazardous areas without gas detection
B2	Improper gas safety operation	X14	No safety condition confirmation
B3	BF body failure	X15	Unclear safety technical disclosure on BFG leakage prevention
B4	Improper emergency behavior	X16	Dispatching personnel without operation qualification
B5	Unsafe behaviors on BFG site	X17	Not dispatching the site supervisor
B6	Inadequate intrinsic safety	X18	Lack of relevant alarm and cut-off devices
B7	Equipment problem	X19	Lack of safety deterrent measure
X1	Valve not closed in place	X20	Failure of personal protective equipment
X2	Violation of operation of valve	X21	Safety device is not repaired in time
X3	Violation of operation of blind flange valve	X22	Working environment does not meet the safety requirements
X4	Use of BFG violates regulations	X23	Equipment itself has defects at the start
X5	Water gas is not cut off before shutdown reduction	X24	Long-term corrosion of BFG equipment and pipeline
X6	Incomplete steam purging	X25	Unreasonable equipment construction and installation
X7	Improper maintenance and inspection of furnace body

.

As the direct causative factors are classified with causality in the 24 model, there are corresponding causal logical relationships between the nodes of each level. Furthermore, these causal relationships can be transferred to the nodes and conditional probability of the Bayesian network. Therefore, the directed acyclic graph of BFG leakage accidents could be drawn according to the corresponding relationship, as shown in Figure 3. The node information in Figure 4 is in Table 2.

3.2.3. Calculate the Prior Probability Parameters in Bayesian Networks

(1)

Prior probability of root nodes

In a BFG leakage accident, each node has two states. When the root node is in “state 1”, the prior probability of the node is the frequency proportion of the direct causative factor occurring in the whole sample. According to the principle that the sum of the probabilities of all states is 1, the prior probability of each node in “state 0” can be obtained. The calculation of the prior probability follows Equation (2).

$$P\left(I_1=i\right)=\frac{m\left(I_1=i\right)}{n}$$

(2)

where P (I₁ = i) is the prior probability of factor I₁ when the state is i. m (I₁ = i) is the frequency of factor I₁ when the status is i. n is the total number of samples. The prior probability of each root node has been calculated in Table 1.

(2)

Prior probability of other nodes

When we know the prior probability of root nodes, the prior probability of other nodes can be calculated according to the causal logical relationship of the 24 model and DAG. There are two kinds of causal logical relationships. One is an “AND gate”, which means when all input events occur, the output event will occur. Taking node B3 as an example, B3 is the output of X7 and X8 with a causal logical relationship of the “AND gate”. According to the Bayesian equation as shown in Equation (1), the conditional probability of B3 is calculated in

Table 3. Example of “AND gate” conditional probability.

Node		Possible State
X7		1		0
X8		1	0	1	0
B3	P (B3 = 1)	1	0	0	0
	P (B3 = 0)	0	1	1	1

. The other is the “OR gate”, which means when any input event occurs, the output event will occur. Taking node B1 as an example, B1 is the output of X1, X2 and X3 with a causal logical relationship of the “OR gate”. Similarly, the conditional probability of B1 is calculated in

Table 4. Example of “OR gate” conditional probability.

Node		Possible State
X1		1				0
X2		1		0		1		0
X3		1	0	1	0	1	0	1	0
B1	P (B1 = 1)	1	1	1	1	1	1	1	0
	P (B1 = 0)	0	0	0	0	0	0	0	1

.

3.2.4. Learning the Parameters of the Bayesian Network

After a Bayesian network was established, 50 BFG leakage accident cases were used to estimate the parameters of the BN, i.e., the conditional probabilities or posterior probabilities, which capture the relationships and interplay among the factors and find the influence of the different factors on BFG leakage accidents. This process is parameter estimation or “learning” [27]. There are four main methods of parameter learning in a Bayesian network.

The first is forward reasoning, which refers to prior probability inference of the parent node and intermediate nodes based on the known probability of root nodes. In Bayesian statistical inference, the prior probability of an uncertain event is the probability distribution of the confidence level of the event before considering some factors [28]. It reflects our understanding of the estimated parameters before the observed data.

The second is reverse reasoning, which refers to posterior probability inference of the intermediate nodes and root nodes. In a Bayesian network, the prior probability is the unsubstantiated event probability without considering the influence of some factors, and the posterior probability is the conditional probability obtained after considering and giving relevant evidence or data. Through the obtained data and considering relevant factors, the value of the prior probability is updated, and a more accurate description of the probability distribution of the estimated parameters is obtained [12,27]. In this study, it refers to how high the probability of the root node is when the parent node is set to be in a certain occurrence state. It reflects the importance of each factor through the posterior probability of the root node, so as to find the key causative factors.

The third is sensitivity analysis, which is an effective method to determine the risk priority by identifying the dependence and sensitivity of different factors on the whole network. The sensitivity of a Bayesian network is suitable for studying the change in probability distribution caused by the change in nodes [27]. Through sensitivity analysis of causative factors, we can understand the effect of each factor on BFG leakage accidents in detail. Sensitivity analysis can also predict the critical point of the accident in the maximum range under the condition of uncertain accident evolution trend.

The fourth is causal chain analysis for finding out the main evolution path of accident occurrence. As the prior probability of the root node is based on the statistical information of accidents, we could observe the change in the “occurrence” state of the parent node caused by the maximum probability state of a single factor, so as to find the main cause chain. It provides a scientific basis for managers to take relevant measures to block the critical path and avoid the occurrence of BFG leakage accidents.

As for the analysis tools of BNs, GeNIe software has been widely used in BN analysis, which has the functions of reverse reasoning, sensitivity reasoning and causal chain reasoning. Through the parameter learning, we can find the key factors and accident cause chain that affect BFG leakage accidents for taking targeted measures.

To sum up, as a qualitative model, the 24 model takes individual and organizational factors into account, allowing for a better and more comprehensive exploration of the causes of BFG leakage, but it has disadvantages in quantitative analysis. A Bayesian network is a good network analysis model, which allows for more quantitative analysis on the risk factors and accident causal chain. Therefore, combining the two not only solves the deficiencies of the 24 model in quantitatively analyzing accident causes and their impact on accident consequences, but also reduces the subjectivity of prior probabilities in BNs.

4. Results

4.1. The Results of Prior Probability Parameters

According to calculated method in Section 3.2.3, the prior probability of each node was obtained as shown in Figure 4. State 0 means that the event of the node does not occur, and state 1 means that the event of the node has occurred.

4.2. The Results of Posterior Probability Parameters

A posteriori probability is a probability estimate that is closer to the actual situation after modifying the original prior probabilities according to the new information. Using the update function of the Bayesian network, we set the state of the parent node T “BFG leakage accident” to occur, that is, node state 1 = 1. At this time, the posterior probability of each node under the condition of the accident can be obtained through reverse reasoning, as shown in Figure 5.

4.3. Sensitivity Analysis

We set the node “BFG leakage accident” as the target node in the GeNIe software, and the sensitivities of other nodes to the target node can be obtained. The results are shown in Figure 6. The darker the color of the node, the more sensitive it is. The nodes that have the most obvious impact on the critical state of the target node are A1, A2, A4, B1, B2, X4, X5, X6. The probability changes in these key points will have a great impact on the probability of BFG leakage accidents.

4.4. Accident Causal Chain Analysis

The main causal chain of the Bayesian network is used to analyze the influence degree between risk factors and the target event. At the same time, it can find the most likely paths leading to an accident. We set the “strength of influence” under the Network function in the GeNIe 2.0 software, and the main causal chains of the Bayesian network can be obtained. The software used in this paper is GeNIe Academic Version 2.3.3828.0 (32-bit), created by BayesFusion, LLC, Pittsburgh, Pennsylvania, USA. The results are shown in

Table 5. Estimated path probabilities on the condition of “strength of influence”.

Child	Parent	Probability	Parent	Child	Probability
A1	T	0.26	X9	B4	0.42
A2		0.20	X10		0.46
A3		0.16	X11		0.46
A4		0.22	X12		0.46
B1	A1	0.38	X13		0.40
B2		0.41	X14	B5	0.20
B3		0.08	X15		0.17
B4	A2	0.38	X16	A3	0.20
B5		0.56	X17		0.20
B6	A4	0.22	X18	B6	0.42
B7		0.19	X19		0.44
X1	B1	0.63	X20		0.44
X2		0.40	X21		0.44
X3		0.25	X22		0.44
X4	B2	0.31	X23	B7	0.48
X5		0.31	X24		0.33
X6		0.31	X25		0.33
X7	B3	0.17
X8		0.17

and Figure 7, where the causal chain with bold arrows is the main evolution path most likely to lead to a BFG leakage accident.

According to Figure 7, among the 25 root nodes, X1 (valve not closed in place), X2 (violation of operation of valve) under B1, X4 (use of BFG violates regulations), X5 (water gas is not cut off before shutdown reduction), X6 (incomplete steam purging) under B2, X18 (lack of relevant alarm and cut-off devices), X19 (lack of safety deterrent measure), X20 (failure of personal protective equipment), X21 (safety device is not repaired in time), X22 (working environment does not meet the safety requirements) under B6 are the strong paths in the first level. B2 (improper gas operation) under A1 (unsafe operation), B4 (improper emergency behavior), B5 (unsafe behaviors on BFG site) under A2 (unsafe behavior) are the strong paths in the second level. A1, A4 under T are the strong paths in the third level.

Considering the above results, three main evolution paths of BFG leakage accidents are obtained. One of them is X1, X2→B1→A1→T, which indicates that unsafe operation related to the valve leads to the valve failures for gas, and further cause BFG leakage. The second is X4, X5, X6→B2→A1→T, which indicates improper gas safety operation such as incomplete steam purging and water gas that is not cut off before shutdown is more likely to cause BFG leakage. The third is X18, X19, X20, X21, X21→B6→A4→T, different from the previous two paths caused by A1 (unsafe operation) to T, as this path is by A4 (unsafe condition), indicating the inadequate intrinsic safety conditions such as the lack of relevant alarm and cut-off devices, lack of safety deterrent measures, failure of personal protective equipment, safety devices not being repaired in time and working environment not meeting the safety requirements.

In addition, there are some sub-chains. For example, though X23, X24, X25 contribute a lot to B7, compared to B6, B7 has less influence on A4. Therefore, control of X23, X24, X25 can only prevent B7, but not B6 to T. The other example is X9, X10, X11, X12 under B4. Compared to B4, B5 contributes more to A2. A2 has a relatively small contribution to T. Therefore, though X9, X10, X11, X12→B4→A2→T is a causal chain, the better way is to control B5 to prevent T from happening.

5. Discussion

5.1. Comparison between Prior Probability and Posterior Probability

We compared the prior probability and posterior probability of each node to explore the change in conditional probability after considering the relevant evidence of 50 cases’ data. It was found that the posterior probabilities of eight nodes have increased, but the increments are small, as shown in Figure 8. Among them, the posterior probability of A1 increased by 0.04, A2 increased by 0.03, A3 increased by 0.01, A4 increased by 0.02. There is also a small increase in B1, B2, B4 and B5. The probability of other nodes basically does not change. Posterior probability reduces the uncertainty of parameters, which is not much different from the prior probability, indicating that the deduction result of the BN model is basically consistent with the actual result, which proves the reliability of the model [29]. In addition, the small increase in A1, A2, A3, A4, B1, B2, B4, B5 indicates that these eight factors contribute more to BFG leakage accidents than we normally see for the direct accident collection.

5.2. Strategies for BFG Leakage Prevention

According to the above results, sensitivity analysis shows that A1, A2, A4, B1, B2, X4, X5, X6 have the most impact on the critical state of BFG leakage accidents. The posterior probability shows that A1, A2, A3, A4, B1, B2, B4, B5 contribute more to BFG leakage accidents than we normally see for the direct accident collection. The main accident causal chains are X1, X2→B1→A1→T, X4, X5, X6→B2→A1→T and X18, X19, X20, X21, X21→B6→A4→T. We list the key nodes in

Table 6. Key factors contributing to the BFG leakage accidents.

Node	Increase in Posterior Probability	Obvious Sensitivity	Belonging to the Main Causal Chains	Description of the Main Causal Chains
A1	■	■	■	X1, X2→B1→A→T X4, X5, X6→B2→A1→T
A2	■	■	□	/
A3	■	□	□	/
A4	■	■	■	X19, X21→B6→A4→T
B1	■	■	■	X1, X2, X3→B1→A1→T
B2	■	■	■	X4, X5, X6→B2→A1→T
B4	■	□	□	/
B5	■	□	□	/
B6	□	□	■	X4, X5, X6→B2→A1→T
X1, X2	□	□	■	X1, X2→B1→A1→T
X4, X5, X6	□	■	■	X4, X5, X6→B2→A1→T
X18, X19, X20, X21, X22	□	□	■	X18, X19, X20, X21, X21→B6→A4→T

PS: ■ means Yes and □ means No.

.

It can be seen from Table 6 that the nodes of A1, A4, B1 and B2 are the four most key points to prevent BFG leakage accidents because they all have “Yes (■)” three times. Additionally, from the perspective of accident cause chain, B1 and B2 belong to the branches of A1. The control of A1 (unsafe operation) should be given priority. In addition, X4, X5, X6 also have relatively high sensitivities. Therefore, X4, X5, X6→B2→A1→T is the most important evolution path. This indicates that control of the unsafe operations (A1) related to gas (B2) and valve (B1) is suggested to be improved. Another important factor is A4 (unsafe condition), which is related to intrinsic safety conditions.

The 3E strategy for safety management refers to education, enforcement and engineering. Considering the causative factors and evolution paths, the key points of the 3E strategy about BFG leakage prevention are suggested as follows. The first is to strengthen education and training to reduce unsafe acts. The key contents of training are suggested to be safe operation related to gas and valve, such as correct opening and closing of gas valves, normalized BFG pipeline purging, etc. The second is to strengthen the enforcement of safe operation procedures related to gas, valve and emergency such as a post-safety responsibility system for operators, application permit for hazardous gas operation, an excellent emergency plan and drill for BFG leakage, etc. The third is to take technical measures to improve intrinsic safety for improving safety conditions such as carrying out continuous detection, early warning and interlock protection for BFG pipeline pressure, oxygen content, etc. Through the key points of the targeted 3E strategy, we can effectively improve the safety awareness of employees and intrinsic safety level of facilities, so as to prevent the occurrence and reduce losses of BFG leakage accidents.

5.3. Limitations

Blast furnace gas leakage accidents are a broad topic involving many complex factors. Accident case analysis is an effective way to study the characteristics of accident generation. As some BFG accident materials have not been published, the sample size of this study does not cover all BFG leakage accidents in China. Although we collect as much information as possible, it is suggested to further strengthen the investigation and publication of BFG leakage accidents. Meanwhile, due to the lack of information about the detailed status at the time of the accident, the study only divides the status of basic events into two types, that are occurrence and non-occurrence, which makes it impossible to analyze more statuses of causative factors and their relations to the accident. More status levels should be considered to further improve the precision of the model and obtain a more reliable probability prediction.

6. Conclusions

In this study, the Accident Causation “2–4” Model and Bayesian network (BN) method are first combined to quantitatively reveal the causative factors and evolution paths of BFG leakage accidents based on the 50 BFG leakage cases from 2000 to 2020 in China. A BN model of BFG leakage was established based on the identification of 25 causative factors by the 24 model.

Results showed that the nodes A1 (unsafe operation), A2 (unsafe behavior), A4 (unsafe condition), B1 (valve failure), B2 (improper gas safety operation), X4 (use of BFG violates regulations), X5 (water gas is not cut off before shutdown reduction), X6 (incomplete steam purging) have relatively high sensitivities, which means the eight nodes have a more obvious impact on the critical state of BFG leakage accidents. The posterior probability of nodes A1 (0.65→0.69) (the numbers in parentheses are the change from the prior probability to the posterior probability), A2 (0.59→0.62), A3 (unsafe command) (0.53→0.54), A4 (0.60→0.62), B1 (0.49→0.51), B2 (0.78→0.79), B4 (improper emergency behavior) (0.85→0.86), B5 (unsafe behaviors on BFG site) (0.52→0.53) increased compared to the prior probability, which means the eight nodes contribute more to the BFG leakage accidents than we normally see for the direct accident collection. Three main accident causal chains are analyzed and obtained, which are X1 (valve not closed in place), X2 (violation of operation of valve)→B1→A1→T (BFG leakage accident), X4, X5, X6→B2→A1→T and X18 (lack of relevant alarm and cut-off devices), X19 (lack of safety deterrent measure), X20 (failure of personal protective equipment), X21 (safety device is not repaired in time), X22 (working environment does not meet the safety requirements)→B6 (inadequate intrinsic safety)→A4→T.

Through comprehensive analysis of these results, it was concluded that A1, A4, B1 and B2 are the four most key points to prevent BFG leakage accidents. As B1 and B2 belong to the branches of A1, the control of A1 (unsafe operation) should be given priority. Because X4, X5, X6 also have relatively high sensitivities, X4, X5, X6→B2→A1→T is the most important evolution path. This indicates that control of unsafe operations (A1) related to gas (B2) and valve (B1) should be improved. Another important factor is A4 (unsafe condition), which is related to intrinsic safety conditions. Lastly, considering the key factors and evolution paths, the key points of the 3E strategy about BFG leakage prevention are suggested.

This study quantitatively reveals the generation characteristics of BFG leakage accidents. Specifically, based on the 50 BFG leakage cases, the causative factors and their prior probabilities in a BN are more factual than subjective expert assignment. The results can provide useful insights for BFG leakage accident prevention and management. It is suggested to strengthen the investigation and publication of BFG leakage report to provide more detailed materials for multi-status BN analysis in the future.