See discussions, stats, and author profiles for this publication at: [Link]

net/publication/380977408

A Comparative Analysis of Support Vector Machine and Decision Tree

Algorithm for Predicting Fault in Uninterruptible Power Supply Systems

Article in International Journal of Innovative Technology and Exploring Engineering · May 2024
DOI: 10.35940/ijitee.F9871.13060524

CITATIONS READS

0 39

3 authors, including:

John Kojo Annan Benjamin Odoi

University of Mines and Technology UMaT
10 PUBLICATIONS 1 CITATION 23 PUBLICATIONS 53 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Benjamin Odoi on 30 May 2024.

The user has requested enhancement of the downloaded file.

 International Journal of Innovative Technology and Exploring Engineering (IJITEE)
ISSN: 2278-3075 (Online), Volume-13 Issue-6, May 2024

A Comparative Analysis of Support Vector

Machine and Decision Tree Algorithm for
Predicting Fault in Uninterruptible Power Supply
Systems
Isaac M. Doe, John K. Annan, Benjamin Odoi

Abstract: Power supply systems can have problems, and Ghana Larger industries utilize uninterruptible power supplies
Gas Limited is not an exception. Ghana Gas Limited uses an (UPS) as an emergency power source to make sure business
intricate Uninterruptible Power Supply (UPS) system which is activities continue as usual in the event that the primary
made up of several parts such as electromechanical components,
PCB boards, and electrolytic capacitors. The majority of power supply fails. When it comes to providing almost
components have technical lifespans that are governed by usage, instantaneous safety from input power failures, a UPS is
operational environment, and working conditions, such as different from a backup or supplemental generator since it
electrical stress, working hours, and working cycles. Most of the uses electricity produced by batteries, ultracapacitors, or
time, these errors affect the integrity and power supply after flywheels [2]. A variety of scenarios need the use of
manufacture. The issue is that it takes longer for the professionals uninterruptible power supply (UPS) to provide consistent and
who operate on this machine to recognize these flaws, which
makes it difficult for them to predict errors quickly or anticipate well-regulated AC voltages for critical workloads, including
the likelihood of faults happening in the system components at an computer servers, medical equipment, air traffic control
early stage for effective corrective action to be performed. Support systems, and communication networks [3]. It has been noted
vector machines (SVM) and decision trees were used in this study in numerous deployments that when system load increases
to anticipate faults for technical data scheduling of uninterruptible over time, an upgraded UPS with a larger capacity is needed
power supply systems for Ghana Gas Limited in an efficient [4]. As power grids grow and loads rise, the primary goal of
manner. Based on a comparative analysis using these two
techniques, faults in Ghana Gas Limited's power supply system a UPS is to ensure the stability and dependability of the
were predicted using a four-hour daily interval dataset on UPS electrical supply. Furthermore, as the system serves as a
recordings, including input voltage, battery voltage, battery precaution or protection against any unanticipated power
current, and alarm, spanning from August 2017 to October 2023. outages that the companies may suffer, it must be reliable,
The findings depicted that the support vector machine was more secure, and less prone to malfunctions [5] and [6]. However,
efficient in detecting the fault locations in the power supply system overloading, overvoltage, power fluctuations, and outdated
with an accuracy of 96.80%, recall of 99.80%, precision of 100 %,
F1-score of 93.15%. The results from the error metrics also components can result in malfunctions in UPS systems,
validate the measures in assessing the predictive ability of the which can cost the industry a lot of money in terms of lost
model with MAE of 0.42%, MSE of 1.18%, RMSE of 4.45%, R 2 of productivity and system replacement [4]. Fault prediction is
99.97%, RMSLE of 0.036%, and MAPE of 0.21%. the process of tracking and evaluating historical data to
Keywords: Power Supply System, Support Vector Machine, identify the presence of a failure in the power system, so that
Decision Tree Algorithm, Precision, Accuracy, Error Metrics actions may be taken to prevent accidents and assure system
recovery [7]. As a result, it is essential to implement fault
I INTRODUCTION prediction models that are efficient and accurate in
determining when faults may occur [1]. Fault prediction is an
The majority of engineers now find that their ability to essential technology and maintenance security technique that
distribute an uninterruptible electrical power supply and is more advanced than fault diagnosis, which is frequently
safeguard their operations is limited by how sophisticated carried out after issues have happened [8]. Making wise
they are at handling power supply outages [1]. decisions to avoid errors and minimize their negative
consequences is made easier with the help of fault prediction.
Manuscript received on 18 April 2024 | Revised Manuscript In order to reduce the frequency and duration of power
received on 03 May 2024 | Manuscript Accepted on 15 May 2024 outages, utility workers can find and eliminate persistent
| Manuscript published on 30 May 2024.
*Correspondence Author(s)
defects with the use of high-accuracy fault prediction in
Isaac M. Doe, Department of Electrical and Electronic Engineering, power systems [9].
University of Mines and Technology, UMaT, Tarkwa, Ghana. E-mail: Previously, UPS system maintenance was reactive, starting
bodoi@[Link], isaacmawulidoe@[Link], ORCID ID: 0009-0000-
8451-0064
when a problem was found. Recent years have seen an
Dr. John K. Annan, Department of Electrical and Electronic Engineering, increase in the popularity of preventive maintenance, which
University of Mines and Technology, UMaT, Tarkwa, Ghana. E-mail: involves changing out components to increase system
jkannan@[Link], ORCID ID: 0000-0001-8056-7880 reliability [10].
Dr. Benjamin Odoi*, Department of Mathematical Sciences, University
of Mines and Technology, UMaT, Tarkwa, Ghana. E-mail:
bodoi@[Link], ORCID ID: 0000-0001-6759-0744
© The Authors. Published by Blue Eyes Intelligence Engineering and
Sciences Publication (BEIESP). This is an open access article under the
CC-BY-NC-ND license [Link]

Published By:
Retrieval Number: 100.1/ijitee.F987113060524 Blue Eyes Intelligence Engineering
DOI: 10.35940/ijitee.F9871.13060524 and Sciences Publication (BEIESP)
Journal Website: [Link] 9 © Copyright: All rights reserved.
 A Comparative Analysis of Support Vector Machine and Decision Tree Algorithm for Predicting Fault in
Uninterruptible Power Supply Systems
The simplest kind of preventive maintenance is to target whether the UPS was, perhaps, in a clean room or in a
consumable parts before they are expected to reach the end of harsher, dustier environment [21]. Costly on-site
their useful life. Fans, AC and DC capacitors, and of course maintenance is carried out irrespective of the device's status
the battery are all consumable parts of UPS systems [11]. and may therefore be too late or too early. The latter situation
Preventative maintenance is the failure prediction based on could result in the servicing of a healthy component, thereby
prediction models created from gathered data or log files in increasing the company's financial costs and decreasing the
more sophisticated systems [12]. Making the right decisions UPS systems’ reliability [22].
to prevent power system problems and estimating their A model for accurate fault prediction and forecasting will
likelihood can both be aided by the analysis of historical data help to improve the level of UPS system reliability and reduce
[13]. In general, utilizing electrical measurement data to its power quality disturbances as well as equipment damages
fullest potential will increase the accuracy of failure [22]. Furthermore, the prediction model will enable Ghana
prediction and guarantee the stability and dependability of the Gas to better manage their engineering resources by
UPS system. Studies utilizing artificial intelligence (AI) and forecasting failures and enabling precautionary actions to be
machine learning have been developed in the past several performed in order to reduce operating expenses caused by
years to predict faults. Right now, it's a worthwhile and unnecessary component replacement and additional charges.
pressing topic [13] [14]. provided a more accurate prediction An important step in understanding the reliability of the UPS
strategy for optimized Artificial Neural Networks (ANN) system as a whole is determining the significance of various
based on multilayer evolutionary algorithms in an effort to UPS parameters [23]. Hence, by monitoring UPS parameters
improve the fault forecasting model's accuracy [15]. such as output voltage, output current, frequency, power
employed ANN to obtain likelihoods of success for five fault factor, working hours, active and reactive powers, a run-time
prediction techniques ranging from 87% to 100% using 33 lifespan calculation of the components could be conducted to
data sets [16]. looked at the use of convolutional neural determine its health state and predict potential UPS failures.
networks (CNNs) to forecast refrigerant charge failures. Two Obviously, to conduct such a task, an intelligent data
classification and regression predictive models were methodology has to be employed. This study emphasis is on
suggested in order to predict the quantitative refrigerant in utilizing machine learning algorithms for fault evaluation and
both cooling and heating applications. In summary, the prediction in the UPS system. The research discusses fault
recommended tasks were finished with a 3.1% mistake rate prediction models with a particular emphasis on UPS
and 99% accuracy. In order to predict power converter operations. Particularly, operational data from UPS
failures, [17] created a fault prediction model utilizing installations has been recorded. In order to create the failure
Markov Chains Analysis based on data collected from several prediction models, data is then processed using Support
UPS installations. Furthermore, classification is an important Vector Machines (SVM), and Decision Tree (DT) algorithms.
part of the fault prediction process. The Support Vector
Machine (SVM) is a hyperplane-configured discriminant II MATERIALS AND METHODS
classifier. SVM-based applications have been shown to be
A. Data Collection
feasible in [18] and [19].
This study aims to estimate the failure rate of the UPS The data for the study were obtained by UPS (recordings
systems by using Support Vector Machine (SVM), and include input voltage, battery voltage, battery current, and
Decision Tree (DT) Algorithms to create prediction models alarm), which records an observation every four hours. The
that the Ghana Gas engineering team would use. The Ghana data was gathered between August 2017 and October 2023.
National Gas Company, also known as Ghana Gas, was The Python programming language was used to do the
established with the responsibility of creating, acquiring, and analysis.
overseeing the natural gas infrastructure required for the B. Support Vector Machine (SVM)
processing, transportation, and marketing of gas to satisfy the Given a set of training data {(𝑥1 , 𝑦1 ), … , (𝑥1 , 𝑦𝑁 )} , where
nation's needs for both household and commercial electricity. 𝑥𝑖 ∈ 𝑅𝐷 are the input vectors and 𝑦𝑖 ∈ {−1,1} are the
Most importantly, a consistent and dependable power source corresponding class labels, an SVM seeks to construct a
is essential to Ghana Gas's operations and activities. hyperplane that separates the data with the maximum margin
Predicting UPS failures is therefore necessary to enhance of separability [24]. 𝑁 is the number of observations, and 𝐷
power supply performance and lower the company's total is the dimension of the input vectors. The decision function
operational expenditure (OPEX). can be written as
𝑠𝑣
The UPS employed by Ghana Gas is a complicated system 𝑓(𝑥) = 𝑠𝑖𝑔𝑛{∑𝑁 𝑠𝑣 𝑠𝑣
𝑗=1 ⬚ 𝛼𝑗 𝑦𝑗 (𝛷(𝑥) ⋅ 𝛷(𝑥𝑗 )) + 𝑏} (1)
consisting of a number of parts, including PCB boards, 𝑠𝑣
Where 𝑥𝑗 are the support vectors, is 𝛷(𝑥) a nonlinear
electromechanical components (such as relays and fans), and
vector function that maps the input vector onto a higher
electrolytic capacitors. The majority of components’
dimensional feature space [25], 𝑦𝑗𝑠𝑣 is the label corresponding
lifespans are determined by their technical attributes and are
influenced by their usage, operational environment, and to the th support vector, 𝑁 𝑠𝑣 is the number of support vectors,
working conditions, that is, working hours, working cycles 𝑏 is a bias term, and 𝛼𝑗 are the Lagrangian multipliers, The
and electrical stress [20]. Currently, preventative inner product is (𝛷(𝑥) ⋅ 𝛷(𝑥𝑗𝑠𝑣 )) called the kernel function.s
maintenance is conducted over a predetermined time period
without considering the level of stress experienced or the
overall health of the UPS system. For instance, fans are
typically replaced every five years without considering
Published By:
Retrieval Number: 100.1/ijitee.F987113060524 Blue Eyes Intelligence Engineering
DOI: 10.35940/ijitee.F9871.13060524 and Sciences Publication (BEIESP)
Journal Website: [Link] 10 © Copyright: All rights reserved.
 International Journal of Innovative Technology and Exploring Engineering (IJITEE)
ISSN: 2278-3075 (Online), Volume-13 Issue-6, May 2024

C. Decision Tree (DT): discussion of the various evaluation and error metrics can be
In the shape of a tree structure, Decision Tree creates checked in the works done by [26]–[33] and [38]-[40].
models. The process gradually creates a decision tree for each
dataset by breaking it down into smaller and smaller sections. III RESULTS AND DISCUSSIONS
A measure used for segmentation is information gain. To This section presents the results and discussion of the
partition the data at the most informative features, we study. The exploratory and the ML models were discussed
establish an objective function. and how they were developed. The results show a novelty
which can be used as an alternative way to detect faults in
|𝑆𝑣 |
𝐼𝐺(𝑆, 𝐴) = ∑⬚
𝑣∈𝑉(𝐴) ⬚ |𝑆|
𝐸𝑛𝑡𝑟𝑜𝑝𝑦(𝑆𝑣 ) (2) uninterruptible power supply systems.
A. Results
𝑆𝑣 is a subset of set S equal to the attribute value of The UPS data includes four-hour time periods for each of
attribute v, and the range of attribute A is represented by the equipment's four shifts. In all, 3912-time frames were
V(A). The measure of impurity or randomness in a dataset is created over the course of three years. The objective is to
called entropy. Entropy is a quantity that is constantly generate statistical features for each window in order to better
between 0 and 1. define each daily signal and to reduce the data's
𝐸𝑛𝑡𝑟𝑜𝑝𝑦(𝑆) = ∑𝑐𝑖=1 ⬚ 𝑃𝑖 𝑙𝑜𝑔 𝑙𝑜𝑔 2𝑃𝐼 dimensionality. To further describe the signal and understand
(3) its evolution over time, eight different characteristics were
Where 𝑃𝑖 is the ratio of the sample number of the subset and extracted from each data frame.
the 𝑖 − 𝑡ℎ attribute value. Maximum, Minimum, Mean, Standard Deviation, Root
D. Evaluation Metrics Mean Square (RMS), Skewness, and Kurtosis Mean Absolute
Deviation (MAD) were features generated. These eight
This study compares many machine learning techniques in
variables were chosen to limit the study's resources and see
an attempt to forecast the location of a wire drawing process
whether they are adequate to achieve an identification. As
problem. Therefore, six (6) well-known assessment measures
none of the variables need frequency analysis, the features are
that are frequently utilized in fault prediction applications
all conducted in the time domain. Furthermore, according to
were used to gauge how well these algorithms performed.
[37], time domain statistical resources provide a high
Since a model's effectiveness cannot be determined by a
performance to characterize trends and changes.
single metric, these evaluation metrics were selected [34]-
[36]. These evaluation metrics conform to literature and a
Table 1: Extracted Statistical Features for the Input Voltage Attribute
Date Max Min Mean Std RMS Skew. Kurt. MAD
04/01/2021 423 418 419.83 1.9407 419.837 0.46611 -1.5259 1.5
05/01/2021 427 409 421.5 7.14843 421.5505 -0.75421 -1.2867 5.6667
06/01/2021 422 416 418.833 2.04124 418.8375 0.185073 -1.3889 1.5
07/01/2021 439 420 428.667 7.76316 428.7252 0.066101 -1.9978 6.6667
08/01/2021 440 417 428.667 9.88418 428.7711 0.014249 -2.0706 9.3333
09/01/2021 441 430 436 4.28952 436.0176 -0.15204 -1.9268 3.6667
10/01/2021 442 430 436.667 4.08248 436.6826 -0.32932 -1.3049 3
11/01/2021 429 421 426.25 3.59398 426.2614 -0.63615 -1.7614 2.625
12/01/2021 431 422 425.5 4.1833 425.5171 0.293685 -2.0338 3.5
04/02/2021 427 421 423.333 2.65832 423.3403 0.181385 -2.0485 2.3333

Table 2: Extracted Statistical Features for the Battery Voltage Attribute

Max Min Mean Std RMS Skew. Kurt. MAD
04/01/2021 255 251 253.50 1.9748 253.506 -0.4544 -1.977 1.6667
05/01/2021 255 254 254.67 0.5164 254.667 -0.5379 -1.958 0.4444
06/01/2021 255 250 253.83 1.9408 253.840 -1.1754 -0.398 1.2778
07/01/2021 254 252 253.50 0.8367 253.501 -0.8537 -1.172 0.6667
08/01/2021 255 235 253 5.6889 253.059 -2.6125 5.3613 3
09/01/2021 255 250 253.17 2.1370 253.174 -0.4858 -1.834 1.7778
10/01/2021 256 254 254.33 0.8165 254.334 1.36083 -0.083 0.5556
11/01/2021 254 251 252.5 1.2910 252.503 0 -2.078 1
12/01/2021 254 250 252.33 1.8619 252.339 -0.0918 -2.180 1.6667
04/02/2021 255 251 253.17 1.7224 253.172 -0.4059 -1.924 1.4444

Tables 1 and 2 show the first ten statistical features extracted from daily input and battery voltage values recorded by UPS,
which are Maximum, Minimum, Mean, Standard Deviation, Root Mean Square (RMS), Skewness and Kurtosis, and Mean
Absolute Deviation (MAD). Figure 1 shows a graphical representation of the features generated for each variable.

Published By:
Retrieval Number: 100.1/ijitee.F987113060524 Blue Eyes Intelligence Engineering
DOI: 10.35940/ijitee.F9871.13060524 and Sciences Publication (BEIESP)
Journal Website: [Link] 11 © Copyright: All rights reserved.
 A Comparative Analysis of Support Vector Machine and Decision Tree Algorithm for Predicting Fault in
Uninterruptible Power Supply Systems

Figure 1. All features for both the Input and Battery Variables
voltage data, and Minimum, Mean, RMS, and Kurtosis for
IV ML MODELS DEVELOPED FAULT battery voltage data. To boost model performance, the
PREDICTION model’s hyper-parameter parameters were fine-tuned. The
Machine learning methods were used to process the input radial basis function kernel was employed since it possessed
voltage and battery voltage statistical features. For the the lowest margin of error and the most iterations
classification task including fault identification, the most (1,000,000). The model’s performance is measured using
popular and appropriate classifiers were used. In the multiple evaluation metrics such as Accuracy, F1-score,
modelling section, the machine learning models will be Recall, and Precision.
trained on the training subset and its performance will be a. Model Evaluation
tested against the unknown testing subset, resulting in Table 3 and 4 illustrate the performance measures derived
confusion matrices and learning curves. from the obtained SVM fault classification for the input
voltage and battery voltage features, respectively. Table 3
A. SVM Model for UPS Fault Prediction
and 4 also demonstrate the performance for the training
The SVM algorithms were designed with the following dataset and the validation or testing dataset, respectively.
features: Mean, RMS, Maximum, and Minimum for input
Table 3: SVM Evaluation Metrics for the Input Voltage Attribute
Training
Model
Accuracy F1-Score Recall Precision
0.9577 0.9211 0.8611 1.0000
SVM - Radial Kernel Testing
0.9624 0.9315 0.8718 1.0000

Table 4: SVM Evaluation Metrics for the Battery Voltage Attribute

Training
Model
Accuracy F1-Score Recall Precision
0.9675 0.9140 0.8738 0.9732
SVM - Radial Kernel Testing
0.9474 0.8679 0.8214 0.9200

Table 3 shows that the radial basis function (rbf) kernel presented in Table 4. were slightly lower than those for the
was calibrated and produced the smallest error margins input voltage dataset.
having Accuracy, F1-Score, Recall, and Precision values of
0.9624, 0.9315, 0.8718, and 1.0000 for the input voltage
dataset. Furthermore, the testing data Accuracy, F1-Score,
Recall, and Precision scores for the battery voltage dataset
Published By:
Retrieval Number: 100.1/ijitee.F987113060524 Blue Eyes Intelligence Engineering
DOI: 10.35940/ijitee.F9871.13060524 and Sciences Publication (BEIESP)
Journal Website: [Link] 12 © Copyright: All rights reserved.
 International Journal of Innovative Technology and Exploring Engineering (IJITEE)
ISSN: 2278-3075 (Online), Volume-13 Issue-6, May 2024

b. Confusion Matrix c. Cross Validation

The confusion matrices for the input voltage and battery Cross validation is a technique for evaluating a machine
voltage datasets are shown in Figure 2 and 3. The matrices learning model’s performance by training it on multiple
show how many correct and wrong predictions the model subsets of data and evaluating it on the remaining data. The
produced for each class. The absolute numbers of models were validated and tested using the k-fold cross-
classification success rates are recorded on the diagonals of validation approach. In this investigation, ten folds or a k
the confusion matrices, while the misclassified samples are value of ten were used. Using this cross-validation approach,
on the parts of the matrix based on the distribution of the dataset is randomly divided into test and training data and
classification errors. For each confusion matrix, the predicted then divided into k groups. The model is validated on one of
label represents the predicted value of the provided sample by the groups, then training is done on the remaining groups.
the trained SVM algorithm, whereas the true label represents
the desired value of that sample.

Figure 4: Plot of SVM Learning Curve for the Input

Voltage Attribute

Figure 2: SVM Confusion Matrix for the Input Voltage

Testing Data
Figure 2 shows that the SVM model number accurately
identified all 94 failure-free data points (True Positive).
Similarly, 34 True Negative means that the model accurately
identified 34 faulty data points while misclassifying 5 of them
as No fault (False Positive).

Figure 5: Plot of SVM Learning Curve for the Battery

Voltage Attribute
Figure 4 and 5 compare a model’s performance on training
and testing data over a range of training instances. Figure 4
shows that the model is well-fitting, as indicated by training
and validation scores that increase to a point of stability with
a small difference between the two final score values. Figure
5 shows that the training score is exceptionally high
Figure 3: SVM Confusion Matrix for the Battery regardless of training instances, and the cross-validation
Voltage Testing Data score grows over time. There is also a considerable variance
between the training and testing scores, indicating that the
In addition, Figure 3 shows that the SVM model accurately
SVM model is overfitting the data.
identified 103 fault free data points (True Positive) and
misclassified 2 fault free data points as faults (False Negative) B. Decision Tree Model for UPS Fault Prediction
for the Battery Voltage dataset. Similarly, 23 True Negative The DT classification was carried out in order to categorize
means that the model accurately identified 23 fault data points the data into binary targets and to construct a classification
while incorrectly classifying 5 fault data points as No fault capable of correctly distinguishing between fault free and
data points (False Positive). Overall, the model appears to be faulty data points.
performing reasonably well, however it generates quite a few
false negatives for various classes.

Published By:
Retrieval Number: 100.1/ijitee.F987113060524 Blue Eyes Intelligence Engineering
DOI: 10.35940/ijitee.F9871.13060524 and Sciences Publication (BEIESP)
Journal Website: [Link] 13 © Copyright: All rights reserved.
 A Comparative Analysis of Support Vector Machine and Decision Tree Algorithm for Predicting Fault in
Uninterruptible Power Supply Systems
The criterion is the parameter that determines how the Tree algorithm are summarized in Tables 5 and 6. Table 5
impurity of a split will be measured. The Gini impurity was shows that the testing Accuracy, F1-Score, Recall, and
the criterion parameter used. Also, “min_samples_split” Precision values for the input voltage dataset are 0.9774,
parameter, which is the minimum number of samples 0.9620, 0.9744, and 0.9500, respectively. In addition, the
required to split an internal node, was set to 20. battery voltage dataset has Accuracy, F1-Score, Recall, and
a. Model Evaluation Precision values of 0.9699, 0.9286, 0.9286, and 0.9286 in
Table 6.
The classification performance results for the Decision
Table 5: DT Evaluation Metrics for the Input Voltage Attribute
Training
Model
Accuracy F1-Score Recall Precision
0.9706 0.9486 0.9444 0.9633
Decision Tree Classifier Testing
0.9774 0.9620 0.9744 0.9500

Table 6: DT Evaluation Metrics for the Battery Voltage Attribute

Training
Model
Accuracy F1-Score Recall Precision
0.9352 0.8175 0.7476 0.9399
Decision Tree Classifier Testing
0.9699 0.9286 0.9286 0.9286

The learning curve in Figure 8 indicates substantial test

variability and a high score up to around 280 instances, but
after this threshold, the model begins to converge on an F1
score of around 0.98. Because the training and test results
have not yet converged, this model could benefit from further
training data. Finally, this model suffers mostly from error
caused by variance (the test data scores are more variable than
the training data), suggesting that the model may be
overfitting.
Figure 9 also shows that the model has a very low training
score at first, which steadily increases as more training
examples are added. Both the training and testing scores
Figure 6: DT Confusion Matrix for the Input Voltage begin to fall at 250 samples, showing that adding more
Attribute examples will aid in the model’s convergence and stability.
The cross-validation graphs reveal the consistency and
stability of the model’s performance.

Figure 7: DT Confusion Matrix for the Battery Voltage

Attribute
Figure 6 shows that the DT model made 92 correct
predictions and 2 wrong predictions for the “No Failure”
class. Figure 8: Plot of DT Learning Curve for the Input
In addition, the model made 38 correct predictions and 1 Voltage Attribute
incorrect prediction for the “Fault” class. Similarly, for the
“No Failure” class, 102 data points were correctly classified
and 2 were incorrectly classified in Figure 7 data points in
the “Fault” class were accurately classified as faults, while 7
were wrongly labeled as No fault.
b. Cross Validation

Published By:
Retrieval Number: 100.1/ijitee.F987113060524 Blue Eyes Intelligence Engineering
DOI: 10.35940/ijitee.F9871.13060524 and Sciences Publication (BEIESP)
Journal Website: [Link] 14 © Copyright: All rights reserved.
 International Journal of Innovative Technology and Exploring Engineering (IJITEE)
ISSN: 2278-3075 (Online), Volume-13 Issue-6, May 2024

and the recorded UPS data from Ghana Gas was used in the
ML modelling in order to create the prediction models, two
variables (input voltage and battery voltage) were evaluated
using data obtained from the UPS during a three-year and
three-month period. Furthermore, the eight statistical features
were derived from both the input voltage and the battery
voltage in order to further characterize the data. In pattern
classification, the capabilities of machine learning algorithms
were used. In addition, the major input parameters used by
the models were the Mean, Min Max RMS, and Skewness.
After modelling, each algorithm’s performance for fault
classification was examined. To assess the efficacy and
capacities of the constructed models, four performance
Figure 9: Plot of DT Learning Curve for the Battery metrics were used: Accuracy, F-1 Score, Recall, and
Voltage Attribute Precision. While the model performed well overall, it was
discovered that it was unable to predict all classes with the
The dependability of any developed model is determined
same level of accuracy using battery voltage data. This
by its capacity to generalize adequately based on test
implies that there might be opportunities to enhance the
outcomes. As a result, a quantitative comparison of the
proposed ML models was performed, as shown in Table 7. model’s functionality.
Individual models were found to be capable of predicting
faults in the UPS. DT outperformed the SVM model in terms RECOMMENDATIONS
of Accuracy, F1-Score, Recall, and Precision. The proposed method’s robustness and accuracy
In Table 7, the DT-based prediction model had an demonstrated its potential for protecting UPS systems in
Accuracy = 0.9774, F1-Score = 0.9620, Recall = 0.9744, and major power companies. It provides companies additional
Precision = 0.9500 for the Input Voltage data, compared to support for equipment reliability decision-making, allowing
SVM, which had an Accuracy = 0.9624, F1-Score = 0.9315, them to be more competitive in the market. It is necessary to
Recall = 0.8718, and Precision = 0.9500. The DT model’s
conduct additional study in order to forecast equipment
Accuracy (0.9774) rating indicates that the model is
failure times as well as to conduct real-time online calibration
extremely dependable in forecasting failures in the TDS UPS
system. Similarly, for the battery voltage data, DT monitoring as data is being gathered. Also, the model enables
demonstrated superior performance to the SVM model. The an analysis of equipment data records, allowing the detection
importance of input voltage and battery voltage-based of faults without prior knowledge of the equipment's status.
variables to fault prediction was also discovered, with input Furthermore, the model is generalizable to any number of
voltage features better predictors for fault prediction. UPS systems. However, to improve the model in the future,
it advised to implement an online algorithm to diagnose and
Table 7a Summary of comparison between ML models
prognosis the equipment during an operation. An artificial
Input data neural network class like the Multi-Layer Perceptron (MLP)
Models
Accuracy F1-Score Recall Precision might be utilized to enhance the performance of the HMM.
SVM 0.9624 0.9315 0.8718 1.0000
DT 0.9774 0.9620 0.9744 0.9500 DECLARATION STATEMENT
Battery Data
SVM 0.9474 0.8679 0.8214 0.9200
Authors declare that there is no conflict of interest or
DT 0.9699 0.9286 0.9286 0.9286
whatsoever. The authors are solely responsible for this
research work and there was no funding from any industry or
Table 7b Error Metrics between SVM and DT Models whatsoever.
Funding No, I did not receive.
Model MAE MSE RMSE R2 RMSLE MAPE
No conflicts of interest to the best of our
Input Data Conflicts of Interest
knowledge.
SVM 0.42% 1.18% 4.45% 99.97% 0.04% 0.21% Ethical Approval and No, the article does not require ethical approval
DT 0.55% 2.25% 12.25% 98.23% 0.03% 0.31% Consent to Participate and consent to participate with evidence.
Yes, Data can be accessed upon request
Battery Data Availability of Data
through email: bodoi@[Link]/
and Material
SVM 0.52% 2.28% 5.26% 90.57% 0.04% 0.51% isaacmawulidoe@[Link].
DT 1.24% 1.84% 6.07% 90.22% 0.13% 0.63% Each author has made an independent
contribution to the article. The individual
contributions of each author are presented
V CONCLUSION below for clarity and transparency. Isaac M.
Authors Contributions
Doe, formulated the research problem,
This study introduced the notion of fault prediction and analysed, and discussed. John K. Annan,
demonstrated how businesses may use it to enhance their Benjamin odoi, supervised the whole research
maintenance cycles. The test result demonstrates that, with an work.
accuracy of 97.74% for input voltage features and 96.99% for
battery voltage features, the proposed decision tree’s fault
classification accuracy is superior to that of SVM. SVM and
Decision tree were employed as trained classification models,

Published By:
Retrieval Number: 100.1/ijitee.F987113060524 Blue Eyes Intelligence Engineering
DOI: 10.35940/ijitee.F9871.13060524 and Sciences Publication (BEIESP)
Journal Website: [Link] 15 © Copyright: All rights reserved.
 A Comparative Analysis of Support Vector Machine and Decision Tree Algorithm for Predicting Fault in
Uninterruptible Power Supply Systems
REFERENCES domain features. IEEE Access, 5, pp. 8682-8690.
[Link]
1. Nan, B., Chen, L., Rodrigo, N.D., Borodin, O., Piao, N., Xia, J., 19. Bigdeli, M., Vakilian, M. and Rahimpour, E., 2012. Transformer
Pollard, T., Hou, S., Zhang, J., Ji, X. and Xu, J., 2022. Enhancing Li+ winding faults classification based on transfer function analysis by
transport in NMC811|| graphite lithium‐ion batteries at low support vector machine. IET electric power applications, 6(5), pp. 268-
temperatures by using low‐polarity‐solvent electrolytes. Angewandte 276. [Link]
Chemie International Edition, 61(35), p.e202205967. 20. Linzen, D., Buller, S., Karden, E. and De Doncker, R.W., 2005.
[Link] Analysis and evaluation of charge-balancing circuits on performance,
2. Lukovic, M., Lukovic, V., Belca, I., Kasalica, B., Stanimirovic, I. and reliability, and lifetime of supercapacitor systems. IEEE transactions
Vicic, M., 2016. LED-based Vis-NIR spectrally tunable light source- on industry applications, 41(5), pp. 1135-1141.
the optimization algorithm. [Link]
Journal of the European Optical Society-Rapid Publications, 12, pp. 1- 21. Litchfield, N.J., Villamor, P., Dissen, R.J.V., Nicol, A., Barnes, P.M.,
12. [Link] A. Barrell, D.J., Pettinga, J.R., Langridge, R.M., Little, T.A.,
3. Guerrero, J.M., Matas, J., de Vicuna, L.G., Castilla, M. and Miret, J., Mountjoy, J.J. and Ries, W.F., 2018. Surface rupture of multiple crustal
2007. Decentralized control for parallel operation of distributed faults in the 2016 M w 7.8 Kaikōura, New Zealand, Earthquake.
generation inverters using resistive output impedance. IEEE Bulletin of the Seismological Society of America, 108(3B), pp. 1496-
Transactions on industrial electronics, 54(2), pp. 994-1004. 1520.
[Link] 22. Lind, H. and Muyingo, H., 2012. Building maintenance strategies:
4. Low, K.S. and Cao, R., 2008. Model predictive control of parallel- planning under uncertainty, Property Management, 30(1), pp. 14-28.
connected inverters for uninterruptible power supplies. IEEE [Link]
Transactions on Industrial Electronics, 55(8), pp. 2884-2893. 23. Aamir, M., Kalwar, K.A. and Mekhilef, S., 2016. Uninterruptible
[Link] power supply (UPS) system. Renewable and sustainable energy
5. Martin, F. and Aguado, J.A., 2003. Wavelet-based ANN approach for reviews, 58, pp. 1395-1410. [Link]
transmission line protection. IEEE Transactions on power delivery, 24. Cherkassky V, Mulier FM. Learning from data: concepts, theory, and
18(4), pp.1572-1574. [Link] methods. John Wiley & Sons; 2007 Sep 10.
6. Sharafi, A., Jafarian, P. and Sanaye-Pasand, M., 2010, May. A [Link]
combined algorithm for high speed transmission-line protection based 25. Suykens JA. Nonlinear modelling and support vector machines.
on traveling-wave. In 2010 9th International Conference on InIMTC 2001. proceedings of the 18th IEEE instrumentation and
Environment and Electrical Engineering, pp. 132-135). measurement technology conference. Rediscovering measurement in
[Link] the age of informatics (Cat. No. 01CH 37188) 2001 May 21 (Vol. 1,
7. Park T, Efros AA, Zhang R, Zhu JY. Contrastive learning for unpaired pp. 287-294). IEEE.
image-to-image translation. In Computer Vision–ECCV 2020: 16th 26. Antonanzas J, Osorio N, Escobar R, Urraca R, Martinez-de-Pison FJ,
European Conference, Glasgow, UK, August 23–28, 2020, Antonanzas-Torres F. Review of photovoltaic power forecasting. Sol
Proceedings, Springer International Publishing, Part IX 16 2020, pp. Energy. 2016;136: pp. 78-111.
319-345. [Link] [Link]
8. Katipamula, S. and Brambley, M.R., 2005. Methods for fault detection, 27. Sharma G, Tripathi V, Mahajan M, Srivastava AK. Comparative
diagnostics, and prognostics for building systems—A review, part II. analysis of supervised models for diamond price prediction. In:
Hvac&R Research, 11(2), pp.169-187. Proceedings of the Conflu 2021 11th Int Conf Cloud Comput Data Sci
[Link] Eng. Published online 2021: pp. 1019-1022.
9. Awadallah, M.A. and Morcos, M.M., 2003. Application of AI tools in [Link]
fault diagnosis of electrical machines and drives-an overview. IEEE 28. Fontana F A, Mäntylä MV, Zanoni M, Marino A. Comparing and
Transactions on energy conversion, 18(2), pp. 245-251. experimenting machine learning techniques for code smell detection.
[Link] Empir Softw Eng Vol. 2016;21: pp. 1143–1191.
10. Giuntini, L. and Brioschi, M., 2015, March. Markov chain analysis for [Link]
failure prediction of power converters. In 2015 IEEE International 29. Muhammad IA , Fabio P, Lin S, Qing W. Machine learning techniques
Conference on Industrial Technology (ICIT), pp. 1336-1341). IEEE. for code smell detection: A systematic literature review and meta-
[Link] analysis. Inf Softw Technol. 2019;108 (1), pp. 15-138.
11. Eusgeld I, Fraikin F, Rohr M, Salfner F, Wappler U. Software [Link]
reliability. Dependability Metrics: Advanced Lectures. 2008:104-25. 30. Ofosu A. R, Odoi B, Asamoah M, “Electricity consumption forecast
[Link] for Tarkwa using autoregressive integrated moving average and
12. Salfner, F. and Tschirpke, S., 2008. Error Log Processing for Accurate adaptive neuro fuzzy inference system”, Serbian Electr Eng.
Failure Prediction. WASL, 8, pp.4. 2021;18(1): pp. 75-94. [Link]
13. Zhang, X., Zhou, X., Lin, M. and Sun, J., 2018. Shufflenet: An 31. Twumasi-Ankrah, S., Odoi, B., Adoma Pels, W. and Gyamfi, E.H.,
extremely efficient convolutional neural network for mobile devices. 2019. Efficiency of imputation techniques in univariate time series.
In Proceedings of the IEEE conference on computer vision and pattern 32. Odoi, B., Brew, L. and Attafuah, C., 2021, “The use of Principal
recognition (pp. 6848-6856). Component Regression and Time Series Analysis to Predict Nitrous
[Link] Oxide Emissions in Ghana”, Energy Res. J., 2021, 12(1), pp.1-12
14. Tao L, Cinquanta E, Chiappe D, Grazianetti C, Fanciulli M, Dubey M, [Link]
Molle A, Akinwande D. Silicene field-effect transistors operating at 33. Odoi B., Boahen A. D. and Brew L. (2024), Predicting Nitrous Oxide
room temperature. Nature nanotechnology. 2015 Mar;10(3): pp. 227- Emissions in Ghana using Long Short-term Memory and Gated
31. [Link] Recurrent Neural Network, eNergetics 2023, 9 th Virtual International
15. Guardado, J.L., Naredo, J.L., Moreno, P. and Fuerte, C.R., 2001. A Conference on Science, Technology and Management in Energy,
comparative study of neural network efficiency in power transformers Conference, 81.
diagnosis using dissolved gas analysis. IEEE Transactions on Power 34. Rathakrishnan V, Bt. Beddu S, Ahmed AN. Predicting compressive
delivery, 16(4), pp. 643-647. [Link] strength of high-performance concrete with high volume ground
16. Eon, C., Breadsell, J., Morrison, G. and Byrne, J., 2019. Shifting home granulated blast-furnace slag replacement using a boosting machine,
energy consumption through a holistic understanding of the home learning algorithms. Sci Rep. 2022;12(1): pp. 1-16.
system of practice. Decarbonising the built environment: Charting the [Link]
transition, pp. 431-447. [Link] 35. Liu Q, Wang X, Huang X, Yin X. Prediction model of rock mass class
6_23 using classification and regression tree integrated AdaBoost algorithm
17. Giuntini, L. and Brioschi, M., 2015, March. Markov chain analysis for based on TBM driving data. Tunn Undergr Sp Technol. 2020; 106: pp.
failure prediction of power converters. In 2015 IEEE International
Conference on Industrial Technology (ICIT), pp. 1336-1341, IEEE.
[Link]
18. Jan, S.U., Lee, Y.D., Shin, J. and Koo, I., 2017. Sensor fault
classification based on support vector machine and statistical time-

Published By:
Retrieval Number: 100.1/ijitee.F987113060524 Blue Eyes Intelligence Engineering
DOI: 10.35940/ijitee.F9871.13060524 and Sciences Publication (BEIESP)
Journal Website: [Link] 16 © Copyright: All rights reserved.
 International Journal of Innovative Technology and Exploring Engineering (IJITEE)
ISSN: 2278-3075 (Online), Volume-13 Issue-6, May 2024

1-15. [Link]
36. Nguyen H, Vu T, Vo TP, Thai HT. Efficient machine learning models Disclaimer/Publisher’s Note: The statements, opinions and data
for prediction of concrete strengths. Constr Build Mater. 2021; 266: contained in all publications are solely those of the individual
pp.1-17. [Link] author(s) and contributor(s) and not of the Blue Eyes Intelligence
37. Saucedo-Dorantes J.J, Arellano-Espitia F, Delgado-Prieto M, Osornio-
Engineering and Sciences Publication (BEIESP)/ journal and/or the
Rios RA. Diagnosis methodology based on deep feature learning for
fault identification in metallic, hybrid and ceramic bearings. Sensors,
editor(s). The Blue Eyes Intelligence Engineering and Sciences
2021 Aug 30;21(17):5832. [Link] Publication (BEIESP) and/or the editor(s) disclaim responsibility for
38. Agyare, S., Odoi, B. and Wiah, E. N. Predicting Petrol and Diesel any injury to people or property resulting from any ideas, methods,
Prices in Ghana, A Comparison of ARIMA and SARIMA Models. instructions or products referred to in the content.
Asian Journal of Economics, Business and Accounting, 2024, 24(5),
pp. 594-608. [Link]
39. Twumasi-Ankrah, Sampson, Benjamin Odoi, Wilhemina Adoma Pels,
and Eric Herrison Gyamfi, "Efficiency of Imputation Techniques in
Univariate Time Series", International Journal of Science,
Environment and Technology, 8 (3), 2019, pp. 430 – 453.
40. Ofosu RA, Odoi B, Boateng DF, Muhia AM. Fault Detection and
Diagnosis of a 3-Phase Induction Motor Using Kohonen Self-
Organising Map. Jurnal Nasional Teknik Elektro. 2023 Mar 31. 12(1),
pp. 30-41.
41. Radhamani, V., & Dalin, G. (2019). Significance of Artificial
Intelligence and Machine Learning Techniques in Smart Cloud
Computing: A Review. In International Journal of Soft Computing and
Engineering (Vol. 9, Issue 3, pp. 1–7).
[Link]
42. Kanani, P., & Padole, Dr. M. (2019). Deep Learning to Detect Skin
Cancer using Google Colab. In International Journal of Engineering
and Advanced Technology (Vol. 8, Issue 6, pp. 2176–2183).
[Link]
43. Nikam, S. S., & Dalvi, Prof. R. (2020). Fake News Detection on Social
Media using Machine Learning Techniques. In International Journal of
Innovative Technology and Exploring Engineering (Vol. 9, Issue 7, pp.
940–943). [Link]

AUTHORS PROFILE
Ing. Isaac M. Doe, PE is an MSc Student in the Department
of Electrical and Electronic Engineering at the University of
Mines and Technology, UMaT, Tarkwa, Ghana. He is a
member of Ghana Institution of Engineering (GhIE), His
research interests include, SCADA Systems, Control
Systems and Networking, Power System Controls, Gas
Metering and Data Transmission, Gas Operations.

Dr. John Kojo Annan is a Senior Lecturer and the Head

of Department of Electrical and Electronic Engineering at the
University of Mines and Technology, UMaT, Tarkwa,
Ghana. He is a member of International Association of
Engineers, and Society of Petroleum Engineers. He has
guided the research work of many students in engineering at the PhD, MPhil,
MSc, and BSc levels. His research interests include, power generation and
utilization, control systems design and analysis, and renewable energy and
resources.

Dr. Benjamin Odoi is a Lecturer, contemporary statistician

and a big data analyst in the Department of Mathematical
Sciences at the University of Mines and Technology,
UMaT, Tarkwa, Ghana. He holds the degrees of BSc from
UMaT, an MSc from University of Peradeniya (UoP), Sri-
Lanka, an MPhil, and a PhD from UMaT. He also holds a
Certificate in Advanced Study in Data Science and Big Data Analytics from
Ghana Institute of Management and Public Administration (GIMPA),
Ghana. He is a member of Laboratory for Interdisciplinary Statistical
Analysis (LISA 2020 Global Network), Colorado, United States of America,
the Officer for Monitoring, Evaluation, and Membership for LISA 2020
Global Network Ghana Region, Nigerian Statistical Association (NSA),
Nigeria, and Institute of Applied Statistics (IASSL), Sri Lanka. He provides
technical advice in areas ranging from sample size calculation, data
collection design, data management procedures, data analysis, statistical and
mathematical modeling, and many more. He has taught many statistical and
related courses at both the undergraduate and postgraduate levels. He has
also guided the research work of many students in applied statistics,
mathematical modeling, public health, financial Statistics, logistics, and
engineering at the PhD., MPhil, MSc, and BSc levels. His computer
programming skills include using SAS, R, Minitab, SPSS, Eviews, XL
Miner, and Pasalade. His research interest and consultancy work cover big
data analytics, applied statistics, climatic changes, financial statistics,
biostatistics, and statistical engineering.

Published By:
Retrieval Number: 100.1/ijitee.F987113060524 Blue Eyes Intelligence Engineering
DOI: 10.35940/ijitee.F9871.13060524 and Sciences Publication (BEIESP)
Journal Website: [Link] 17 © Copyright: All rights reserved.
View publication stats