Received 13 April 2024, accepted 20 April 2024, date of publication 2 May 2024, date of current version 10 May 2024.

Digital Object Identifier 10.1109/ACCESS.2024.3395927

An Ensemble Deep Learning Model for Vehicular

Engine Health Prediction
ISINKA JOSEPH CHUKWUDI1 , NAFEES ZAMAN 2 , MD ABDUR RAHIM 3,

MD ARAFATUR RAHMAN 1 , (Senior Member, IEEE),

MOHAMMED J. F. ALENAZI 4 , (Senior Member, IEEE),
AND PRASHANT PILLAI 1 , (Senior Member, IEEE)
1 Schoolof Mathematics and Computer Science, University of Wolverhampton, WV1 1LY Wolverhampton, U.K.
2 Facultyof Electrical Engineering and Computing, University of Zagreb, 10000 Zagreb, Croatia
3 Facultyof Mechanical and Automotive Engineering Technology, Universiti Malaysia Pahang Al-Sultan Abdullah, Pekan, Pahang 26600, Malaysia
4 Department of Computer Engineering, College of Computer and Information Sciences, King Saud University, Riyadh 11451, Saudi Arabia

Corresponding author: Md Arafatur Rahman ([email protected])

This work was partially supported by the University of Wolverhampton, UK, UK-Saudi Challenge Fund 2023-24 and Researcher
Supporting Project number (RSPD2024R582), King Saud University, Riyadh, Saudi Arabia.

ABSTRACT Predictive maintenance has gained importance across various industries, including the
automotive sector. It is very challenging to detect vehicle failures in advance due to the intricate composition
of various components and sensors. The vehicle’s reliability is of utmost importance for ensuring the absence
of fatalities or malfunctions to foster economic development. This study introduces an innovative method
for developing a predictive framework for vehicle engines with faster and higher decision accuracy. The
framework is specifically designed to recognize patterns and abnormalities that may suggest prospective
engine problems in real-time and allow proactive maintenance. We assessed the performance of the
developed vehicular engine health monitoring systems using a deep learning model based on essential
measures like root mean square error, root mean square deviation, mean absolute error, accuracy, confusion
matrix, and area under the curve. In this case, the deep learning models are developed by following ensemble
techniques using the most prominently used machine learning techniques. Significantly, Stacked Model 1
outperformed other stacked models (Models 2 and 3) and achieved an impressive AUC value of 0.9702 with
a low root mean square error (RMSE) of 0.3355, a high accuracy rate of 0.9470, and a precision of 0.9486.
It happens due to the effective incorporation of different approaches into Stacked Model 1, which signifies
a significant advancement in predicting vehicular engine failures. The model can be used in real-time
monitoring systems to continuously monitor the health of vehicular engines and provide early warnings
of potential failures, thereby reducing maintenance costs and improving safety.

INDEX TERMS Vehicular engine health monitoring system, machine learning, deep learning, ensemble
stacking, vulnerability assessment, decision strategy, micro services.

I. INTRODUCTION faults are essential to prevent vehicle breakdowns and

The use of artificial intelligence (AI) and other data-driven reduce maintenance costs. Traditional methods of monitoring
methods to realize Industry 4.0 is increasingly being used vehicle health involve scheduled inspections or reactive
in the automotive industry around vehicle fault diagnosis maintenance after a failure has occurred, which can be
systems [1]. The reliability and performance of vehicle expensive and time-consuming. The emergence of AI and
engines are critical factors for a safe and effective trans- the Internet of Things (IoT) has paved the way for the
portation system. Early detection and diagnosis of engine real-time collection and analysis of substantial sensor data
from vehicles. This is called an AI-enabled vehicle health
The associate editor coordinating the review of this manuscript and monitoring system (VHMS) [2]. This capability presents
approving it for publication was Khursheed Aurangzeb. prospects for predictive maintenance and fault diagnosis,
2024 The Authors. This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
VOLUME 12, 2024 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ 63433
 I. J. Chukwudi et al.: Ensemble Deep Learning Model for Vehicular Engine Health Prediction

offering a proactive approach to identify potential problems • Analyzed and validated the performance of the proposed
and ensure the efficient functioning of vehicles [3]. However, models based on root mean square error, root mean
analyzing such large datasets and detecting complex patterns square deviation, mean absolute error, and area under
and relationships in the data require sophisticated machine the curve with computation time.
learning, deep learning techniques, and an end-to-end data Furthermore, achieving these goals can be notably
architecture framework. One of the main challenges in enhanced through the utilization of a Stacked Ensemble
vehicular health prediction is the need for high-quality and Model. This model proves effective by leveraging the
diverse training data. Acquiring and categorizing extensive strengths of various machine learning models to generate
data that precisely represents the diverse conditions of a predictions that are both more accurate and dependable.
vehicle might pose a challenge due to its inherent difficulty The ensemble model integrates predictions from multiple
and time-intensive nature [2]. In this case, a novel VHMS was individual models, thereby mitigating errors and biases and
presented in [4]. Also, deep learning has demonstrated signif- ultimately enhancing the overall precision of predictions.
icant promise in scrutinizing time-series data and forecasting The subsequent sections of this study are organized as
future values. It is important to note that no singular deep follows: Section II delves into past related works; Section III
learning model can encompass all the intricacies inherent elaborates on the methodology; Section IV presents the
in engine sensor data. Therefore, there is a need to explore results and findings; and Section V draws out the conclusion
ensemble deep-learning models that can combine multiple and future research direction.
models to improve prediction accuracy and robustness [5].
However, for the complete vehicular system, developing II. RELATED WORK
a complete VHMS using ensemble deep-learning models Vehicles have been gaining tremendous popularity due to
is challenging due to its complexities and massive data their excellent transport capacity, fast, efficient, flexible,
management, which should be moved forward by considering pleasant journey, minimal physical effort, and substantial
several subsections like the engine system, transmission economic effect [4]. According to Industry 4.0, it is essential
system, or vehicular chassis. Then, the fusion of individual to develop a system that monitors and informs the structural
outcomes from different subsystems could be used to make condition of a vehicle intelligently so that maintenance
better and more robust VHMS decisions [4]. As part of expenses can be minimized and longevity can be increased
VHMS, Rahim et al. proposed an AI-based vehicular engine significantly [6]. As a part of these, [7] presented a
health monitoring system (VEHMS) using deep learning data-driven predictive maintenance approach for a vehicle
(MLP) techniques [2]. powertrain using machine learning algorithms. The approach
However, the decision accuracy of this proposed method involves collecting sensor data and using machine learning
for predicting vehicular engine health was within 81%. algorithms to process and analyze the data to predict
Additionally, the performance of the proposed models when maintenance is needed. Similarly, [8] implemented a
was analyzed based on decision accuracy only instead of real-time data processing framework for VHMS that involves
an extensive discussion by considering root mean square collecting sensor data, pre-processing the data, and using
error, root mean square deviation, mean absolute error, machine learning algorithms to analyze and interpret the data.
confusion matrix, and area under the curve with computation The framework is designed to be scalable and can handle
time. These are significant factors in implementing the large volumes of data in real-time. Also, [9] suggested a big
VEHMS in real-time to meet Industry 4.0 requirements. data analytics approach for VHMS that involves collecting
Thus, to overcome the addressed problems, we developed a and processing data from various sources, such as sensors and
stacked ensemble combining Random Forest, support vector maintenance records. The approach involves using machine
machine,Gradient Boosting, Decision Tree and K-Nearest learning algorithms as AI to analyze and interpret the data
Neighbors as the final estimator for vehicular engine health to predict when maintenance is needed. Additionally, [10]
prediction. We found that this model can effectively analyze presented a data-driven framework for prognostics and health
large amounts of engine sensor data and provide accurate management of power electronics in electric vehicles. The
and reliable predictions of engine health, enabling proactive framework involves collecting sensor data, pre-processing the
maintenance and avoiding costly breakdowns in real-time. data, and using machine learning algorithms to analyze and
Therefore, the contributions outlined in this paper include the interpret the data to predict when maintenance is needed.
following: In the afore-discussed VHMS, various deep learning
• Developed a real-time model that can accurately predict techniques were employed in addition to machine learning
potential failures or issues with a vehicle’s engine techniques. Reference [11] work presents a predictive
and classify them into Good, Minimal, Moderate, and maintenance framework for VHMS that uses deep learning
Critical. algorithms to predict potential vehicle failures based on
• Introduced a data pre-processing technique and scaled it sensor data. The framework was designed to improve
by using sklearn, label encoder, and standard scaler to failure prediction accuracy and reduce maintenance costs.
get better outcomes. More so, [11] proposed a predictive maintenance approach

63434 VOLUME 12, 2024

I. J. Chukwudi et al.: Ensemble Deep Learning Model for Vehicular Engine Health Prediction

for electric vehicles that involves using machine learning III. METHODOLOGY
algorithms to predict potential failures based on sensor data. A. VEHMS DECISION STRATEGY
The approach was designed to improve vehicle reliability The individual components/points Severity Value of a
and reduce maintenance costs. Reference [12] suggested a vehicular engine from [1] is
predictive maintenance approach for heavy-duty vehicles 
m

X 
that involves using machine learning algorithms to predict SVi (t) = Si (t)Wi {1 − λ}k Ij × IRIj (1)
potential failures based on sensor data. The approach is  
j=1
designed to reduce maintenance costs and improve vehicle
uptime. However, [11] implemented a predictive maintenance where Si is data from sensors of individual components or
approach for autonomous vehicles that uses machine learning point, λ is the individual components or points degradation
algorithms to predict potential failures based on sensor data. performance for the kth 10,000 Km vehicle movement. Wi is
The approach is designed to improve vehicle safety and the weightage of these components or points, obtained as
reduce maintenance costs. RSi
Wi = P (2)
Due to the complexities of vehicular data and com- RSi
putational time constraints, several studies have recently using relative significance RSi based onPValue Focused
explored the effectiveness of stacking ensemble models Thinking (VFT) from 0 to 10 [2], and RSi = 1. Ij
across diverse domains, encompassing finance, healthcare, is the individual value of various intensifying factors that
automotive, computer vision, and more. Nevertheless, the may affect individual component points, obtained using the
performance of these models can hinge on several factors, Vulnerability Assessment Questionnaire Technique based on
including the selection of base models, the quality of training an intuitive questionnaire and CSVS Qualitative Severity
data, and the complexity of the specific problem at hand. Evaluation Scale [3], [4], with CSVS Qualitative Severity
For example, [13] compared the performance of different Evaluation Scale [5] for rating and indexing as impact
ensemble models, including stacking, bagging, and boosting, severity on engine performances. Here, the rating and
for predicting stock prices. The results showed that the indexing scale is 0.0 for No impact, 0 < Low impact ≥ 3.99
stacking ensemble model outperformed the other methods for, 3.99 < Medium impact ≥ 6.99, 6.99 < High impact
in terms of both accuracy and robustness. In another study, ≥ 8.99, and 8.99 < Critical impact ≥ 10. Finally, the
[14] proposed a stacked autoencoder ensemble model for intensity of relative importance (IRIj ) of intensifying factors
predicting the risk of heart disease. The result showed that of individual components is given by
the model achieved higher accuracy and AUC. RIj
In the field of computer vision, [15]proposed a novel IRIj = P . (3)
RIj
approach for object detection using a deep learning-based
stacking ensemble model. The findings indicated that the Pj is calculated based on VFT [2] from 0 to 10, and
RI
model proposed attained higher accuracy in comparison Wi = 1. is the ratio between the relative importance of
to other cutting-edge models. In a review article by [16], an intensifying factor to the total relative importance of all
the authors highlighted the advantages of using stacking intensifying factors
P of individual VC/VP.
ensemble models in healthcare applications. They argued that In addition,  mj=1 Ij × X ×  IRIj is vulnerability affecting
Pm
this approach can improve the accuracy and reliability of value, and Wi× j=1 Ij × IRIj is the VC/VP’s vulnerability
disease diagnosis, drug discovery, and personalized medicine. value. However, equation (1) helps to define the health of an
Lastly, to predict vehicular engine health in real-time as a individual VC/VP by following the stated condition with the
part of VHMS, [2] proposed using the conventional stacked threshold as follows for the t time interval.
ensemble technique, but the decision accuracy bounded 
Critical, if SVi (t) ≥ THC
within 80.3 % compared with our approach which achieved


if THC < SVi (t) ≥ THM

higher accuracy and AUC.
 Moderate,
(4)
Also, recently employed conventional stacked ensemble 

 Minor, if THM < SVi (t) ≥ THMN
techniques are presented in Table 2, and Table 1 shows a if SVi (t) < THMN

Good,
comparative analysis of the existing approaches with our The Threshold (TH ) as THC for Critical problem, THM
proposed scheme. for Moderate problem, THMN for Minor problem and
However, to meet the requirements of Industry 4.0, it is THG for Good condition. Since the above-stated boundary
necessary to enhance the performance of the vehicular engine limits define the individual VC/VP health conditions in the
health monitoring systems (VEHMS) Decision strategy by engine subsystem and for t time, the following matrix from
raising decision accuracy and computational time with equation (1) provides the decision about whole engines by
other performance criteria. In this case, a stacked ensemble aggregating column-wise of stated elements of Table 3, i.e.
VEHMS combining Random Forest, Support vector machine, X X X
Gradient Boosting, Decision Tree, and K-Nearest Neighbors VEHMSD (t) = SViC + SViM
addressed the limitations of the previous study and showed X X
+ SViMN + SViG (5)
its novelty applicability in the real world.

VOLUME 12, 2024 63435

 I. J. Chukwudi et al.: Ensemble Deep Learning Model for Vehicular Engine Health Prediction

TABLE 1. Comparative analysis of the existing approaches with the proposed scheme.

TABLE 2. Recently employed various conventional stacked ensemble techniques.

TABLE 3. Complete engine decision matrix. individual number of VC/VP, and j denotes the number of
intensifying factors for i-th VC/VP.

B. VEHMS DATA PROCESSING AND ANALYTIC MODEL

ARCHITECTURE
The architecture of the VEHMS data life-cycle for predicting
real-time vehicle engine health using a newly developed
stacked ensemble combining KNN, SVM, RF, Ada, and XGB
where, engine condition is is presented in Fig.1. However, the proposed architecture is
 X presented as follows:

 Critical, if VEHMSD (t) ≥ THDC
 X
if THDC <

 Moderate, VEHMSD (t) ≥ THDM C. VHMS DATA FLOW
X


 Minor, if THDM < VEHMSD (t) ≥ THDMN Overheating, piston, misfire, starter, and lubricant infor-
mation and baseline conditions need to be recorded from
X
VEHMSD (t) < THDMN

Good, if

vehicle engine sensors and communicated to IoT Hub via a
Here, the Threshold (TH ) as THDC for Critical, THDM cellular-enabled device, which was discussed in detail in [2].
for Moderate, THDMN for Minor and THDG for Good However, in this case, we introduced Stream Analytics which
condition of engine. However, this Decision strategy can could receive the message from IoT Hub in real-time, process
decide whether the engine health as good condition or has a it according to the business logic, and transmit the data to
minor, moderate, or critical problem by utilising the stated the serving layer for storage. Depending on the data, several
thresholds. Another notable matter is that i denotes the databases are used. The messages are stored in Azure Cosmos

63436 VOLUME 12, 2024

I. J. Chukwudi et al.: Ensemble Deep Learning Model for Vehicular Engine Health Prediction

FIGURE 1. Vehicle engine prediction architecture.

TABLE 4. Dataset head before LabelEncoder.

TABLE 5. Dataset tail before LabelEncoder.

DB, whereas SQL DB contains relational data and serves as impact the model’s effectiveness. In this case, Sample Size
a data source for the display and action layers. of 3003 of 7 instances, the dataset size may be relatively
small for training ensemble deep learning models effectively.
Adequate sample size is essential to prevent overfitting and
D. DATA EXTRACTION ensure model generalization to unseen data. Techniques like
Vehicle engine performance analysis and predictive main- data augmentation or transfer learning may help address this
tenance necessitate access to accurate and relevant data limitation.
from engine systems [23]. Quickly extracting data from car
engine systems is critical for generating useful insights for
performance analysis and predictive maintenance [11]. There E. DATA TRANSFORMATION
are various methods and techniques for data extraction from The process of data transformation plays an important role
vehicle engine systems, such as onboard diagnostics (OBD) in both data analysis and machine learning processes. This
ports, Controller Area Network (CAN) bus, and sensor- is because the quality and suitability of the data for analysis
based methods; however, this paper uses a sensor-based significantly influence the accuracy and reliability of the
method considering data quality, data availability, and data results obtained. [25]. Proper data transformation techniques
security [24]. can improve data quality, reduce noise, deal with missing
In implementing an Ensemble Deep Learning Model for data, and allow data to be successfully processed for decision-
Vehicular Engine Health Prediction, the quality and quantity making [26]. In this case, we used StandardScaler as a data
of the extracted data play a pivotal role. The DataFrame transformation and data pre-processing technique to scale
dimensions (3003, 7) indicate that the dataset comprises and normalized the dataset features to a standard scale to
3003 rows with 7 columns (features). However, several ensure that each feature has a mean of 0 and a standard
potential limitations in the data collection process could deviation of 1. It is a popular pre-processing procedure

VOLUME 12, 2024 63437

 I. J. Chukwudi et al.: Ensemble Deep Learning Model for Vehicular Engine Health Prediction

TABLE 6. Dataset head after converting to dummy variables.

TABLE 7. Dataset tail after converting to dummy variables.

FIGURE 2. VHMS stakeholder management framework.

used before applying ML/DL algorithms to numerical data categorical variable is represented as a set of binary columns,
analysis. Also, we Encoded the Targeted Variable (Decision) where each column indicates the presence or absence of a
with the LabelEncoder. This is appropriate for classification particular category.
tasks where the target variable is categorical as shown in
Table 4 and Table 5. It is a pre-processing technique that F. FEATURE ENGINEERING
assigns a unique numerical label to each distinct category in This stage holds substantial significance within the machine
a categorical feature, making it easier for ML/DL algorithms learning pipeline, as the excellence and pertinence of features
to interpret such data [27]. furthermore after encoding can markedly influence the effectiveness and interpretability
the target variable, the categorical variables in the feature of ML/DL models. Properly built features can lead to more
matrix are being converted into dummy variables using accurate predictions, enhanced model interpretability, and
pd.get_dummies(), see Table 6 and 7 This ensures that each faster model training times. However, this study considered

63438 VOLUME 12, 2024

I. J. Chukwudi et al.: Ensemble Deep Learning Model for Vehicular Engine Health Prediction

FIGURE 3. Plotting training set accuracies.

FIGURE 4. Decision strategy (A).

FIGURE 5. Decision strategy (B).

only 7 features in the DataFrame see Fig. 10 shows the

feature prediction’s scatter matrix for each input variable models (also known as weak models or base learners) to get
for the feature prediction [28]. Given the constraint of a a more accurate and robust prediction [29]. The outputs of
small number of features, it’s crucial to carefully analyze the the base models are utilized as input features for a higher-
existing features and potentially engineer new ones that might level model, known as the meta-model, which subsequently
provide additional predictive power. produces the ultimate prediction [30].
Assume we have N base models represented by h1, h2 . . . ,
G. DATA MODELLING AND PREDICTION hN and a meta-model represented by H . X represents the
In stack ensemble techniques, data modeling and prediction input data used to train the basic models, and y represents
often entail integrating the predictions of numerous base the associated output (i.e., target variable). The base models’

VOLUME 12, 2024 63439

 I. J. Chukwudi et al.: Ensemble Deep Learning Model for Vehicular Engine Health Prediction

FIGURE 6. Plotting test set accuracies.

predictions for a given input X denoted by y1, y2 . . . , yN Root square error, Accuracy, Confusion Matrix, and AUC
respectively. These predictions are then combined to generate respectively, see Table11.
a new feature matrix, designated by Xmeta, which is fed into
the meta-model H . The final prediction is made by the meta-
model H , which is denoted by Ymeta. J. DEFINITION OF EVALUATION METRICS USED
The equations are presented below: Root Mean Square Error(RMSE): is a metric that quantifies
 the average magnitude of the discrepancies between projected
 y1 = h1 (X ), y2 = h2 (X ), . . . , yN = hN (X )
 and actual values. The algorithm computes the square root of
Xmeta = [y1 , y2 , . . . , yN ] (6) the mean of the squared discrepancies between the expected
Ymeta = H (Xmeta , y) and actual values. Smaller RMSE values suggest superior


model performance, as they correspond to less prediction
errors.
H. STAKEHOLDER Root Mean Square variation(RMSD): is a metric that
A vehicle health monitoring system could benefit from their quantifies the average variation between projected values and
support and input if stakeholders are managed properly. actual values, similar to RMSE (Root Mean Square Error).
Stakeholders can be classified based on their function and The statement describes how the measure calculates the
interest [31]. Also, effective stakeholder management is a extent to which data points deviate from the regression line.
vital success factor in the automotive business. It entails iden- Smaller RMSD values suggest superior model performance,
tifying stakeholders, prioritising their problems, designing as they reflect reduced variability in prediction accuracy.
a communication strategy, and measuring stakeholder sat- Mean Absolute Error (MAE): The Mean Absolute Error is a
isfaction. [32]. Regular stakeholder meetings, collaborative metric that quantifies the average absolute deviation between
decision-making, and effective communication channels can projected and actual data. It offers a more understandable
assist in overcoming stakeholder management difficulties and and precise measure of error in comparison to RMSE.
accelerate industrial innovation, which is presented in Fig 2. Smaller Mean Absolute Error (MAE) values imply superior
model performance, as they correspond to reduced average
I. PERFORMANCE EVALUATION prediction errors.
We evaluated individual ensemble approaches using a range Accuracy: Accuracy is a metric that quantifies the ratio
of performance metrics, including RMSE, RMSD, MAE, of accurately predicted instances to the total number of

63440 VOLUME 12, 2024

I. J. Chukwudi et al.: Ensemble Deep Learning Model for Vehicular Engine Health Prediction

FIGURE 7. Comparative accuracies of different machine learning.

FIGURE 9. Classification report for ML: Confusion matrix.

FIGURE 8. The architecture of the proposed stacked ensemble technique. actual results for several classes. The framework is comprised
of four quadrants, namely true positive, true negative, false
positive, and false negative. It aids in assessing the model’s
instances. It is frequently employed in classification tasks effectiveness in terms of its ability to accurately classify and
and offers a comprehensive evaluation of model performance. detect any instances of misclassification or biases.
Greater accuracy values indicate superior model performance Area Under the Curve (AUC): The AUC is a metric
in accurately classifying engine health statuses. that quantifies the performance of a classification model
confusion matrix: A confusion matrix is a comprehensive across all possible categorization levels. The plot illustrates
representation of a model’s predictions compared to the the relationship between the true positive rate and the

VOLUME 12, 2024 63441

 I. J. Chukwudi et al.: Ensemble Deep Learning Model for Vehicular Engine Health Prediction

FIGURE 10. Scatter-matrix for each input variable.

false positive rate, yielding a single numerical value that [35]. By considering a range of metrics, this study ensure a
quantifies the model’s capacity to differentiate across classes. thorough evaluation of the ensemble deep learning model’s
AUC values that are higher suggest superior discrimination effectiveness in predicting engine health states accurately and
and overall performance of the model.The logic for the reliably.
selection of these particular metrics was based on their
ability to provide a thorough evaluation of the model’s 1) EXPERIMENTAL SETUP
effectiveness in predicting the health of vehicle engines. Figure 5 and 5 described the variables utilised in this
These metrics offer extensive insights into many aspects of study and, further categorised them into Critical-0, Good-1,
model accuracy, precision, and resilience. RMSE, RMSD, Minor-2 and Moderate-3 which are encoded into dummy
and MAE quantify the size of prediction errors, accuracy variable see Table 6 and 7. We started by testing each machine
gauges the overall performance of classification, confusion learning algorithm to understand their predictive capacity etc,
matrix offers specific classification outcomes, and AUC see Fig. 3, Fig. 6, Fig. 7 and Fig. 9. The result indicated
quantifies the discriminatory power of the model, [33], [34], that SVM, K-NN, and Gradient Boosting demonstrate strong

63442 VOLUME 12, 2024

I. J. Chukwudi et al.: Ensemble Deep Learning Model for Vehicular Engine Health Prediction

FIGURE 11. Stacked model 1 learning curve. FIGURE 12. Stacked model 1 confusion matrix.

performance and generalization, while Decision Tree and prediction accuracy based on sample vehicle engine features
Gaussian Naive Bayes exhibited lower performances in to identify good, minimal, moderate, and critical components,
capturing the underlying patterns. To improve the overall see Fig 5.
predictive performance, this further necessitates exploring The mathematical formula for ensemble stacking in
an ensemble stacked method. The integration of each ML VHMS thus can be represented as follows:
to form (stacked models 1, 2, and 3 respectively) an Let’s say we have N base models denoted as M 1, M 2, . . . ,
ensemble method was randomly sampled and based on expert Mn and one meta-model denoted as S. Let’s denote the
experience. Stacked model 1: is a combination of Random training dataset as D which consists of input features denoted
Forest, support vector machine, Gradient Boosting, Decision as X and corresponding target labels denoted as y.
Tree, and K-Nearest Neighbors algorithm. Stacked model 2:
is a combination of Logistic Regression, support vector 2) TRAINING PHASE
machine, Linear Discriminant Analysis, Gradient Boosting, For each base model M 1(1 <= i <= N : Train the base
and AdaBoost algorithm. Stacked model 3: is a combination model M 1 using the training dataset D to generate predictions
of Logistic Regression, K-Nearest Neighbors, support vector and estimate the heteroscedastic uncertainties. Denote the
machine, Linear Discriminant Analysis, Gradient Boosting, predictions as Pi = Mi(X ) and the estimated uncertainties
AdaBoost, Decision Tree, Random Forest, and Gaussian asUi = Ui(X ), where Uiis a vector of uncertainties for each
Naive Bayes algorithm. However, the ensemble model uti- prediction. Combine the predictions and uncertainties from
lizes a weighted combination of predictions from individual all the base models into a new feature matrix denoted as
machine-learning models. By combining the predictions from P = [P1, P2, . . . PN ] = and U = [U 1, U 2, . . . UN ],where
multiple models, the ensemble aims to improve overall pre- each column ofPrepresents the predictions from one base
dictive performance and robustness. The evaluation metrics model, and each column of U represents the estimated
of the ensemble, such as accuracy, precision, recall, and uncertainties from one base model.Train the meta-model S
AUC, are similar to or better than those of individual models, using the combined predictions P, the target labels y, and the
indicating the effectiveness of the ensemble approach. It is estimated uncertainties U to create a meta-model denoted as
worth mentioning that the integration of different machine S = S(P, y, U ).
learning techniques into the ensemble model contributes to
improved predictive performance and reliability compared to 3) PREDICTION PHASE
using a single model alone. For making predictions on a new dataset D, Where each
The Ensemble Stacking technique is promising and base model Mi(1 <= i <= N ): is considered. Generate
utilized in this study to increase prediction accuracy for predictions denoted as Pi = Mi(X ), where X represents the
the vehicular engine health monitoring Decision strategy. input features of the new dataset D. Combine the predictions
It acquires the ability to integrate predictions from two from all the base models into a new feature matrix denoted as
or more fundamental machine-learning algorithms through P = [P1, P2, . . . PN ], where each column of P represents the
a meta-learning approach, in this case, the three primary predictions from one base model. Use the trained meta-model
classes of ensemble learning methods are stacking, bagging, S to generate the final prediction denoted as y = S(P). The
and boosting as Stacked Model 1, Stacked Model 2, and design of the proposed ensemble technique is illustrated in
Stacked Model 3, respectively. These three ensemble stacked Fig. 8.
models were designed in addition to the base classifiers The selection of the meta-model and the hyper-parameters
such as logistic regression, AdaBoost, RF, and XGBoost for of the base models and the meta-model are critical

VOLUME 12, 2024 63443

 I. J. Chukwudi et al.: Ensemble Deep Learning Model for Vehicular Engine Health Prediction

considerations in VHMS ensemble stacking. Furthermore, TABLE 8. Evaluation of stacked model 1.

precisely measuring the heteroscedastic uncertainty for each
base model’s predictions is critical in VHMS ensemble
stacking. However, based on the existing literature [2], in this
study, three layered models are developed for the prediction
of vehicle engine health monitoring systems. A stacking
ensemble architecture, named Stacked Model 1, is con-
structed by combining Random Forest, AdaBoost, and Gradi-
ent Boosting, and Stacked Model 2 is created by merging the
Logistic Regression Classifier, Random Forest Classifier, and
Gradient Boosting Classifier, while Stacked Model 3 is the
combination of Logistic Regression, KNeighborsClassifier,
SVC, Random Forest Classifier, AdaBoost Classifier and
Gradient Boosting Classifier. Moreover, Logistic Regression
is used as the final estimator. The base estimators comprise
Logistic Regression, K-Nearest Neighbors (KNN), Decision
Tree, Support Vector Machine (SVM), and Naive Bayes. The
evaluation includes various techniques such as Regression
Metrics (RMSE, RMSD, MAE, R2 ), Classification Metrics
(Accuracy, Confusion Matrix, Precision, Recall, AUC), and
Cross-validation is conducted using RepeatedStratifiedKFold
with 10 splits and 3 repeats to ensure reliable model
evaluation. Additionally, learning curves are generated to
visualize the model’s performance.

K. EXPERIMENT AND PERFORMANCE ANALYSIS

1) STACKED MODEL 1
The Stacked Model 1 obtained a mean accuracy of 92.93
%, with a low standard deviation of 1.46 %, suggesting a
consistent and stable predicting ability.
The model demonstrates a high level of accuracy,
see Fig. 11, as evidenced by its low regression met-
rics. The model’s performance metrics are as follows:
RMSE = 0.3355, MAE = 0.0728, R2 = 0.9021. The FIGURE 13. Stacked model 2 learning curve.

metrics show that the model effectively reduces pre-

diction errors, leading to accurate decision predictions.
The confusion matrix shows the model’s accurate predic- 2) STACKED MODEL 2
tions for all four classes, with minimal misclassifications As depicted in Fig. 13, the accuracy of Stacked Model 2
in Fig. 12. in recognizing occurrences across different engine health
The performance of VEHMS is important since it requires statuses is 94.70 %, making it important for precise decisions
accurate identification of various engine health states to in VEHMS.
make informed decisions. The model also obtained a The model’s efficacy in reducing prediction errors is
precision score of 94.86 % and a recall exceeding 94.70 % evident from the low values of significant metrics–Mean
in Fig. 11. Absolute Error 0.0795, Root Mean Squared Deviation
These measures prioritize the model’s precision in accu- 0.1325, and Root Mean Squared Error 0.3639, respectively.
rately detecting positive occurrences, which is essential for These metrics illustrate the model’s capacity to reduce
reliable decision-making in vehicular situations. This model’s disparities between feature and target values, hence ensur-
ability to distinguish between different classes is measured ing accurate decision predictions. The model’s robustness
by the Area Under the Curve (AUC) score, which achieves is shown in its high R-squared value of 0.8849 which
a value of 97.02 %. The Model has a higher capacity indicates its ability to explain a significant proportion of
to accurately differentiate between classes in VEHMS, the variation in engine health data patterns. This is crucial
which is important for identifying small variations in engine for facilitating decision-making in the VEHMS system.
health states. However, despite such achievement of Stacked The confusion matrix, in contrast, offers comprehensive
model 1, Table 8 further suggests its limitations, and area of data regarding the classification accuracy of the model, see
improvement. Fig. 14.However, Table 9, underscores the examination of

63444 VOLUME 12, 2024

I. J. Chukwudi et al.: Ensemble Deep Learning Model for Vehicular Engine Health Prediction

TABLE 9. Evaluation of stacked model 2.

FIGURE 15. Stacked model 3 learning curve.

FIGURE 14. Stacked model 2 confusion matrix.

FIGURE 16. Stacked model 3 confusion matrix.

the model highlighting it strength, Weaknesses and area

of improvement. Despite this indications, the model had a
precision of 94.93 % and a recall of 94.70 % in successfully The confusion matrix Fig. 16 provides a comprehensive
predicting class labels. Accurate detection of positive cases evaluation of the model’s classification performance across
in vehicular scenarios requires an unbiased examination of several classes. The recall and precision rates are 94.86 %
the actual positive rates. However, the Area Under the Curve and 94.70 % respectively. The result further demonstrates
(AUC) score of 0.9665 emphasized the model’s capacity to the model’s proficiency in precisely recognizing positive
accurately identify subtle engine health issues in VEHMS, instances—a crucial aspect in intricate traffic scenarios. The
which is important for accurate identification. AUC value of 0.9610 provides suggestive evidence of the
model’s ability to accurately categorize data. A high Area
3) STACKED MODEL 3 Under the Curve (AUC) of 0.9653, indicates that the model
The Stacked Model 3 in this experiment achieved a class label is capable of effectively differentiating labels across different
prediction accuracy of 93.01 % as shown in Fig. 15. classes and identifying even minor vehicle health problems.
The outcome additionally demonstrates a decrease in The findings underscore the model’s proficiency in pre-
predicted mistakes through the utilization of low-key indi- cisely recognizing positive instances—a crucial aspect in
cators such as RMSE (0.3355), RMSD (0.1126), and MAE intricate traffic scenarios.
(0.0728), which indicates a high level of precision. The
accuracy of VEHMS decisions relies on the model’s ability to IV. DISCUSSION
minimize differences between projected and actual outcomes. The comprehensive evaluation of three Stacked Ensemble
The model’s strong R-squared score of 0.9021 demonstrates Models, namely Stacked Model 1, Stacked Model 2 and
its effectiveness in explaining a significant amount of Stacked Model 3 within the context of Vehicle Engine Health
variance and its ability to accurately capture and show data Monitoring Systems (VEHMS) provided valuable insights
patterns. This is particularly essential for making informed into their effectiveness for decision accuracy prediction as
decisions regarding engine health monitoring. showed in Table 11.

VOLUME 12, 2024 63445

 I. J. Chukwudi et al.: Ensemble Deep Learning Model for Vehicular Engine Health Prediction

TABLE 10. Evaluation of stacked model 3. accuracy in VEHMS with a commendable accuracy of
94.70%. Its precision in minimizing prediction errors is
evident through low values of RMSE, RMSD, and MAE.
The R-squared score of 0.9021 reinforces its effectiveness
in capturing data patterns crucial for decision correctness.
High recall 94.70% and precision 94.86% values in the
confusion matrix 16 highlight its reliable recognition of
positive events. The AUC score of 0.9653 further supports
its ability to handle classification tasks, emphasizing its
capacity to capture subtle vehicle health issues. Stacked
Model 3 is validated as a precise prediction tool, enhancing
decision-making precision and dependability in VEHMS.
Finally, the decision to adopt a particular Stacked Ensemble
TABLE 11. Comparison of stacked models result. Model into VEHMS should align with specific application
goals. For precision-centric applications, Stacked Model 1
outperformed the rest of the model in terms of decision
accuracy prediction. Stacked Model 2, with its superior
discriminative ability, is the preferred choice for applications
prioritizing discrimination. Stacked Model 3, with consistent
performance and high accuracy, serves as a reliable option for
general-purpose applications. The integration of the selected
model into VEHMS ensures robust decision accuracy predic-
tion, contributing to more effective and reliable engine health
monitoring and risk management systems.
In summary, each stacked model demonstrates high accu-
racy, precision, and robustness in predicting engine health
states. The comprehensive evaluation through performance
metrics, confusion matrices, and AUC scores enhances
Stacked Model 1 demonstrates a consistent mean accuracy transparency and credibility by offering a comprehensive
of 92.98%, with a low standard deviation of 1.47%, indicating insight into the prediction process of the models and
stable predicting ability. Its low regression metrics RMSE, their dependability in real-world automotive situations.
MAE, and R2 underscore its efficiency in reducing prediction Additionally, acknowledging the models’ limitations and
errors, ensuring highly accurate decision predictions. The areas for improvement further strengthens transparency
confusion matrix Fig. 12 reveals high accuracy in predicting and trustworthiness by promoting critical examination and
all four classes, emphasizing its suitability for VEHMS. continuous refinement of the predictive models.
Remarkable precision of 94.86% and recall of 94.70%
further highlight its capacity for accurate identification of A. COMPARISON OF STACKED MODEL 1 WITH
positive occurrences. The AUC score of 97.02% indicates EXISTING APPROACH
exceptional discriminative ability, crucial for nuanced engine Again, result from the experiment appears that the Stacked
health monitoring. Stacked Model 1 emerges as a precise Model 1 outperformed the current methods found in [2] for
prediction instrument, especially for applications prioritizing predicting the health of vehicle engines in multiple important
precision. In contrast, Stacked Model 2 showcases noteworthy aspects as shown below.
performance metrics, including an accuracy of 94.70%, High Accuracy and Stability: The Stacked Model 1 demon-
demonstrating its ability to identify occurrences across strates exceptional precision and consistency, with a mean
multiple classes. Its low values of MAE, RMSD, and accuracy of 92.93%, surpassing the conventional approach’s
RMSE signify precise decision prediction, essential for accuracy of 80.3%. Furthermore, the fact that it has a low
reliable vehicular health monitoring. The R-squared value standard deviation of 1.46% demonstrates a reliable predic-
of 0.8849 emphasizes its robustness in explaining large tive capability, which is essential for real-time monitoring
variations. The confusion matrix affirms its accuracy in pre- when consistent performance is required.
dicting class labels with a precision of 94.93% and recall of Minimized Forecasting Inaccuracies: The regression met-
94.70%. The exceptional AUC score of 0.9665 highlights its rics, with RMSE = 0.3355, MAE = 0.0728, and R2 = 0.9021,
superior discriminative ability, making it a preferred choice indicate that the model successfully minimizes prediction
for applications prioritizing discrimination. Stacked Model errors. This demonstrates a significant degree of accuracy in
2 proves to be a dependable forecasting tool for enhancing its forecasts, which is crucial for making precise decisions.
decision prediction accuracy in VEHMS. However, Stacked Precise Detection of Engine Health Conditions: The
Model 3 contributes significantly to decision prediction confusion matrix and precision-recall scores indicate that

63446 VOLUME 12, 2024

I. J. Chukwudi et al.: Ensemble Deep Learning Model for Vehicular Engine Health Prediction

the Stacked Model 1 achieves high prediction accuracy for computational resources, including high-performance com-
all engine health states, with minimum misclassifications. puting (HPC) infrastructure and specialized hardware accel-
Precise accuracy is essential for promptly detecting different erators. Limited access to such resources may hinder the
engine health conditions, guaranteeing dependable decision- scalability and accessibility of the predictive framework.
making.The model demonstrates a high discriminative capac- Stacked model 3 in particular highlighted a warning message
ity, as evidenced by its AUC score of 97.02%. This suggests which suggests that the lbfgs solver used in logistic regression
its exceptional ability to reliably distinguish between various failed to converge, reaching the maximum number of itera-
classes of engine health conditions. The ability to distinguish tions, see Table 10. This warning is particularly relevant to
between different engine health states is crucial for promptly computational resources. Increasing the number of iterations,
maintaining and preventing potential failures, particularly as suggested, would demand more computational power and
when it comes to recognizing minor deviations. time during the training phase.
The unique advantages of the ensemble approach are: The
improved performance of Stacked Model 1 is presumably due
to its ensemble character. Through the integration of various 1) ADDRESSING IMPLEMENTATION CHALLENGES
models, it is possible to capture a wide range of patterns Data Quality Assurance: Implementing robust data quality
and interconnections within the data, resulting in improved assurance procedures, including data validation, cleansing,
forecast accuracy and the ability to apply knowledge to and preprocessing, was essential to ensure the reliability
new situations. Ensemble methods are recognized for their and integrity of the input data. Leveraging automated data
capacity to address overfitting and decrease variation, which validation tools, domain knowledge, and expert insights
may account for the model’s stability and consistent perfor- helped identify and address data quality issues effectively.
mance.The Stacked Model 1 provides a substantial enhance- Additionally, conducting sensitivity analyses and sensitivity
ment compared to current methods for predicting the health of tests helped assess the robustness of the predictive models to
vehicle engines. This is mainly because of its high accuracy, variations in data quality and input parameters.
stability, exact identification of engine health conditions, Model Complexity Management: Managing model com-
and strong ability to differentiate between different states of plexity and interpretability was critical to ensure practical
engine health. The advantages mentioned above demonstrate utility and deployment of the predictive framework in
the efficacy of ensemble deep learning methods in improving real-world settings. Employing model regularization tech-
predicted accuracy for intricate systems such as automotive niques, feature selection methods, and model simplification
engines. strategies helped mitigate overfitting and improve model
interpretability. Moreover, conducting model sensitivity anal-
B. IMPLEMENTATION CHALLENGES yses and sensitivity tests helped assess the impact of model
Data Quality Assurance: One of the primary challenges complexity on predictive performance and generalization.
encountered during the implementation of the experiment Optimizing Computational Resources: Optimizing the
was ensuring the quality and reliability of the input data. utilization of computational resources, including parallel
Vehicular engine health data are subject to various sources computing, distributed computing, and cloud computing,
of noise, outliers, and inconsistencies, which can affect the helped mitigate the computational challenges associated
performance of predictive models. Ensuring the accuracy, with training and evaluating the Stacked Ensemble Models
completeness, and consistency of the dataset required exten- 3. Leveraging scalable and efficient algorithms, model
sive data preprocessing and cleaning efforts. Additionally, parallelism, and hardware accelerators (e.g., GPUs) helped
integrating heterogeneous data sources and formats posed accelerate model training and inference tasks, improving
challenges in standardizing and harmonizing the data for scalability and efficiency. Researchers and practitioners
analysis. should consider optimizing the computational resources
Model Complexity: Implementing and evaluating three allocated for model training to ensure convergence and
Stacked Ensemble Models within the context of Vehicle improve overall efficiency. Additionally, scaling the data as
Engine Health Monitoring Systems (VEHMS) involved recommended in the warning message can also affect compu-
dealing with complex and high-dimensional data. The tational requirements, as preprocessing steps may introduce
complexity of the models, including the integration of mul- additional computational overhead. Therefore, the avail-
tiple machine learning algorithms and ensemble techniques, ability and optimization of computational resources play a
added computational overhead and resource requirements. crucial role in effectively implementing and training machine
Balancing model complexity with interpretability was crucial learning models like Stacked Model 3. By addressing these
to ensure practical utility and deployment in real-world implementation challenges and limitations related to data
settings. quality, model complexity, and computational resources,
Computational Resources: The computational resources researchers and practitioners can develop more robust,
required for training, evaluating, and deploying the Stacked scalable, and effective predictive frameworks for vehicular
Ensemble Models posed significant challenges. Deep learn- engine health monitoring, contributing to improved reliability
ing models and ensemble techniques often require substantial and safety in automotive systems.

VOLUME 12, 2024 63447

 I. J. Chukwudi et al.: Ensemble Deep Learning Model for Vehicular Engine Health Prediction

C. REAL WORLD APPLICATION AND SCALABILITY automated testing frameworks, the seamless deployment
The developed model holds promise for various real-world and maintenance of the predictive maintenance solution are
applications and scalability. In automotive manufacturing ensured. Edge computing is another aspect, it involves imple-
plants, the model can be integrated into quality control menting models at the edge, such as onboard car systems,
processes to identify faulty engines before installation. in situations where there is limited network connectivity or a
In fleet management systems, it can enable predictive main- need for low latency. This approach improves responsiveness
tenance scheduling, optimizing vehicle uptime and reducing and minimizes dependence on centralized data processing
maintenance downtime. Moreover, the model can adapt infrastructure.
to different types of vehicles or engine configurations by
retraining on data specific to those vehicles or configurations. 4) INTEGRATION WITH EXISTING MAINTENANCE
Its scalability allows for deployment across diverse vehicle PROCESSES
fleets, ranging from passenger cars to heavy-duty trucks, Organizational Alignment: Integrating predictive mainte-
enhancing safety and reliability across the transportation nance solutions with existing maintenance processes requires
sector. Although this paradigm shows promise, its adoption in organizational buy-in and alignment across departments.
the automobile sector has significant practical obstacles that Collaborating with maintenance teams, providing training on
must be solved to ensure feasibility and efficacy. new technologies, and demonstrating the value proposition of
predictive maintenance enhance acceptance and adoption.
1) DATA ACQUISITION
5) ORGANIZATIONAL BUY-IN
Data Quality, Availability and Security The utilization
of sensor-based methodology guarantees the accuracy and Stakeholder Engagement: Organizational buy-in refers to
accessibility of data. However, it is crucial to ensure broad the level of support and commitment that an organization
coverage of pertinent sensor data from various vehicle demonstrates towards a certain idea, initiative, or decision.
kinds and configurations. Standardization efforts may be Engaging stakeholders is crucial for the success of the
necessary to provide compatibility and integration with predictive maintenance effort. This involves obtaining sup-
different data sources, such as onboard diagnostics (OBD) port and agreement from important stakeholders, such as
ports and Controller Area Network (CAN) bus. Data security management, operations, and IT departments. Conveying
is of utmost importance in safeguarding critical vehicle data the advantages, return on investment (ROI), and how the
from unauthorized access or cyber attacks. To ensure data solution aligns strategically with company goals encourages
privacy and security, it is imperative to have strong data support and dedication. Promoting a culture that values
encryption, access restrictions, and compliance with industry data-driven decision-making and innovation is essential for
requirements such as GDPR, CCPA and to establish explicit fostering the adoption and ongoing enhancement of practices.
standards and guidelines to oversee the collection, storage, Facilitating cultural transformation can be achieved by
and utilization of vehicle data, with a strong focus on ensuring implementing programs that offer incentives, recognition,
transparency and obtaining user consent. and training opportunities to increase data literacy and
encourage experimenting with new technologies.
To tackle these difficulties, a multidisciplinary strategy
2) BIG DATA PROCESSING MECHANISMS
is necessary, which involves the participation of data sci-
Scalability The ability to process huge quantities of sensor entists, engineers, domain specialists, and stakeholders. The
data from various vehicle fleets necessitates the use of suggested predictive maintenance system has the potential to
scalable big data processing mechanisms. Utilizing dis- improve safety, reliability, and efficiency in the automobile
tributed computing frameworks such as Apache Spark or sector by utilizing sensor-based methodologies, ensuring data
harnessing cloud-based technologies can effectively manage quality, availability, and security, and addressing the practical
data processing needs. Real-time Processing: The immediate obstacles of implementation.
examination of continuously flowing sensor data allows
for prompt detection of engine health problems. Utilizing 6) ETHICAL AND SOCIETAL IMPLICATIONS
stream processing frameworks, such as Apache Kafka, Impact on Employment: Predictive maintenance technologies
and implementing real-time analytics models can expedite have the potential to enhance vehicle performance and
decision-making in automobile manufacturing and fleet minimize the occurrence of unforeseen failures. However,
management. they may also cause significant changes to the conventional
employment arrangements within the automotive repair and
3) MODEL DEPLOYMENT maintenance industry. With the increasing capabilities of
Operationalization: The process of deploying predictive automobiles to perform self-diagnosis and preventative main-
maintenance models into production environments neces- tenance, there may be a reduced need for human mechanics
sitates the smooth integration with pre-existing systems and technicians. This change has the potential to result in job
and procedures. By implementing resilient model deploy- displacement for workers in these sectors, emphasizing the
ment pipelines, version control systems using GitHub, and necessity for retraining initiatives and assistance for those

63448 VOLUME 12, 2024

I. J. Chukwudi et al.: Ensemble Deep Learning Model for Vehicular Engine Health Prediction

impacted to successfully move into different positions or conducted in the real world. The study’s findings, however,
industries. could serve as a guide for the automotive industry and a
Environmental Sustainability: Predictive maintenance can standard procedure for enhancing prediction accuracy.
enhance the longevity of vehicles and decrease emissions Future research should focus on increasing the efficiency
by preventing unnecessary repairs and replacements. How- of the novel technique by exploring alternative ensemble
ever, it also prompts concerns regarding the environmental methods. Techniques such as gradient boosting, random
consequences associated with the production and disposal forests, and ensemble stacking offer opportunities to enhance
of advanced automotive components. With the increasing prediction accuracy, scalability, and computational effi-
complexity of cars due to improved sensors and electrical ciency. Comparative studies evaluating the performance of
systems, it is necessary to evaluate the environmental different ensemble methods can provide valuable insights into
impact of manufacturing and disposing of these components. their effectiveness for vehicular engine health monitoring.
Furthermore, it is important to take into account the energy Another avenue for future research is to test the prediction
usage linked to data processing and analysis for predictive accuracy of the predictive framework using real-world
maintenance within the broader framework of environmental vehicle datasets from heterogeneous sources. Incorporating
sustainability initiatives. diverse datasets representing a wide range of vehicle types,
Biases in Data and Models: Predictive maintenance mod- engine configurations, and operating conditions can help
els are trained using historical data, which may contain biases evaluate the robustness and generalization capabilities of the
that were present during the data gathering process or biases model. Additionally, investigating the impact of additional
that are ingrained in the underlying systems due to societal sensor data, such as temperature, pressure, or vibration,
factors. For instance, if the training data predominantly com- on prediction accuracy can further enhance the model’s
prises automobiles from specific demographics or geographic effectiveness in real-world scenarios. Optimizing model
regions, the prediction models may not effectively apply to hyperparameters, including ensemble composition and tun-
different populations. Additionally, if diagnostic algorithms ing parameters, to improve prediction performance and
are not thoroughly verified across varied datasets, biases in generalization should be other area for future research
these algorithms could have a disproportionate impact on spe- Furthermore, exploring the integration of domain knowledge
cific demographic groups. To tackle these biases, it is crucial or expert systems into the predictive framework can enhance
to meticulously focus on the quality, diversity, and fairness of interpretability and reliability. Incorporating domain-specific
data at every stage of developing and implementing predictive insights and constraints can help tailor the model to specific
maintenance technology. Ultimately, predictive maintenance application scenarios, improving its practical utility and
technologies provide substantial advantages in terms of adoption in real-world settings. By addressing these future
enhancing vehicle performance and minimizing maintenance research directions, researchers can advance the field of
expenses. However, they also give rise to crucial ethical and predictive maintenance for vehicular engines, addressing
societal concerns. To ensure ethical and equitable deployment emerging challenges in automotive reliability and safety,
of these technologies, it is crucial to address employment, and ultimately improving the performance and longevity of
privacy, environmental sustainability, and biases in data and automotive systems.
models.
REFERENCES
V. CONCLUSION [1] R. S. Peres, X. Jia, J. Lee, K. Sun, A. W. Colombo, and J. Barata,
‘‘Industrial artificial intelligence in Industry 4.0–systematic review,
In conclusion, a vehicle engine prediction architecture is pre- challenges and outlook,’’ IEEE Access, vol. 8, pp. 220121–220139, 2020.
sented in the paper. This study utilized a computer-generated [2] Md. A. Rahim, Md. A. Rahman, Md. M. Rahman, N. Zaman, N. Moustafa,
dataset to execute an improved machine and deep learning and I. Razzak, ‘‘An intelligent risk management framework for monitoring
vehicular engine health,’’ IEEE Trans. Green Commun. Netw., vol. 6, no. 3,
model that alerts the user promptly and with the utmost pp. 1298–1306, Sep. 2022.
priority of fault in a vehicular engine in real time. The [3] L. Decker, D. Leite, L. Giommi, and D. Bonacorsi, ‘‘Real-time anomaly
effectiveness of machine learning algorithms and ensemble detection in data centers for log-based predictive maintenance using an
approaches was evaluated using measures including RMSE, evolving fuzzy-rule-based approach,’’ in Proc. IEEE Int. Conf. Fuzzy Syst.
(FUZZ-IEEE), Jul. 2020, pp. 1–8.
RMSD, MAE, R2 , Accuracy, Confusion Matrix, and AUC. [4] Md. A. Rahman, Md. A. Rahim, Md. M. Rahman, N. Moustafa, I. Razzak,
Stacked Model 1 outperformed the individual algorithms T. Ahmad, and M. N. Patwary, ‘‘A secure and intelligent framework for
and the other stacked models 2 and 3. The AUC of vehicle health monitoring exploiting big-data analytics,’’ IEEE Trans.
Intell. Transp. Syst., vol. 23, no. 10, pp. 19727–19742, Oct. 2022.
0.9702 was impressive. The ensemble model combined [5] M. R. Sarkar, S. G. Anavatti, T. Dam, M. M. Ferdaus, M. Tahtali,
Random Forest, support vector machine,Gradient Boosting, S. Ramasamy, and M. Pratama, ‘‘GATE: A guided approach for time
Decision Tree and K-Nearest Neighbors for prediction. A low series ensemble forecasting,’’ Expert Syst. Appl., vol. 235, Jan. 2024,
Art. no. 121177.
root mean square error (RMSE) of 0.3355, a high accuracy [6] H. Nordal and I. El-Thalji, ‘‘Modeling a predictive maintenance manage-
rate of 0.9470, and a precision of 0.9486 shows that the ment architecture to meet Industry 4.0 requirements: A case study,’’ Syst.
model performs well in decision prediction. The confusion Eng., vol. 24, no. 1, pp. 34–50, Jan. 2021.
[7] F. Arena, M. Collotta, L. Luca, M. Ruggieri, and F. G. Termine, ‘‘Predictive
matrix proves engine health issue categorization works. The maintenance in the automotive sector: A literature review,’’ Math. Comput.
drawback of this study is that the results may differ if Appl., vol. 27, no. 1, p. 2, Dec. 2021.

VOLUME 12, 2024 63449

 I. J. Chukwudi et al.: Ensemble Deep Learning Model for Vehicular Engine Health Prediction

[8] K. Kanagaraj and S. Geetha, ‘‘Data analytics framework based on cloud [29] N. El-Rashidy, S. El-Sappagh, T. Abuhmed, S. Abdelrazek, and
environment, integration of cloud computing with Internet of Things: H. M. El-Bakry, ‘‘Intensive care unit mortality prediction: An improved
Foundations,’’ Analytics Applications, pp. 251–275, 2021. patient-specific stacking ensemble model,’’ IEEE Access, vol. 8,
[9] K. L. Tsui, Y. Zhao, and D. Wang, ‘‘Big data opportunities: pp. 133541–133564, 2020.
System health monitoring and management,’’ IEEE Access, vol. 7, [30] M. Garouani, A. Ahmad, M. Bouneffa, and M. Hamlich, ‘‘Autoencoder-
pp. 68853–68867, 2019. kNN meta-model based data characterization approach for an automated
[10] W. Chen, L. Zhang, K. Pattipati, A. M. Bazzi, S. Joshi, and E. M. Dede, selection of AI algorithms,’’ J. Big Data, vol. 10, no. 1, pp. 1–40, Feb. 2023.
‘‘Data-driven approach for fault prognosis of SiC MOSFETs,’’ IEEE Trans. [31] J. Linåker, B. Regnell, and D. Damian, ‘‘A method for analyzing stake-
Power Electron., vol. 35, no. 4, pp. 4048–4062, Apr. 2020. holders’ influence on an open source software ecosystem’s requirements
[11] A. Theissler, J. Pérez-Velázquez, M. Kettelgerdes, and G. Elger, ‘‘Predic- engineering process,’’ Requirements Eng., vol. 25, no. 1, pp. 115–130,
tive maintenance enabled by machine learning: Use cases and challenges Mar. 2020.
in the automotive industry,’’ Rel. Eng. Syst. Saf., vol. 215, Nov. 2021, [32] S. Singh, S. C. Misra, and F. T. S. Chan, ‘‘Establishment of critical success
Art. no. 107864. factors for implementation of product lifecycle management systems,’’ Int.
[12] S. Katreddi, S. Kasani, and A. Thiruvengadam, ‘‘A review of applications J. Prod. Res., vol. 58, no. 4, pp. 997–1016, Feb. 2020.
of artificial intelligence in heavy duty trucks,’’ Energies, vol. 15, no. 20, [33] U. Mohseni, C. B. Pande, S. Chandra Pal, and F. Alshehri, ‘‘Prediction
p. 7457, Oct. 2022. of weighted arithmetic water quality index for urban water quality using
[13] M. Jiang, J. Liu, L. Zhang, and C. Liu, ‘‘An improved stacking framework ensemble machine learning model,’’ Chemosphere, vol. 352, Mar. 2024,
for stock index prediction by leveraging tree-based ensemble models and Art. no. 141393.
deep learning algorithms,’’ Phys. A, Stat. Mech. Appl., vol. 541, Mar. 2020, [34] Z. Rahmatinejad, T. Dehghani, B. Hoseini, F. Rahmatinejad, A. Lotfata,
Art. no. 122272. H. Reihani, and S. Eslami, ‘‘A comparative study of explainable ensemble
[14] Q. Wei, Q. Zhang, H. Gao, T. Song, A. Salhi, and B. Yu, ‘‘DEEPStack- learning and logistic regression for predicting in-hospital mortality in the
RBP: Accurate identification of RNA-binding proteins based on autoen- emergency department,’’ Sci. Rep., vol. 14, no. 1, p. 3406, Feb. 2024.
coder feature selection and deep stacking ensemble classifier,’’ Knowl.- [35] V. Villa, G. Bruno, K. Aliev, P. Piantanida, A. Corneli, and D. Antonelli,
Based Syst., vol. 256, Nov. 2022, Art. no. 109875. ‘‘Machine learning framework for the sustainable maintenance of building
[15] D. Cao, H. Xing, M. S. Wong, M.-P. Kwan, H. Xing, and Y. Meng, ‘‘A facilities,’’ Sustainability, vol. 14, no. 2, p. 681, Jan. 2022.
stacking ensemble deep learning model for building extraction from remote
sensing images,’’ Remote Sens., vol. 13, no. 19, p. 3898, Sep. 2021.
[16] C. Patino-Alonso, M. Gómez-Sánchez, L. Gómez-Sánchez, E. Rodríguez-
Sánchez, C. Agudo-Conde, L. García-Ortiz, and M. A. Gómez-Marcos,
‘‘Diagnosing vascular aging based on macro and micronutrients using
ensemble machine learning,’’ Mathematics, vol. 11, no. 7, p. 1645,
Mar. 2023.
[17] F. Li, L. Zhang, B. Chen, D. Gao, Y. Cheng, X. Zhang, Y. Yang, K. Gao,
and Z. Huang, ‘‘An optimal stacking ensemble for remaining useful life
estimation of systems under multi-operating conditions,’’ IEEE Access, ISINKA JOSEPH CHUKWUDI received the
vol. 8, pp. 31854–31868, 2020. Bachelor of Science degree in computer science
[18] Y. Patel, S. Patel, R. Gupta, M. S. Obaidat, R. Kakkar, N. K. Jadav, from the University of Greenwich, U.K., in 2014,
and S. Tanwar, ‘‘Blockchain and AI-based engine fault detection scheme and the master’s degree from the University
for autonomous vehicles,’’ in Proc. Int. Conf. Commun., Comput., of Coventry, U.K. He is currently pursuing
Cybersecurity, Informat. (CCCI), Oct. 2023, pp. 1–8. the Ph.D. degree in computer science with the
[19] S. Mun, J. Hwang, L. Bian, T. Falls, and W. G. Bond, ‘‘Ensemble learning University of Wolverhampton, U.K. His research
for fault condition prediction and health status monitoring in military interests include big data analytics, machine, and
ground vehicles,’’ in Proc. IIE Annual Conf., 2023, pp. 1–6. deep learning, predictive modeling, electric and
[20] R. Khoshkangini, P. Mashhadi, D. Tegnered, J. Lundström, and autonomous vehicles, the Internet of Things (IoT),
T. Rögnvaldsson, ‘‘Predicting vehicle behavior using multi-task ensemble cloud computing, distributed computing, and computer programming. He is
learning,’’ Expert Syst. Appl., vol. 212, Feb. 2023, Art. no. 118716. a member of the IET and has a professional experience in IT consulting.
[21] I. Ullah, K. Liu, T. Yamamoto, M. Zahid, and A. Jamal, ‘‘Electric vehicle
energy consumption prediction using stacked generalization: An ensemble
learning approach,’’ Int. J. Green Energy, vol. 18, no. 9, pp. 896–909,
Jul. 2021.
[22] I. Tessaro, V. C. Mariani, and L. D. S. Coelho, ‘‘Machine learning models
applied to predictive maintenance in automotive engine components,’’ in
Proc. 1st Int. Electron. Conf. Actuator Technol., Mater., Devices Appl.,
Nov. 2020, p. 26.
[23] D. Giordano, F. Giobergia, E. Pastor, A. La Macchia, T. Cerquitelli,
E. Baralis, M. Mellia, and D. Tricarico, ‘‘Data-driven strategies for NAFEES ZAMAN received the B.Sc. degree in
predictive maintenance: Lesson learned from an automotive use case,’’ computer science from International Islamic Uni-
Comput. Ind., vol. 134, Jan. 2022, Art. no. 103554.
versity Malaysia, in 2020. He is currently pursuing
[24] R. Young, S. Fallon, P. Jacob, and D. O’Dwyer, ‘‘Vehicle telematics and
the joint master’s degree in intelligent field robotic
its role as a key enabler in the development of smart cities,’’ IEEE Sensors
systems with the University of Girona, Spain, and
J., vol. 20, no. 19, pp. 11713–11724, Oct. 2020.
the University of Zagreb, Croatia. He is also a
[25] V. Kolisetty and D. Rajput, ‘‘A review on the significance of machine
Visiting Researcher with the KTH Royal Institute
learning for data analysis in big data,’’ Jordanian J. Comput. Inf. Technol.,
vol. 6, no. 1, pp. 1–12, 2019. of Technology, Stockholm, Sweden. Previously,
[26] T. Emmanuel, T. Maupong, D. Mpoeleng, T. Semong, B. Mphago, and he was a Machine Learning Research Intern with
O. Tabona, ‘‘A survey on missing data in machine learning,’’ J. Big Data, IBM CoE Malaysia and VICOROB, Spain. His
vol. 8, no. 1, pp. 1–37, Oct. 2021. research interests include intelligent systems (robotics), computer vision, 3D
[27] R. Thomas, ‘‘A novel ensemble method for detecting outliers in categorical deep learning, and computer graphics. He was a recipient of the Erasmus
data,’’ Int. J. Adv. Trends Comput. Sci. Eng., vol. 9, no. 4, pp. 4947–4953, Mundus Scholarship (2021–2022) and other awards, including the Best
Aug. 2020. Innovation Award in Malaysia Technology Expo Special Edition, in 2020,
[28] J.-W. Chen, ‘‘Analysis of subsampled image size for detection and Gold Medal in National Innovation and Invention Competition Malaysia,
identification of brake pad contours by using deep learning,’’ in Proc. IEEE in 2019, and Second Runner-Up in Nasa Space App Challenge Malaysia,
Eurasia Conf. IoT, Commun. Eng. (ECICE), Oct. 2020, pp. 174–176. in 2019.

63450 VOLUME 12, 2024

I. J. Chukwudi et al.: Ensemble Deep Learning Model for Vehicular Engine Health Prediction

MD ABDUR RAHIM received the B.Sc. degree in MOHAMMED J. F. ALENAZI (Senior Member,
mechanical engineering from KUET, Bangladesh, IEEE) received the B.S., M.S., and Ph.D. degrees
and the M.Sc. degree in automotive engineer- in computer engineering from the University of
ing from UMP, Malaysia. His research inter- Kansas, in 2010, 2012, and 2015, respectively.
ests include AI, traditional, connected, electric, He is currently an Associate Professor in computer
autonomous vehicles, ITS, AI-based SHM, the engineering with King Saud University and a
Internet of Vehicles (IoV), intelligent robotic reviewer of several international journals. His
systems, big-data analytics, machine learning, research interests include cybersecurity, focusing
deep learning, renewable energy, and ICE. on network security, encryption, vulnerability
analysis, and machine learning, where he applies
AI to enhance network security and performance. He also works on the
design and analysis of resilient networks, network routing, and mobile ad-hoc
MD ARAFATUR RAHMAN (Senior Member,
network (MANET) protocols. A member of ACM, his work contributes to the
IEEE) received the Ph.D. degree in electronic
intersection of cybersecurity and machine learning for developing adaptive
and telecommunications engineering from the
and threat-resistant systems.
University of Naples Federico II, Naples, Italy,
in 2013. He has around 15 years of research and
teaching experience in the domain of computer
and communications engineering. Currently, he is
a Reader in cyber security with the School of Engi-
neering, Computing & Mathematical Sciences,
University of Wolverhampton, U.K. He was an
Associate Professor with the Faculty of Computing, Universiti Malaysia
Pahang. He was a Postdoctoral Research Fellow with the University of
Naples Federico II, in 2014, and a Visiting Researcher with the Sapienza
University of Rome, in 2016. He has coauthored around 150 prestigious
IEEE and Elsevier journals [e.g., IEEE TRANSACTIONS ON INDUSTRIAL
INFORMATICS, IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS,
IEEE TRANSACTIONS ON GREEN COMMUNICATIONS AND NETWORKING, IEEE
TRANSACTIONS ON SERVICES COMPUTING, IEEE Communications Magazine,
JNCA (Elsevier), and FGCS (Elsevier)]; and conference publications (e.g.,
IEEE Globecom and IEEE DASC). His research interests include the
Internet of Things (IoT), wireless communication networks, cognitive radio
networks, 5G, vehicular communication, cyber-physical systems, big data, PRASHANT PILLAI (Senior Member, IEEE) is
cloud-fog-edge computing, machine learning, and security. He has served currently an Indian-Born British Computer Scien-
as a Specialty Chief Editor for IoT Theory and Fundamental Research tist. He is also the Pro Vice-Chancellor of Research
(specialty section of Frontiers in the Internet of Things); an Advisory Board and Knowledge Exchange with the University of
Member; an Editor Computers (MDPI); a Lead Guest Editor (IEEE ACCESS Wolverhampton, U.K., where he was the Associate
and Computers); an Associate Editor of IEEE ACCESS and Patron; the General Dean of Research and Knowledge Exchange and a
Chair; an Organizing Committee Member; the Publicity Chair; the Session Professor in cyber security. He is also the Director
Chair; a Programme Committee Member; and a member of Technical of Cyber Quarter -Midlands Centre for Cyber
Programme Committee (TPC) in numerous leading conferences worldwide Security. He specializes in security and privacy
(e.g., IEEE Globecom, IEEE DASC, IEEE iSCI, and IEEE ETCCE) and for communication networks. He has led several
journals. His name was enlisted inside the World’s Top 2% Scientists list national/international research projects in the areas of networking protocols
released by Stanford University under the category of ‘‘Citation Impact and cyber security. He has been invited for talks at various international
in Single Calendar Year, in 2019 and 2021.’’ He was awarded a Higher conferences and has published over 100 papers within peer-reviewed journals
Education Academy (HEA) Fellowship from the U.K. He was endorsed and conferences. He is a fellow of British Computer Society (FBCS) and
by the Royal Academy of Engineering, U.K., as Global Talent under the the Higher Education Academy (FHEA) and a member of the Institution of
category of ‘‘Exceptional Talent,’’ in 2022. He is the Founder of RITECHS, Engineering and Technology (MIET).
a U.K.-based research organization.

VOLUME 12, 2024 63451