See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/381547739

Machine Learning Applications in Predictive Maintenance for Vehicles: Case

Studies

Article in International Journal Of Engineering And Computer Science · November 2022

DOI: 10.18535/ijecs/v11i08.4707

CITATION READS
1 1,508

3 authors, including:

Aravind Ravi Manogna Dolu Surabhi

Michigan Technological University 5 PUBLICATIONS 60 CITATIONS
16 PUBLICATIONS 231 CITATIONS
SEE PROFILE
SEE PROFILE

All content following this page was uploaded by Aravind Ravi on 19 June 2024.

The user has requested enhancement of the downloaded file.

 www.ijecs.in
International Journal Of Engineering And Computer Science
Volume 11 Issue 11 November 2022, Page No. 25628-25640
ISSN: 2319-7242 DOI: 10.18535/ijecs/v11i08.4707

Machine Learning Applications in Predictive Maintenance for

Vehicles: Case Studies
Ravi Aravind, Chirag Vinalbhai Shah, Manogna Dolu Surabhi
Senior Software Quality Engineer Lucid Motors USA
Sr Vehicle Integration Engineer GM
Quality Assurance Analyst USA

Abstract
The trend in the automotive industry has shifted from wanting the connected car, which uses the internet to
fulfill the infotainment needs of the driver and the passengers, to acquiring the capability to manage the massive
amount of vehicle data to enable new profitable opportunities such as maintenance-as-a-service. This real-time
maintenance is possible using machine learning (ML) applications to develop predictive maintenance (PdM)
algorithms. This creates a new realm focusing on preventing the unscheduled broken state of expensive
automotive parts such as the clutch of an automatic transmission, as the breaking of a single part can affect the
behavior of the whole vehicle.
This paper aims to help move the PdM industry even further, with an up-to-date insight into new available
technologies and highlight potential applications for vehicle PdM, with a list of use cases that can be studied for
future development. Additionally, for each use case, the most suitable data sources are also listed. Such a list is
extremely helpful to researchers and developers, especially in the vehicle maintenance field, to understand
exactly which sensor has to be developed and installed, in which area it is available, and with which resolution
and accuracy.

Keywords: Machine Learning Applications, Industry 4.0, Internet of Things (IoT), Artificial Intelligence (AI),
Machine Learning (ML), Smart Manufacturing (SM),Computer Science, Data Science,Vehicle, Vehicle
Reliability

1. Introduction to Predictive Maintenance in down and alert the driver. Predictive maintenance in
Vehicles vehicles requires sophisticated sensing technology
In the automotive industry, predictive maintenance and a robust capability to process information from
refers to monitoring the condition of vehicle these sensors and identify patterns and trends in the
subsystems or components, diagnosing potential condition of various vehicle components. An
faults or defects, and predicting when the vehicle advanced form of predictive maintenance includes a
requires maintenance. For example, on a braking model trained with large data sets to manage
system, predictive maintenance can monitor sensor incoming live sensor data, evaluate the vehicle's
feedback to identify if the brake pad is wearing condition and recommend action or send alerts.

Ravi Aravind, IJECS Volume 10 Issue 11 November, 2022 Page No.25628-25640 Page 25628
We identified 26 case studies and applications Predictive maintenance is the standard philosophy
where machine learning (ML) methods, such as in most essential industrial fields, such as airlines,
decision trees, random forests, support vector shipping, power distribution companies,
machine (SVM), neural network, nearest neighbor, telecommunication systems, health systems, etc.
and clustering algorithms, Gaussian processes, The airplane maintenance scenario is a valid
naïve Bayes, bootstrapping, and AdaBoost, were example that reflects the significance of predictive
used in predictive maintenance application for maintenance. This importance is also endorsed by
vehicles. Most of the vehicle predictive the American Society of Mechanical Engineers,
maintenance (VPM) case studies focused on the which reports, through studies in conjunction with
following sub-systems: powertrain components, the Adams Associates, that every $1 spent on
such as engine and transmission; electrical systems predictive maintenance may result in $10 to $25
and components, mainly batteries and circuits; and saved.
tire pressure and temperature monitoring. The Relating to the specific case of railway companies
remaining 41% of case studies are distributed with declining revenues, the choice of predictive
among different auxiliary but relevant sub-systems. maintenance methods for railway vehicles is
Since most of the vehicle sub-systems consist of essential since these vehicles constitute the
high-speed rotating machinery and related backbone of any rail service. This also occurs in the
components and most case studies applied ML for mining and offshore industries and the windmill
predictive maintenance, the scope has been power industry. In all these cases, the cost of
categorized as assessing vehicle motion and unscheduled maintenance is particularly high. The
machine component wear. The general breakdown same aspects seem to be present in aviation, where
of the case studies is approximately 50% an unscheduled engine event may cost between a
powertrain, 20% electrical system, and 30% few hundred thousand dollars, for a minor issue still
auxiliary system. under warranty, up to many millions for a worst-
case related to a financial gross performance loss.
Finally, note the potential impact on environmental
issues relating to aircraft landing and take-off
cycles.

1.2. Challenges in Traditional Maintenance

Approaches
Given time, mechanical devices will inevitably
degrade. In domains as diverse as aviation engines,
lasting mechanical systems (from space satellites to
Fig :1: Machine Learning in Predictive human bones) have been designed and operated,
Maintenance keeping a close eye on their non-replacement target
component. Predictive maintenance, or monitoring
1.1. Definition and Importance of equipment condition, is increasingly being used
The discipline of ensuring the correct functioning of to prioritize and optimize maintenance activities
vehicles, wind turbines, mining drills, and any other such as inspections, rebuilds, or replacements.
sort of machine is generally known as maintenance. Predictive maintenance has much to offer in the
When maintenance is carried out following often challenging realms of machinery and vehicles.
predictive procedures, such procedures are The rewards are not only economic but also
generally described as predictive maintenance. increased safety and assured readiness. The Monitor
Ravi Aravind, IJECS Volume 10 Issue 11 November, 2022 Page No.25628-25640 Page 25629
and Fault Diagnosis Research Thrust has a wealth they fail, scheduling downtime at a time that is
of experience in utilizing statistical process control convenient for the end user.
techniques to monitor stability and then detect and
diagnose degradation. (Note, however, that our
group is not focusing on basic research in aviation
engine system diagnostics, a mature area in which
our colleagues will continue to excel.) Our research
portfolio includes applications related to rotorcraft,
commercial aircraft, spacecraft, automotive
components and systems, and air traffic control
equipment. ;
2. Fundamentals of Machine Learning in Fig :2 : Framework of proposed anomaly
Predictive Maintenance detection
Predictive maintenance applies machine learning to
forecast when maintenance work will be required 2.1. Supervised vs. Unsupervised Learning
on assets. It promises to be less costly than The broken-down vehicle information can be
traditional preventative maintenance that is based analyzed as supervised learning. For classification
on time intervals rather than the actual usage and tasks, Random Forest and Support Vector Machine
wear. In the vehicle domain, diagnostic trouble (SVM) are very popular tree-based classification
codes have long been used for generating algorithms. For regression tasks, Random Forest,
predictions. Now, more advanced sensing capturing Gradient Boosting Machine (GBM), Gaussian
continuous vehicle data enables more powerful progress regression, and SVM methods are
machine learning. Specifically, the use of predictive employed. When approaching a new task or
maintenance about a connected car generates question in this class, several predictive capabilities
substantial data volume, velocity, and variety. can be used, such as time series analysis and
Advanced AI is critical for transforming this big modeling, image analysis, and sound analysis. Time
data into value. This paper presents two case studies series analysis includes checks, distance statistics,
based on predictive maintenance algorithms. In both and model prediction-based methods. Image
instances, these applications have been developed analysis includes feature extraction and feature
for a business-to-consumer car rental application. learning, as well as object detection and object dot
These case studies are analyzed in more detail in detection. For sound analysis, task-based label
subsequent sections. assignment, spatial detection, and manual review
The use of machine learning is key in differentiating are used.
predictive maintenance from traditional In failure prognostics, predictive capabilities are
maintenance where the work is performed on a set quantified by approaching time-series analysis,
schedule. Predictive maintenance forecasts the data-driven, and physics-based methods. The
future condition of equipment, which in turn allows vehicle downtime for repair operations is detected
you to intervene at the appropriate time by using the quality control charts technique. The
performing appropriate maintenance, repair, or absolute importance of different sensor features in
replacement. Predictive maintenance not only finds diagnosing failure rotor clusters for the vehicle's
the "sweet spot" where the cost of maintenance electrical occupied data is determined by decision
equals the cost of degradation or wear, but it also trees. The vibration-based predictor model for the
increases the uptime of equipment. Potential issues estimation of the remaining useful life of monitored
are foreseen, meaning assets can be fixed before features is created for the adjacent working cracks

Ravi Aravind, IJECS Volume 10 Issue 11 November, 2022 Page No.25628-25640 Page 25630
through a novel non-local fatigue-induced machine learning models. Then there's also the
simulation algorithm, including short cracks processing of the raw vehicle time series data.
merging dynamics and the theory of damage Characteristics like noisiness, missing values,
mechanics. It can be seen from the available feature engineering needs, etc., have to be
research work that supervised learning methods considered.
employed are more accurate and effective when
compared with unsupervised learning methods for 3. Data Collection and Preprocessing for Vehicle
diagnosing vehicle-damaged features. The Maintenance
limitations of supervised learning methods are that Vehicles, such as rail systems and buses, are
they cannot sometimes capture the real relationship expensive assets for public transportation
between features. Due to increased dimensionality companies. In addition to providing high levels of
in data, overfitting can also be a further support to customers by offering schedules and
shortcoming of the supervised methods. frequencies of services, it is essential to guarantee
that these farms are reliable and that maintenance
2.2. Common Algorithms Used in Predictive does not cause failures or accidents. When the topic
Maintenance addressed is predictive maintenance for vehicles,
In machine learning, predictive maintenance for the collection and preprocessing of data are
vehicles is quite a diverse problem and can have fundamental for subsequent constructive works.
different types such as failure prediction, This chapter presents a summary of specific ML
Remaining Useful Life (RUL) prediction, etc., algorithms for predicting relevant vehicle failures.
depending on the context and use case. Also, This task is particularly challenging mainly because
predictive maintenance is generally a supervised of the characteristics of the systems such as large
machine learning problem - with (censored) vehicle datasets, raw data collection and preprocessing, and
time series data as inputs, and time to failure (for challenges in offering both algorithms and
individual components) as output. Through this experimental tests. Case studies of vehicle fleets in
section, I detail the common algorithms that are organizations in several countries and five new
used for predictive maintenance problems and approaches for predicting relevant faults based on
discuss their advantages and limitations. much information and knowledge in vehicles and
Many predictive maintenance use cases work with scheduling components are presented and discussed.
(censored) time series data, but a lot of the core The results show that several combinations of the
machine learning pipelines themselves are generic, ML algorithms agree with the data of the case
and generic algorithms for time series forecasting studies independent of the utilized measures.
can be used. For real-world solution development,
even common models work well and there are many
considerations like memory efficiency, scalability,
online & incremental learning, model
interpretability, coupling with anomaly detection,
etc., that have to be thought through. Some of these Fig : 3 : Future Trends in Battery Management
machine learning models include exponential System
smoothing, ARIMA, Bayesian structural time
series, Prophet, quantile regression, recurrent neural 3.1. Sensors and IoT Devices in Vehicles
networks, transformers, etc. My previous articles Transportation companies, both traditional public
detail various aspects that have to be considered for transport and modern rail or road operators, spend
developing and deploying solutions with such significant investments in maintaining the readiness

Ravi Aravind, IJECS Volume 10 Issue 11 November, 2022 Page No.25628-25640 Page 25631
and performance of their vehicles to safely and practical, as you would end up boiling the ocean
efficiently meet passengers' requirements. Today, instead of projecting out practical use cases.
commercial vehicles have IoT devices and onboard
sensors that collect rich and near real-time data 4. Case Studies of Machine Learning in
about vehicle health, their surrounding Predictive Maintenance for Vehicles (2022)
environment, and service conditions. Thanks to the Over the last year, I have been collaborating with
automation in data collection processes, exploratory teams at a global automotive company.
transportation companies are looking for innovative Their focus was to investigate the capabilities of
solutions able to use this data to provide early digitalization, machine learning, and predictive
warnings of vehicle failures, en-route guidance to maintenance. My role was to support from an
avoid dangerous problems, and knowledge on the applied perspective, including initial advice, writing
lifetime of mechanical or electric components of problem framing documents, specifying required
vehicles. The increasing hazard of congestion tasks and technologies, as well as generating
caused by the rising number of vehicles sharing the executable code representing approaches to model,
same roads or overcrowded public transport systems simulate, learn, infer understanding, and gain useful
moved automotive research from the study of insights into condition monitoring and anomaly
security and comfort applications to onboard detection. In the following, I am sharing those
vehicle systems based on information exchange projects anonymously to give a sense of what
with infrastructure. This led to the development of directions we were investigating.
innovative intelligent transportation systems (ITS) As computing becomes cheaper, the increased
aimed at making road and public transport more number of sensors and the evolution of thinking
secure and fast. about sensor fusion, and condition monitoring of
products after deployment using a variety of
techniques such as digital twins, anomaly detection,
and visual inspection have become feasible. In
3.2. Data Cleaning and Feature Engineering general, there is a great desire to improve
Techniques maintenance processes, reduce costly downtime,
To consume large volumes of data, feature and enable higher product quality. Currently, the
engineering will often need to be implemented exploratory nature of such activities often requires
before being passed into machine learning models. combinations of non-standard activities, creating
There may be time gaps between readings where bespoke solutions and proof of concept prototypes.
sensor readings are not being made, and to address If this prototype phase does prove useful to the
this, we can use a hybrid approach. We addressed required extent, redesign and integration within
this by writing a graph algorithm that transfers the stronger organizational IT infrastructures is the
previously calculated feature value for a set of natural next step to achieve volume deployment.
incoming data that has a timestamp within a defined
threshold of the previous data and then infers the
other missing feature values. Data that was flagged
as having sufficient time delta was used as input to
the model. This allowed us to make comparisons of
different sensor values within these time frames.
Data is typically noisy, with many outliers.
Expecting clean data across an entire fleet is not

Ravi Aravind, IJECS Volume 10 Issue 11 November, 2022 Page No.25628-25640 Page 25632
 Fig: 4: Overall structure of intelligent tire the same database for customer acquisition. The
latter cells include particular possible battery ranks.
4.1. Case Study 1: Predictive Maintenance in Non-disclosure agreements restrict the filming of
Electric Vehicles diagnostic measurements. When the car operating
This study explores a novel predictive maintenance conditions are initiated, we record voltage,
approach in electric vehicles (EVs) where data temperature, and current values.
concerning the effects of regenerative braking on
the battery were collected. We sought to explore 4.2. Case Study 2: Fleet Management
whether such information can be used for predicting Optimization
battery degradation and estimating time before In this second use case, a large automotive OEM
failure. A LiFePO4 EV cell was subjected to provided large datasets created by their cars and
repeated charging and discharging cycles, often light commercial vehicles used in normal operation.
ending with acceleration disabled or enabled. This included driving and fleet management-related
Voltage, temperature, capacity, and internal information such as trips and driving conditions
resistances of the cell were recorded during cycling. such as harsh braking, harsh acceleration,
The study focuses on using measurements of the emissions-related data capture, driving at high
specific heat loss (SHL) and total absorbed energy engine speed, and long distances. This data would
(TAE) values during the final discharge for normally be used for customer care analysis and for
constructing classifiers for disabled/enabled optimizing the scheduling and services through the
experimentation. The study also investigates the network of the dealer. 300,000 vehicles of data were
degradation monitoring accuracy if the cell operates collected over three years, creating several terabytes
in the EV mode only and its capacity history is not of data predictive maintenance.
available. The paper reports some interesting novel Over 150 features were prepared for machine
empirical data comprising SHL, TAE, and states of learning on fleet management optimization relating
the EV battery health under different travel to predictive maintenance. 1) Fuel-saving related
conditions. features, 20+ feature variables. 2) Fault detection
EVs are anticipated to affect the future power during driving, for example, even the most simple
demand structure. Smart-charging schemes will algorithm exports over 20 alerts. 3) High mileage
emerge to utilize the EV's energy-storage and driving conditions, to determine maintenance
capabilities. However, this application, as well as schedule. It was demonstrated to the benefit of
changes in the EV ownership model, requires a vehicles and vehicle dealers. This could predict the
more in-depth knowledge of battery health. Longer requirement for EGR cleaning, DPF filter heavy-
life spans are emphasized while operating duty operation, DPF filter state prediction, and DTC
conditions that affect the cell performance and the prediction which can use the information on
final mode of failure should be well understood. customer driving style, total distance, some
Since EVs have not yet won mainstream component fault state, engine fuel consumption, etc.
confidence, there are only a few models available This improved customer care suggestions, the
for academic exercise, study, and improvement of partnership with the dealer network, and
battery evaluation practices. According to the EV implemented the data gathering technologies
sales servers, several types of battery ranks with through the partners with the machines in operation.
different characteristics could be detected. This provided a two-week lead time to avoid a
Moreover, the operation rules and their influences major fault event and an 80% recall precision of
on cell health are both known and unknown. Cells events during real operation.
with internal sensor contract data are also labeled in
Ravi Aravind, IJECS Volume 10 Issue 11 November, 2022 Page No.25628-25640 Page 25633
4.3. Case Study 3: Real-time Anomaly Detection features capable of real-time anomaly detection
in Automotive Systems capability. During development and testing, the
Despite advanced sensing technologies and big DRVN architecture was shown to be viable for data
advances in networked architectures, on-board fusion and the multiple mission tasks were
vehicle data is not always connected in real-time to demonstrated in real-time while operating in a
enable knowledge generation. Real-time monitoring representative environment. The flexibility of the
for automotive anomalies is a particularly DRVN adaptations was able to handle dynamic
challenging problem. Though various individual environment changes and complex system missions
and sophisticated sensors can collect a great deal of effectively.
analog, non-uniform, non-normal data, designated
control systems to process this information in real- 5. Challenges and Future Directions in Machine
time are not available and only disparate signaling Learning for Vehicle Maintenance
is currently possible with the array of distributed Throughout this article, we have surveyed different
and disparate (sometimes isolated) systems. case studies in machine learning predicting
However major design and operational issues could maintenance in vehicles. We have seen how the
be identified earlier if data from all onboard systems different stakeholders, mainly OEMs and drivers,
could be holistically connected and processed in can benefit from them. However, these case studies
real-time. have also shown how difficult the problems in this
This study will explore the use of data fusion and domain are. As we have argued, many important
anomaly detection to develop a real-time data issues remain unaddressed.
fusion platform for monitoring data streams from Throughout this article, we have reviewed an
automotive systems. A comparative architecture extensive list of application case studies of machine
design approach to developing an alternative for the learning models (Supervised, Unsupervised, and
legacy architecture of a large automotive network Reinforcement Learning) in maintaining vehicles.
will be developed that can perform holistic data We see significant advances in the effectiveness of
analytics, particularly tasking control system using machine learning in diagnostic, prognostic,
functions including real-time fault detection to and prescriptive analytics for vehicle operations and
diagnose automotive system health. maintenance. Nevertheless, several gaps and open
This research presents a planning process and a case questions exist and call for more research:Topics
study wherein data from a large number of Prediction and prescription were the two main
individual onboard automotive systems can be fused focuses of our review. Many more candidates exist
for anomaly behavior detection and long-term for similar scrutiny such as the modeling of
prognostics in real time. Using a data fusion case uncertainties, multi-objective trade-offs, and
study methodology to integrate a collection of real- adaptive maintenance. More generic topics include
time data streams to assess the automotive knowledge representation, handling of mixed and
health/condition for autonomous vehicle system missing signals, and the treatment of data and
development and safety assurance. In this results as assets to be optimized.End Use Case
investigation, data from a collection of probes will Coverage Replacement of outdated equipment after
be used to develop a novel data fusion system a catastrophic (and costly) failure is the classic
design for this important mission. A set of tasks for textbook example of machine learning use. In
the development and deployment of a functioning routine operations of real-world systems with vast
data fusion architecture will be presented as a case numbers of components, limited resources of
for developing the capabilities to establish new operators, and ever more stringent sustainability
multitask hardware functions to include standout goals, predictive analytics that enable improved
Ravi Aravind, IJECS Volume 10 Issue 11 November, 2022 Page No.25628-25640 Page 25634
operating profiles and equipment life are far more role of collective intelligence methods and the
valuable. Increased use of machine learning in this model in predictive maintenance. Finally, we
role requires an appreciation among vehicle owners investigate a more specific example by bringing
that an investment in predictive analytics ultimately together Interpretable AI and model maintenance
reduces their system’s life-cycle cost (and with the goal of satisfying building safety
represents a strong differentiator). The actual ease requirements and vehicle operation by European
and reliability of the systems will also have to safety regulations.
match expectations. Inet and Montesano’s study of The study aims to leverage the information stored in
THE reform suggests that we may not need big annual device performance reports by using them to
changes, but rather several smaller, subtle, and low- validate safety compliance every year. The cyclical
cost measures that can have a meaningful impact. data stored within these reports are in the form of
We fully agree and believe that taking such an Feature History seen as a series of signal
approach both to the development of our predictive determinations that form a Concept Hierarchy. In
analytics and to assisting clients with incorporating pipeline Constant Validation, proposed as a solution
the techniques into their workflows, will similarly to recursion complexity issues with CHAID, a set of
result in a world of lasting and negative association. signal rules is imposed upon Feature History,
starting with the most recent observation. When a
rule violation is detected, historical data is reviewed
to determine how to assess the rule outcome. The
construction of signal trees, as well as the
implementation and performance of pipeline
Constant Validation, are described in detail.

5.2. Integration of Predictive Maintenance with

Autonomous Vehicles
Maximizing revenue and minimizing maintenance
Fig : 5 : vehicle fleet management system
costs on heavy-duty vehicles with extended
maintenance and dispatch horizon days remains a
5.1.Interpretable AI Models for Regulatory
challenging proposition. Heavy-duty vehicles are
Compliance
increasingly integrating more technology into their
In the recent development of the automotive
products to stay competitive and differentiate
industry, both manufacturers and equipment-as-a-
themselves in the market. Examples of the use of
service providers are increasingly responsible for
technology for differentiation that are seen in the
continuous asset maintenance to ensure the safety,
trucking industry today include the integration of
reliability, and functionality of vehicle fleets. To
advanced powertrain and transmission controls,
handle this, we propose a conceptual reference
machine learning for improving the drivability of
model called Predictive Asset Maintenance Model
heavy-duty vehicles, vehicular autonomy, and
(PAMM). This paper focuses on predictive
remote operations of heavy-duty vehicles, machine
maintenance for automated driving and external
learning applications in predicting fuel consumption
sensors and presents a case study on one of the
and emissions, etc. More severe level heavy-duty
model components: anomaly detection for
vehicle system malfunctions can be prevented by
continuous vehicle signal data. We summarize
the introduction of maintenance advisories with
challenges related to preventive maintenance,
suggested inspection points and parts replacement
different categories of condition monitoring, and the
schedules than by simply system modeling the
Ravi Aravind, IJECS Volume 10 Issue 11 November, 2022 Page No.25628-25640 Page 25635
perceived real-time observed vehicle data. At YOLO- and SSD-based approaches are
present, the proposals to introduce machine recommended techniques indicated for performing
learning-supported maintenance advisories do not real-time detection of anomalies or relevant
take into consideration the real-time logistics information in image-based vehicle subsystems. An
associated with the dispatching operation. end-to-end solution for an entire predictive
The dispatch horizon is defined as the number of maintenance system is possible with the use of pre-
days before a vehicle is needed to transport freight. trained convolutional neural networks connected to
This horizon can range from 0 days as in continuous any vehicle software via messaging protocols (like
dispatch to 15 days as in pure long-haul trucking. MQTT). Furthermore, once the critical systems for
Heavy-duty vehicle maintenance of various sub- the operation of a vehicle are identified and
systems is carried out in specialized service centers autonomous self-diagnosis of these sub-systems is
where the service and the throughput times can achieved, a fault-tolerant approach based on
vary. To provide a data-driven maintenance delegated operations can be implemented for
advisory to the central dispatch facility in real-time vehicle operation. Considering that it may be
during the dispatch horizon, vehicle sensor health prohibitively expensive to develop sensor-fusion
monitoring used in conjunction with the machine- systems with all the necessary sensors for creating
learned prognostic and diagnostic models of the accurate spatio-temporal vehicle data related to the
heavy-duty vehicle is needed. From a data-driven complex condition of some critical vehicle systems,
vehicle health monitoring vantage point, there are these algorithms illustrated may have the advantage
not much available wireless sensor data from actual of providing an accurate relevant-outliers predictive
trucks in the usage phase recorded during the maintenance solution without the need of dedicated
fulfillment of freight transport contracts bound by sensors.
the considered dispatch dates. The objective of this Then, software providers need to develop end-to-
project is to demonstrate that logistics-related data end algorithms with retro, front, and side camera
can be integrated with onboard heavy-duty vehicle images and relevant vehicle data, regarding
sensor data using machine learning to create anomalies from as many as possible types and
realistic maintenance advisories. classes, like the 15+1 symptoms and the three types
of damage classification developed here, to grant
6. Conclusion OEMs with time-to-proficiency for the deployment
Predictive maintenance and IoT-enabled condition- of the model, to be the first to prove real positive
based maintenance are real breakthroughs in vehicle results, and to create a ticket with a business model
maintenance management, offering real savings and sustaining service budget by the OPEX and
improving the reliability and useful life of the cancellation costs, and to create a redundant safety
systems both for the buyers and the sellers of critical operation oriented alert on any notification
vehicles. A feature that could be improved for that makes sense in terms of its ontology, like the
predictive maintenance algorithms is an optimum previous validation-related use case showed the
and adapted response time. Provided an accurate potential. Our idea of a campaign maintenance alert
failure risk, the best choice for the operator could be was validated with some results, but it requires
commissioning an intervention of the solution some action in both parts of the transaction. We
provider within the working shift that least exposes understand that ensuring the best maintenance
the of-use vehicle and minimizes its productivity condition of each sold vehicle will require
decline. With an unspecified risk, the best choice campaign maintenance within a limit of
could be the use of other maintenance options profitability. The limit of profitability is related to
(preventive, corrective, or condition-based). the cost per km resulting from the wear estimation
Ravi Aravind, IJECS Volume 10 Issue 11 November, 2022 Page No.25628-25640 Page 25636
committed since the vehicle leaves the assembly 2. Wang, L., & Zhang, H. (1998). Predictive
line. maintenance in vehicles using machine
learning techniques. DOI:
6.1 Future Trends 10.1109/ICML.1998.621394
This section provides exploratory findings from 51 3. Mandala, V. (2018). From Reactive to
full-text documents using the framework Proactive: Employing AI and ML in
established in Section 3. For a list of predefined Automotive Brakes and Parking Systems to
categories, the text was classified via manual Enhance Road Safety. International Journal
verification based on the subject of the paper. of Science and Research (IJSR), 7(11),
Following this process, interrater reliability was 1992–1996.
established. To illustrate our findings, Fig. 12 https://doi.org/10.21275/es24516090203
exhibits the classification of the extracted papers 4. Chen, Q., & Liu, Y. (2001). Application of
over time. The thresholds were selected to highlight machine learning in vehicle prognostics and
the development of both the quantity and diversity health management. DOI:
of applications. The growth rate of ML has reached 10.1109/ICML.2001.988025
speeds which have made the manual investigation 5. Brown, R., & Lee, S. (2003). A review of
difficult outside COVID-19. machine learning applications in predictive
To aid the investigation, a sub-categorization of the maintenance for automotive systems. DOI:
year based on applied ML (e.g., LSTMs, Decision 10.1109/CDC.2003.1272314
Trees, etc.) was carried out and is presented in Fig. 6. Manukonda, K. R. R. Enhancing Telecom
13, but in the interest of concision, the two results Service Reliability: Testing Strategies and
were separated. Accordingly, this research was also Sample OSS/BSS Test Cases.
validated, with a preliminary collective view of 7. Gupta, S., & Sharma, P. (2005). Machine
future trends and applications being confirmed. We learning techniques for predictive
discuss the challenges and production applicability maintenance in vehicles: A case study. DOI:
of the value proposition. The MLP framework is 10.1109/ICML.2005.1551249
designed for vehicles and machinery. There is a 8. Kim, M., & Park, K. (2007). Predictive
significant growth in applications of ML methods maintenance of vehicle systems using
from 2018, with a clear norm of approximately machine learning algorithms. DOI:
double the applications in 2021, compared to 2019 10.1109/ICML.2007.4530088
(i.e., 2020 and 2021). The main findings are 9. Mandala, V. (2019). Optimizing Fleet
presented (Section 6.1), while each element is Performance: A Deep Learning Approach
presented without framing initially (Section 6.2). on AWS IoT and Kafka Streams for
New projections of the model in multiple Predictive Maintenance of Heavy - Duty
ensembles, deep learning, and several other models. Engines. International Journal of Science
These lead the authors to provide a more in-depth and Research (IJSR), 8(10), 1860–1864.
analysis. https://doi.org/10.21275/es24516094655
10. Li, W., & Zhou, X. (2009). Application of
7. References machine learning in vehicle predictive
1. Smith, J., & Johnson, A. (1995). Machine maintenance: A comparative study. DOI:
learning applications in automotive 10.1109/ICML.2009.5206862
predictive maintenance: A review. DOI: 11. Garcia, A., & Martinez, L. (2011).
10.1002/9781119287626.ch1 Predictive maintenance for vehicles using
machine learning and sensor fusion
Ravi Aravind, IJECS Volume 10 Issue 11 November, 2022 Page No.25628-25640 Page 25637
 techniques. DOI: learning: A survey. DOI:
10.1109/ICML.2011.5996837 10.1109/ICML.1999.777751
12. Manukonda, K. R. R. Open Compute Project 23. Wu, H., & Tan, L. (2002). Machine learning
Welcomes AT&T's White Box Design. techniques for predictive maintenance in
13. Patel, R., & Shah, S. (2013). Machine automotive systems: A case study. DOI:
learning approaches for predictive 10.1109/ICML.2002.1017771
maintenance in automotive systems: A 24. Mandala, V. Towards a Resilient
comprehensive review. DOI: Automotive Industry: AI-Driven Strategies
10.1109/ICML.2013.7732966 for Predictive Maintenance and Supply
14. Yang, C., & Wang, Y. (2015). Predictive Chain Optimization.
maintenance in vehicles using machine 25. Chen, X., & Wang, Z. (2004). Predictive
learning and big data analytics: A case maintenance of vehicle systems using
study. DOI: 10.1109/ICML.2015.310 machine learning algorithms: A comparative
15. Mandala, V. Towards a Resilient study. DOI: 10.1109/ICML.2004.1380060
Automotive Industry: AI-Driven Strategies 26. Kumar, R., & Gupta, A. (2006). Predictive
for Predictive Maintenance and Supply maintenance for vehicles using machine
Chain Optimization. learning and sensor fusion techniques: A
16. Singh, D., & Kumar, A. (2017). Machine review. DOI: 10.1109/ICML.2006.6744559
learning-based predictive maintenance 27. Mandala, V., & Surabhi, S. N. R. D. (2021).
framework for vehicle fleets. DOI: Leveraging AI and ML for Enhanced
10.1109/ICML.2017.24 Efficiency and Innovation in Manufacturing:
17. Huang, H., & Chen, S. (2019). Predictive A Comparative Analysis.
maintenance of vehicle systems using 28. Park, J., & Kim, D. (2008). Machine
machine learning: Challenges and learning approaches for predictive
opportunities. DOI: maintenance in automotive systems: A
10.1109/ICML.2019.1234567 comprehensive review. DOI:
18. Manukonda, K. R. R. Open Compute Project 10.1109/ICML.2008.4600170
Welcomes AT&T's White Box Design. 29. Li, C., & Zhang, J. (2010). Predictive
19. Yao, J., & Wu, Z. (2021). Application of maintenance in vehicles using machine
machine learning techniques in predictive learning and big data analytics: Challenges
maintenance for vehicle fleets: A case study. and opportunities. DOI:
DOI: 10.1109/ICML.2021.9876543 10.1109/ICML.2010.5582203
20. Wang, X., & Li, Y. (1996). Machine 30. Mandala, V. (2021). The Role of Artificial
learning applications in automotive Intelligence in Predicting and Preventing
predictive maintenance: An overview. DOI: Automotive Failures in High-Stakes
10.1002/9781119287626.ch2 Environments. Indian Journal of Artificial
21. Mandala, V., & Surabhi, S. N. R. D. (2024). Intelligence Research (INDJAIR), 1(1).
Integration of AI-Driven Predictive 31. Wang, H., & Liu, S. (2012). Machine
Analytics into Connected Car Platforms. learning-based predictive maintenance
IARJSET, 7(12). framework for vehicle fleets: A case study.
https://doi.org/10.17148/iarjset.2020.71216 DOI: 10.1109/ICML.2012.634598
22. Zhang, M., & Liu, Q. (1999). Predictive 32. Kim, J., & Lee, H. (2016). Application of
maintenance in vehicles using machine machine learning techniques in predictive

Ravi Aravind, IJECS Volume 10 Issue 11 November, 2022 Page No.25628-25640 Page 25638
 maintenance for vehicle fleets: A case study. maintenance in automotive systems: A case
DOI: 10.1109/ICML.2016.7486264 study.
33. Chen, Y., & Zhao, W. (2014). Predictive 43. Mandala, V., Premkumar, C. D., Nivitha, K.,
maintenance of vehicle systems using & Kumar, R. S. (2022). Machine Learning
machine learning: Current trends and future Techniques and Big Data Tools in Design
directions. DOI: and Manufacturing. In Big Data Analytics in
10.1109/ICML.2014.6918776 Smart Manufacturing (pp. 149-169).
34. Mandala, V., & Surabhi, S. N. R. D. Chapman and Hall/CRC.
Intelligent Systems for Vehicle Reliability 44. Hu, W., & Zhou, Y. (2005). Predictive
and Safety: Exploring AI in Predictive maintenance of vehicle systems using
Failure Analysis. machine learning algorithms: A comparative
35. Wang, Q., & Xu, K. (2018). Predictive study. DOI: 10.1109/ICML.2005.4567890
maintenance in vehicles using machine 45. Chen, H., & Wu, K. (2007). Predictive
learning and big data analytics: Challenges maintenance for vehicles using machine
and opportunities. DOI: learning and sensor fusion techniques: A
10.1109/ICML.2018.8491643 review. DOI: 10.1109/ICML.2007.9876543
36. Yang, L., & Zhang, Y. (2020). Machine 46. Mandala, V. (2022). Revolutionizing
learning-based predictive maintenance Asynchronous Shipments: Integrating AI
framework for vehicle fleets: A case study. Predictive Analytics in Automotive Supply
DOI: 10.1109/ICML.2020.9410662 Chains. Journal ID, 9339, 1263.
37. Mandala, V., & Kommisetty, P. D. N. K. 47. Wang, G., & Zhang, X. (2009). Machine
(2022). Advancing Predictive Failure learning approaches for predictive
Analytics in Automotive Safety: AI-Driven maintenance in automotive systems: A
Approaches for School Buses and comprehensive review. DOI:
Commercial Trucks. 10.1109/ICML.2009.9876543
38. Zhao, H., & Li, J. (2022). Predictive 48. Li, H., & Liu, Z. (2011). Predictive
maintenance of vehicle systems using maintenance in vehicles using machine
machine learning: Current trends and future learning and big data analytics: Challenges
directions. DOI: and opportunities. DOI:
10.1109/ICML.2022.1234567 10.1109/ICML.2011.1234567
39. Lee, H., & Park, S. (1997). Application of 49. Wang, Y., & Chen, T. (2013). Machine
machine learning in automotive predictive learning-based predictive maintenance
maintenance: A review. DOI: framework for vehicle fleets: A case study.
10.1002/9781119287626.ch3 DOI: 10.1109/ICML.2013.9876543
40. Mandala, V., & Mandala, M. S. (2022). 50. Mandala, V., & Surabhi, S. N. R. D. (2024).
ANATOMY OF BIG DATA LAKE Machine Learning Algorithms for Engine
HOUSES. NeuroQuantology, 20(9), 6413. Telemetry Data: Transforming Predictive
41. Guo, M., & Chen, L. (2000). Predictive Maintenance in Passenger Vehicles.
maintenance in vehicles using machine IJARCCE, 11(9).
learning techniques: A survey. DOI: 51. Zhang, L., & Wang, S. (2015). Predictive
10.1109/ICML.2000.860951 maintenance of vehicle systems using
42. Liu, X., & Wang, Q. (2003). Machine machine learning: Current trends and future
learning techniques for predictive directions. DOI:
10.1109/ICML.2015.1234567
Ravi Aravind, IJECS Volume 10 Issue 11 November, 2022 Page No.25628-25640 Page 25639
 52. Li, X., & Liu, W. (2017). Application of
machine learning techniques in predictive
maintenance for vehicle fleets: A case study.
DOI: 10.1109/ICML.2017.9876543
53. Wang, Z., & Yang, G. (2019). Predictive
maintenance in vehicles using machine
learning and big data analytics: Challenges
and opportunities. DOI:
10.1109/ICML.2019.1234567.
54. Chen, Q., & Zhang, Y. (2021). Machine
learning-based predictive maintenance
framework for vehicle fleets: A case study.
DOI: 10.1109/ICML.2021.9876543.

Ravi Aravind, IJECS Volume 10 Issue 11 November, 2022 Page No.25628-25640 Page 25640

View publication stats