Proceedings of 2024 International Conference on Cognitive Robotics and Intelligent Systems (ICC - ROBINS 2024)

IEEE Xplore Part Number: CFP24UD1-ART; ISBN: 979-8-3503-7274-8

Predictive Vehicle Maintenance using Deep Neural

Networks
2024 International Conference on Cognitive Robotics and Intelligent Systems (ICC - ROBINS) | 979-8-3503-7274-8/24/$31.00 ©2024 IEEE | DOI: 10.1109/ICC-ROBINS60238.2024.10533924

Neil Johnson S.E. Vinodh Ewards Dr. Salaja Silas Dr. G. Jaspher W Kathrine
Computer Science and Computer Science and Computer Science and Computer Science and
Engineering Engineering Engineering Engineering
Karunya Institute of Technology Karunya Institute of Technology Karunya Institute of Technology Karunya Institute of Technology
and Sciences and Sciences and Sciences and Sciences
Coimbatore, India Coimbatore, India Coimbatore, India Coimbatore, India
[email protected] [email protected] [email protected] [email protected]

Abstract— The integration of smart sensors and Internet of data-driven insights, fostering a more sustainable, secure, and
Things (IoT) devices within vehicles is transforming the economically feasible future for the automotive industry and its
automotive industry, generating a wealth of data with immense stakeholders.
potential for predictive maintenance systems. This research study
explores a subset of these sensors such as, engine temperature, oil Therefore, this research study presents machine learning
pressure, battery voltage, fuel level, odometer reading, tier techniques to analyse sensor data and classify a car's condition
pressure, brake fluid level, air filter status, transmission fluid into four distinct categories: Excellent (0), Good (1), Needs
level, and coolant level. This study proposes a novel predictive Service (2), and Very Bad Condition (3). This fine-grained
vehicle health maintenance system that leverages real-time sensor classification enables targeted and timely maintenance
data to predict potential issues and optimize servicing schedules, interventions, preventing unnecessary service while ensuring
minimize breakdowns, and enhance the overall vehicle reliability. optimal vehicle performance and longevity. The proposed
Using a sample of a dataset with readings and states of the system offers several key contributions:
mentioned sensors, a deep neural network was trained and the
results show a prediction accuracy of 97%. • Comprehensive Sensor Utilization: We leverage a
broader range of onboard sensors compared to traditional
Keywords—Predictive vehicle health maintenance, IoT, Machine approaches, capturing more aspects of vehicle health for
Learning, Data Science, Efficient vehicle service. a more holistic assessment.
I. INTRODUCTION • Advanced Machine Learning: We employ sophisticated
Maintaining optimal vehicle health throughout its lifespan is machine learning algorithms, namely ANNs and
crucial for both economic and environmental reasons. Random forest, to effectively analyse the complex sensor
Traditionally, the scheduled maintenance based on mileage or data and accurately classify vehicle condition.
time intervals often leads to unnecessary services or component • Actionable Insights: By providing real-time and
failures without considering the vehicle's true condition. To actionable insights, our system empowers owners and
address this challenge, this research work proposes a novel fleet managers to make informed maintenance decisions,
predictive maintenance system which is specifically designed optimizing costs and extending vehicle lifespan.
for cars, leveraging the rich data stream from onboard sensors to
accurately classify vehicle health and predict service II. LITERATURE SURVEY
requirements. In [1], Bhakat et al. delve into the synergy between Internet
As vehicles grow increasingly sophisticated, more sensor of Things (IoT) and machine learning algorithms, showcasing
data emerges, necessitating a paradigm shift in conservation. their combined potential for real-time accident detection and
Traditional, fixed- interval schedules are inefficient, constantly emergency response systems. Their research underscores the
leading to unnecessary repairs or abrupt breakdowns. This growing significance of leveraging sensor data and AI
research study proposes a predictive vehicle health maintenance algorithms for various vehicle health and safety applications.
system, which uses real-time sensor data to anticipate issues, This aligns directly with the core principles of the proposed
ensuring precise service when demanded. predictive vehicle health maintenance system presented in this
research work. Both approaches share a fundamental reliance on
In addition to safety and environmental considerations, the extracting actionable insights from sensor data for slightly
profitable implications of predictive maintenance are different purposes. While Bhakat et al. [1] focus on immediate
substantial. Unplanned maintenance events can lead to incident detection and response and explored the realm of
expensive repairs, vehicle time-out, and disturbances to proactive maintenance by predicting potential issues based on
transportation services. By proactively associating implicit sensor readings. This shared foundation of exploiting the power
issues through sensor data analysis, service providers can of sensor data and AI algorithms highlights the
streamline their operations, allocate resources more efficiently, interconnectedness of various efforts within the broader goal of
and offer an advanced level of reliability to guests. This design enhancing vehicle safety and performance.
represents a forward- looking initiative to harness the power of

979-8-3503-7274-8/24/$31.00 ©2024 IEEE 322

Authorized licensed use limited to: GITAM University. Downloaded on August 05,2024 at 04:40:41 UTC from IEEE Xplore. Restrictions apply.
Proceedings of 2024 International Conference on Cognitive Robotics and Intelligent Systems (ICC - ROBINS 2024)
IEEE Xplore Part Number: CFP24UD1-ART; ISBN: 979-8-3503-7274-8

In [2], Hassan Ali emphasizes the versatility of Artificial that utilizes machine learning algorithms to analyze real-time
Intelligence (AI) in condition monitoring and fault diagnosis sensor data from engine, battery, and fluid-related systems. This
across diverse industries. He underscores the immense potential focused exploration allows us to develop a system tailored for
of AI in the automotive sector, specifically for predictive predicting potential issues and recommending optimal
maintenance tasks. This aligns directly with the focus of this maintenance schedules, ultimately contributing to enhanced
paper, which explores the development of a novel predictive vehicle reliability and reduced breakdowns.
vehicle health maintenance system. Both share a common
ground in recognizing the benefits of using AI to analyse sensor By acknowledging the valuable review provided by
data and predict potential issues. While Hassan Ali provides a Pourrahmani et al. [4] and highlighting the specific area of focus
broader overview of AI applications in various domains, this within our project, we demonstrate a nuanced understanding of
paper delves deeper into the automotive sphere, proposing a the broader IoT landscape in the automotive industry and how
specific system utilizing machine learning algorithms to analyse our work contributes to advancements in predictive maintenance
real-time sensor data and recommend optimal maintenance solutions.
schedules. This comparative approach highlights the growing Building upon the foundation laid by Rai et al. [5], our work
adoption of AI in predictive maintenance within the automotive further explores the application of machine learning algorithms
industry, showcasing its potential to enhance vehicle reliability in IoT-driven predictive maintenance specifically within the
and minimize breakdowns. automotive domain. Their study effectively demonstrates the
In [3], Theissler et al. provide valuable insights into the value of using sensor data and machine learning for predicting
challenges and opportunities in machine learning-based machine failures and optimizing maintenance schedules. This
predictive maintenance (PdM) for automotive systems. They resonates directly with our own investigation into leveraging
emphasize the crucial role of labeled data, multi-source data real-time sensor data from vehicles to predict potential issues
integration, and model interpretability aspects directly relevant and ensure timely interventions. While Rai et al. [5] utilize a
to the methodological considerations within this project. Their broader scope encompassing various machines and sensor types,
work resonates with our approach in several ways: Shared our research delves deeper into the automotive sector, focusing
Challenges: Both projects acknowledge the need for sufficient on specific sensor data relevant to vehicle health (e.g., engine
labeled data for training effective PdM models. Theissler et al. temperature, battery voltage, oil pressure) and utilizing tailored
discuss the limitations of existing datasets, which aligns with our machine learning techniques for this unique context. This
exploration of strategies to create and leverage labeled data focused approach allows us to develop a system specifically
efficiently. Multi-source Data Integration: Both studies designed for predicting vehicle breakdowns and optimizing
recognize the advantages of incorporating data from multiple maintenance plans, contributing to enhanced vehicle reliability
sources beyond individual sensors. Similar to Theissler et al.'s and reduced downtime.
emphasis on combining various data streams, our project Like our research, Mohammad et al. [6] recognize the
integrates information from engine, battery, fluid-related limitations of traditional scheduled maintenance and propose an
sensors, and other relevant sources to create a comprehensive IoT-based predictive maintenance system for logistic vehicles.
picture of vehicle health. Interpretability Importance: The need Their focus on a specific industry aligns with our understanding
for interpretable models is highlighted by both Theissler et al.'s of the need for tailored solutions. While they utilize a specific
work and our project. While they delve into technical aspects of set of sensors and data visualization methods, their project
interpretability techniques, our project shares this concern and demonstrates the effectiveness of sensor data collection and
considers methods to ensure transparency and understanding of analysis for predicting maintenance needs. This resonates with
the predictions made by the PdM system. our own approach, which leverages a broader range of sensors
By acknowledging the valuable contributions of Theissler et and advanced machine learning algorithms for more
al. [3] and drawing parallels with our own methodology, this comprehensive vehicle health prediction. By acknowledging
comparison strengthens the project's foundation and their work and its focus on a specific application, we can
demonstrates a deep understanding of the current landscape in showcase the broader potential of predictive maintenance
automotive PdM research. systems across various vehicle types and domains.
Similar to our research on predicting vehicle health issues,
Pourrahmani et al. [4] offer a comprehensive review of
Internet of Things (IoT) applications within the automotive Sharma and Kalra [7] explore the application of machine
industry, with particular emphasis on areas like battery and learning for predictive maintenance in commercial vehicles,
engine monitoring. This directly aligns with the core focus of focusing on tire wear. Their work demonstrates the effectiveness
our project, which revolves around leveraging sensor data from of using past and current vehicle and tire condition data to
various sources for proactive vehicle health maintenance. Both predict the remaining useful life of tires. This aligns with our
approaches share a fundamental recognition of the immense own approach of leveraging sensor data and machine learning
potential held by sensor data in optimizing vehicle performance algorithms for early detection and prediction of potential
and minimizing downtime. problems. While their work focuses specifically on tires, it
highlights the broader value of predictive maintenance for
While Pourrahmani et al. [4] provide a broader overview of different vehicle components. Additionally, their consideration
various IoT applications within the automotive realm, our of factors like load conditions, road conditions, and driver
project takes a more targeted approach. We delve deeper into the behaviour resonates with our understanding of the need for
realm of predictive maintenance, proposing a specific system comprehensive data analysis to achieve accurate predictions. By

979-8-3503-7274-8/24/$31.00 ©2024 IEEE 323

Authorized licensed use limited to: GITAM University. Downloaded on August 05,2024 at 04:40:41 UTC from IEEE Xplore. Restrictions apply.
Proceedings of 2024 International Conference on Cognitive Robotics and Intelligent Systems (ICC - ROBINS 2024)
IEEE Xplore Part Number: CFP24UD1-ART; ISBN: 979-8-3503-7274-8

comparing their regression model with other prediction methods

and demonstrating its effectiveness in a real-world fleet
management scenario, they support the feasibility and benefits
of data-driven approaches for commercial vehicle maintenance.
This strengthens the case for similar applications in other areas
of vehicle health prediction, as explored in our own research.
III. PROPOSED METHODOLOGY
Before This study aims to develop a predictive vehicle
maintenance system using sensor data and a Deep Neural
Network (DNN) model. The proposed system aims to classify
the vehicle's condition into four categories: excellent, good,
needs service, and very bad. This information can be used by
drivers to make informed decisions about their vehicle's
maintenance and potentially avoid breakdowns.
Fig. 2 A layer by layer break down of the Deep Neuarl Network with total nodes
in each layer.

Fig 2 depicts the architecture of the deep neural network

(DNN) model used for predicting vehicle health in this study.
This sequential model employs five fully connected dense
layers with a decreasing number of neurons (64, 32, 16, 8, 4) as
they progress towards the output layer. ReLU activation
functions are used between layers to introduce non-linearity,
enabling the model to learn complex relationships within the
data. The final layer consists of four neurons with Softmax
activation, corresponding to the four predicted vehicle health
categories: excellent, good, needs service, and very bad. The
data that is being provided for the training and testing of the
model has 11 attributes excluding the target attribute. These 11
attributes correspond to the sensors reading. The model
identifies patterns in the values and based on the patterns of the
values classifies the vehicle into one of the previously discussed
classes.

IV. RESULTS
The proposed DNN model achieved an impressive overall
accuracy of 97.68%, demonstrating its effectiveness in
classifying vehicle health into the four categories. A detailed
breakdown of the classification performance for each class is
provided in the confusion matrix and classification report.

Fig. 1 A flow diagram for initial development of the predictive model.

Fig 1 illustrates the methodology of the proposed predictive

vehicle health maintenance system, including data acquisition,
preprocessing, model selection, training, and evaluation. After
training, the model can be deployed on a smartphone
application accessible to drivers. This application utilizes
sensor data from the vehicle to make predictions about its health
and displays this information to the driver, enabling them to
make informed decisions regarding maintenance or roadside
assistance.
The model is a sequential model whose structure can be seen
in Fig 2

Fig. 3 Confusion Matrix for the deep neural network showcasing predictive
performance for all classes.

979-8-3503-7274-8/24/$31.00 ©2024 IEEE 324

Authorized licensed use limited to: GITAM University. Downloaded on August 05,2024 at 04:40:41 UTC from IEEE Xplore. Restrictions apply.
Proceedings of 2024 International Conference on Cognitive Robotics and Intelligent Systems (ICC - ROBINS 2024)
IEEE Xplore Part Number: CFP24UD1-ART; ISBN: 979-8-3503-7274-8

The confusion matrix in Fig 3 reveals that the model

accurately classified most instances, particularly in the
"excellent" and "needs service" categories, with precision and
recall values exceeding 96%. Minor confusion occurred
between the "good" and "very bad" classes, as reflected in the
lower precision (93%) for "good" and slightly lower recall
(96%) for "very bad." This suggests the model might benefit
from further refinement to better distinguish these two
conditions.

TABLE I. CLASSIFICATION REPORT FOR DNN MODEL

Class Precision Recall F1-Score Support
0 0.96 0.87 0.91 89
1 0.93 0.96 0.94 349
2 0.99 0.96 0.97 1915
3 0.98 1.00 0.99 2647
Accuracy 0.98 5000
Fig. 5 Training and validation loss curve of the predictive model
Macro Avg 0.96 0.94 0.95 5000 Analysis of the loss and accuracy curves reveals key insights
Weighted
Avg 0.98 0.98 0.98 5000 about the training process. The steadily decreasing loss curve
Classification report for the DNN model with the evaluated metrics.
exemplifies efficient learning, effectively minimizing the error
The classification report as seen in TABLE I confirms the between predicted and actual values. Meanwhile, the stable
strong performance across all classes, with F1-scores accuracy curve signifies strong generalization ability,
consistently above 0.94. Notably, the "needs service" and "very demonstrating that the model performs well on unseen data
bad" classes achieved F1-scores exceeding 0.98, indicating the without overfitting to the training set. This further corroborates
model's ability to reliably identify vehicles requiring the close alignment between training and validation accuracy,
maintenance or experiencing critical issues. potentially indicating the effectiveness of implemented
When it come to the training accuracy and validation regularization techniques in preventing overfitting.
accuracy the curves can be seen in Fig 4, and the training and A secondary implementation of another well-known
validation loss can be observed in Fig 5. classification algorithm the random forest algorithm was also
tested. After fitting the same data to a random forest model, it
shows a suboptimal accuracy of 91.24%. A detailed breakdown
of the model can be seen thorough the confusion matrix below
in Fig.6

Fig. 4 Training and validation accuracy curve of the predictive model

Fig. 6 Confusion Matrix for the random forest showcasing predictive

performance for all classes.

The confusion matrix in Fig 6 shows that the text

classification model performs well on neutral and negative
sentences, with high accuracy rates, but it sometimes

979-8-3503-7274-8/24/$31.00 ©2024 IEEE 325

Authorized licensed use limited to: GITAM University. Downloaded on August 05,2024 at 04:40:41 UTC from IEEE Xplore. Restrictions apply.
Proceedings of 2024 International Conference on Cognitive Robotics and Intelligent Systems (ICC - ROBINS 2024)
IEEE Xplore Part Number: CFP24UD1-ART; ISBN: 979-8-3503-7274-8

misclassifies positive and sarcastic sentences, indicating By providing real-time insights into vehicle health and
potential areas for improvement. empowering car owners with timely information, the proposed
PVMS system offers a promising avenue for improving car
TABLE II. CLASSIFICATION REPORT FOR RANDOM FOREST maintenance practices and fostering a more proactive approach
Class Precision Recall F1-Score Support
to vehicle care. Further research and development efforts
0 0 0 0 89 directed towards user interface design, service integration, and
1 0 0 0 349 data security are crucial steps towards realizing the full potential
2 0.81 1.00 0.90 1915 of this technology and transforming the after-sales service
3 1.00 1.00 1.00 2647 landscape for car owners.
Accuracy 0.91 5000
REFERENCES
Macro Avg 0.45 0.50 0.47 5000
Weighted
Avg 0.85 0.91 0.87 5000 [1] A. Bhakat, N. Chahar and V. Vijayasherly, "Vehicle Accident Detection
& Alert System using IoT and Artificial Intelligence," 2021 Asian
Classification report for the random forest model with the evaluated metrics.
Conference on Innovation in Technology (ASIANCON), PUNE, India,
2021.
The classification report from TABLE II shows decent [2] Yasir Hassan Ali 2018, Artificial Intelligence Application in Machine
results in overall accuracy but when considered with respect to Condition Monitoring and Fault Diagnosis.
the individual classes the accuracy is very suboptimal and does [3] Andreas Theissler, Judith Pérez-Velázquez, Marcel Kettelgerdes, Gordon
Elger, Predictive maintenance enabled by machine learning: Use cases
not show any promise. For this reason, it is not considered for and challenges in the automotive industry, Reliability Engineering &
the task of predictive analysis for the vehicle. System Safety, Volume 215, 2021.
Overall, the results demonstrate the proposed DNN model's [4] Hossein Pourrahmani, Adel Yavarinasab, Rahim Zahedi, Ayat
promising potential for accurate and reliable vehicle health Gharehghani, Mohammad Hadi Mohammadi, Parisa Bastani, Jan Van
herle, The applications of Internet of Things in the automotive industry:
prediction. The high overall accuracy, detailed class-wise A review of the batteries, fuel cells, and engines, Internet of Things,
performance, and positive interpretation of the confusion Volume 19, 2022.
matrix and classification report highlight the model's [5] A. K. Rai et al., "IOT Driven Predictive Maintenance Using Machine
effectiveness. While minor improvements could be made for Learning Algorithms," 2023 6th International Conference on
differentiating between "good" and "very bad" conditions, the Contemporary Computing and Informatics (IC3I), Gautam Buddha
Nagar, India, 2023.
overall performance suggests this approach can be valuable for
[6] S. Mohammad, M. A. A. Masuri, S. Salim and M. R. A. Razak,
informing drivers about their vehicle's health and promoting "Development of IoT Based Logistic Vehicle Maintenance System," 2021
proactive maintenance. IEEE 17th International Colloquium on Signal Processing & Its
Applications (CSPA), Langkawi, Malaysia, 2021.
V. CONCLUSION [7] N. Sharma and M. Kalra, "Predictive Maintenance for Commercial
Vehicles Tyres Using Machine Learning," 2022 13th International
This study investigated the feasibility and potential benefits Conference on Computing Communication and Networking
of incorporating a predictive vehicle maintenance system Technologies (ICCCNT), Kharagpur, India, 2022.
(PVMS) into after-sales services for car owners. By leveraging [8] S. Vasavi, K. Aswarth, T. Sai Durga Pavan, A. Anu Gokhale, Predictive
sensor data and a deep neural network (DNN) model, the analytics as a service for vehicle health monitoring using edge computing
proposed system aimed to classify vehicle health into four and AK-NN algorithm.
categories: excellent, good, needs service, and very bad. This [9] J. Przepióra, P. Bałazy, P. Książek, P. Knap, P. Pasławski and M.
Pietrzak, "Smart IoT platform for Big Data analysis in predictive
information could empower car owners to make informed and maintenance," 2023 24th International Carpathian Control Conference
timely service appointments, potentially extending the lifespan (ICCC), Miskolc-Szilvásvárad, Hungary, 2023.
of their vehicles and minimizing potential breakdowns. [10] J. Wang et al., "Research on Construction of the Smart Internet of
The evaluation demonstrated promising performance, with Vehicles," 2021 IEEE International Conference on Power, Intelligent
an overall accuracy of [insert accuracy value] achieved by the Computing and Systems (ICPICS), Shenyang, China, 2021.
DNN model. Further analysis of class-wise metrics revealed [11] K. S. Hoong Ong, D. Niyato and C. Yuen, "Predictive Maintenance for
Edge-Based Sensor Networks: A Deep Reinforcement Learning
strong performance in identifying vehicles requiring Approach," 2020 IEEE 6th World Forum on Internet of Things (WF-IoT),
maintenance or experiencing critical issues. While minor New Orleans, LA, USA, 2020.
improvements could be made for differentiating between [12] Yaguo Lei, Bin Yang, Xinwei Jiang, Feng Jia, Naipeng Li, Asoke K.
certain health states, the overall results suggest the proposed Nandi, Applications of machine learning to machine fault diagnosis: A
review and roadmap, Mechanical Systems and Signal Processing,
PVMS holds significant potential for real-world applications. Volume 138, 2020.
Looking towards the future, integrating the PVMS with a [13] V. R. Kanakala, K. Jagan Mohan and V. K. Reddy, "A Survey on Efficient
user-friendly smartphone application can further enhance its Methodologies & optimization techniques for vehicle maintenance using
practicality and user adoption. Such an application could IoT and Transfer Learning," 2022 International Conference on Smart
leverage push notifications to alert car owners about their Technologies and Systems for Next Generation Computing (ICSTSN),
Villupuram, India, 2022.
vehicle's health and guide them to nearby service centers when
[14] W. Zhang, D. Yang and H. Wang, "Data-Driven Methods for Predictive
necessary. Additionally, incorporating features for handling Maintenance of Industrial Equipment: A Survey," in IEEE Systems
minor issues and facilitating roadside assistance could create a Journal, vol. 13, no. 3, pp. 2213-2227, Sept. 2019.
comprehensive solution catering to various vehicle
maintenance needs.

979-8-3503-7274-8/24/$31.00 ©2024 IEEE 326

Authorized licensed use limited to: GITAM University. Downloaded on August 05,2024 at 04:40:41 UTC from IEEE Xplore. Restrictions apply.