Bulletin of Electrical Engineering and Informatics

Vol. 14, No. 3, June 2025, pp. 2188~2197

ISSN: 2302-9285, DOI: 10.11591/eei.v14i3.9398  2188

Revving up insights: machine learning-based classification of

OBD II data and driving behavior analysis using g-force metrics

Siddhanta Kumar Singh1, Anand Sharma2

1
Department of Computer and Communication Engineering, School of Computer Science and Engineering, Manipal University Jaipur,
Jaipur, India
2
Department of Computer Science and Engineering, School of Engineering and Technology, Mody University of Science and
Technology, Lakshmangarh, India

Article Info ABSTRACT

Article history: This research work uses machine learning (ML) approaches to classify on-
board diagnostics II (OBD II) data and g-force measures to provide a
Received Oct 3, 2024 thorough analysis of driving behavior. The research paper effectively
Revised Nov 30, 2024 demonstrates the classification of driving behaviours using OBD II and g-
Accepted Dec 25, 2024 force data. Driving behaviours are analyzed by using ML algorithms such as
random forest (RF), AdaBoost, and K-nearest neighbors (KNN). The
analysis goes beyond a summary by discussing how OBD II data, g-force
Keywords: metrics, and the algorithms interrelate to classify ten distinct driving
behaviors (e.g., weaving, swerving, and sideslipping). The RF classifier
Controller area network achieved the highest accuracy, which reinforces the strength of the chosen
G-force models. The inclusion of comparisons with other techniques supports
On-board diagnostics II arguments about the model's performance. The related works section
Sideslipping connects the references to the central topic by highlighting prior approaches
Swerving and research studies related to OBD II and driver behaviour analysis. The
Weaving goals of this study are improving the accuracy of driving behaviour
classification, with implications for traffic safety, driver education, and
insurance sectors.
This is an open access article under the CC BY-SA license.

Corresponding Author:
Siddhanta Kumar Singh
Department of Computer and Communication Engineering, School of Computer Science and Engineering
Manipal University Jaipur
Jaipur, India
Email: [email protected]

1. INTRODUCTION
The number of vehicles grows annually due to the rapid rising economy and government
liberalization policy for foreign automakers. The number of non-professional drivers is also rising quickly at
the same time. The primary causes of traffic accidents are now the individual drivers, as most inexperienced
drivers lack driving experience, are ignorant of the state of the vehicles, and have low knowledge of traffic
safety. Therefore, it is utmost important to determine the driving behavior so that the local authority can
retrieve the vehicle information and analyze and then take proper action.
On-board diagnostics (OBD) is a standard protocol [1] for vehicles that monitors various aspects of
vehicle performance and health. It emerged in the mid-1990s as a significant advancement over the original
OBD I system, which was introduced in the 1980s. OBD II's primary purpose is to check the engine's major
components and alert the driver and concerned person if there is any malfunction, thus aiding in the
maintenance and repair of vehicles. OBD II systems consist of a standardized digital communications port,
known as the data link connector (DLC), which is typically located under the driver side dashboard. This port

Journal homepage: http://beei.org

Bulletin of Electr Eng & Inf ISSN: 2302-9285  2189

allows external devices, such as scan tools and diagnostic software, to interface with the vehicle's computer
system. OBD-II has basically four communication protocols for interfacing the OBD-II port. But the
preference of protocol is dependent on the vehicle manufacturer. The key work here lies in a comprehensive
model development that categorizes behavior of driving with high accuracy using OBD II and G-force data.
The research work introduces a method that contrasts with previous approaches by incorporating multiple
weak classifiers into a robust classification system via AdaBoost. Additionally, the paper's application of
random forest (RF) and its comparison with other algorithms like Naive Bayes and logistic regression
demonstrate new findings in the classification of driving behaviours.

2. RELATED WORKS
Various diagnostic methods including Autel Maxidiag (elite series) and Launch X 431 were used [2]
to enhance the problems and malfunction identification. Research by Ramai et al. [3] offers the foundation
for an inexpensive method of online EV monitoring. Two Hyundai Ioniq EV had a Raspberry Pi ZeroW and
additional parts fitted in order to connect to them via the OBD-II connector. Wen et al. [4] conducted the
security analysis of wireless OBD-II scanners in this study. They designed and built DONGLESCOPE, an
automated program that tests these dongles on a real car in real time, covering all possible assault phases.
Options for monitoring important vehicle performance were described in this study by Yadav and Pathak [5],
along with a synopsis of the sensors used to retrieve these parameter values.
Shaikh et al. [6] developed an Android application that monitors driving behavior and notifies the
user of any discrepancy in driving habits in an effort to avert a catastrophic event. Vaiti et al. [7] research
proposed a data driven method for cluster emission calculation based on vehicle parameters related to
emissions. Ameen et al. [8] propose a way to categorize four driving behaviors: dangerous, aggressive,
secure, and typical behavior, with the goal of reducing the chance of accidents. Three light-duty passenger
cars (LDPVs) were tested by Zheng et al. [9] utilizing a laboratory dynamometer and the NEDC as a type-
approval cycle.
The platform presented in this work Peppes et al. [10] combines open-source technology with
machine and deep learning techniques to collect, store, process, analyze, and correlate data coming from cars.
The results of a driving behavior literature review are discussed by Hermawan and Husni [11]. This study
covers methods to collect OBD II data, and analyze, model, and assess it. Gharbins [12] assessed the degree
of proficiency among technicians in utilizing the OBD II instrument for standard maintenance on
automobiles containing electronic components, in addition to the degree of diagnostic equipment and
reference materials available in nearby repair shops.
Big data analysis requires the use of several languages and technologies, including Hadoop, Python,
Spark, R, and MATLAB, which are all covered by Meenakshi et al. [13]. The real-world statistics of a mild
hybrid car might differ depending on a number of factors, such as the vehicle, engine cycle, and powertrain,
as examined by Barbier et al. [14]. By examining the signals from the electronic control unit's PIDs,
Campoverde et al. [15] created an algorithm that can identify two typical driving behaviors, such as braking
to slow down and disengaging to shift gears. Subscription-based car maintenance options were recommended
by Maalik and Ponnampalam for people who don't have the time for repairs and upkeep [16]. Hamed et al.
[17] employ machine learning (ML) to improve the fuel consumption forecast accuracy model to decrease
consumption of fuel. OBD data for vehicle dynamics analysis and forecasting is examined by Navali et al.
[18]. According to the results, the OBD may provide data for a range of real-time and offline applications.
Using a transformer neural network (TNN) ML technique, Fernández et al. [19] established a way for
creating accurate speed correction data from OBD II data. Using OBD, Song and Kim [20] propose a method
for determining CAN specifications linked to important vehicle metrics.
Research by Kim and Baek [21] present a method that automatically extracts private in-vehicle data
by correlating sensor data with the sought information. A portable system to monitor mobile use while
driving and, if required, take control of a driver's phone whenever the car attains a speed limit (>10 km/h)
was proposed by Khandakar et al. [22]. A mathematical, graphic, and analytical approach for examining
customer driving behavior is provided by Navneeth et al. [23]. Through the OBD interface, Kumar and Jain
[24] suggested method that gathers vital performance data of vehicle, such as RPM, speed, position of
accelerator paddle, determined motor load, and other characteristics. The suggested approach categories
driver behavior using ML algorithms including AdaBoost, support vector machine (SVM) and RF. This
paper’s main goal is on DB analysis methods, and it is rendered in an elaborated manner [25].

3. METHOD
All the real time OBD data with the attributes like device time, absolute throttle position (ATP) (%),
accelerator pedal position (APP) (%), air fuel ratio, average trip speed (whilst stopped or moving)(km/h),
Revving up insights: machine learning-based classification of OBD II data … (Siddhanta Kumar Singh)
2190  ISSN: 2302-9285

engine load (%), revolutions per minute, fuel flow (FF) rate (gal/min), intake manifold pressure (psi), fuel
trim bank (FTB) 1 short term (%), kilometers per litre (long term average) (kpl), City driving (%), idle
driving (%), run time since engine start (s), speed (OBD II) (km/h), trip average KPL (kpl) were collected.
The wATP, wAPP, wFFR, wFTB1L, wIMP, wTPM are the normalized weights ATP, APP, FF rate, FTB 1
long, IMP and TP (manifold) respectively. So, the sum of the weights is given by (1):

Sum = wATP + wAPP + wFFR + wFTB1L + wIMP + wTPM (1)

where the normalized weights given by (2) to (7):

wATP
w ′ ATP = (2)
Sum

wAPP
w ′ APP = (3)
Sum

wFFR
w ′ FFR = (4)
Sum

wFTB1L
w ′ FTB1L = (5)
Sum

wIMP
w ′ IMP = (6)
Sum

wTPM
w ′ TPM = (7)
Sum

the final model for engine load can be written as (8):

𝐸𝑛𝑔𝑖𝑛𝑒 𝐿𝑜𝑎𝑑 = 𝑤 ′𝐴𝑇𝑃 . 𝐴𝑇𝑃 + 𝑤 ′𝐴𝑃𝑃 . 𝐴𝑃𝑃 + 𝑤 ′ 𝐹𝐹𝑅 ∗ 𝐹𝐹𝑅 +

𝑤 ′ 𝐹𝑇𝐵1𝐿 ∗ 𝐹𝑇𝐵1𝐿 + 𝑤 ′ 𝐼𝑀𝑃 ∗ 𝐼𝑀𝑃 + 𝑤 ′ 𝑇𝑃𝑀 ∗ 𝑇𝑃𝑀 (8)

This equation involves fuel usages which contrast with the existing (13). It has been found that fuel
consumption negatively affects driving scores and is closely tied to driver conduct. In (9) and (10) calculate
average fuel consumption, which is obtained by averaging all instantaneous fuel consumption measurements:
lit
𝑙𝑖𝑡 Fuel Flow(hr)
𝐹𝑢𝑒𝑙 𝐶𝑜𝑛𝑠𝑢𝑚𝑝𝑡𝑖𝑜𝑛( )= km (9)
𝑘𝑚 Speed( )
hr

MAF
𝐹𝑢𝑒𝑙 𝐹𝑙𝑜𝑤 = (10)
λ∗AFR∗ρ

where λ is the OBD II parameter which denotes air and fuel ratio and has standard value 1, AFR denotes the
stoichiometric ratio having standard value 14.7, MAF denotes mass air flow rate in gm/sec from MAF
sensor, ρ is the petrol density which is typically 770 gm/liter. It is measured using (11):

Speed
𝐹𝑢𝑒𝑙 𝐶𝑜𝑛𝑠𝑢𝑚𝑝𝑡𝑖𝑜𝑛 = (11)
Mass Air Flow

IDL_ENG indicates that the engine is not running. This suggests that fuel is being wasted, which will
eventually lower the driving score as indicated (12).

−1, 800 ≤ rpm ≤ 1000 and Gear = N and Speed = 0

Score = { (12)
1, Otherwise

Engine load parameter reading at idle is around 20%, while a reading of 100% indicates that the
engine is under full load. In general, a load parameter reading of 70% to 80% during normal driving mode is
considered optimal for both performance and fuel efficiency. An engine load >80% or <70% negatively
affects driving behavior negatively. Engine load as given in terms of air flow in (13):
𝐶𝑢𝑟𝑟𝑒𝑛𝑡 𝐴𝑖𝑟 𝐹𝑙𝑜𝑤
𝐸𝑛𝑔𝑖𝑛𝑒 𝐿𝑜𝑎𝑑 = Barometric Pressure 298
(13)
𝑀𝑎𝑥 𝐴𝑖𝑟 𝐹𝑙𝑜𝑤(𝑟𝑝𝑚). .√
29.92 Tamb+273

Bulletin of Electr Eng & Inf, Vol. 14, No. 3, June 2025: 2188-2197
Bulletin of Electr Eng & Inf ISSN: 2302-9285  2191

High speed braking is a type a braking when a vehicle applies brake suddenly as in Figure 1, and
acceleration on y-axis abruptly decreases and keeps even negative for some time while acceleration on x-axis
remains flat. Normal acceleration and normal deceleration are in the range of 0.1 m/sec2<normal acceleration
<2.74 m/sec2 and -0.1 m/sec2<normal deceleration<-2.74 m/sec2 respectively.
Brakes applied abruptly negatively affects driving-score if a<2.74 m/sec2 and 𝜎<2.05 and brake was
applied at v>55 km/hour. We can define the driving score S as given by (14):

𝑆 = 𝑓(𝑎, 𝜎, 𝑣) (14)

To represent the negative effect on driving scores when the conditions are met, we can introduce a
penalty function P as:

𝑘, 𝑖𝑓 𝑎 < 2.74 𝑎𝑛𝑑 𝜎 < 2.05 𝑎𝑛𝑑 𝑣 > 55

𝑃={ (15)
0, 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒

where k is a positive constant that represents the penalty value. Then, the driving score S can be modeled as
shown in (16):

𝑆 = 𝑆0 − 𝑃(𝑎, 𝜎, 𝑣) (16)

where, S0 is the initial driving score before the penalty.

If in the first 1.5 seconds, a force of 56.69 kg is experienced, then hard braking will be experienced.
Sudden unintended acceleration (SUA) is described as an unplanned, sudden, high-power acceleration from a
standing start, or a very slow starting speed combined with what appears to be a loss of braking efficiency as
shown in Figure 2.
When a vehicle speeds up suddenly, acceleration (ax) on x-axis remains flat while acceleration (ay)
on y-axis sharply goes up. Thus, the standard deviation (σ𝑎𝑥 ) and value range of acceleration on x-axis are
small. And ax(t) is x-axis accel as a time-function, ay(t) is y-axis accel as a time-function, σ𝑎𝑥 is denotes
standard deviation of ax and R 𝑎𝑥 is value range of ax. Given that ax remains flat, we can represent as (17) a
constant (C):

𝑎𝑥 (𝑡) = 𝐶 (17)

given that ay sharply increases, we can model it as a step function or an exponential function. A step function
(S) is a simple way to represent a sudden increase:

0, 𝑖𝑓 𝑡 < 𝑡0
𝑆={ (18)
𝐴, 𝑖𝑓 𝑡 ≥ 𝑡0

where A is the sharp increase in acceleration at time t0. Alternatively, an exponential function as given below
can represent a sharp increase more smoothly:

𝑎𝑦 (𝑡) = 𝐴(1 − 𝑒 −𝜆(𝑡−𝑡0) ) (19)

where λ is the rate at which ay increases, and t0 is the time when the vehicle starts to speed up suddenly.
Since ax remains constant, therefore,
The standard deviation σ𝑎𝑥 ≈ 0 𝑎𝑛𝑑 The value range R 𝑎𝑥 ≈ 0 because ax does not vary. Hard
acceleration is considered if acceleration >2.74 m/sec2. Some of the visual impressions of SUA could be car
with blurred background, leaving tire marks, front end lifted and emitting smoke. when we rev up, fuel is
squandered when the engine is revved up again without accomplishing any productive activity. This is seen
as negative in our proposed approach for calculating driving score.
The negative driving score S can be:

𝑆 = −𝑘. 𝑓(𝑅). 𝛿(𝐺). 𝛿(𝑉) (20)

where R, V, and G are the rpm, speed and gear respectively and k is a constant factor that determines the
severity of the penalty.
The indicator functions are given in (21) and (22).

Revving up insights: machine learning-based classification of OBD II data … (Siddhanta Kumar Singh)
2192  ISSN: 2302-9285

𝛿(𝐺) = 1 𝑖𝑓 𝐺 = 0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒 𝛿(𝐺) = 0 (21)

𝛿(𝑉) = 1 𝑖𝑓 𝑉 = 0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒 𝛿(𝑉) = 0 (22)

Penalty function: ƒ(R)=max (0, R−900). This function returns the amount by which RPM exceeds 900. If
RPM is 900 or less, the function returns 0. Therefore, the final score is given by (23).

𝑆 = −𝑘. max (0, 𝑅 − 900). 𝛿(𝐺). 𝛿(𝑉) (23)

The weaving pattern, as depicted in Figure 3, exhibits a sharp fluctuation in acceleration along the x-
axis, which persists for a certain amount of time. A negative score is awarded, if the SD (x-axis data) is large
and the Range (x-axis data) is large and acceleration (y-axis data) is Smooth. Smoothness of acceleration
(SmoothAy) can be evaluated as shown below using the mean absolute deviation (MAD) of the acceleration,
a lower MAD indicates smoother acceleration.
1
SmoothAy = ∑𝑛𝑖=1|𝐴𝑦,𝑖 − 𝐴̅𝑦 | (24)
𝑛

where 𝐴𝑦,𝑖 is the individual acceleration data point and 𝐴̅𝑦 is the mean acceleration. For negative award
function, the score should be high when SDx and R x are high and SmoothAy is low (indicating smooth
acceleration) as shown in (25):

𝑆𝐷𝑥 𝑅𝑥 ̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
SmoothAy
𝑆𝑐𝑜𝑟𝑒 = 𝑤1 . ( ) + 𝑤2 . ( ) − 𝑤3 . ( ) (25)
max (𝑆𝐷𝑥 ) max (𝑅𝑥 ) max (SmoothAy)

where w1 , w2 , and w3 are weights that determine the relative importance of each term, and the terms are
normalized by their maximum values to ensure they are comparable. Therefore, we can rewrite the model as:

𝑆𝐷𝑥 𝑅𝑥 ̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
SmoothAy
𝑆𝑐𝑜𝑟𝑒 = 𝑤1 . ( ) + 𝑤2 . ( ) − 𝑤3 . (̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ ) (26)
𝑆𝐷𝑥,𝑚𝑎𝑥 𝑅𝑥,𝑚𝑎𝑥 SmoothAy𝑚𝑎𝑥

where, SDx is standard deviation of x-axis data; 𝑆𝐷𝑥,𝑚𝑎𝑥 is maximum standard deviation observed in the
̅̅̅̅̅̅̅̅̅̅̅̅̅ is mean
dataset; R x is range of x-axis data; R x,max is maximum range observed in the dataset; SmoothAy
̅̅̅̅̅̅̅̅̅̅̅̅̅
absolute deviation of acceleration along y-axis; SmoothAymax is maximum mean absolute deviation
observed in the dataset; and w1 , w2 , w3 are weights to balance the components (can be tuned based on
empirical data or specific requirements).
Swerving is an enormous peak in acceleration on the x-axis is observed when swerving takes place,
as illustrated in Figure 4. The negative score S can be formulated as (27):

𝑆 = 𝑤1 . 𝑝𝑒𝑎𝑘(𝑎𝑥 ) + 𝑤2 . 𝑟𝑎𝑛𝑔𝑒(𝑎𝑥 ) + 𝑤3. 𝜎(𝑎𝑥 ) + 𝑤4 . |𝜇(𝑎𝑥 )| − 𝑤5 . 𝑓𝑙𝑎𝑡𝑛𝑒𝑠𝑠(𝑎𝑦 ) (27)

where, w1 , w2 , w3 , w4 , and w5 are weights that can be adjusted based on the importance of each factor,
flatness(𝑎𝑦 ) is inversely proportional to the variability of 𝑎𝑦 . A potential measure could be the reciprocal of
the standard deviation of 𝑎𝑦 :

1
𝑓𝑙𝑎𝑡𝑛𝑒𝑠𝑠(𝑎𝑦 ) = (28)
𝜎(𝑎𝑦 )+𝜀

where ϵ is a small constant to avoid division by zero.

Sideslipping is shown in Figure 5, sideslipping causes a rapid decline in y-axis acceleration. The
driving score S can be written as(29).
𝑑𝑎𝑦
𝑆 = 𝑤1 . max (| |) + 𝑤2 . (−𝑚𝑖𝑛(𝑎𝑦 )) + 𝑤3 . (−𝑎̅𝑦 ) + 𝑤4 . (max(𝑎𝑦 ) − min(𝑎𝑦 )) + 𝑤5 . |𝑎̅𝑥 | (29)
𝑑𝑡

To balance the importance of each component, we might choose w1 , w2 , w3 , w4 , and w5 as

.1,.5,.5,.3 and .2 respectively. These weights can be adjusted based on empirical data or specific application
𝑑𝑎
needs. max (| 𝑦 |) denotes sharp fall in ay can quantify the sharp fall by looking at the second derivative
𝑑𝑡
(jerk) or by defining a threshold for the rate of change.

Bulletin of Electr Eng & Inf, Vol. 14, No. 3, June 2025: 2188-2197
Bulletin of Electr Eng & Inf ISSN: 2302-9285  2193

𝑚𝑖𝑛(𝑎𝑦 )denotes minimum value of 𝑎𝑦 : min(ay)<0

𝑎̅𝑦 denotes mean value of 𝑎𝑦 : 𝑎̅𝑦 <0
max(𝑎𝑦 ) − min(𝑎𝑦 ) denotes range of 𝑎𝑦 and ensures this value is large.
|𝑎̅𝑥 |denotes mean value of 𝑎𝑥 not near zero, indicating significant sideways motion.
A fast U turn is the case when a driver makes a sudden U-turn shown Figure 6, to the right or left,
x-axis acceleration increases rapidly to a very high value or decreases rapidly to a very low value,
respectively. The driving score S can be modeled as a function of 5 tuples as given in (30):

𝑆 = 𝑓 (𝜇𝑎𝑥 , 𝜎𝑎𝑥 , 𝑅𝑎𝑥 , 𝜇𝑎𝑦 , 𝑇) (30)

where 𝜇𝑎𝑥 is the mean of 𝑎𝑥 , 𝜎𝑎𝑥 is the standard deviation of 𝑎𝑥 ,𝑅𝑎𝑥 is the range of 𝑎𝑥 , 𝜇𝑎𝑦 is the mean of 𝑎𝑦
and T is the time duration of the maneuver. A form (31) for the driving score could be:

|𝜇𝑎𝑦 |
𝑆 = 𝑘1 . |𝜇𝑎𝑥 | + 𝑘2 . 𝜎𝑎𝑥 + 𝑘3 . 𝑅𝑎𝑥 + 𝑘4 . (1 − ) + 𝑘5 . 𝑇 (31)
max(|𝑎𝑦 |)

where 𝑘1 , 𝑘2 , 𝑘3 , 𝑘4 , and 𝑘5 are weighting coefficients that determine the importance of each term. These
coefficients can be adjusted based on empirical data or specific requirements for the driving score.

Figure 1. Sudden braking Figure 2. Sudden unintended Figure 3. Weaving movement

acceleration

Figure 4. Swerving movement Figure 5. Sideslipping movement Figure 6. Fast U turn movement

Various mathematical models were developed, and the dataset was analyzed using classification
algorithms in Python ML. The supervised algorithms applied include AdaBoost, RF, K-nearest neighbor
(KNN), Naive Bayes, and logistic regression. The AdaBoost algorithm, an instance of adaptive boosting
methodology, serves as a ML algorithm employed for classification, by amalgamating numerous weak
classifiers to form a robust classifier while an ensemble learning technique called RF constructs several
decision trees and aggregates their predictions to increase accuracy. KNN uses the majority label of their
closest neighbours to categorise data points. Based on Bayes' Theorem, a probabilistic classifier assumes that
Revving up insights: machine learning-based classification of OBD II data … (Siddhanta Kumar Singh)
2194  ISSN: 2302-9285

characteristics are conditionally independent. For each training sample (xi , yi ) is assigned a weight 𝑤𝑖 .
Initially, all weights are set equally.
1
𝑤𝑖 = 𝑓𝑜𝑟 𝑎𝑙𝑙 𝑖 (32)
𝑁

For each iteration t=1…up to T (T being the total number of iterations), a weak classifier ht(x) is
trained using the weighted training-data and compute the classification error ϵt of ht as (33):

∈ 𝑡 = ∑𝑁
𝑖=1 𝑤𝑖 . 𝟏(ℎ𝑡 (𝑥𝑖 ) ≠ 𝑦𝑖 ) (33)

where, 1 is the indicator function.

Then calculate the weight αt for the weak classifier as (34):

1 1−∈𝑡
𝛼𝑡 = 𝑙𝑛 ( ) (34)
2 ∈𝑡

(𝑡+1)
𝑤𝑖 = 𝑤𝑖𝑡 exp (−𝛼𝑡 𝑦𝑖 ℎ𝑡 (𝑥𝑖 )) (35)

the weight is normalized using (36):

(𝑡+1)
(𝑡+1) 𝑤𝑖
𝑤𝑖 = 𝑁 (𝑡+1) (36)
∑𝑗=1 𝑤𝑗

the final strong classifier is (37).

𝐻(𝑥) = 𝑠𝑖𝑔𝑛(∑𝑇𝑡=1 𝛼𝑡 ℎ𝑡 (𝑥𝑖 )) (37)

The parameters 𝛼𝑡 play a crucial role in determining the impact of each weak classifier on the
ultimate decision, giving priority to classifiers that exhibit strong performance on the weighted training
dataset. Drivers’ ranks are awarded on the scale of 10 with 10 being excellent driving behavior. For every
mistake committed, negative points will be rewarded in total score. Table 1 shows the score determinant.

Table 1. Driving score determinant

Driving parameters Effect on driving score Impact on driving score
Fuel_cons -ve HIGH
Idle_Eng -ve HIGH
Eng_Load +ve MODERATE
HIGH_SPEED_BRAKING -ve HIGH
SUA -ve HIGH
REV_ENGINE -ve HIGH
Weaving -ve HIGH
Swerving -ve HIGH
Sideslipping -ve HIGH
Fast U Turn -ve HIGH

4. RESULTS AND DISCUSSION

Feature classification offers a visual representation of derived parameters for various driver-classes
from ten drivers from D1 to D10 on honda brio at different terrain for 10 kms. Figure 7 reveals that driver
classes D1, D3, and D7 achieve the highest driving scores, each incurring a single negative penalty for idle
engine, high-speed braking, and idle engine, respectively. Similarly, the remaining parameters provide a clear
visualization of the data, facilitating the development of a model to classify drivers. The result is based on the
those driving classes mentioned which is uniqueness of this work and the accuracy is also highest as
mentioned below.
The driving behavior analysis [24] made and the accuracy was assessed using training and test
dataset. The accuracy rates were as follows: AdaBoost at 77%, Naive Bayes at 88%, KNN at 98%, logistic
regression at 99%, and RF at 100%. RF is substantially slower than all other classification techniques
because it uses multiple decision trees for predictions and hence for speedy prediction we can assume logistic
regression and KNN which also gives nice predictions.

Bulletin of Electr Eng & Inf, Vol. 14, No. 3, June 2025: 2188-2197
Bulletin of Electr Eng & Inf ISSN: 2302-9285  2195

Driving behavior
10
Driver score

0
D1 D2 D3 D4 D5 D6 D7 D8 D9 D10

-5
Driver class
Avg_Fuel IDLE_ENGINE Engine_Load
HIGH_SPEED_BRAKING SUA REV_ENGINE
Weaving Swerving Sideslipping

Figure 7. Driving score

5. CONCLUSION
ML techniques such as Ada Boost, RF, KNN, Naive Bayes, and logistic regression, were used to
develop and validate a model for classifying driving behavior. The suggested techniques were simple to use
and quite accurate, with random forest's maximum accuracy of 100%. This approach is still beneficial and
can aid the traffic police, insurance company, local government, and claim processing. The findings of this
research may also have a direct bearing on the development of driving assistance and classification systems.
Although contemporary tools and algorithms are employed in this research project, there is still much room
for the occasional implementation of further contemporary software tools and algorithms in accordance with
future requirements. The method that is being given is not exclusive to cars with internal combustion
engines; it may also be applied to contemporary hybrid and electric vehicles. The recommended course of
action is doable and adaptable to new technology. Delays in data gathering and storage have an impact on the
suggested method's results; as a result, classification of behavior is not possible for short road journeys,
particularly when the route or vehicle are different.

ACKNOWLEDGEMENTS
We are thankful to the department and school for giving us opportunities and ample time and help
whenever required.

FUNDING INFORMATION
No funding is involved for this paper.

AUTHOR CONTRIBUTIONS STATEMENT

This journal uses the Contributor Roles Taxonomy (CRediT) to recognize individual author
contributions, reduce authorship disputes, and facilitate collaboration.

Name of Author C M So Va Fo I R D O E Vi Su P Fu
Siddhanta Kumar ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓
Singh
Anand Sharma ✓ ✓ ✓ ✓ ✓

C : Conceptualization I : Investigation Vi : Visualization

M : Methodology R : Resources Su : Supervision
So : Software D : Data Curation P : Project administration
Va : Validation O : Writing - Original Draft Fu : Funding acquisition
Fo : Formal analysis E : Writing - Review & Editing

Revving up insights: machine learning-based classification of OBD II data … (Siddhanta Kumar Singh)
2196  ISSN: 2302-9285

CONFLICT OF INTEREST STATEMENT

All the authors do not have any conflict of interest.

INFORMED CONSENT
No personal or patient information is used for scientific reasons.

DATA AVAILABILITY
The data that support the findings of this study are available from the corresponding author,
[SKSingh], upon reasonable request.

REFERENCES
[1] S. K. Singh and A. K. Singh, “Driving pattern analysis to determine driver behaviours for a local authority based on the cloud
using OBD II,” International Journal of Electrical and Computer Engineering Systems, vol. 13, no. 10, pp. 937–944, Dec. 2022,
doi: 10.32985/ijeces.13.10.14.
[2] M. M. Oluwaseyi and A. M. Sunday, “Specifications and Analysis of Digitized Diagnostics of Automobiles: A Case Study of on
Board Diagnostic (OBD II),” International Journal of Engineering Research & Technology (IJERT), vol. V9, no. 01, Jan. 2020,
doi: 10.17577/ijertv9is010045.
[3] C. Ramai, V. Ramnarine, S. Ramharack, S. Bahadoorsingh, and C. Sharma, “Framework for Building Low-Cost OBD-II Data-
Logging Systems for Battery Electric Vehicles,” Vehicles, vol. 4, no. 4, pp. 1209–1222, Oct. 2022, doi: 10.3390/vehicles4040064.
[4] H. Wen, Q. A. Chen, and Z. Lin, “Plug-N-Pwned: Comprehensive vulnerability analysis of OBD-II dongles as a new over-the-air
attack surface in automotive IoT,” Proceedings of the 29th USENIX Security Symposium, pp. 949–965, 2020.
[5] P. Yadav and P. K. Pathak, “On Board Diagnostics (OBD): A DTC (Diagnostic Trouble Code) Study,” in International
Conference on Intelligent Technologies & Science - 2021 (ICITS-2021), 2021, pp. 1–4.
[6] M. K. Shaikh, S. Palaniappan, F. Ali, and M. Khurram, “Identifying Driver Behaviour Through Obd-Ii Using Android
Application,” Palarch’s Journal of Archaeology of Egypt/Egyptology, vol. 17, no. 7, pp. 13636–13647, 2020.
[7] T. Vaiti, L. Tišljaric, T. Erdelic, and T. Caric, “Traffic Emissions Clustering Using OBD-II Dataset Based on Machine Learning
Algorithms,” Transportation Research Procedia, vol. 64, pp. 364–371, 2022, doi: 10.1016/j.trpro.2022.09.040.
[8] H. A. Ameen, A. K. Mahamad, S. Saon, M. A. Ahmadon, and S. Yamaguchi, “Driving behaviour identification based on OBD
speed and GPS data analysis,” Advances in Science, Technology and Engineering Systems, vol. 6, no. 1, pp. 550–569, Jan. 2021,
doi: 10.25046/aj060160.
[9] X. Zheng et al., “Real-world fuel consumption of light-duty passenger vehicles using on-board diagnostic (OBD) systems,”
Frontiers of Environmental Science and Engineering, vol. 14, no. 2, pp. 1–10, Apr. 2020, doi: 10.1007/s11783-019-1212-6.
[10] N. Peppes, T. Alexakis, E. Adamopoulou, and K. Demestichas, “Driving behaviour analysis using machine and deep learning
methods for continuous streams of vehicular data,” Sensors, vol. 21, no. 14, p. 4704, Jul. 2021, doi: 10.3390/s21144704.
[11] G. Hermawan and E. Husni, “Acquisition, Modeling, and Evaluating Method of Driving Behavior Based on OBD-II: A Literature
Survey,” IOP Conference Series: Materials Science and Engineering, vol. 879, no. 1, Jul. 2020, doi: 10.1088/1757-
899X/879/1/012030.
[12] J. Gharbins, Research on assessing knowledge level of mechanics on OBD,
https://www.academia.edu/8632320/Research_on_assessing_knowledge_level_of_mechanics_on_OBD 2022. (Date Accessed Jul
2024).
[13] Meenakshi, R. Nandal, and N. Awasthi, “Obd-ii and big data: A powerful combination to solve the issues of automobile care,”
Advances in Intelligent Systems and Computing, 2021, pp. 177–189, doi: 10.1007/978-981-15-7907-3_14.
[14] A. Barbier, J. M. Salavert, C. E. Palau, and C. Guardiola, “Analysis of Real-Driving Data Variability for Connected Vehicle
Diagnostics,” IFAC-PapersOnLine, vol. 55, no. 24, pp. 45–50, 2022, doi: 10.1016/j.ifacol.2022.10.260.
[15] P. A. M. Campoverde, N. D. R. Campoverde, G. P. N. Quirola, and A. K. B. Naula, “Characterization of Braking and Clutching
Events of a Vehicle Through OBD II Signals,” in Advances in Intelligent Systems and Computing, 2021, pp. 134–143, doi:
10.1007/978-3-030-59194-6_12.
[16] U. Maalik, P. Ponnampalam, and P. Pirapuraj, “Intelligent Vehicle Diagnostic System for Service Center using OBD-II and IoT,”
International Conference of Science and Technology, pp. 209–214, 2021.
[17] M. A. Hamed, M. H. Khafagy, and R. M. Badry, “Fuel Consumption Prediction Model using Machine Learning,” International
Journal of Advanced Computer Science and Applications, vol. 12, no. 11, pp. 406–414, 2021, doi:
10.14569/IJACSA.2021.0121146.
[18] N. Navali, L. Vanajakshi, and D. M. Bullock, “Application of On-Board Diagnostics (OBD) Data for Vehicle Trajectory
Prediction,” in Lecture Notes in Civil Engineering, 2023, pp. 319–328, doi: 10.1007/978-981-19-4204-4_19.
[19] A. F. Fernández, E. S. Morales, M. Botsch, C. Facchi, and A. G. Higuera, “Generation of Correction Data for Autonomous
Driving by Means of Machine Learning and On-Board Diagnostics,” Sensors, vol. 23, no. 1, pp. 1–28, Dec. 2023, doi:
10.3390/s23010159.
[20] H. M. Song and H. K. Kim, “Discovering CAN Specification Using On-Board Diagnostics,” IEEE Design and Test, vol. 38, no.
3, pp. 93–103, Jun. 2021, doi: 10.1109/MDAT.2020.3011036.
[21] B. Kim and Y. Baek, “Sensor-based extraction approaches of in-vehicle information for driver behavior analysis,” Sensors, vol.
20, no. 18, pp. 1–20, Sep. 2020, doi: 10.3390/s20185197.
[22] A. Khandakar et al., “Portable system for monitoring and controlling driver behavior and the use of a mobile phone while
driving,” Sensors (Switzerland), vol. 19, no. 7, pp. 1–17, Mar. 2019, doi: 10.3390/s19071563.
[23] S. Navneeth, K. P. Prithvil, N. R. S. Hari, R. Thushar, and M. Rajeswari, “On-board diagnostics and driver profiling,” in
Proceedings of the 2020 International Conference on Computing, Communication and Security, ICCCS 2020, IEEE, Oct. 2020,
pp. 1–6, doi: 10.1109/ICCCS49678.2020.9277449.
[24] R. Kumar and A. Jain, “Driving behavior analysis and classification by vehicle OBD data using machine learning,” Journal of

Bulletin of Electr Eng & Inf, Vol. 14, No. 3, June 2025: 2188-2197
Bulletin of Electr Eng & Inf ISSN: 2302-9285  2197

Supercomputing, vol. 79, no. 16. pp. 18800–18819, Dec. 13, 2023, doi: 10.1007/s11227-023-05364-3.
[25] M. Malik and R. Nandal, “A framework on driving behavior and pattern using On-Board diagnostics (OBD-II) tool,” Materials
Today: Proceedings, vol. 80, pp. 3762–3768, 2023, doi: 10.1016/j.matpr.2021.07.376.

BIOGRAPHIES OF AUTHORS

Siddhanta Kumar Singh has 25 years of work experience in the fields of

software training in engineering colleges, university and multinational companies in IT
technologies in India and China. He is working as an Assistant Professor at Manipal
University Jaipur. He did his B.E. (Computer Science and Engineering), M.Tech. (Computer
Science and Engineering) and Ph.D. (Computer Science and Engineering). The areas of
reserach interest are on board diagnostic II, driving behavioural analysis, ML, and IoT. He can
be contacted at email: [email protected].

Anand Sharma is an Associate Professor in the Department of Computer Science

and Engineering, Mody University of Science and Technology, Laksmangarh, India. His
qualifications are BE (IT), MTech (IT) and Ph.D. He has more than 17 years of experience. He
has more than 156 papers published, and 37 conference papers presented, 8 books written, and
7 patents published. He also received the best researcher award, best young scientist award,
best academician award. His area of research is ML, network security, and OBD. He can be
contacted at email: [email protected].

Revving up insights: machine learning-based classification of OBD II data … (Siddhanta Kumar Singh)