Algorithm performance comparison for earthquake signal recognition on smartphone’s accelerometer

TELKOMNIKA Telecommunication, Computing, Electronics and Control
Vol. 18, No. 5, October 2020, pp. 2505~2516
ISSN: 1693-6930, accredited First Grade by Kemenristekdikti, Decree No: 21/E/KPT/2018
DOI: 10.12928/TELKOMNIKA.v18i5.14708  2505
Journal homepage: http://journal.uad.ac.id/index.php/TELKOMNIKA
Algorithm performance comparison for earthquake signal
recognition on smartphone’s accelerometer
Hapsoro Agung Nugroho1
, Haryas Subyantara Wicaksana2
, Hariyanto3
, Rista H. Virgianto4
1,2,3
Department of Instrumentation for MKG, School of Meteorology, Climatology, and Geophysics, Indonesia
4
Department of Climatology, School of Meteorology, Climatology, and Geophysics, Indonesia
Article Info ABSTRACT
Article history:
Received Nov 29, 2019
Revised Jan 13, 2020
Accepted Apr 10, 2020
Micro-electro-mechanical-system accelerometer is able to detect acceleration
signal caused by earthquake. Such type of accelerometer is also used by
smartphones. There are few algorithms that can be used to recognize the type
of acceleration signal from smartphone. This study aims to find signal
recognition algorithm in order to consider the most proper algorithm for
earthquake signal detection. The initial stage of designing the recognizer is
data collection for each type of signal classification. The next step is to apply
a highpass filter to separate the signals collected from the gravitational
acceleration signal. The signal is divided into several segments. The system
will extract features of each signal segment in the time and frequency domain.
Each signal segment is then classified according to the type of signal using
the classifier through a series of training data processes. The classifier which
has the highest accuracy value is exported into the new input signal modeling.
As the result, fine K-NN algorithm has the highest level of accuracy in
the classification. The fine K-NN algorithm has an accuracy rate of 99.75% in
the classification of human activity signals and earthquake signals with a
memory capacity of 6,044 kilobytes and processing time of 15.93 seconds.
This algorithm has the best classifier criteria compared to decision tree,
support vector machine and linear discriminant analysis algorithms.
Keywords:
Accelerometer
Earthquakes
Human activity
K-NN algorithm
Signal recognition
This is an open access article under the CC BY-SA license.
Corresponding Author:
Hapsoro Agung Nugroho,
Department of Instrumentation for Meteorology Climatology and Geophysics (MKG),
School of Meteorology Climatology and Geophysics,
I/5 Perhubungan St., Pondok Betung, Pondok Aren, 15221, Banten, Indonesia.
Email: hapsoro.agung@bmkg.go.id
1. INTRODUCTION
Indonesia is situated in the cross zone of three tectonic plates, namely Eurasia, Indo-Australia and
Pacific. Based on this condition, Indonesia becomes a dangerous area of earthquake zones [1, 2]. Thousands
of strong motion earthquakes with more than 6.0 moment magnitude impacted Indonesia in 1976-2006 [3, 4].
Law number 31 Year 2009 of Indonesian Republic states that the Meteorology, Climatology and Geophysics
Agency (BMKG) is an agency authorized to publish information on earthquake events, both pre-earthquake
and post-earthquake cases [5]. Confirmation of the incident at the earthquake site is needed to support
the BMKG performance. This certainly involves the active role of the community.
Crowdsourcing is now a prominent terminology to collect, process, and distribute information
to society. It possibly becomes a robust idea to overcome the small number of seismic monitoring networks
in Indonesia. One of the media for crowdsourcing is smartphone. Nowadays, smartphones are embedded

 ISSN: 1693-6930
TELKOMNIKA Telecommun Comput El Control, Vol. 18, No. 5, October 2020: 2505 - 2516
2506
by digital accelerometer. This component has potential to detect ground strong motion parameters.
The Quaqe-Catcher Network (QNC) and Community Seismik Network (CSN) utilize Micro electro mechanical
system (MEMS) accelerometer to detect building vibration caused by earthquake [6]. Such type of
accelerometer is also used by smartphones. However, the noise of smartphone’s accelerometer faces big
obstacles, mainly due to human activities [7].
Anguita et al. from the University of Genova produced a set of accelerometer data to identify six types
of human movement. These movements include standing, walking, lying down, sitting, going up
and down the stairs [8]. Bayat, Pomplun and Tran from the University of Massachusetts examined
the use of a single triaxial accelerometer sensor on smartphones based multilayer perceptron with an accuracy
of 90% [9]. Chen et al. succeeded in designing a HAR based on long-short term memory (LSTM) with
an accuracy of 92.1% [10]. However, it has not been improved for seismology. This study aims to compare
signal recognition algorithm in order to determine the most proper algorithm for earthquake signal detection.
2. RESEARCH METHOD
Smartphone accelerometer sensor type is LSM6DSL. It is a triaxial accelerometer with 16-bit
resolution and 50 Hz sampling frequency [11]. Several types of MEMS-type accelerometer signal inputs
include gravity acceleration signals, acceleration due to human body movements, offsets and noise [12, 13].
The initial stage of designing the recognizer is to collect data for each type of signal classification. The next
step is to apply a highpass filter to separate the signals collected from the gravitational acceleration signal.
The signal is divided into several segments. The system will extract features of each signal segment in
the time and frequency domain. Each signal segment is then classified according to the type of signal using
the classifier through a series of training data processes. The classifier which has the highest accuracy value is
exported into the new input signal modeling. The classification test is done by using Python 3 language program
in Linux Ubuntu terminal.
Data are collected by recording smartphone accelerometer signals on activities carried out by
10 subjects. Each subject was instructed to sit, stand, lie down, walk and run. The activities were carried out
at varying speed and gestures according to the subject's habits. The activities were carried out when
the smartphone is placed in shirt pocket and trouser pocket with a variety of types according to the subject's
clothing during the study. Samples of earthquake signals were taken from BMKG accelerograph signals
that record earthquake events in Lombok and Palu. The amount of earthquake signal raw data is 214 data, while
the subject activity raw data are 2545 data in the shirt pocket and 2430 data in the trouser pocket.
Highpass filters are designed to separate the linear acceleration signal of the subject's movement from
the gravitational acceleration signal. This filter is a Butterworth type 3 highpass filter with a cutoff frequency
of 0.1 Hz. Order 3 is considered to be quite effective to reduce gravitational acceleration signals with dominant
frequency ranging from 0.1-0.5 Hz [8]. Data extraction was undertaken by collecting all sample data into a
data set. Data windowing uses non-overlapping techniques with frame duration of one second, or 50 raw data
in one signal type. The signal was extracted to obtain the signal features in the time domain and frequency
domain [14]. Figure 1 shows a flowchart of the process used in this study.
2.1. Decision tree (DT) algorithm
Decision tree is a classifier that works by arranging decision trees based on predictors or features
that exist to determine the class of objectives. The parameters required are entropy and information gain
of each signal feature. Information gain is a measure of the effectiveness of a feature in classifying data.
The information gain equation is formulated as follows [15, 16]:
𝐼𝐺 (𝑆, 𝑓) = 𝐸𝑛𝑡𝑟𝑜𝑝𝑦(𝑆) − ∑
|𝑆 𝑉|
|𝑆|
𝐸𝑛𝑡𝑟𝑜𝑝𝑦(𝑆 𝑉)
𝑣∈𝑣𝑎𝑙𝑢𝑒(𝑓)
IG (S, f) is the information gain value for a particular feature or predictor. Entropy (S) is the value of
overall data entropy. V is a possible value for predictor f, while value (f) is a set of possible values for predictor
f. | Sv | is the number of samples for values V and | S | is the sum of all data samples. Entropy (SV) is the value
of entropy for samples that have a value of V. The predictor that has the highest IG will be the root of the other
predictors in the decision tree. The remaining IG predictor value is recalculated to become the second root,
and so on. Root continues branching down, therefore it will appear a destination class at the top of the decision
tree [10].

TELKOMNIKA Telecommun Comput El Control 
Algorithm performance comparison for earthquake signal recognition… (Hapsoro Agung Nugroho)
2507
Input accelerometer data
(X,Y,Z-Axis)
Gravitation signal filter
Signal segmentation
time domain feature
extraction
Fourier transformation
Training classifier signal
(DT,LDA, KNN, SVM)
accuracy
(DT,LDA, KNN, SVM)
export classifier modeling
Output signal classification
Sitting, Standing, Laying, Walking, Running, Earthquake
there is an
earthquake signal?
Save
modelling
finish
start
Y
Human activity filter
N
Figure 1. Flowchart activity signal and earthquake signal
2.2. Linear discriminant analysis (LDA) algorithm
Linear discriminant analysis is to reduce the dimensionality of input data predictors. LDA aims
to separate the two destination classes by dividing the decision regions [17]. LDA consists of three steps of
work [18]. The first step is to calculate separability between classes using the value of the interclass variance.
The second step is to calculate the distance between the class average against each sample in the class itself or
intraclass variance. The third step is to reduce the dimensional space between classes hence the value of
interclass variance is increased while the value of intraclass variance is decreased [12]. The LDA algorithm is
coherently described as follows [19]:
− Determine a data matrix with row numbers M and column numbers N
− Calculate the average of each class (1 x M)
− Calculate the average of all classes (1 x M)
− Calculate interclass variance:
𝑆 𝐵 = ∑ 𝑛𝑖(𝜇𝑖 − 𝜇)𝑐
𝑖=1 (𝜇𝑖 − 𝜇) 𝑇
− Calculate intraclass variance:
𝑆 𝑊 = ∑ ∑ (𝑥𝑖𝑗 − 𝜇 𝑗)
𝑛 𝑗
𝑖=1
𝑐
𝑗=1 (𝑥𝑖𝑗 − 𝜇 𝑗) 𝑇
− Calculate variance matrix:
𝑊 = 𝑆 𝑊
−1
𝑆 𝐵

 ISSN: 1693-6930
2508
− Calculate Eigen value and eigen vector from matrix W
− Arrange Eigen vector based on Eigen value. First Eigen vector is used as lower dimensional space (Vk)
− Project all samples to Vk
− Calculate the distance of new sample to data average value which are projected by Vk for each class.
The smallest distance value is included into certain class.
2.3. Support vector machine (SVM) algorithm
Support vector machine is to determine the best hyperplane as a separator between destination
classes [20]. Hyperplane is a delimiter that divides a vector space into two parts, therefore two different classes
can be separated. The best characteristic of hyperplane is the maximum margin value. Margin is the distance
between the hyperplane and the nearest predictor vector of the two classes. This predictor vector is called
support vector. Hyperplane can be used to separate linear and nonlinear data [21]. The predictor is denoted as
xi and the class is denoted as yi which contains of two classes that assumed to be -1 and 1, while the w vector
is a support vector. The hyperplane equation for linear data is formulated as follows:
𝑦𝑖⃗⃗ = 𝑤⃗⃗ . 𝑥𝑖⃗⃗⃗ + 𝑏 Hyperplane equation
𝑤⃗⃗ . 𝑥𝑖⃗⃗⃗ + 𝑏 ≤ −1 for class -1
𝑤⃗⃗ . 𝑥𝑖⃗⃗⃗ + 𝑏 ≥ 1 for class 1
The largest margin can be determined by maximizing the value of the hyperplane's distance to
the vector at its closest point. If the margin value has greater value, then it has better classification. Hyperplane
for nonlinear data is designed using the Kernel function, which transforms a 2-D vector field into a 3-D vector.
Nonlinear predictors are easier to separate in 3-D vector space. The commonly used kernel functions are
the polynomial, gaussian and sigmoid kernels [22].
− Polynomial kernel transformation equation:
𝐾(𝑥𝑖,⃗⃗⃗⃗ 𝑥𝑗⃗⃗⃗ ) = (𝑥𝑖,⃗⃗⃗⃗⃗⃗ 𝑥𝑗,⃗⃗⃗⃗ + 1) 𝑝
, p is the highest number of exponent
− Gaussian kernel transformation equation:
𝐾(𝑥𝑖,⃗⃗⃗⃗ 𝑥𝑗⃗⃗⃗ ) = exp (−𝜎‖𝑥𝑖⃗⃗⃗ − 𝑥𝑗⃗⃗⃗ ‖
2
), 𝜎 is variance of vector x
− Sigmoid kernel transformation equation:
𝐾(𝑥𝑖,⃗⃗⃗⃗ 𝑥𝑗⃗⃗⃗ ) = tanh (𝛼𝑥𝑖⃗⃗⃗ 𝑥𝑗⃗⃗⃗ + 𝛽), alpha and beta are sigmoid constant
The hyperplane equation applied to nonlinear data by utilizing the Kernel function is calculated as follows:
𝑦𝑖⃗⃗ = 𝑤⃗⃗ . 𝐾(𝑥𝑖,⃗⃗⃗⃗ 𝑥𝑗⃗⃗⃗ ) + 𝑏
2.4. K-nearest neighbor (K-NN) algorithm
K-Nearest Neighbor algorithm is to calculate the distance of a new data input against the K-data
learning model [23]. This algorithm is also useful to search for the nearest neighbor from a new data input.
The proximity of new input data to the model data is generally calculated using the Euclidean Distance [24]
as follows:
𝐸(𝐴, 𝐵) = ∑ √(𝐴𝑖 − 𝐵𝑖)2𝑁
𝑖=1
A is the new input data, while B is the learning model data. The new input data is tested against
each learning data point, then against the neighbor sequence with the smallest distance value according to
the K number. K-NN has several types of processing algorithms [25], namely:
− Fine K-NN, using the closest neighbor (K=1).
− Medium K-NN, using ten closest neighbors (K=10).
− Coarse K-NN, using one hundred closest neighbors (K=100).
− Cosine K-NN, using distance calculation of the closest neighbor based on cosine distance matrix:

2509
𝑑 = 1 −
𝑢⃗ . 𝑣
|𝑢⃗ |. |𝑣|
− Cubic K-NN, using distance calculation of the closest neighbor based on cubic distance matrix:
𝑑 = √∑|𝑢𝑖 − 𝑣𝑖|3
𝑛
𝑖=1
3
In general, the K-NN algorithm steps can be described as follows:
a. Determine K value or number of the closest neighbor
b. Calculate new input distance to all learning models.
c. Arrange the new input distance from the closest to far distance
d. Detemine closest neighbor category based on K value.
e. Use majority category class of closest neighbor to predict the result of new input data.
3. RESULTS AND ANALYSIS
Table 1 shows the results of the smartphone accelerator signal extraction for various activities in
the time domain. Table 1 proves that significant difference occurs in characteristics between human activity
signals and earthquake signals in the time domain. Linear acceleration signals due to human activity reported
greater value for all features compared to earthquake signals. The signal distribution of acceleration of human
activity tends to be leptokurtic, while earthquake signals tend to be platykurtic when viewed from the value of
signal kurtosis. The skewness value of each signal type does not show a significant difference.
Table 1. Signal’s feature extraction result in time domain
Feature (m/s2
) Sitting Standing Laying Walking Running Earthquake
meanX 1.65 1.41 1.88 0.80 0.97 0.04
meanY 2.77 1.77 2.31 2.98 3.39 0.10
meanZ 1.88 1.97 2.58 0.82 0.98 0.01
median 1.68 1.41 1.87 0.83 1.17 0.04
median 2.83 1.79 2.39 2.97 4.49 0.10
medianZ 1.95 1.95 2.61 0.93 1.25 0.01
madX 0.93 0.78 1.20 1.52 3.74 0.02
madY 1.26 0.90 1.33 2.70 7.61 0.04
madZ 1.16 1.02 1.32 1.67 3.60 0.02
rmsX 2.17 1.84 2.58 2.26 5.13 0.05
rmsY 3.41 2.25 3.03 4.97 10.48 0.12
rmsZ 2.57 2.49 3.28 2.51 5.29 0.02
stdevX 1.15 0.97 1.54 2.03 4.98 0.02
stdevY 1.53 1.11 1.69 3.52 9.43 0.04
stdevZ 1.42 1.26 1.68 2.26 5.11 0.02
maxX 2.77 2.17 3.80 5.19 13.04 0.05
maxY 4.00 2.55 4.00 8.23 19.55 0.13
maxZ 3.34 3.10 4.30 6.22 16.14 0.02
minX -2.42 -2.19 -3.24 -5.20 -13.07 -0.02
minY -2.94 -2.38 -3.32 -7.93 -19.03 -0.08
minZ -2.81 -2.40 -3.56 -6.12 -15.40 -0.02
SMA 728.42 582.28 779.85 776.74 1640.87 18.87
skewnessX -0.06 0.01 0.06 0.01 -0.01 -0.06
skewnessY 0.11 0.05 0.03 0.02 0.01 -0.11
skewnessZ 0.06 -0.06 0.04 -0.04 0.10 -0.02
kurtosisX 3.37 3.15 3.71 4.19 4.20 1.19
kurtosisY 2.94 3.33 3.46 3.98 3.37 1.27
kurtosisZ 3.41 3.40 3.62 4.88 6.12 1.10
Table 2 shows that the activity signal energy is much greater compared to the earthquake signal
energy, while the dominant frequency of earthquake signal tends to be greater than the dominant frequency of
non-locomotor activity signal. Running and running activity signals hold more unique information content than
other activity signalsdue to their high entropy values, therefore it is more easily identified. Earthquake signals
also demonstrate quite different feature values compared to human activity signals. Prominent features to
classify human activity signals and earthquake signals include maximum amplitude, high school values,
dominant frequency, spectral centroids and signal energy.

 ISSN: 1693-6930
2510
Table 2. Signal’s feature extraction result in frequency domain
Frequency Feature Sitting Standing Laying Walking Running Earthquake
fdomX (Hz) 1.11 1.15 1.28 2.97 4.17 2.36
fdomY (Hz) 1.23 1.49 1.39 2.34 3.02 2.30
fdomZ (Hz) 1.00 1.01 1.21 5.33 12.24 2.70
energyX (kJ) 78.98 56.90 93.55 64.57 343.12 0.48
energyY (kJ) 153.25 72.85 120.60 333.08 1240.18 3.86
energyZ (kJ) 85.44 83.50 140.17 79.12 345.94 0.02
centroidX (Hz) 5.63 5.55 6.86 9.36 11.56 41.69
centroidY (Hz) 4.04 4.46 6.01 7.16 8.75 39.98
centroidZ (Hz) 5.07 4.86 6.56 10.18 14.26 44.73
entropyX (bits) 1.67 1.89 1.51 2.21 1.66 1.21
entropyY (bits) 1.04 1.56 1.17 1.37 1.16 1.26
entropyZ (bits) 1.41 1.36 0.98 2.12 1.68 1.17
meanfreqX (Hz) 1.70 1.68 2.19 4.33 5.93 45.12
meanfreqY (Hz) 1.18 1.48 1.89 2.86 3.81 43.75
meanfreqZ (Hz) 1.61 1.42 1.93 5.01 8.94 47.09
3.1. Signal classification using decision tree (DT) algorithm
The decision tree classification is built based on a decision tree scheme. The types of DT algorithms
used include simple DT, medium DT and complex DT. The difference in the three types of DT is based on
the number of branches/decision tree nodes. The number of DT medium nodes is greater than that of simple DT,
while the number of DT complex nodes is greater than the number of DT medium nodes. The Simple DT decision
tree has 7 nodes. The signal feature with the highest information gain (IG) is the mean absolute deviation on
the z axis. This feature is used by the DT algorithm to classify earthquake signals and human activity signals.
Other features used are spectral power and mean absolute deviation on the y axis. The simple DT algorithm is
unable to classify signals of sitting, standing and lying activity due to the limited number of nodes.
The Medium DT decision tree has 34 nodes. Features with high IG values used by the medium DT
include mean absolute deviation, spectral power, spectral centroids, root mean square, mean frequency,
standard deviation and dominant frequency. DT medium algorithm is able to classify almost all HAR data,
except signal from sitting activity. Sitting activity signal cannot be classified properly because its features have
a small IG value, therefore it is classified as other activity signals. The complex DT decision tree has
193 nodes. All features are used in the complex DT algorithm. These features can classify signals of sitting,
standing, lying, walking, running and earthquake signals. All signal features, both time and frequency domain
features, are still used in the complex DT, hence this algorithm is more complex compared to other DT
algorithms and this algorighm is able to map nodes with a small IG.
Table 3 shows the results of testing the simple DT, medium DT and complex DT. The complex DT
algorithm has higher level of classification accuracy compared to the medium DT, both when the smartphone
is in shirt pocket and trouser pocket, with accuracy values of 84.36% and 83.49%, respectively. This proves
that the number of decision tree nodesis directly proportional to the accuracy of the DT algorithm. An increase
in the number of decision tree nodes can increase the classification ability of the DT algorithm.
Table 3. The performance of DT algorithm
Smartphone Position Type of DT True (TP+TF) False (FP+FN) Accuracy (%)
Trouser Pocket Simple DT 1753 792 68.88
Medium DT 1877 668 73.75
Complex DT 2147 398 84.36
Shirt Pocket Simple DT 1534 896 63.13
Medium DT 1696 734 69.79
Complex DT 2029 401 83.49
3.2. Signal classification using linear discriminant analysis (LDA) algorithm
LDA algorithm classifies signals using vectors in the form of linear lines. Table 4 shows the accuracy
of the LDA algorithm. The accuracy of this algorithm is lesser than the accuracy of the decision tree algorithm.
Walking, running and earthquake signals have low predictive error rates in the LDA algorithm. Figure 2 shows
the cluster of signal data based on features with the highest IG (information gain) value. These features include
mean absolute deviation on the z axis and y axis and spectral power on the y axis. Figure 2 demonstrates that
the three types of signals can be separated using linear lines vector to be classified properly. Sitting, standing and
lying activity signals have overlapping clusters. Therefore, if these activities are plotted based on features with
the highest IG, it will be difficult to separate using linear line vectors. This causes the signal of sitting, standing
and lying activity cannot be classified properly by the LDA algorithm. Figure 3 shows the proof of such statement.

2511
Table 4. The performance of LDA algorithm
Smartphone Position True (TP+TF) False (FP+FN) Accuracy (%)
Trouser Pocket 1900 645 74.67
Shirt Pocket 1745 685 71.81
Figure 2. Cluster data of locomotor activity signal and earthquake signal
Figure 3. Cluster data of nonlocomotor activity signal
3.3. Signal classification using support vector machine (SVM) algorithm
SVM algorithm uses hyperplane to classify signal classes. Types of SVM algorithms tested include linear
SVM, quadratic SVM, cubic SVM, fine SVM, medium SVM and coarse SVM. These types are obtained based on

 ISSN: 1693-6930
2512
different types of hyperplane. Quadratic SVM and cubic SVM use polynomial kernels. This kernel projects
hyperplane fields with nonlinear functions in three-dimensional space. Quadratic SVM uses polynomial power of 2,
while cubic SVM has polynomial power of 3. The classification results of SVM algorithm with polynomial kernel
report good accuracy, which is higher than 80%. Polynomial kernels are designed for nonlinear data and made
hyperplane to be flexible in separating between classes of activity signal data, therefore the classification accuracy
is relatively high.
Fine SVM, medium SVM and coarse SVM use gaussian kernels with tiered kernel scale values. Kernel
scale type fine SVM is 2.9; Medium SVM type is 9.1; and Coarse SVM type is 36. Kernel scale is a multiplier factor
in the Gaussian kernel. An increase in the value of the kernel scale generally causes a decrease in the accuracy of
the Gaussian type SVM algorithm as shown in Table 5, both for the location in the trouser pocket and shirt pocket.
The increase in kernel scale decreses the hyperplane margins, therefore the classification between classes of activity
data becomes more ambiguous. The confusion in classification results in suboptimal accuracy.
Linear SVM does not utilize the kernel because it considers the activity signal data as linearly correlated
data. SVM Linear algorithm is formed in the form of linear vector in two-dimensional space. Activity signal data is
nonlinear data, therefore the results of linear SVM classification show lower accuracy compared to other types of
SVM algorithm, except Coarse SVM. Figure 4 shows a cluster of activity signal data in the time domain feature.
Figure 5 shows a cluster of activity signal data in the frequency domain feature. Both figures prove that activity
signal data is difficult to classify using linear functions, because they overlap in the same plot area.
Table 5. The performance of SVM algorithm
Smartphone Position Type of SVM True (TP+TF) False (FP+FN) Accuracy (%)
Trouser Pocket Linear 1887 658 74.14
Quadratic 2435 110 95.67
Cubic 2233 312 87.74
Fine 2367 181 92.89
Medium 2084 496 80.77
Coarse 1785 759 70.16
Shirt Pocket Linear 1727 703 71.07
Quadratic 2168 262 89.22
Cubic 2380 50 97.94
Fine 2232 198 91.85
Medium 1878 552 77.28
Coarse 1588 842 65.35
Figure 4. Cluster data of activity acceleration signal in time domain

2513
Figure 5. Cluster data of activity acceleration signal in frequency domain
3.4. Signal classification using K-nearest neighbor (K-NN) algorithm
Types of K-NN algorithms tested include fine K-NN, Medium K-NN, Coarse K-NN, Cosine K-NN
and Cubic K-NN. The fine K-NN, Medium K-NN and Coarse K-NN algorithms are distinguished based on
the number of nearest neighbors, while Cosine K-NN and Cubic K-NN are distinguished according to
the provisions of the calculation of the distance to the nearest neighbor. Table 6 shows the results of testing
the signal data classification using the K-NN algorithm. Fine K-NN only considers 1 nearest neighbor to
classify data. Medium K-NN considers the 10 closest neighbors, while Coarse K-NN considers 100 closest
neighbors in a data cluster. Table 6 demonstrates that the increase of the number of nearest neighbors as a
consideration of the K-NN algorithm actually causes a decrease in the level of classification accuracy.
The number of nearest neighbors as predictors actually adds to the level of algorithm ambiguity in determining
the signal input class. Nonlinear HAR data has a cluster area with overlapping data classes, therefore it is
heterogeneous. The criteria to increase the number of nearest neighbors expands the coverage of the cluster
area, causing data classes to become more heterogeneous, and resulting in confusion over the results of
classification predictions. The Cosine K-NN algorithm uses cosine metric distance to determine the distance
to the nearest neighbor, while Cubic K-NN employs cubic metric distance. The accuracy of these two distance
rules are not lower than the Euclidean distance used by fine K-NN and Medium K-NN to classify signals.
Cosine K-NN and Cubic K-NN accuracy levels are lower than fine K-NN and Medium K-NN.
Table 6. The performance of K-NN algorithm
Smartphone Position Type of K-NN True (TP+TF) False (FP+FN) Accuracy (%)
Trouser Pocket 2543 2 99.92 99,92
1936 609 76.07 76,07
1742 803 68.45 68,45
1911 634 75.09 75,09
1903 642 74.77 74,77
Shirt Pocket 2420 10 99.59 99,59
1781 649 73.29 73,29
1631 799 67.12 67,12
1771 659 72.88 72,88
1729 701 71.15 71,15
3.5. Algorithm performance comparation result
Each algorithm has a different level of classification accuracy. The algorithm selection criteria do not
only include the level of accuracy, but also the speed of data processing and memory capacity. Table 7 shows

 ISSN: 1693-6930
2514
a comparison of all classifier algorithms based on these criteria. The accuracy of the algorithm is obtained from
the average accuracy of the smartphone activity signal classification in trouser pocket and shirt pocket.
Table 7. The performance comparison of the algorithms
Algorithm Accuracy (%) Processing Time (seconds) Memory Capacity (kilobytes)
Simple DT 66.01 5.17 12
Medium DT 71.77 5.27 22
Complex DT 83.93 6.91 54
LDA 73.10 4.82 387
Linear SVM 72.61 14.12 6,529
Quadratic SVM 92.44 21.97 21,213
Cubic SVM 92.84 25.57 16,564
Fine SVM 92.37 81.2 65,411
Medium SVM 79.03 32.05 29,087
Coarse SVM 67.76 33.63 29,641
Fine K-NN 99.75 15.93 6,044
Medium K-NN 74.70 13.36 6,044
Coarse K-NN 67.78 15.9 6,044
Cosine K-NN 73.98 13.24 6,044
Cubic K-NN 72.96 428.24 6,045
DT algorithm has fast processing time with the smallest memory capacity, but the level of accuracy is
lower compared to the fine K-NN algorithm and some types of SVM algorithms. The processing time and memory
capacity of the DT algorithm increases proportionally with the number of nodes and the level of complexity of
the decision tree it builds. The LDA algorithm has the fastest processing time with a small memory capacity, but
the level of accuracy is lower than the average accuracy of DT, SVM and K-NN algorithms. SVM algorithm has
the largest memory capacity with a longer processing time compared to the DT algorithm, LDA and some K-NN.
The design of hyperplane in the SVM algorithm produces greater number of variables compared to the design of
the decision tree on DT, linear lines on the LDA and calculation of the nearest neighbor distance on K-NN.
The number of variables affects the processing time and memory capacity of the algorithm. The K-NN algorithm
has a processing time of 13 to 16 seconds, except cubic K-NN. The calculation of cubic metrix distance is more
complex than Euclidean distance or cosine metrix distance, thus cubic-KNN requires long processing time.
The average memory capacity of the K-NN algorithm is 6,044 kilobytes. This capacity is far greater compared to
DT and LDA, but smaller than SVM. The fine-KNN algorithm has the highest accuracy value, medium processing
time and medium memory capacity.
4. CONCLUSION
Fine K-NN algorithm has the highest level of accuracy in the classification of human activity signals
with lower memory capacity and shorter processing time. The number of fine K-NN's closest neighbors is only
one, therefore it does not cause ambiguity in the classification process. The fine K-NN algorithm has
an accuracy rate of 99.75% in the classification of human activity signals and earthquake signals with a memory
capacity of 6,044 KB and a processing time of 15.93 seconds. This algorithm has the best classifier criteria
compared to DT, SVM and LDA algorithms.
REFERENCES
[1] R. Harris, J. Major, “Waves of destruction in the East Indies: the Wichmann catalogue of earthquakes and tsunami
in the Indonesian region from 1538 to 1877,” Geol Soc London, vol. 441, no. 1, pp. 9-46, 2017
[2] J. F. Di Leo, J. Wookey, J. O. S. Hammond, et al., “Mantle flow in regions of complex tectonics: Insights from
Indonesia,” Geochemistry, Geophys Geosystems, vol. 13, no. 12, 2012.
[3] T. Lay, H. Kanamori, C. J. Ammon, et al., “Depth‐varying rupture properties of subduction zone megathrust faults,”
J Geophys Res Solid Earth, vol. 117, no. B4, 2012.
[4] T. M. Rasyif, S. Kato, “Development of accurate tsunami estimated times of arrival for tsunami-prone cities in Aceh,
Indonesia,” Int J Disaster Risk Reduct., vol. 14, pp. 403-410, 2015.
[5] BMKG, “Law 31 Year 2009 of Indonesian Republic,” Jakarta: Sekretariat Negara, 2009.
[6] M. D. Kohler, T. H. Heaton, M. H. Cheng, “The Community Seismic Network and Quake-Catcher Network: Enabling
structural health monitoring through instrumentation by community participants,” Sensors and Smart Structures
Technologies for Civil, Mechanical, and Aerospace Systems 2013, International Society for Optics and Photonics,
pp. 86923X, 2013.

2515
[7] J. Han, E. Owusu, L. T. Nguyen, A. Perrig, J. Zhang, Accomplice: Location inference using accelerometers on
smartphones,” 2012 Fourth International Conference on Communication Systems and Networks (COMSNETS 2012),
pp. 1-9, 2012.
[8] D. Anguita, A. Ghio, L. Oneto, X. Parra, and J. Ortiz, “A Public Domain Dataset for Human Activity Recognition
Using Smartphones,” ESANN 2013 proceedings, European Symposium on Artificial Neural Networks,
Computational Intelligence and Machine Learning, pp. 437-442, 2013.
[9] Bayat, Pomplun and Tran, “A Study on Human Activity Recognition Using Accelerometer Data from Smartphones,”
The 11th International Conference on Mobile Systems and Pervasive Computing (MobiSPC-2014), pp. 450-457, 2014.
[10] Y. Chen, K. Zhong, J. Zhang, Q. Sun, and X. Zhao, “LSTM Networks for Mobile Human Activity Recognition,”
International Conference on Artificial Intelligence: Technologies and Applications (ICAITA 2016), pp. 50-53. 2016:
[11] G. M. Guzman-Acevedo, G. E. Vazquez-Becerra, J. R. Millan-Almaraz, H. E. Rodriguez-Lozoya, A. Reyes-Salazar,
J. R. Gaxiola-Camacho, et al., “GPS, Accelerometer, and Smartphone Fused Smart Sensor for SHM on Real-Scale
Bridges,” Advances in Civil Engineering, vol. 2019, pp. 1-15, 2019.
[12] A. Khan, M. Siddiqi, S. W. Lee, “Exploratory data analysis of acceleration signals to select light-weight and accurate
features for real-time activity recognition on smartphones,” Sensors, vol. 13, no. 10, pp. 13099-13122, 2013.
[13] M. Arif, M. Bilal, A. Kattan, S. I. Ahamed, “Better physical activity classification using smartphone acceleration
sensor,” J. Med. Syst., vol. 38, no. 9, pp. 95, 2014;
[14] J. Suto, S. Oniga, P. P. Sitar, “Feature Analysis to Human Activity Recognition Raw data pre-processing and
classification Feature extraction,” Int. J. Comput. Commun. Control, vol. 12, no. February, pp. 116-130, 2017.
[15] Y. Zhang, S. Lu, X. Zhou, M. Yang, L. Wu, B. Liu, et al., “Comparison of machine learning methods for stationary
wavelet entropy-based multiple sclerosis detection: decision tree, k-nearest neighbors, and support vector machine,”
Simulation, vol. 92, no. 9, pp. 861-871, 2016.
[16] M. Amarnath, V. Sugumaran, H. Kumar, “Exploiting sound signals for fault diagnosis of bearings using decision
tree,” Measurement, vol. 46, no. 3, pp. 1250-1256. 2013;
[17] R. Nathan, O. Spiegel, S. Fortmann-Roe, R. Harel, M. Wikelski, W. M. Getz, “Using tri-axial acceleration data to
identify behavioral modes of free-ranging animals: general concepts and tools illustrated for griffon vultures,” J. Exp.
Biol., vol. 215, no. 6, pp. 986-996, 2012.
[18] U. Raghavendra, P. K. Mahesh, A. Gudigar, “A novel face recognition method using pca, lda and support
vector machine,” International Conference on Computer Science and Information Technology, pp. 241-249, 2012.
[19] A. Tharwat, T. Gaber, A. Ibrahim, A. E. Hassanien, “Linear discriminant analysis: A detailed tutorial,” AI Commun.,
vol. 30, no. 2, pp. 169-190, 2017.
[20] A. A. Matkan, M. Hajeb, S. Sadeghian, “Road extraction from lidar data using support vector machine classification,”
Photogramm Eng Remote Sens., vol. 80, no. 5, pp. 409-422, 2014;
[21] H. Bhavsar, M. H. Panchal, “A review on support vector machine for data classification,” Int J Adv Res Comput Eng
Technol., vol. 1, no. 10, 2012.
[22] X. Ma, Y. Zhang, Y. Wang, “Performance evaluation of kernel functions based on grid search for support vector
regression,” 2015 IEEE 7th international conference on cybernetics and intelligent systems (CIS) and IEEE
conference on robotics, automation and mechatronics, pp. 283-288, 2015.
[23] B. Yu, X. Song, F. Guan, Z. Yang, B. Yao, “K-Nearest neighbor model for multiple-time-step prediction of
short-term traffic condition,” J Transp Eng., vol. 142, no. 6, pp. 4016018, 2016.
[24] L. H. Lee, C. H. Wan, R. Rajkumar, D. Isa, “An enhanced Support Vector Machine classification framework by using
Euclidean distance function for text document categorization,” Appl Intell., vol. 37, no. 1, pp. 80-99, 2012.
[25] J. Vitola, F. Pozo, D. Tibaduiza, M. Anaya, “A sensor data fusion system based on k-nearest neighbor pattern
classification for structural health monitoring applications,” Sensors, vol. 17, no. 2, p. 417, 2017.
BIOGRAPHIES OF AUTHORS
Hapsoro Agung Nugroho is an Assistant Professor at the Department of Instrumentation
for Meteorology, Climatology and Geophysics (MKG), School of Meteorology Climatology
and Geophysics, Jakarta, Indonesia. He received Bachelor Degree in Electrical Engineering
from Udayana University, Bali, Indonesia. He received Master Degree in Instrumentation
and Control from Bandung Institute Technology, Bandung Indonesia, in 2012. His interest
area are machine learning, intelligent instrument, and pattern recognition.

 ISSN: 1693-6930
2516
Haryas Subyantara Wicaksana, is a data analist in Agency for Meteorology, Climatology
and Geophysics, Jakarta, Indonesia. He received Bachelor Degree in Department of
Instrumentation for Meteorology, Climatology and Geophysics (MKG), School of
Meteorology Climatology and Geophysics, Jakarta, Indonesia, in 2019. His interest area are
human activity recognition, intelligent instrument, and machine learning.
Hariyanto is an Assistant Professor at the Department of Instrumentation for Meteorology,
Climatology and Geophysics (MKG), School of Meteorology Climatology and Geophysics,
Jakarta, Indonesia. He received Bachelor Degree in Electrical Engineering from Institute of
Technology Sepuluh November, Surabaya, Indonesia. He received Master Degree in
Instrumentation and Control from Bandung Institute Technology, Bandung Indonesia, in
2014. His interest area are control system, meteorology instrument, and signal recognition.
Rista Hernadi Virgianto is an Assistant Professor at the Department of Climatology,
School of Meteorology Climatology and Geophysics, Jakarta, Indonesia. He received
Bachelor Degree in Climatology from School of Meteorology Climatology and Geophysics,
Jakarta, Indonesia. He received Master Degree in Applied Meteorology and Climatology
from University of Birmingham, England, in 2016. His interest area are data science, air
quality instrument, and deep learning.

Algorithm performance comparison for earthquake signal recognition on smartphone’s accelerometer

More Related Content

What's hot (18)

Similar to Algorithm performance comparison for earthquake signal recognition on smartphone’s accelerometer (20)

More from TELKOMNIKA JOURNAL (20)

Recently uploaded (20)

Algorithm performance comparison for earthquake signal recognition on smartphone’s accelerometer