A Fuzzy Inference Engine based Alarm Generation

for Critical Condition in Industry Environment

A Sudhanva Anuritha Juturu Arbeena Hassan
Department of Computer Science and Department of Computer Science and Department of Computer Science and
Engineering Engineering Engineering
Global Academy of Technology Global Academy of Technology Global Academy of Technology
RR Nagar, Bengaluru, Karnataka, India RR Nagar, Bengaluru, Karnataka, India RR Nagar, Bengaluru, Karnataka, India
[email protected] [email protected] [email protected]

Likith Raj V Dr.R Kanagavalli Dr.Roshan Joy Martis

Department of Computer Science and Department of Computer Science and Department of Computer Science and
Engineering Engineering Engineering
Global Academy of Technology Global Academy of Technology Global Academy of Technology
RR Nagar, Bengaluru, Karnataka, India RR Nagar, Bengaluru, Karnataka, India RR Nagar, Bengaluru, Karnataka, India
[email protected] [email protected] [email protected]

Abstract - In this paper, fuzzy set-based inference systems are desensitized to frequent false alerts, can result in critical
explored as decision support system for application in critical warnings being overlooked, thus compromising safety and
alarm generation in an industrial plant setup. The parameters efficiency. To overcome these issues, fuzzy inference systems
used in alarm generation are Temperature, Pressure and Water (FIS) are being implemented. Fuzzy logic handles imprecise
Level. The fuzzy systems explored are Mamdani, Sugeno and
ANFIS. The three fuzzy systems are evaluated on simulated data.
and noisy data effectively, allowing for dynamic thresholds that
The evaluation employs both 3-Fold and 10-Fold cross-validation adapt to changing process conditions. By incorporating expert
techniques to ensure robustness. Performance metrics, including knowledge into fuzzy rules, FIS can mimic human reasoning,
Accuracy, Sensitivity, and Specificity, are utilized to assess the providing more nuanced and accurate alarm generation. This
effectiveness of each system. Results indicate that the Mamdani results in reduced false alarms, improved detection accuracy,
and ANFIS systems consistently outperform Sugeno, with the and overall enhanced reliability of the alarm systems, thereby
ANFIS system exhibiting better results across all metrics. significantly improving safety and operational efficiency in
Specifically, for the 3-Fold cross-validation, ANFIS achieved an industrial environments.
accuracy of 88.68% ± 1.27, a sensitivity of 91.09% ± 1.59, and a Fuzzy logic-based inference systems have become highly
specificity of 85.93% ± 1.70. For the 10-Fold cross-validation,
ANFIS achieved an accuracy of 88.78% ± 1.25, a sensitivity of
effective instruments in the field of advanced control
91.11% ± 1.60, and a specificity of 86.13% ± 1.86 showing the techniques, especially when it comes to managing the
better performance and accurate alarm calls in the system. imprecision and uncertainties present in many industrial and
Additionally, the provided standard deviations highlight the environmental contexts. Of the many fuzzy inference models
reliability and consistency of the observed results. These findings available, the Mamdani and Sugeno models are two of the more
outweigh the potential of fuzzy set-based inference systems, well-known options, with unique benefits and approaches. The
particularly ANFIS which stands better followed by Mamdani, as objective of this study is to implement an automated alarm
reliable decision support tools for critical alarm generation in system in various industries of critical environments using
industrial plant setups, offering enhanced precision and reliability fuzzy inference models namely Sugeno, Mamdani and ANFIS.
in detecting abnormal conditions.
Keywords - Fuzzy logic, Mamdani, Sugeno, ANFIS, Industrial
The implementation of automated alehavesystems to
alarm automation, Consistency, Reliability. regulate vital parameters, such water flow, has become essential
in many industrial and environmental domains to guarantee
I. INTRODUCTION efficiency, safety, and legal compliance. Fuzzy logic-based
techniques have become well-liked instruments in the growing
Many traditional alarm systems in industrial environments field of advanced control methodologies because they can
have failed due to their reliance on fixed thresholds, which do effectively manage the inherent complexity and uncertainty of
not adapt to the dynamic nature of industrial processes. This these systems. On the other hand, choosing the right fuzzy
rigidity often leads to frequent false alarms or missed inference model is essential to guaranteeing the effectiveness
detections, as these systems are sensitive to noise and minor and dependability of these systems in practical settings. Fuzzy
fluctuations that are not indicative of critical issues. logic was first proposed by a professor at the University of
Additionally, traditional alarm systems lack contextual California of Berkeley named Lotfi A.Zadeh (1965). Zadeh
awareness, meaning they cannot account for variability in proposed the theory that the objects of fuzzy sets have no exact
process conditions, which further reduces their reliability. This boundaries and are not necessarily right or wrong and are
problem of "alarm fatigue," where operators become
expressed in degrees. Fuzzy systems are based on the notion of imprecise conditions in industrial processes. Unlike traditional
a fuzzy set, which maps the domain of the input to the domain binary logic, fuzzy logic accommodates the inherent vagueness
of the output. The fundamental difference between classical set of real-world data, enabling a more nuanced representation of
theory sets and fuzzy sets is the output value of the set. In system states and behaviours. This flexibility makes fuzzy
classical sets, there are only two output values of zero and one. inference engines well-suited for dynamic and complex
In fuzzy sets, there are many output values with highest degree industrial environments, where conditions can vary widely, and
of membership. conventional alarm systems may fail to provide timely and
In recent years, the Adaptive Neuro-Fuzzy Inference accurate warnings. The use of a fuzzy inference engine-based
System (ANFIS) has emerged as a formidable tool for modeling approach for alarm generation holds several advantages. By
complex systems, blending the adaptability of neural networks incorporating linguistic variables and fuzzy sets, it enables the
with the interpretability of fuzzy logic. ANFIS offers a unique formulation of rules that capture expert knowledge and
synergy, harnessing the nonlinear modeling capabilities of experience, enhancing the system's ability to recognize and
neural networks while retaining the intuitive rule-based respond to critical conditions effectively. Moreover, fuzzy
reasoning of fuzzy logic. This hybrid approach enables ANFIS inference engines can adapt and learn from historical data,
to effectively capture intricate relationships within data, making allowing for continuous refinement and improvement of alarm
it particularly well-suited for applications in critical industrial generation strategies over time. This paper proposes the
environments where uncertainties abound. By seamlessly development and implementation of a fuzzy inference engine-
integrating neural network learning mechanisms with fuzzy based alarm generation system tailored specifically for
logic inference rules, ANFIS not only achieves high predictive industrial environments. By leveraging the principles of fuzzy
accuracy but also provides insights into the underlying logic, this system aims to enhance safety, efficiency, and
decision-making process. Its ability to adapt to changing data reliability in industrial processes by providing timely and
patterns and learn from experience renders ANFIS invaluable accurate alarms in response to critical conditions. Through a
in scenarios where traditional modeling techniques may fall comprehensive analysis and evaluation, the effectiveness and
short. Thus, ANFIS stands at the forefront of advanced control practicality of the proposed approach will be demonstrated,
methodologies, offering a powerful solution for tackling the offering insights into its potential applications and benefits in
complexities inherent in real-world systems. In recent years, the real-world industrial settings.
Adaptive Neuro-Fuzzy Inference System (ANFIS) has emerged
as a formidable tool for modeling complex systems, blending B. LITERATURE SURVEY
the adaptability of neural networks with the interpretability of Most of the natural phenomena are subjective and not
fuzzy logic. ANFIS offers a unique synergy, harnessing the objective. In the natural course of decision range of values used
nonlinear modeling capabilities of neural networks while is not precise. Hence, there is a need to model imprecision. In
retaining the intuitive rule-based reasoning of fuzzy logic. This 1965, L A Zadeh [1] proposed fuzzy sets and fuzzy extension
hybrid approach enables ANFIS to effectively capture intricate principle to model the imprecise data. There are two major
relationships within data, making it particularly well-suited for fuzzy inference engines [2] viz, Takagi-Sugeno-Kang model
applications in critical industrial environments where (briefly referred as Sugeno model) and Mamdani model. Deepa
uncertainties abound. By seamlessly integrating neural network et al [3], surveyed fuzzy inference models and artificial
learning mechanisms with fuzzy logic inference rules, ANFIS intelligence for automation and alarm generation in industry. In
not only achieves high predictive accuracy but also provides a recent study, [4] a supervised learning model, Adaptive Neuro
insights into the underlying decision-making process. Its ability Fuzzy Inference System (ANFIS) is used to optimize wind
to adapt to changing data patterns and learn from experience energy in solar photovoltaics.
renders ANFIS invaluable in scenarios where traditional Having successful usage of fuzzy models in different
modeling techniques may fall short. Thus, ANFIS stands at the applications in the current study authors have developed a
forefront of advanced control methodologies, offering a prototype of fuzzy inference system using Mamdani, Sugeno,
powerful solution for tackling the complexities inherent in real- and ANFIS for predicting hazardous alarms in an industrial
world systems. setup.

A. BACKGROUND C. MOTIVATION
In industrial environments, ensuring safety and efficiency Modern industrial settings rely on automated monitoring
is paramount, often requiring real-time monitoring and systems to identify potential dangers. However, traditional
response to critical conditions. Traditional alarm systems in threshold-based alarms often struggle with the complexity of
these settings rely on rigid thresholds and rules, which may not these dynamic environments. These systems frequently trigger
adequately capture the complexity and uncertainty inherent in false alarms or miss critical events, leading to unnecessary
industrial processes. As a result, there is a growing interest in downtime, safety risks, and reduced efficiency. This research is
adopting more adaptive and robust approaches for alarm driven by the need for more accurate alarm systems in the face
generation, particularly leveraging the capabilities of fuzzy of changing external factors and operating parameters.
inference engines. Fuzzy inference engines offer a promising Conventional threshold-based alerts often produce inaccurate
solution by allowing for the modelling of uncertain and results, causing disruptions and compromising safety. There's a
pressing need for intelligent alarm generation systems that can
identify subtle patterns and deviations hinting at potential B. Data Used
catastrophes. Furthermore, the ever-evolving nature of To conduct a comprehensive assessment of the FIS models,
industrial processes demands a shift towards more advanced we meticulously generated datasets that closely emulated real-
alarm generation methods. Traditional systems struggle to world environmental conditions. Leveraging the powerful
adapt to fluctuating operating conditions, potentially delaying computational tools available in the MATLAB environment,
critical warnings. Increased automation and efficiency across we employed the normrnd function to produce normally
sectors necessitate intelligent alarm systems with real-time distributed random data. This function allowed us to simulate
analysis capabilities to make informed decisions. This research realistic fluctuations in key variables, such as temperature,
proposes a novel approach using fuzzy inference engines to pressure, and water level, by specifying appropriate mean
bridge the gap in existing alarm generation techniques. Fuzzy values and standard deviations.
logic excels at capturing the inherent complexity and The FIS models are subsequently subjected to diverse testing
uncertainty of industrial processes. By incorporating fuzzy scenarios to gain valuable insights into their robustness and
logic, this research aims to enhance the accuracy, reliability, reliability, facilitating informed decisions regarding their
and flexibility of alarm systems. This will empower industries deployment in real-world industrial applications. This
to proactively manage risks and optimize efficiency. The comprehensive evaluation allows for the identification of
successful application of fuzzy logic across various industrial potential weaknesses or limitations within the FIS models,
domains highlights its potential. However, a deep dive into guiding further refinement and optimization efforts.
using fuzzy inference engines specifically for critical condition Additionally, by simulating rare but critical events, such as
detection in industrial settings remains lacking. This research equipment malfunctions or process deviations, the testing
addresses this gap by proposing a systematic methodology for process ensures that the FIS models exhibit resilience and
designing and implementing a fuzzy inference engine-based effectiveness in detecting and responding to anomalous
alarm generation system tailored to the needs of industrial situations.
applications. The goal is to refine current alarm generation
technologies, paving the way for safer, more robust, and C. Dataset Preparation:
efficient industrial operations in our increasingly This section outlines the comprehensive process for creating the
interconnected and complex world. This research project aims dataset that was utilized to assess the Mamdani, Sugeno, and
to contribute significantly to the advancement of alarm ANFIS models' performance.
generation technologies, ultimately promoting safer, more We created datasets for temperature, pressure, and water level
reliable, and efficient industrial processes. using MATLAB's ‘normrnd’ function in order to replicate
actual environmental conditions.
D. CONTRIBUTION AND ORGANIZATION OF PAPER
This paper introduces a fuzzy inference engine-based Table 1. Dataset Description
approach for alarm generation in industrial environments. Our
contributions include proposing a more adaptable and accurate PARAMETER LOW MEDIUM HIGH HAZARDOUS
method compared to traditional systems, improving precision Temperature(T) 40±12 70±10 100±20 180±40
to reduce false alarms, and emphasizing context-awareness for Pressure(P) 10±5 50±20 100±30 160±40
informed decision-making. The paper is structured to provide Water level(W) 10±5 20±8 32±6 -
an overview of challenges, a review of relevant literature,
details on methodology and experiments, analysis of results,
and concluding remarks. Through this organization, we aim to For temperature, pressure, and water level, we devised
comprehensively present our novel approach and its linguistic phrases to describe the subtle fluctuations in
implications for industrial automation and safety. environmental conditions. The four linguistic labels low,
medium, high, and hazardous were used to categorize each
II. MATERIALS AND METHODOLOGY metric. According to their mean (m) and standard deviation (s)
(expressed as (m+s)), these terms were represented by suitable
A. Study Design membership functions.
The primary objective of this study was to rigorously Temperature(T)
evaluate the performance of fuzzy inference systems (FIS) in Low: 𝑇 ∼𝑁 (𝑚low , 𝑠𝑙ow )
predicting alarms for water flow control systems. By leveraging Medium: 𝑇 ∼ 𝑁 ( 𝑚medium , 𝑠medium)
the capabilities of FIS models, specifically the Mamdani and High: 𝑇 ∼𝑁 (𝑚high , 𝑠high)
Sugeno approaches, we aimed to develop an effective alarm Hazardous: T~N(mhazardous, shazardous)
prediction system that could accurately forecast potential issues Pressure (P)
in water flow management based on environmental factors. Low: 𝑃 ∼ 𝑁 (𝑚low , slow)
This research endeavor was driven by the need to enhance the Medium: 𝑃 ∼ 𝑁 ( 𝑚medium, 𝑠medium )
reliability and proactive monitoring of water flow systems, High: 𝑃 ∼ 𝑁 ( 𝑚high , 𝑠high )
ultimately contributing to improved operational efficiency and Hazardous: 𝑃 ∼ 𝑁(mhazardous , 𝑠hazardous )
risk mitigation strategies. Water Level (W)
Low: W~N (mlow, slow)
 Medium: W~N (mmedium, smedium) extracted from data using techniques such as clustering or rule
High: W~N (mhigh, shigh) induction.
Hazardous: W~N( mhazardous, shazardous)
c) Flow Diagram:
The standard deviation (s) and mean (m) values were carefully The procedure of creating datasets was essential to
selected to represent the diversity within each category as well assessing how well fuzzy inference systems (FIS)
as normal environmental conditions. These membership performed in water flow control systems for the purpose of
functions played a key role in the fuzzy inference systems' predicting alarms. To replicate realistic environmental
fuzzification, which allowed the models to manage the inherent conditions, datasets were carefully created using the
unpredictability and uncertainty in the input data. MATLAB environment.
A dataset of 5810 samples with three features,
D. Overall Methodology temperature, pressure, and water level was considered. The
{normrnd} function's parameters, which comprise mean
values and standard deviations that represent normal
environmental fluctuations, were carefully selected to
replicate the variability found in real-world circumstances.
These datasets allowed for a thorough assessment of the
Mamdani and Sugeno FIS models' effectiveness in water
flow control system alert prediction through training and
testing.

Fig. 1. The Overall Methodology

Methodology followed is depicted in Fig 1

• Benchmarking Analysis: Once both FIS models are

trained, their performance is assessed using
benchmarking analysis. This analysis will allow us to
compare the two models and determine which one
performs better for your specific application

E. Fuzzy Inference System

The methods followed are:

Fig 1a: Structure of Fuzzy Logic Controller
a) Fuzzification:
In the fuzzification stage, crisp input variables are d) Fuzzy Inference:
transformed into fuzzy sets using membership functions. This 1. Sugeno: The Sugeno method, also known as the Takagi-
process allows for the representation of imprecise or uncertain Sugeno-Kang (TSK) model, is a type of fuzzy inference
data in a fuzzy manner, enabling the fuzzy inference engine to system where the output is represented by a crisp value,
make decisions based on linguistic variables rather than typically a linear combination of the input variables. This
precise values. method starts with fuzzification, converting crisp inputs
(𝑥−𝑐) 2
− (1) into degrees of membership within fuzzy sets. The rule
𝑀=𝑒 2𝜎 2
base consists of IF-THEN rules, but unlike Mamdani, the
𝑀𝐿 𝑀𝑀 𝑀𝐻 𝑀𝐻𝑧 (2) consequent part of each rule is a precise mathematical
𝑇 ={ + + + }
𝑇𝐿𝑜𝑤 𝑇𝑀𝑒𝑑𝑖𝑢𝑚 𝑇𝐻𝑖𝑔ℎ 𝑇𝐻𝑎𝑧𝑎𝑟𝑑𝑜𝑢𝑠 function, often linear. The inference mechanism employs
the weighted average method to combine the outputs of all
𝑀𝐿 𝑀𝑀 𝑀𝐻 𝑀𝐻𝑧 (3) rules, weighted by their respective firing strengths. This
𝑇𝑁 = 𝑎𝑣𝑔𝑚𝑎𝑥 { + + + }
𝑇𝐿 𝑇𝑀 𝑇𝐻 𝑇𝐻𝑧
approach yields a crisp output directly, enhancing
computational efficiency. The Sugeno model is
b) Fuzzy Rules: particularly well-suited for systems requiring precise
Fuzzy rules form the basis of the inference process in a control actions and can seamlessly integrate with adaptive
fuzzy inference engine. These rules define the relationship techniques, making it ideal for complex and dynamic
between input variables and output variables in linguistic terms. systems. Its ability to produce exact numerical outputs
They are typically formulated based on expert knowledge or
 makes it highly effective for applications in control,
optimization, and modelling.
2. Mamdani: The Mamdani method is a widely used approach A. Software Framework:
to fuzzy inference, where the output is a fuzzy set defined We used matlab for simulation and modelling. The choice
by linguistic terms. This method begins with fuzzification, of python allows for flexibility, ease of implementation, and
converting crisp input values into degrees of membership access to powerful numerical computation tools.
within predefined fuzzy sets. The rule base comprises IF-
THEN statements that capture the relationship between B. Input Variables:
input variables and output fuzzy sets. The inference We considered three critical input variables:
mechanism applies these rules using fuzzy logic operations • Pressure: Measured in pascals (Pa).
like MIN (for AND) and MAX (for OR) to compute the • Temperature: Recorded in degrees Celsius (°C).
degree of membership for each output linguistic variable. • Water Level: Represented as a percentage (0% to 100%).
The resulting fuzzy outputs are then aggregated and
subjected to defuzzification, typically using the centroid C. Membership Functions:
method, to obtain a crisp value. Mamdani systems are Each input variable was associated with linguistic terms
celebrated for their simplicity and interpretability, closely (e.g., “low,” “medium”,” high” and “hazardous”) defined by
mirroring human reasoning processes. This makes them appropriate membership functions. We used gaussian
highly effective in applications requiring intuitive membership functions to capture the gradual transitions
decision-making and control, such as industrial automation, between these terms.
environmental control, and consumer electronics.
3. ANFIS: Adaptive Neuro-Fuzzy Inference Systems (ANFIS) D. Alarm Thresholds:
combine the adaptability of neural networks with the To trigger the alarm, we set predefined thresholds for
interpretability of fuzzy systems. ANFIS models are each variable:
trained using a hybrid learning algorithm that adjusts the • Pressure: High pressure beyond a certain limit.
parameters of fuzzy membership functions based on input- • Temperature: Extremely high or low temperatures.
output data pairs. This approach offers improved accuracy • Water Level: Abnormally low or high or hazardous
and generalization capabilities. water levels.
• The data model in the image shows a typical ANFIS
architecture with five layers. 5. Defuzzification:
• The data model in the image shows a typical ANFIS Defuzzification is the process of converting fuzzy output
architecture with five layers. sets into crisp values for decision-making or control actions.
• Input Layer: Takes raw data (like speed and load). Various methods, such as centroid, weighted average, or
• Fuzzification: Converts data into fuzzy values (like maximum membership, can be used for defuzzification,
"high speed" or "low load"). depending on the specific requirements of the application.
• Rule Layer: Applies fuzzy rules based on the
• fuzzified data. ∫ 𝑀𝑅 𝑁 (𝑇).𝑇.𝑑𝑇
• Normalization: Combines the outputs of the fuzzy
Z* = (4)
∫ 𝑀𝑅 𝑁 (𝑇)𝑑𝑇
rules.
• Output Layer: Produces a crisp final output (a decision 6. k-fold cross-validation:
or prediction). In our assessment of the fuzzy inference engine's
effectiveness and ability to generalize, we implemented both 3-
fold and 10-fold cross-validation techniques. In the 3-fold
strategy, the dataset was divided into three distinct subsets: one
for training the model, another for parameter optimization
through validation, and the final subset for evaluating overall
performance during testing. Similarly, in the 10-fold method,
the dataset underwent division into ten subsets, ensuring equal
representation of each class across all folds. Throughout
classification, one fold was exclusively reserved for testing
purposes, while the remaining folds were amalgamated for
training the classifier. This process iterated ten times,
Fig 2– Data Model for ANFIS systematically rotating through each subset for testing.
Subsequently, we calculated the average accuracy, sensitivity,
4. Simulation Environment and specificity across the ten folds to gauge the efficiency of
We conducted our experiments using a custom-built our proposed system.
simulation environment. Here are the key components:
7. Performance Evaluation: 7. Natural language is the foundation of fuzzy logic.
The performance of the fuzzy inference engine is evaluated Systems of fuzzy inference (FIS) are systems that use fuzzy
based on several metrics, including accuracy, precision, recall, logic to execute reasoning similar to that of a human. There are
and F1-score. These metrics provide insights into the system's a few known varieties of FIS, including Sugeno, Tsukamoto,
ability to accurately detect critical conditions while minimizing and Mamdani. The following are the three FIS’s differences:
false alarms. Additionally, real-world testing in industrial Any consequent in the Tsukamoto technique is represented
environments may be conducted to validate the practical by fuzzy sets whose membership functions are monotone. Each
effectiveness of the proposed alarm generation system. inference rule’s output, which is represented as crisp sets and is
Generally speaking, a fuzzy logic controller is made up of determined by the predicate. A weighted average is utilized to
four fundamental parts. The elements of fuzzy logic determine the ultimate outcome. Using the composition rule
controllers—fuzzification, rules base, inference, and with MAX, the Mamdani approach applies the MIN function
defuzzification—are depicted in Fig. 1 implications. MAX-MIN method is another name for the
Mamdani method. Determine a certain crisp value as the output
Within the fuzzification interface are the following by inferring from the fuzzy number’s output. The Sugeno
functions: method and the Mamdani method are comparable; however, the
A function known as "fuzzification" is used to transform input latter’s output is a constant or a set of linear equations rather
data into appropriate linguistic values that can be thought of as than a fuzzy set.
labels for fuzzy sets.
B. Scale mapping is used to transfer the range of values of input E. Mamdani Model Simulation
variables into corresponding universes of discourse. a) Rule Base:
C. fuzzification is the process of measuring the values of input We defined linguistic rules based on expert knowledge. For
variables. instance:
• If temperature is not hazardous and pressure is low
The application domain and control goal are the two or normal, then the output is OFF.
fundamental components of the rules. A database and linguistic • If temperature is high, then the output is ON.
fuzzy control rules base make up this system: • If pressure is hazardous, then the output is ON.
i) The data base offers the definitions required to build fuzzy • If water level is high, then the output is ON.
data manipulation and linguistic control rules. • If temperature is high and pressure is high and water
ii) The rules base uses a collection of linguistic control rules to level is low, then the output is ON.
represent the control goals and control strategy of the domain • If pressure is low, then the output is ON.
experts. • If temperature is low or medium or hazardous and
pressure is high and water level is medium or high,
In general, fuzzy logic is an extension of Boolean logic that
then the output is ON.
deals partially with the concept of truth. Classic logic states that
everything can be represented as a binary number (e.g., 0, 1,
b) Fuzzy Inference Engine:
black, white, yes, or no). In fuzzy logic, Boolean truth is
• The Mamdani model computed intermediate fuzzy
replaced by truth level. Membership value from 0 to 1, grey
output based on the input variables and the rule
level, black, and white, as well as the linguistic form of
base.
uncertainty such as “a little,” “medium,” and “very” are
• We used the centroid defuzzification method to
examples of truth levels in fuzzy logic.
obtain crisp alarm activation levels.
The advantages of fuzzy approach are related to results
associated with human cognitive characteristics, particularly in
F. Sugeno Model Simulation
situations involving concepts, patterns, and decision making in
Input Transformation:
an uncertain or unclear environment.
• Unlike the Mamdani model, the Sugeno model
Several explanations for why fuzzy logic is so popular include: directly mapped crisp input values to linear or
I. Fuzzy logic is a simple idea to comprehend. nonlinear functions.
Fuzzy reasoning’s underlying mathematical ideas are quite Output Calculation:
straightforward and simple to comprehend. • Weighted averages of output functions were
2. The flexibility of fuzzy logic is great. computed.
3. Inaccurate data can be tolerated by fuzzy logic. • The Sugeno model provided crisp output value
4. Complex nonlinear functions can be modelled by fuzzy
logic. G. Implementation of ANFIS:
5. Expert experiences can be directly established and applied The setup of ANFIS training utilizes MATLAB’s ANFIS
by fuzzy logic, bypassing the need for training. toolbox to facilitate the training process. Input variables,
6. Conventional control methods can be utilized in conjunction including temperature, pressure, and water level, are selected
with fuzzy logic. based on their relevance to the alarm prediction system and the
availability of data. Linguistic terms such as “low,” “medium,”
and “high” are defined for each input variable, along with provides average of sensitivity, specificity and accuracy
corresponding membership functions. These membership computed over all the folds.
functions are carefully tailored to capture the variability in input When comparing the sensitivity and specificity of the
data accurately. ANFIS training involves iterative optimization models, FIS Sugeno exhibited the highest sensitivity of 96.78%,
of fuzzy inference system parameters and neural network indicating its ability to correctly identify true positives.
weights to minimize the error between predicted and actual However, FIS Mamdani showed a higher specificity of 84.28%
outputs. compared to FIS Sugeno's specificity of 69.66%, implying its
capacity to accurately detect true negatives. ANFIS, with its
The trained ANFIS model is then validated and tested using superior accuracy, offers a balanced performance in terms of
separate datasets to evaluate its performance in terms of sensitivity and specificity, making it a promising approach for
accuracy, sensitivity, and specificity, ensuring robustness and alarm prediction in critical industrial environments.
reliability in critical industrial applications. Once validated and The ANFIS model demonstrated superior accuracy compared
tested, the trained ANFIS model stands ready for integration to both FIS Mamdani and FIS Sugeno, achieving an average
into industrial systems for real-time alarm prediction. In accuracy of 88.68% and 88.78% for 3-fold and 10-fold cross-
operation, the model continuously receives input data from validation techniques, respectively. This improvement in
sensor networks monitoring various parameters essential to accuracy suggests that ANFIS effectively captures the complex
system health. Leveraging its fused fuzzy inference system and relationships within the dataset, resulting in more precise alarm
neural network architecture, the ANFIS model swiftly predictions for water flow control systems.
processes incoming data streams. Through this, it provides Overall, the results suggest that ANFIS outperforms
timely assessments of potential alarm conditions, serving as an traditional FIS Mamdani and FIS Sugeno models in terms of
early warning system for operators. These predictive insights accuracy, highlighting its potential for enhancing alarm
empower proactive decision-making, enabling operators to systems' effectiveness in water flow control and other industrial
swiftly intervene and prevent potential disruptions or hazards applications.
within the industrial environment. As the system operates and The ability of ANFIS to adaptively learn from data and
accumulates additional data, periodic reassessment and capture complex nonlinear relationships contributes to its
refinement of the ANFIS model ensure its continued efficacy in enhanced performance. In real-world applications, ANFIS
maintaining operational safety and efficiency. offers a more reliable and precise alarm system in critical
industrial environments, improving operational efficiency and
IV. EXPERIMENTAL RESULTS risk mitigation strategies.

In this study, we explore the effectiveness of three fuzzy

inference system (FIS) models namely Sugeno, Mamdani, and
Adaptive Neuro-Fuzzy Inference System (ANFIS) for Table 2: Rules considered for fuzzy inference process
generating automated alarm outputs efficiently in critical
industrial environments. In Sugeno and Mamdani inference
systems, rules are manually prepared based on empirical RULE
TEMPERATURE PRESSURE
WATER
ALARM
observation. However, in ANFIS, rules are adaptively derived NO. LEVEL
such that error in alarm generation is minimum most. The input 1 Hazardous Low Medium ON
parameters such as T, P and W are subjected to all three systems 2 Not Hazardous Low Medium OFF
and the performance is observed on the dataset generated as 3 Not Hazardous Medium Low OFF
described in section II.B. The objective here is to identify the
4 Not Hazardous Medium Low OFF
system which provide best possible reliability and accuracy in
alarm generations in the industrial environment. The three 5 Hazardous - - ON
systems are implemented in MATLAB 2018a. The data is split 6 - Hazardous - ON
into training and testing sets using 3-fold and 10-fold cross 7 - - High ON
validation and the performance (sensitivity, specificity and
8 High - - ON
accuracy) is computed over all the folds and the average is
9 - Low - ON
reported.
As noted in Table 1 and Eq. (1) Gaussian membership 10 Low High Medium ON
function is used for fuzzification of all the input variables. The 11 Low High High ON
membership functions for T, P, W and Alarm as are depicted in 12 Medium High Medium ON
Figure 2. For Sugeno and Mamdani systems the rule base (input 13 Medium High High ON
output relation between variables) is as shown in Table 2. For
14 High High Medium ON
ANFIS, rule base is derived based on convergence of epochs in
15 High High High ON
multiple iterations using the principle of minimization of error
between data label and obtained output in iterations. Table 3
 Figure 2.a Membership functions for temperature
Figure 2.b Membership functions for pressure

Figure 2.c Membership functions for temperature

Figure 2.d Membership function for alarm

V. DISCUSSION

In this study, Mamdani inference provides higher

sensitivity, specificity and accuracy in comparison to Sugeno
inference. However, Sugeno inference provides consistent
performance as standard deviation of sensitivity, specificity and
Table 3: Performance Metrics of Fuzzy Inference Systems accuracy is less compared to Mamdani inference. It is observed
that ANFIS provides higher performance compared to
Mamdani and Sugeno inference. Performance of ANFIS is
FUZZY SENSITIVITY SPECIFICITY ACCURACY
more consistent than Mamdani inference as the standard
CLASSIFIERS (%) (%) (%) deviation between different folds is less. In all the three systems
sensitivity is higher and specificity is relatively lower. It
SUGENO indicates that whenever there is a true alarm it will be reported
(3-Fold cross- 96.61±0.62 67.92±1.97 83.21±1.46
with higher performance and whenever there is a false alarm, it
validation)
is reported with a lesser performance. It means if there is a real
SUGENO alarm it is not missed and the risk of catastrophe of the industrial
(10-Fold cross- 96.62±0.69 67.93±2.81 83.21±1.00
validation)
plant, loss of investment, facility and risk of life is prevented.
However, if there is a false alarm and if it is missed it doesn’t
MAMDANI cause any risk as no harm will be caused to the plant and the
(3-Fold cross- 91.05±1.02 85.91±2.00 88.66±1.29
validation) human resource working in it.
MAMDANI Following are the key contributions of this study
(10-Fold cross- 91.12±2.27 85.99±2.45 88.66±1.73
validation) 1. Proposition of a fuzzy set based inference in alarm
generation in critical industrial environments. To protect
ANFIS
(3-Fold cross- 91.09±1.59 85.93±1.70 88.68±1.27
tragedy, loss of life, damage to infrastructure and
validation) subsequent loss of business and reputation.
2. Among fuzzy models logical analysis on which model can
ANFIS
(10-Fold cross- 91.11±1.60 86.13±1.86 88.78±1.25 provide an improved performance is systematically studied
validation) and reported, the same best performing model can be used
in more complex industrial plants leading to a new
approach.

VI. CONCLUSION AND FUTURE SCOPE

To sum up, this research has provided an extensive analysis

of the Mamdani and Sugeno fuzzy inference models in the
setting of an automated water flow control system warning. We
have shown that these models are effective in correctly
predicting the alert status based on temperature, pressure, and in the development and deployment of complex control systems
water level data by applying fuzzy logic principles. According in practical settings.
to our research, the Mamdani model had greater specificity than
the Sugeno fuzzy inference system, although the former
demonstrated better sensitivity. Nevertheless, overall VII. ACKNOWLEDGEMENT
performance for both versions was comparable, and the The authors gratefully acknowledge the support provided
decision between them came down to the particular demands by GLOBAL ACADEMY OF TECHNOLOGY for their
and specifications of the application. assistance, which enabled the completion of this research.
Our fuzzy logic-based alarm system's success highlights fuzzy Additionally, we extend our thanks to Dr. Roshan Joy Martis
logic's promise in control systems where human knowledge and for their valuable guidance, insightful feedback, and technical
experience are crucial. Compared to traditional control systems, assistance throughout the duration of this study. Their
our approach provides improved resilience, adaptability, contributions have significantly enriched the quality of this
trustworthiness, and efficiency by efficiently managing work.
uncertainty and imprecision in input and output values.
VIII. REFERENCES
In the future, more improvements to the system's
functionality might be made by adding more input variables, [1] L.A. Zadeh Fuzzy sets. Department of Electrical Engineering and
Electronics Research Laboratory, University of California, Berkeley,
honing fuzzy rules, and thoroughly testing the system in a California, USA , June 1965.
variety of settings. Furthermore, our suggested system's [2] Naseer Sabri, S. A. Aljunid, M. S. Salim, R. B. Badlishah, R.
scalability makes it applicable to a wider range of control Kamaruddin, M. F. Abd Malek July 2013 Fuzzy Inference System: Short
applications than just managing water flow, such as controlling Review and Design International Review of Automatic Control
(I.RE.A.CO.), Vol. 6, N. 4 ISSN 1974-6059 July 2013.
room temperature, traffic light control, and automotive systems. [3] R Deepa et al Fuzzy Logic based control system and Artificial
We contribute to the development of intelligent automation and Intelligence in Industrial Automation 2020 IOP Conf. Ser.: Mater. Sci.
decision-making processes by increasing the use of fuzzy logic Eng. 995 012034.
in control systems, making systems safer, more effective, and [4] K. Rajkumar, K. Ambika, K. Abhinaya, A. Nafeeza Begum, K. Teena
and U. Viveka, "Hybrid Solar and Wind Power Generation System Using
more versatile in a variety of contexts. In the end, our research ANFIS Algorithm," 2023 Intelligent Computing and Control for
highlights the importance of fuzzy logic as a useful instrument Engineering and Business Systems (ICCEBS), Chennai, India, 2023.