Artificial Intelligence Improving Safety and Risk Analysis: A Comparative
Analysis for Critical Infrastructure
A. Guzman1, S. Ishida3, E. Choi4, A. Aoyama2.
1,2,3,4
Graduate School of Technology Management, Ritsumeikan University, Ibaraki, Japan
(gr0263vi@ed.ritsumei.ac.jp)
Abstract - Recently, the sustainability of traditional
technologies employed in critical infrastructure brings a
serious challenge for our society. In order to make decisions
related with safety of critical infrastructure, the values of
accidental risk are becoming relevant points for discussion.
However the challenge is the reliability of the models
employed to get the risk data. Such models usually involve
large number of variables and deal with high amounts of
uncertainty. The most efficient techniques to overcome those
problems are built using Artificial Intelligence (AI).
Therefore, this paper aims to investigate and compare AI
algorithms for risk assessment. These algorithms are
classified mainly into Expert Systems, Artificial Neural
Networks and Hybrid intelligent Systems. This paper
explains the principles of each classification system, as well
as its applications in safety. Lately, this paper performs a
comparative analysis of three representative techniques,
such as Fuzzy-Expert System, Neural Networks, and
Adaptive Neuro Fuzzy Inference System.
Keywords – Risk Assessment, Safety, Artificial
Intelligence, ANFIS
I. INTRODUCTION
The sustainability of traditional technologies related with
critical infrastructure constitutes one of the major
challenges that today’s societies and firms are facing.
Besides that, the adaptation of those technologies to the
effects of the climate change in sensible environments
represents a critical concern for safety and risk
management. Regarding this issue, reference [1] argue
that social consequences of catastrophic risks are
increasing rapidly, due mainly to the concentration of
people and critical infrastructure in hazard-prone areas,
aggravated by the lack of knowledge about the risks.
Additional to the social consequences described
above, and considering the industrial sector as critical
infrastructure due to its large impact to the economy in
case of a failure the relevance of industrial safety has
become a critical issue for the current society [2]. Then,
regarding the safety concern, industrial operators and
regulators have been performing risk assessments in
attempts to evaluate accurately probabilities of failure of
the infrastructure, and consequences associated with those
failures.
However, estimating accidental risks in critical
infrastructure involves a substantial effort and costs due to
number of variables involved, complexity and lack of
information. Therefore, this paper aims to investigate and
compare artificial intelligence algorithms for risk
assessment, which could be capable to deal efficiently
with the complexity and uncertainty.
The advantage point of the artificial intelligence
methods is that they have been largely employed in risk
assessment as efficient engineering tools to solve
978-1-5090-3665-3/16/$31.00 ©2016 IEEE
nonlinear problems in complex environments of
uncertainty. For example, the findings of a comparative
study presented in [3] suggest that the accuracy of the
artificial intelligence methods is superior to the traditional
statistical methods regarding nonlinear patterns in risk
assessment. The basic idea of using artificial intelligence
(AI) methods in safety is focused in the objective of
imitating the human expertise scoring risks and setting
tolerance levels.
As stated by the AI pioneer, Herbert Simons [4], the
AI have introduced new techniques that enhance the
procedural rationality of decision-makers, supporting
them in handling complexity and great masses of
information.
The most popular AI methods found in the literature, and
employed in risk assessment, are mainly classified as
Expert systems, Artificial Neural Networks and Hybrid
Intelligent Systems.
Hence, this study intends to investigate the AI
methods, and compare its performance in evaluating risks,
specifically evaluating risks related with the safety of
critical infrastructure.
The body of this paper is divided in four sections. The
first section describes the Expert Systems and their
applications in safety. The following section defines the
Artificial Neural Networks, and presents their related
applications. The third section describes the Hybrid
intelligent Systems and presents their uses in safety.
Lastly, the algorithms are implemented and a comparative
analysis is discussed.
II. METHODOLOGY
The methodology employed by this study follows a
hybrid technique of empirical research, which combines
qualitative and quantitative analysis. The qualitative
analysis is focused in the investigation of risk assessment
methodologies using artificial intelligence in the field of
process safety. On the other hand, regarding the
quantitative analysis, a comparison of risk assessment
methodologies is performed using simulation analysis.
III. EXPERT SYSTEMS FOR SAFETY ANALYSIS
Expert systems are broadly defined as a structured
methods that contain a body of knowledge that emulates
human expert decision-making skills in a specific domain
of expertise [3]. Hence, the concept behind expert systems
is basically the process of transferring the expertise from
a human to a computer program [5]. Then, the knowledge
is stored in a database of the computer, and by using a set
of inference procedures, operators can make complex
decisions that might require substantial human expertise
for their solution. As shown in Fig. 1, a basic expert
471
 Proceedings of the 2016 IEEE IEEM
system comprises three main components: A user
interface, a knowledge database and an inference engine.
Expert Knowledge
Data Base  extracted from conceptual risk matrixes. A typical rule
Input 1 Output 
Inference Engine
Interface
Operations  Interface
Fig. 1. Framework of a basic expert system
Thus, as presented in the framework of Fig. 1, expert
knowledge constitutes the core component of the whole
process. Here, the knowledge can be represented using a
variety of methods like semantic nets, frames, or
predicated logic. However, the most common method for
knowledge representation is the conditional rule method
(if-then), also known as the production rule method [3].
The inference engine is then referred as the part that
processes input attributes and searches for rules in the
knowledge base to achieve a final solution [6].
There are several sub-categories of expert systems that
could be found in the literature. These sub-categories
basically differ by the nature of the inference engine.
Among the most relevant categories for safety, these
methods include: Rule-based systems (RB), knowledge-
based systems (KB), fuzzy expert systems (FB), and case-
based reasoning (CB).
In general, the main advantages of this type of artificial
intelligence are established by its capacity to deal with
uncertain or missing data, and its ability to combine
quantitative and qualitative data [7]. Because of these
advantages, expert systems have been widely employed in
safety domains. Regarding safety of critical infrastructure,
expert systems have been employed successfully in risk
assessment, knowledge base maintenance, probabilistic
fault diagnosis, engineering failure analysis, and incident
management. A sample of these applications are
categorized in Table I.
TABLE I
Sample of application of Expert Systems (ES) in Safety
Sub-
Application Authors
Category
Rule base maintenance RB Higa and Lee (1998)
Probabilistic fault diagnosis RB Leung (2000)
Engineering failure diagnosis KB Graham-Jones (1995)
Plant incident management KB Finch and Lees (1997)
Fault diagnosis CB Yang et al. (2004)
Fault diagnosis FB Lee et al. (2000)
*
Source: [5]
A. Risk assessment using expert systems
An artificial system for accidental risk assessment
could be considered as a procedure that emulates the
decision-making of an expert regarding the safe operation
of certain infrastructure [6]. Defining expert as the person
who is able to solve safety-related problems that the
general population cannot, or a person who is able to
solve those problems much more efficient.
One common way of represent expert knowledge in the
safety field is as a collection of conditional rules of the
form if-then, which are extracted mainly from risk
assessment matrixes. Therefore, as stated in [8], in an
472
expert system for safety, the relationship between the
input variables (Probability and Consequence) and the
output variable (Risk) is defined by conditional rules
can be read as: If the Probability of Failure is Low and the
Consequence of Failure is Very High, Then the Risk of
Failure is Medium.
These rules also can be built to deal with uncertainty
through certainty factors and probability-based
algorithms, such as approaches that employ fuzzy logic
[9]. For this reason fuzzy inference systems are
commonly applied in risk determination.
A simple illustration of a fuzzy inference system for
risk assessment is depicted in Fig. 2. In this
representation, the crisp inputs for probability (P) and
Consequence (L) are mapped into fuzzy triangular sets,
which belong to a particular rule for risk determination.
If Probability is: And Consequence is:  Then Risk is:
Low Very High  Medium
Rule 
P L  
Fig. 2. Fuzzy inference system for risk assessment
B. Expert-base algorithm for Risk Assessment
In order to compare the performance of safety
analysis using AI techniques, this paper intend to extract
representative algorithms from each of those techniques.
For the case of expert systems, the algorithm evaluates
accident attributes based on inputs for probabilities and
consequences of failure. A representative algorithm,
presented in [7], emulates the thinking process of experts
in the field of pipelines safety. This expertise is combined
with deterministic values in a knowledge engine. This
process is schematically depicted in Fig.3.
Inference engine
Knowledge Base
Rules (Risk Matrix)
Probability
Inference Process
Fuzzy Logic Operations
Consequence Risk
Fig. 3. Expert-base algorithm for risk assessment [7]
The inference process is implemented following an
aggregative inference process and a centroid method of
defuzzification, typically computed as follows:
∗
∙
1
Where ∗ is the crisp value for the fuzzy risk
distribution ( , and is the membership function of
the risk distribution.
In summary, the B. Expert-base algorithm tests the
attributes of probabilities and consequences, and based on
the fuzzy inference process, assess whether the risk is
tolerable or not. Then, the critical characteristics of an
accidental failure can be analyzed in detail to implement
effective safety measures.
 Proceedings of the 2016 IEEE IEEM
IV. ARTIFICIAL NEURAL NETWORKS IN SAFETY
DOMAINS
A large number of techniques in AI have been
inspired by biological mechanisms with the aim of
emulate human learning and bio-evolution [10]. One of
the most representative techniques inspired by the
biological structure of the brain is the artificial neural
network, which consists of a large amount of
interconnected neurons that cooperate each other to
produce an output stimulus.
Artificial neural network (ANN) could be defined as
information processing systems which use learning and
adaptive capabilities [3]. This systems has been applied
extensively in computational modeling of subjective
information, decision making and forecasting
applications.
The basic design of an ANN consists of a number of
nodes connected through directional links. Each node
symbolizes an information processing unit, and the arrows
symbolize the causal link between connected nodes [11].
A common architecture of an ANN is shown in Fig. 4, in
which the nodes are represented by circles, some of which
serve as input nodes, output nodes, and hidden nodes.
X1 Input Layer Hidden Layer Output Layer
K
Y1
X2
K K
K value on the output vector. These hidden neurons are
Y2
K K
Xn K
Fig. 4. Common architecture of ANN
The algorithm of an ANN could be described as the
information flow between inputs ( , , … ) and
outputs ( , ) through a pattern of connectivity of
nodes. This connectivity is characterized by the
connection strength between nodes, its weight, and the
layers arrangement. Each node has an activation level,
which is computed by using the activation rule (function
that combines the inputs) and produces an output
according with the output function [10].
The learning process of an ANN is referred as the
iterative process of adjustment applied to its connection
weights with the aim of adapting to its real environment.
Then, as stated in [11], the learning process takes part
when the free parameters of an ANN are modified
through a process of stimulation by the environment in
which the network is set. One of most accurate learning
methods for ANNs is described by a regression analysis
called group method of data handling (GMDH), which
consists in the determination of the optimal configuration
of the ANN and its parameters employing polynomial
theory.
The applications of ANN can be classified in two
major generic classes: Static pattern recognition class, and
dynamic control and monitoring [12]. Regarding
applications in safety domains, these are mainly
associated with hazard identification, diagnostic systems,
monitoring systems, safety alarms, supervisory systems,
chemical sensors programming, risk assessment, and
473
maintenance planning. A sample of these applications are
categorized in Table II.
TABLE II
Sample of application of ANN in Safety of Infrastructure
Application Sub-Category Authors
Fault diagnosis Static Pattern Hung et al. (2004)
Hazard identification Static Pattern Ruiz (2000)
Monitory systems Dynamic Pattern Seker (2003)
Alarm systems Dynamic Pattern Pehrsson (2000)
Risk assessment Static Pattern Sii (2000)
Fault diagnosis GMHD Witczak (2005)
A. Risk assessment using ANNs
Along with the learning abilities of the ANNs, other
advantage that potentiates their application in safety and
risk assessment is related with the capability of analyzing
large quantities of data to determine patterns under
uncertain environments. Therefore, ANNs serve as
powerful tools to deal in situations where quantitative
relationships are not available or inaccurate [13].
As described in [14], the risk assessment method
employing ANNs consists basically of a set of sensory
units that constitute the input layer, a hidden layer of
computation, and an output layer of computation. The
neurons in the input layer compute the information
regarding Probability of failure , and Consequence of
failure . The neurons in the hidden layer extract the data
from the input vector and use them to estimate the risk
responsible for the reasoning part of the process.
Regarding the learning stage, this is carried out by
following either a back propagation analysis or a feed
forward methodology, such as the GMDH method.
B. ANN algorithm for Risk Assessment
A representative algorithm to evaluate risks using
ANNs is presented in Fig. 5. Same as the expert-base
algorithm, the ANN algorithm estimates the risk based on
inputs for probabilities and consequences of failure.
Then, as depicted in Fig. 5, the input values are processed
in the first layer by the neurons with the probability and
consequence functions and . Later on, in the
hidden layer each neuron has an activation value into it,
obtained from the neurons of the previous layer; these
activation values are multiplied by the strength of the
associated weight. These products are summed and passed
through a transfer function that fractionize the summed
inputs into a scale from 0 to 1. Finally, this values are
computed in the output neuron in order estimate the risk
value. When the final value obtained from the output
layer cannot meet the acceptable error criteria, it is
necessary to adjust the weights in an iterative process.
Fig. 5. ANN algorithm for risk assessment
 Proceedings of the 2016 IEEE IEEM
Then from Fig. 5, processing an input vector
k produces an output denoted by:
, ,
Lately, the error is obtained comparing the output
value with a target output , which is denoted by:
V. HYBRID INTELLIGENT SYSTEMS FOR SAFETY
ANALYSIS
Hybrid Intelligent systems (HIS) are defined as
efficient and robust systems that synergistically combine
complementary strengths of diverse computational
intelligence techniques, and overcome the weakness of
the processing capabilities to solve difficult learning tasks
[15]. HIS not only represents the interaction of multiple
intelligent techniques but also integrates intelligent
techniques with current computer systems and databases
[3]. Then, single intelligent technologies serve as the base
structure to build algorithms that may emulate human
decision-making to support complex processes.
As stated in [16], the research in the HIS field work
has been focused mainly in areas like micro-level
integration of fuzzy-expert systems and neural networks
among others. The reason for the integration of these
intelligent techniques is established by their
complementary attributes in learning capability,
brittleness and explanation. The learning capability is
referred to the ability of the model to adapt itself to its
current environment. Brittleness is referred to the inability
of an intelligent system to deal with lack of information,
inexact variables or inconsistent knowledge. And the
explanation property is referred to ability to provide users
with explanations of the reasoning process employed [16].
Table III shows a comparative scoring of these
complementary attributes.
TABLE III
Complementary Attributes of Intelligent Systems
Attributes Score (1-5)
Technique
Learning Coping with Explanation
ability Brittleness
ES (Conventional) 1 1 5 computes the overall output as a summation of values
ES (Fuzzy) 1 5 4
ANN 5 5 1 coming into this node. This could be expressed as:
*Source: [16], Scale 1: weak, 5: Strong)
Although the application of HIS in safety is relatively
new, most of this related applications are associated with
infrastructure health monitoring, structural damage,
seismic hazards and risk analysis. A sample of these
applications are categorized in Table IV.
TABLE IV
Sample of application of HIS in Safety of Infrastructure
Application Methodology Authors
Infrastructure monitoring Multi-Agent Liang (2015)
Structural damage Neuro-Fuzzy Wang (2008)
Seismic hazards Neuro-Genetic Liu (2014)
Risk analysis Neuro-Fuzzy Fragiadakis (2013)
A. Risk Assessment and HIS
Regarding risk evaluations, HIS has been recently
applied because their synergistic features coping with
uncertainty and learning ability. Besides that, most of
applications related with safety and risk assessment are
, a neuron nonlinear complex systems embedded in dynamic
environments, which means that the information is
2 difficult to capture and the global optimal solution is hard
to obtain with a single intelligent algorithm [17].
Therefore, risk assessment techniques using HIS tend to
3
be a synergy between fuzzy expert systems and neural
networks. In most of these applications, HIS techniques
use fuzzy numbers to represent the data of probabilities
and consequences of failure. Later on, this techniques
construct a neural network based on inference rules from
the risk matrix. Lately, a process of calibration or training
of the network is performed in order to adjust the learning
pattern using fuzzy numbers.
As stated in [3], the main advantages of this risk
assessment techniques using HIS are the ability of dealing
with linguistic representations, qualitative aggregation of
the knowledge base, and the learning ability of the neural
networks. These techniques are known in the literature as
Neuro-Fuzzy systems, Fuzzy Adaptive Networks (FAN),
or Adaptive Neuro-Fuzzy Inference Systems (ANFIS).
B. HIS algorithm for Risk Assessment
Among the recent applications of hybrid techniques
in safety, the most representative architecture is
characterized by the ANFIS structure, which is depicted
in Fig. 6.
The ANFIS algorithm is defined as a multilayer network
which uses neural network learning structure and fuzzy
reasoning with the aim of mapping inputs, such as
probabilities and consequences of failure, into an output
value, such as the risk.
As can be seen from Fig. 6, the ANFIS structure consists
of five layers of processing. In the first layer all nodes are
adaptive nodes that generate membership grades from the
inputs. In the second layer, the nodes are fixed nodes with
a multiplication function denoted by Π. The third layer
consists of fixed nodes with a normalization function
denoted by N. The fourth layer consists of adaptive nodes
that compute the consequent parameters of the fuzzy
process with the normalized values. Lastly, the fifth layer
4
Where, represents the number of inputs ( and ),
represents the number of conditional rules, and , , and
are the consequent parameters of the fuzzy process.
Fig. 6 ANFIS structure for two inputs.
474
 Proceedings of the 2016 IEEE IEEM

VI. ALGORITHMS IMPLEMENTATION AND
RESULTS
Regarding the comparison analysis of the algorithms
presented in sections III, IV and V of this paper, the
dataset employed was taken from a pipeline study-case
presented in [7]. Then, the objective was to take three
samples of information and calculate the mean absolute
percentage error (MAPE) of each intelligent algorithm.
In order to make a fair comparison of the artificial
intelligence algorithms, the MAPE is calculated based on
same inputs and trained with previous failure data over a
100 iterations.
A. Performance Analysis
The inputs for the ES, ANN, and ANFIS algorithms
were the probability of failure, ranging from 0 to 1, and
the consequence of failure, which is estimated in
monetary units using a logarithmic scale. Moreover, ES
and ANFIS algorithms employed a matrix of 25
conditional rules with a triangular membership function.
The three algorithms were implemented using Matlab.
Table V and Fig. 7 show the results for each intelligent
algorithm after 100 iterations.
100
90
Accuracy (100-MAPE)
error in less quantity of iterations. Therefore, ANFIS
might constitute a suitable choice to estimates accidental
risks for critical infrastructure.
Although the selection of the risk assessment
approach depends on the complexity of the environment
and the quality of the information available, the most
efficient and accurate assessment could be achieved by a
synergetic combination of artificial intelligent techniques.
The limitations established by these artificial
techniques, which are object of study for future research,
are related with the subjectivity of the defuzzification
methods and commonly insufficient data for activation
and training the ANNs.
ACKNOWLEDGMENT
The corresponding author would like to extend his
gratitude to the Otsuka Toshimi Scholarship Foundation
for contribute to the realization this research paper.
REFERENCES
[1] Ermolieva, T., & Erm0liev, Y. (2005). Catastrophic Risk
Management: Flood and. Applications of Stochastic Programming.
[2] Groth, M., & Cortekar, J. (2015, April). Is there a need for
government interventions to adapt energy infrastructures to climate
change? A German case study. EGU General Assembly
Conference
[3] Bahrammirzaee, A. (2010). A comparative survey of artificial
intelligence applications in finance: artificial neural networks,

80 expert system and hybrid intelligent systems. Neural Computing

70 and Applications, 19(8), 1165-1195.
60 [4] Simon, H. A. (1996). The sciences of the artificial. MIT press.
50 Expert System  [5] Liao, S. H. (2005). Expert system methodologies and applications a
40 (Fuzzy) decade review from 1995 to 2004. Expert systems with
30
ANN applications, 28(1), 93-103.
[6] Nakatsu, R. (2009). Rule ‐ Based Reasoning with Diagrams.
20 ANFIS Diagrammatic Reasoning in AI, 188-227.
10 [7] Guzman, A. & Aoyama A. (2016), Measuring the Benefit of Invest
0 in Pipeline Safety using Fuzzy Risk Assessment. Journal of Loss
0 20 40 60 80 100 Prevention in the Process Industries,
Number of Iterations [8] Jamshidi, A., Yazdani-Chamzini, A., Yakhchali, S. H., &
Khaleghi, S. (2013). Developing a new fuzzy inference system for
Fig. 7 Performance of AI techniques in Risk Assessment
pipeline risk assessment. Journal of loss prevention in the process
TABLE V industries.
Performance of AI techniques in Risk Assessment [9] Bose, B. K., Patel, N. R., & Rajashekara, K. (1997). A neuro-
MAPE Accuracy Iterations for fuzzy-based on-line efficiency optimization control of a stator flux-
Technique
(%) (%) Convergence oriented direct vector-controlled induction motor drive. Industrial
ES (Fuzzy) 16 84 1 Electronics, IEEE Transactions on, 44(2), 270-273.
ANN 5.8 94.2 35 [10] Ekbia, H. R. (2010). Fifty years of research in artificial
ANFIS 5.5 94.5 17 intelligence. Annual review of information science and technology,
44(1), 201
[11] Kantardzic, M. (2011) Artificial Neural Networks. Data Mining:
VII. CONCLUSIONS AND DISCUSSION Concepts, Models, Methods, and Algorithms, Second Edition, 199
It has been demonstrated that artificial intelligence [12] Lisboa, P. J. (2001). Industrial use of safety-related artificial neural
methods constitute an important source of methodologies networks. HSE Books.
[13] Sing Sii, H., Wang, J., & Ruxton, T. (2001). Novel risk assessment
that certainly improve the accuracy of risk assessment of techniques for maritime safety management system. International
critical infrastructure. Therefore, the implementation of Journal of Quality & Reliability Management, 18(9), 982-1000.
this techniques might contribute for future decisions [14] Ramamoorti, S., Bailey, J., & Traver, R. O. (1999). Risk
regarding the safety of the society and sustainability of assessment in internal auditing: a neural network approach.
International Journal of Intelligent Systems in Accounting, Finance
industrial infrastructure. Management
Although the literature review show that most of the [15] Taha, I., & Ghosh, J. (1997). Hybrid intelligent architecture and its
applications of artificial intelligence methods in the field application to water reservoir control. INT J SMART ENG SYST
of safety are concentrated in expert systems and artificial DESIGN, 1(1), 59-75.
[16] Zili Zhang, Chengqi Zhang (2004) Agent-Based Hybrid Intelligent
neural networks, the development of hybrid techniques Systems, Basics of Hybrid Intelligent Systems Volume 2938 of the
constitute a big opportunity to create knowledge in this series Lecture Notes in Computer Science pp 13-28.
field. [17] Liu, M., He, Y., Wang, J., Lee, H. P., & Liang, Y. (2015). Hybrid
From the performance results of the artificial intelligent algorithm and its application in geological hazard risk
assessment. Neurocomputing, 149, 847-853.
intelligence algorithms, it is clear that ANFIS perform
better that their counterparts achieving smaller values of

475