1. Introduction
There have been more than 260 accidents in chemical laboratories in the United States, and most accidents have caused casualties since 2001, according to statistics. Most of these accidents occurred in school laboratories and should have more attention paid to them, and they need to be analyzed deeply. School laboratories have complex environments and different kinds of risks, including fires, explosions, electric shocks, leaks, etc., where the leakage of toxic and harmful gases is difficult to detect and prevent. After accidents happen, hazard identification or risk analysis are always missing. We have analyzed some cases of gas leakage accidents in laboratories in the past ten years. The details of these cases show that the causes of gas leakage are similar, and the consequences of the accident are more serious if not handled properly. In 2015, during the replacement of a gas cylinder in a laboratory of the Shanghai Jiao Tong University in China, the H2S in the cylinder leaked and poisoned one worker due to inhalation of H2S. Therefore, it is necessary to conduct risk assessments of the leakage of toxic gases in school laboratories, which is of great significance to ensure personnel safety.
In the past ten years, numerous studies have focused on s the safety of laboratories [
1,
2,
3,
4,
5,
6,
7]. Many models and tools are used to identify laboratory hazards, such as the ‘bowtie diagram’ and ‘Assessment and Classification of Hazards in Laboratories’ (ACHiL) [
8,
9,
10]. Some studies have studied the risks of fire and explosion in laboratories, as well as the emergency management model, hoping to minimize the possibility of casualties in the future [
11]. Not only that, but some studies have used process hazard analysis (PHA) and vulnerability assessment methodology for chemical facilities (VAM-CF) methods to assess the risks of chemical facilities [
12,
13]. Furthermore, a large number of studies have focused on risk assessment of laboratories. Leggett described a straightforward technique designed to identify and assess the hazards of conducting a chemical synthesis in the research environment. He also discussed the relationship between the hazards and consequences of an upset event, the likelihood of the upset happening, and the resulting risk to personnel, property, and the environment [
14,
15]. Ouédraogo et al. proposed a new approach named laboratory assessment and risk analysis—LARA to assess risks in the research/academic environment. The core of this methodology relies on defining adequate role player factors to assess risks in the research environment and their mathematical combination to proceed quantitative risk assessment [
16,
17]. Research on gas leaks has focused on gas pipeline leak hazards, leak detection, monitoring indoor air quality, and compressed gas treatment [
18,
19,
20,
21,
22]. However, there are few studies on the impact of personnel, equipment, management, and other factors on the probability and consequences of gas leakage in laboratories. Therefore, we attempted to analyze the factors affecting the probabilities and consequences of gas leaks in laboratories based on the Bayesian network, and construct a risk assessment model to undertake dynamic quantitative risk assessment to analyze the risk of gas leakage in school laboratories.
There is not a well-accepted definition of the concept of risk. There are many different aspects to understand and illustrate the definition of risk. Some definitions are based on probability, chance, or expectation, some are illustrations of unexpected consequences or dangers, and others rely on uncertainty. Some consider risk to be subjective and cognitive, depending on the knowledge available, while others separate the ontological state of risk from the evaluator. Aven thoroughly discussed and summarized these definitions, their principles, advantages and disadvantages, and recent development trends [
23]. Not only that, Aven also put forward some novel understandings of risks [
24,
25]. In this study, risk is the combination of probability of an event and its consequences [
26]. The assessment of the risk of gas leakage in the laboratory is mainly about the probability of gas leakage and the consequences caused by gas leakage to express uncertainty in terms of probability. Through the probability of gas leakage, and the severity of consequences obtained, the risk of gas leakage was evaluated to improve the safety management level of gas leakage in the laboratory.
There are various qualitative risk assessment methods, which are easily applied and rely more on experts instead of data and equations. Commonly used quantitative risk assessment methods, such as fault tree analysis and event tree analysis, are visualized and computationally simple, but these methods tend to ignore the causal relationship between risk factors and cannot update dynamically. When multiple risk factors are managed at the same time, it is difficult to realize the linked calculation and comparative analysis of their respective risks for multiple scenarios with common features due to the complicated evaluation operation. The complexity and dynamic characteristics of the risk of gas leakage in school laboratories have provided the possibility for Bayesian networks to be applied to the risk assessment of this scenario. Although the structure of the bow-tie diagram (BT) is clear, it cannot describe the evolution of the scenario and the results of risk analysis when multiple causes occur simultaneously. The connection between the Bayesian network and the risk of gas leakage in the laboratory is mainly reflected in three aspects: (1) When a certain factor changes, the Bayesian network can adjust other factors affected by it in time. (2) The Bayesian network can handle the overall risk state when certain risk factors remain the same in different stages of the same scene or in different scenes with the same nature. (3) The Bayesian network has a low requirement on the known information of the evaluation object, which can be used for reasoning in the case of incomplete and uncertain data. By combining expert experience with sample data, key points of contact between information can be captured, and major contradictions highlighted.
The Bayesian network has been widely used in the safety and security field because of intuitive appeal with available software. In terms of urban security, Tang et al. established a Bayesian network to analyze the risk of an urban dirty bomb attack [
27]. Wu et al. established a comprehensive model based on the Bayesian network (BN) and the Delphi method for the rapid and dynamic assessment of the fire evolution process and consequences, in underground subway stations [
28]. In terms of natural disasters, Han et al. proposed an earthquake disaster chain risk evaluation method that couples the Bayesian network and Newmark model based on natural hazard risk formation theory with the aim of identifying the influence of earthquake disaster chains [
29]. In addition, the Bayesian network is also used in rural security, accident severity analysis, aviation safety, protection systems, etc. [
30,
31,
32,
33]. The literature above shows that the Bayesian network works well in solving uncertainty problems.
This paper aims to analyze the factors affecting the probabilities and consequences of gas leaks using Bayesian networks based on expert experience, Dempster–Shafer theory, field investigations, and case studies. A risk assessment model is established based on a Bayesian network to quantitatively assess the risks of gas leakage in school laboratories. The result can be used to guide the establishment of a toxic gas leakage warning system, which helps improve the safety management level of gas in the school laboratory, and reducing the possibility of gas leakage posing a threat to personal safety.
3. Results and Discussion
In this study, the Bayesian network can also perform predictive analysis by obtaining the state of certain root nodes for a given accident scenario. Through predictive analysis, we can quantitatively simulate the different states after gas leakage and the consequences of the accident scenario. Meanwhile, the main factors affecting the probability of gas leakage in laboratories were examined based on sensitivity analysis, assessing the impacts of different factors on the consequences of accidents.
3.1. Critical Threats Identification and Analysis
Sensitivity analysis refers to an uncertainty analysis technology that identifies sensitive factors that have a significant impact on the object from several influential factors. The analysis software Netica was used in this study to achieve the sensitivity analysis function, which can be used to analyze the influence of various factors on “Gas Leakage” and “Casualties”. In Netica, click the node you want to analyze, such as “Gas Leakage”, and select “Sensitivity to findings” in “Network” to generate a sensitivity analysis report.
The sensitivity analysis results of other nodes affecting “Gas Leakage” are listed in
Table 3. As can be seen from the data in the table, the node “Gas Leakage” was mainly affected by “Unsafe behavior of personnel”, which conformed to an actual situation, since, according to accident cases, most accidents are caused by improper behavior of personnel. Therefore, to reduce the possibility of gas leakage, personnel misconduct should be avoided as much as possible.
The sensitivity analysis results of the node “Unsafe behavior of personnel” to the root nodes are listed in
Table 4. As shown above, the node “Unsafe behavior of personnel” is mainly affected by “Obey the experimental specifications” and “Familiar with the experimental content”. Reducing the probability of “Unsafe behavior of personnel” requires attention to these two aspects. Before the experiment begins, it should be ensured that the relevant personnel of the experiment have mastered the relevant content and specifications of the experiment, and implemented the safety responsibility system of the laboratory. The safety supervisor should supervise the behavior and operation of the experimenter throughout the process to avoid the occurrence of unsafe behavior.
In addition, the data in
Table 3 shows that the sensitivity of the node “Gas Leakage” to “Safety management defect” is small, and this phenomenon indicates that the probability of gas leakage directly caused by “Safety management defect” is low. Although the immediate cause of gas leakage is mainly human, environmental, and equipment problems, it is often accompanied by safety management defects. The scenario combinations of these four nodes and the probability of gas leakage are shown in
Table 5. As can be seen from the data in the table, the probability of gas leakage increases rapidly when there is a problem with people, the environment, or equipment and is accompanied by safety management defects.
3.2. Impacts of Different Factors on the Consequences
To analyze the impact of each safety node on the consequences of the accident, the sensitivity analysis of the node “Casualties” is carried out. The results are shown in
Table 6. As can be seen from the data in the table, in addition to the node “Gas Leakage”, the node “Casualties” is mainly affected by “Toxic and harmful gas concentration”. Among these nodes, the node “Reaction conditions” is limited by objective factors, such as the state of the laboratory, which is difficult to control by humans, and the node “Emergency response” is affected by the node “Forecast and warning”. Therefore, this part mainly discusses the impact of toxic and harmful gas concentrations, personnel protection, and forecasting and warning on the consequences of accidents. The combination of several states of these nodes are given as listed in
Table 7, and the estimated probability of consequences is shown in
Figure 4.
As shown in
Figure 4, when “Toxic and harmful gas concentration” transfers from “Critical point not reached” to “Reach the critical point”, the probability of “Yes” of “Safety” drops from 0.913 to 0.001, the probability that the “Critical state” is in the state “Yes” increases from 0.04 to 0.799, and the probability of “Yes” of “Casualties” increases from 0.035 to 0.17. The results show that the concentration of toxic and harmful gases in the laboratory has a great impact on the consequences of the accident and should be controlled properly. Therefore, in the daily operation of the laboratory, the ventilation state should be well maintained to reduce the probability that the concentration of toxic and harmful gases reaches the critical point when the gas leaks, and mitigate the consequences of the accident.
Similarly, as can be seen from
Figure 4, when “Personal protection” transfers from “Yes” to “No”, the probability of “Yes” of “Casualties” is significantly increased. Personal protective measures can protect the user, reduce the probability of potential damage to the human body caused by the reaction. In addition, “Forecast and warning” can also reduce the severity of the consequences of the accident. After successful “Forecasting and warning”, personnel in the laboratory can take emergency measures to control the leaked parts in time, and also remind relevant personnel to evacuate quickly to reduce the probability of casualties.
3.3. Validity of Risk Assessment
Quantitative risk analysis differs from other areas of applied science as it attempts to model events that are unlikely to occur, so the validity of the risk assessment model needs to be verified. Borg et al. summarized two major risk interpretations and made recommendations for the validity of risk assessments based on different risk interpretations [
37]. Aven explained the general definitions of different types of validity in more detail and summarized the scope and methods of different validity definitions [
38].
The effectiveness of the risk assessment is defined in the following categories:
The degree to which the produced risk numbers are accurate compared to the true underlying risk (V1).
The degree to which the assigned probabilities adequately describe the assessor’s uncertainties of the unknown quantities considered (V2).
The degree to which the epistemic uncertainty assessments are complete (V3).
The degree to which the analysis addresses the right quantities (V4).
(V1) and (V4) are suitable for classical methods, and (V2), (V3), and (V4) are suitable for Bayesian prediction methods. Analysis of the methods and contents of this study shows that (V2) is suitable for verifying the validity of the risk assessment in this study. It is not straightforward to verify that the validity requirement (V2) is met. Some important principles and procedures are involved, as follows.
(i) Coherent uncertainty assessments are achieved by using the rules of probability, including Bayes’ theorem, for updating of assessments in the case of new information.
(ii) Comparisons are made with relevant observed relative frequencies if available.
(iii) Training in probability assignments is required to make assessors aware of heuristics as well as other problems of quantifying probabilities, such as superficiality and imprecision.
(iv) Using models, including probability models, to simplify the assignment process.
(v) Using procedures for incorporating expert judgments.
(vi) Accountability: The basis for all probability assignments must be identified.
These principles and procedures provide a basis for establishing a standard for the probability assignments; the aim being to extract (elicit) and summaries knowledge about the unknown quantities (parameters), using models, observed data, and expert opinions. It seems reasonable to say that the requirement (V2) is met provided that this standard is followed.
In this study, the risk assessment model is based on the Bayesian network and satisfies condition (i). Since there are no relevant observed relative frequencies, but a large number of case studies were provided to the experts. With the help of experts’ background knowledge, the error between subjective data and objective data can be reduced to meet conditions (iii) and (vi). Expert experience was obtained through a questionnaire. After obtaining the scores of the experts, the Dempster–Shafer evidence theory was used to process the data from different experts, and finally, the probability distribution was obtained, so the conditions (iv) and (v) are met. In addition, the research had no interest or conflict with the selected experts, ensuring the motivation of the experts. In summary, in this study, the above criteria were followed in the process of a probability distribution, (V2) was met, and the validity of risk assessment was verified.
3.4. Accident Scenario Predictive Analysis
In this study, the model was demonstrated as an accident scenario through the “7.3” incident at Zhejiang University. In the accident, two teachers mistakenly passed carbon monoxide gas to another laboratory, caused one fatality. According to the accident investigation, some root nodes with certain states are shaded in gray in
Figure A2. According to the accident cause investigation, “Familiar with the experimental content” and “Obey the experimental specifications” were assigned the “No” state, “Safety education” was assigned the “Bad” state, and “Safe operation procedures” and “Drug management system” were assigned the “No” state. According to the serious consequences, “Reaction condition” was assigned the “Yes” state, “Personnel protection” was assigned the “No” state, and “Forecast and warning” was assigned the “Failure “state. As shown in
Figure A2, the occurrence probability of gas leakage was 95.3%. “Casualties” had the highest probability of accident consequence at 52.1%. The accident scenario shows that the predictive results of this model are consistent with reality.
4. Conclusions
To comprehensively represent and assess the risk of gas leakage in school laboratories, this study applied an integrated risk assessment model for rapid and dynamic modeling gas leaks in school laboratories based on the Bayesian network. The model was used to analyze how the environment, personnel behavior, equipment, and safety management affect the probability of gas leakage and the effects of toxic and harmful gas concentrations on the consequences of the accident. The main conclusions are:
(1) The behavior of personnel has a significant impact on the probability of gas leakage. If there are unsafe behaviors of personnel, the probability of gas leakage would still reach 39.3%, even though the environment, equipment, and management are in good condition. In terms of personnel behavior, it is necessary to pay attention to the experimenter’s compliance with the experimental specifications.
(2) The probability of gas leakage directly caused by safety management defects is extremely low, but problems with people, environment, or equipment are often accompanied by safety management defects, which increase the probability of gas leakage.
(3) The concentration of toxic and harmful gases has the greatest impact on the consequences of accidents. The ventilation state of the laboratory should be controlled strictly to reduce the probability of toxic and harmful gases reaching the critical point.
(4) Effective personnel protection, successful forecasting, and early warning can effectively mitigate the consequences of accidents.
With the lack of accident data in laboratories, the probability distribution of Bayesian nodes was obtained based on expert experience, but it could present comparatively quantitative risk distribution for reference. In addition, the model can provide guidance for risk analysis of other accidents related to laboratory safety, such as explosions and leakage of liquid reagents, by simply adjusting some nodes, node relationships, and conditional probabilities in the model. However, the model has little guiding effect on the risk analysis of safety issues with strong human initiative, such as suicide. Due to the convenience of the probability update of this model, with the constructed network, quantitative risk assessment can be performed quickly before each experiment to assist the safety management and decrease losses. In future work, with real-time monitoring data, such as toxic and harmful gas concentrations, this model can be dynamically mobilized to achieve real-time dynamic quantitative risk assessment and early warning of gas leakage in the laboratory to support the emergency decision and treatment.