Microsoft Word - 1.docx


 CHEMICAL ENGINEERING TRANSACTIONS  
 

VOL. 77, 2019 

A publication of 

 
The Italian Association 

of Chemical Engineering 
Online at www.cetjournal.it 

Guest Editors: Genserik Reniers, Bruno Fabiano 
Copyright © 2019, AIDIC Servizi S.r.l. 
ISBN 978-88-95608-74-7; ISSN 2283-9216 

Developing Data-based Bayesian Network for Human 
Operation Analysis during Hot Work  

Yanchun Wanga, Yi Liub,*, Wei Wanga, Jian Chena, Hanhai Dongb, Dongfeng 
Zhaob,Faisal Khanc 
a College of Mechanical and Electronic Engineering, China University of Petroleum (East China), Qingdao 266580, China 
b College of Chemical Engineering, China University of Petroleum (East China), Qingdao 266580, China 
c Centre for Risk, Integrity and Safety Engineering (C-RISE) , Faculty of Engineering and Applied Science, Memorial 
University, St. John’s, NL, A1B 3X5, Canada  

liuyi@upc.edu.cn 

Hot work has caused numerous accidents during the inspection and maintenance process in recent years in 
China. This paper collected the hot work accidents in China since 2007 and integrated them into a database.  
The database contained the information such as the time, site, sequence, source, type, cause and 
consequence. Through analysing the database, most accidents were relevant to human operations. We 
classified unsafe and irregular operations caused by human factors during hot work. The causes of hot work 
accidents involved a series of problems during the approval, supervision and construction. Integrating hot 
work procedures, we established the main safety evaluation framework containing three factors, i.e., approval, 
execution environment, and protective measures. The method of Bayesian Network (BN) aiming at solving 
uncertain problems was utilized to identify unsafe operations quantitatively during hot work. Eventually, eight 
key proposals were presented to avoid hot work accidents. The proposed framework was developed to assess 
human operations in chemical plants during hot work. Causes of hot work accidents in chemical plants mainly 
consisted of failure to handle permits of hot work, production system without isolation, unremoved flammable 
materials, and failure to install fire prevention measures. The results can provide compelling support for 
human management and training during hot work. 

1. Introduction 

In the production process of enterprises, especially the chemical plants determine the characteristics of high 
temperature and high pressure, flammable, explosive and easy poisoned because of their production 
characters. The system of production must carry out the necessary maintenance after a certain period of 
operation, among which the hot work maintenance operations are frequent (Li, 2017). According to the safety 
regulations for hot work of chemical production units, hot work refers to non-routine operations and can 
directly or indirectly generate open flames. Hot work is a technical, demanding, risky and special operation. 
Workers are at risk when doing hot work, including oil and gas, hazardous chemicals production and storage 
industries (Tang and He, 2015). Due to various reasons, dangerous accidents often occur during hot work. 
The types of accidents include fire and explosion, burning, mechanical damage, etc., causing huge losses and 
seriously threatening the lives of employees. Therefore, it is very important to take measures to prevent 
accidents. 
In recent years, many hot work accidents have occurred in China, most of which occurred during the 
inspection and maintenance process, and the consequences are serious. Hot work has attracted the attention 
of safety management personnel, but there is no reasonable way to prevent it. In the past research on the hot 
work accidents, most of them focused on the process and consequences of a single accident, such as the 
U.S. Chemical Safety and Hazard Investigation Board (CSB) investigated the Partridge Raleigh oil storage 

                                

 
 

 

 
   

                                                  
DOI: 10.3303/CET1977123 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Paper Received: 23 February 2019; Revised: 16 April 2019; Accepted: 14  July  2019 

Please cite this article as: Wang Y., Liu Y., Wang W., Chen J., Dong H., Zhao D., Khan F., 2019, Developing Data-based Bayesian Network for 
Human Operation Analysis during Hot Work, Chemical Engineering Transactions, 77, 733-738  DOI:10.3303/CET1977123  

733



tanks explosion and proposed seven key preventive measures for hot work in the tank fields. No one has 
conducted statistics on recent hot work accidents. After the accident statistics and analysis, we found that 
most of the accidents were related to human operations. Therefore, it is important to combine the specific 
steps of the hot work and gradually find out the common causes of accidents, standardize the operation of 
relevant operators, strengthen safety management, do a good job of safety precautions, and avoid accidents. 
Zhang (2005) put forward the steps of hot work: determination of the location of hot work; on-site analysis and 
implementation of safety measures; on-site risk notification; approval of hot work permits and confirmation of 
safety measures; on-site monitoring and implementation of operations; completion of inspection and 
acceptance; cleaning up the site . 
Researchers have dedicated effort to understanding how and why accidents occur. Katsakiori et al. (2009) 
postulated the theory and model of accident causation, including linear models and complex nonlinear models. 
Mawuli et al. (2017) summarized the models used in recent decades, showed potential tools for accident 
modeling, and how to implement BN, Fault Tree, FRAM, and other probabilistic modeling methods. After the 
accomplishment of the hot work accident database statistics, the paper aims to use the BN to form an 
evaluation framework to identify the key factors causing the hot work accidents. The main framework includes 
approval, execution environment and protective measures. The method of BN aims at forecasting possible hot 
work accident scenarios from past accidents data. In this methodology, the probabilities can be updated as 
new information becomes available. The BN can be used flexibly to consider the interdependence and 
condition of the factors involved in the hot work process. 

2. Bayesian Network and Sensitivity Analysis 

Friis-Hansen (2000) compared the outcome of the proposed model using BN and the output from an Event 
Tree Analysis. The study also applied the Bayesian network to diagnose misfire and leakage in a marine 
diesel engine and attempted to combine BN with structural reliability methods for requalifying a pipeline in the 
North Sea. Martin et al. (2014) used Bayesian networks to analyze the factors affecting the performance of 
tasks that involve a high risk of falls from ladders or other auxiliary equipment. Another application of BN was 
made by Myles et al. (2015), who applied BN to a model of risk propagation in a supply network. Shubharthi et 
al. (2016) proposed to develop a dynamic fault tree for a chemical process system/sub-system based on BN 
and provided the dynamic Bayesian network to demonstrate the dynamic operational risk assessment 
methodology. 
BN is a graphical network based on probabilistic reasoning, which mainly describes the uncertainty of 
knowledge (Ding et al., 2009). BN was first proposed by Pearl (1988), which is an effective tool for applying 
probability theory to uncertainty reasoning. BN follows a qualitative part (Directed Acyclic Graph) and a 
quantitative part (Conditional Probability Table), represented by N=<G, P>. G is represented as a directed 
acyclic graph of <V, E> having N nodes. V={X1,...,Xn} represents a set of variables, and the directed edges E 
represent a probability causality relationship between nodes. P is the quantitative part, which refers to the 
conditional probability of each variable under the parent node, and is represented by specifying a conditional 
probability table (CPT) on each node. The basis of BN is the Bayes theory, which is expressed as: 

 
(1) 

Pearl (1988), Zhang and Thai (2016), and Pristrom et al. (2016) have explained in detail the principle of BN. 
Uncertainties exist in the use of BN to model scenarios. It may be in the form of the probabilities used, 
including those derived from expert opinions by Mawuli et al. (2017). Finding the critical factors from BN is the 
primary task of modeling. Lliwg (2015) suggested that the most effective way to identify the most important 
parameters in a BN model is to implement sensitivity analysis. To determine the most important factors in a 
hot work accident scenario is one of the main goals of this present study. 
The basic reasoning of a BN is to calculate the posterior probability distribution of an investigated variables set 
for a given observations variable set V. BN inference is mainly based on the Bayesian network's expression of 
the joint probability distribution of variables, and the process of quickly calculating the probability value to be 
evaluated. Researchers have proposed many effective algorithms for reasoning, including precise and 
approximation algorithms. These algorithms have been developed into software packages to facilitate the 
application of BN. The GeNIe software developed by the University of Pittsburgh Decision Systems Laboratory 

P( | ) = ( | ) ( )( )  

734



greatly simplifies the calculation process and enables evaluations and diagnostic results to be very intuitive 
from the diagram. This paper uses GeNIe to model and reason hot work accident scenario. 

3. Hot work accident scenario model 

3.1 Collection of past hot work accidents 

It has been found some literature describing hot work accidents that occurred during facilities life cycle. 
However, a complete record of accidents, as well as occurrence sequence, its probable causes and 
consequences still need to be added effectively. The procedure for establishing a database of hot work 
accidents is similar to that of the database of accidents and incidents at fuel ethanol facilities. There are four 
main sources of data, including internet, literatures, reports, and newspapers. Due to limited resources, only 
reported incidents were collected. The information collected mainly contains the time, location, process, 
causes and consequences of the accident. A total of 67 hot work accidents have occurred since 2007, 58 of 
which occurred in petroleum and chemical plant areas and caused dangerous consequences of fire and 
explosion. In addition to this, there are many less serious hot work accidents, no specific place and time, so it 
is not listed. Figure 1 shows the number of accidents per year from 2007 to 2018. According to the collected 
hot work database, the causes of accidents are analysed primarily for the unsafe operations of people. The 
model established has three main frameworks: approval, operating environment and protection measures.  
 

 

Figure 1: The number of hot work accidents occurring each year 

3.2 Classify potential human operational factors and build BN model 

This step performs hazard identification in the scene of hot work. To identify the potential factors for each step 
in conjunction with the operational procedures and the hot work safety precautions required by the National 
Fire Protection Association (NFPA) standards. Any hot work spot must have a permit. The approval process is 
the entire procedure of implementing the safety measures, including the certification, isolation of production 
systems, purging, replacement, removal of combustibles, gas detection and fire-fighting measures, etc. 
Secondly, the person in charge of the project must provide a safe disclosure to the executors of the hot work. 
Establishing a BN model based on potential human operational factors to obtain the probability of an accident 
scenario. Each node of the network has two states: “yes” or “no”.  “Yes” indicates the positive affirmation of 
cause for a particular variable, while “no” is a negative indication. In this paper, the source of prior probability 
has three channels involved reports, expert opinions, and statistical accident database. In the reports of fire 
and explosion fault trees in petrochemical industries such as oil tanks and liquefied petroleum gas, most 
causes of the accidents originated from combustibles and open flames generated by illegal hot work. In this 
type of fault tree analysis, illegal hot work appeared as intermediate events rather than specific to every 

735



human operation. Therefore, the expert scoring method was adopted. To effectively reduce the prior 
probability error, the multiple interpolation method is used to refit the prior probability of the expert opinion, 
reducing the adverse subjective interference in the human evaluation and improving the credibility of the 
expert opinion. At the same time, combined with the frequency of the hot work accidents in the past ten years, 
these accidents database can be used as a guide for experts to choose reasonable probability. The essence 
of this study is not to require quantitative accuracy, but to illustrate the proposed method and have a 
quantitative way of how to make decision-making opinions when similar incidents occurred. 
The BN model for common causes of hot work accidents is shown in Figure 2. The causes are divided into 
three aspects: approval, execution environment and protective measures. The main causes of approval failure 
process include the contractor, hot work certificate, and signature work without risk identification. The 
contractor and the hot work certificate were further broken down. The poor working environment may be due 
to the presence of flammable and explosive substances, ineffective supervisor and no communication, and the 
causes of flammable and explosive substances contain four contents. The protective measure's failure can be 
attributed to no fire-fighting measures, improper labor insurance supplies, improper emergency handling, and 
the use of non-explosion-proof tools. 

 

 

Figure 2: The BN for hot work accident 

3.3 Sensitivity analysis 

The causes of the accident are composed of multiple factors. Finding the most critical factors require 
sensitivity analysis, which can provide information about resource allocation for decision making and identify 
the most critical variables or factors in an accident scenario. In the process of performing sensitivity analysis, 
the basis for the calculation of the change ratio is Eq. (2). The sensitivity analysis results are shown in Figure 
3. The probability of the most sensitive factor is higher than the others. Only the most significant changes are 
shown in the figure. Sensitivity analysis results are subjective and guided by the probabilities of other 
variables. It’s highly dependent on decision makers. Therefore, it is crucial to use this methodology as the first 
step to decision making and choose the variables for prioritization when preventing scene changes. It is also 
important for reducing the impact of an accident should it occur. 

 
(2) 

    ℎ  1 −  

736



No f
ire-f

ighti
ng m

easu
res

Und
ocum

ented

With
out r

isk i
dent

ifica
tion

Unre
mov

ed fl
amm

able
 mat

erial
s

Prod
uctio

n sys
tem 

with
out i

solat
ion

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

 

C
ha

ng
e 

ra
tio

Parameters

 

Figure 3: Change ratio of the casual factors of hot work accidents 

4. Conclusions and discussion 

The application of Bayesian Network methodology for the analysis of causes during hot work has been 
presented. The use of Bayesian network provides analysts with the opportunity to establish interdependencies 
between casual factors, which is not possible in conventional methods such as Fault Tree. Sensitivity analysis 
is performed to define the critical factors leading hot work accidents, it is the best way identify the most 
contributory factors. The result is key to making a decision on the construction and supervision of resources 
for accident prevention during hot work. 
The result in Figure 3 showed that in the cause of human operations in hot work accidents, variables with 
significant change mainly include failure to install firefighting measures, failure to handle permits of hot work, 
without risk identification, unremoved flammable materials and production system without isolation. Among 
them, the setting of firefighting measures and the approval process of hot work permit are the major reasons. 
During the operations in the process of construction, the importance of human beings cannot be neglected. 
The occurrence of fires and explosions was caused by the staff's failure to carry out the above procedures, so 
education and training managers and operators are particularly important. According to the consequences of 
sensitivity analysis, eight key measures were proposed to prevent hot work accidents: 

1. Written permission— A professional who conducts a hazard identification, reviews and approves all hot 
work, issues a hot work permit, and clarifies the content of the work. 

 2. Risk Assessment— Conducting a risk assessment carefully before starting a hot work. Identify 
potential risks and propose risk control measures based on the job site, operating procedures, and the 
surrounding environment. 

3. Emergency— Preparing emergency plans and setting fire prevention measures. If a fire is triggered, it 
can be put out in time. 

4. Isolation— Adding blind plates to equipment, pipes, etc. to prevent leakage, and parking operations. 
5. Elimination—Cleaning up the combustibles, and the flammable materials under and below the hot point 

must be cleaned up. Cleaning and replacement are required when working in the equipment.   
6. Gas Monitoring—Gas monitoring of the working area using a flammable gas detector before and during 

the hot work. 
7. Training—Safety management personnel and operators should receive safety training for hot work, and 

hold a certificate. 
8. Supervising the contractor—Selecting a qualified contractor to inform the operational risk and the 

safety supervisor was in place. 

737



These recommendations and factors require conducting and monitoring and should be given more attention to 
prevent hot work accident occurrence. The above recommendations also can apply to other special work, 
such as confined space operations. 
The hot work accidents collected in this article are far from enough. If you want to study the hot work accidents 
better, you will collect accidents at home and abroad through relevant departments and conduct detailed 
analysis. The management of people in hot work needs to start with the suggestions, but in actual operation, 
unexpected situations often occur, including human mental state and physical condition, etc. These factors are 
not considered in this paper. More precise research needs to consider these aspects. 
The BN has the ability to express uncertainty and conduct uncertainty reasoning, and it can flexibly predict 
and diagnose the system. This paper combined Bayesian Network to effectively assemble expert knowledge 
and experience to analyse and discuss the causes of hot work accidents. The data used might have errors, 
and the estimation of top event probability was not accurate. Accurate research needs to collect more data 
and consider the complexity of the operating scheme. It can be simulated using the Taylor series and Monte 
Carlo method mentioned by Liwag (2015), and the obtained values will be more accurate. 

References 

Afenyo, M., Khan, F., Veitch, B., & Yang, M., 2017, Arctic shipping accident scenario analysis using Bayesian 
network approach, Ocean Engineering, 133, 224–230. 

Barua, S., Gao, X., Pasman, H., & Mannan, M.S., 2016, Bayesian network based dynamic operational risk 
assessment, Journal of Loss Prevention in the Process Industries, 41, 399–410. 

Ding, S.B., Ru, Y., 2009, Airline flight safety system evaluation based on Bayesian Network, Journal of Civil 
Aviation University of China, 27, 30–36. 

Friis-Hansen, A., 2000, Bayesian network as a Decision Support Tool in Marine Application, Ph.D. Thesis, 
Technical University of Denmark, Copenhagen, Denmark. 

Garvey, M.D., Carnovale, S., & Yeniyurt, S., 2015, An analytical framework for supply network risk 
propagation: a Bayesian network approach, European Journal of Operational Research, 243, 618–627. 

Katsakiori, P., Sakellaropoulous, G., Manatakis, E., 2009, Towards an evaluation of accident investigation 
methods in terms of their alignment with accident causation models, Safety Science, 47, 1007–1015. 

Liwåg, H., 2015, Survivability of an ocean patrol vessel-analysis approach and uncertainty treatment, Marine 
Structures, 43, 1–21. 

Li, W.J., 2017, Accident analysis and safety control measures in hot work, Petrochemical Safety Technology, 
6. 

Martin, J.E., Rivas, T., Matias, J.M., Taboada, J., Argüeelles, A., 2009, A Bayesian network analysis of 
workplace accidents caused by falls from a height, Safety Science, 47, 206–214. 

Pearl, J., 1988, Probabilistic reasoning in Intelligent Systems: Networks of Plausible 
Inference, Morgan Kaufmann Publishers, Inc., 70, 1022–1027. 
Pristrom, S., Yang, Z., Wang, J., Yan, X., 2016, A novel flexible model for piracy and robbery assessment of 

merchant ship operations, Reliability Engineering & System Safety, 155, 196–211. 
Tang, B., He, C. (Ed), 2015, Partridge Raleigh tank fire accident explosion accident gives us enlightenment, 

Tianjin Gu'an Enterprise Management Consulting Co., Ltd, China. 
Zhang, G., Thai, V.V., 2016, Expert elicitation and bayesian network modeling for shipping accidents: a 

literature review, Safety Science, 87, 53–62. 
Zhang, X.Y., 2005, Hazard analysis and measures taken for hot work accidents, 21, 27–29. 

738