Applied Acoustics 113 (2016) 116–120

Contents lists available at ScienceDirect

Applied Acoustics
journal homepage: www.elsevier.com/locate/apacoust

Infrasound signal classification based on spectral entropy and support

vector machine
Mei Li a, Xueyong Liu b,⇑, Xu Liu a
a
School of Information Engineering, China University of Geosciences, Beijing, China
b
School of Humanities and Economic Management, China University of Geosciences, Beijing, China

a r t i c l e i n f o a b s t r a c t

Article history: The operation speed of the algorithm is the critical factor in the real-time monitoring of infrasound
Received 21 September 2015 signals. The existing methods mainly focus on how to improve the accuracy of classification and can’t
Received in revised form 11 January 2016 be used in real time monitoring because of their slow running speed. We adopt spectral entropy into
Accepted 21 June 2016
the feature extraction of infrasound signals. Combined with the support vector machine algorithm, the
Available online 29 June 2016
proposed method effectively extracted the signal features meanwhile greatly improved the operation
efficiency. Experimental results show that the running speed of the proposed method is 1.0 s, which is
Keywords:
far less than 4.7 s of the comparison method. So the proposed method can be applied in real-time
Feature extraction
Spectral entropy
monitoring of earthquakes, tsunamis, landslides and other infrasound events.
Infrasound signal Ó 2016 Elsevier Ltd. All rights reserved.
Support vector machines
Pattern recognition

1. Introduction type of infrasound signals, Chilo proposed a method for filtering

Many natural phenomena and human activities such as earth-
quakes, volcanic eruptions, tsunamis, landslides, explosions, light-
ning, and rocket launching bring about infrasound signals. The
frequency of infrasound signals typically ranges from 0.01 Hz to
20 Hz which is lower than that of audible sounds. Infrasound sig-
nals attenuate slowly and can propagate hundreds of kilometers.
So infrasound signal detecting is very useful in disaster warning
system, such as the Comprehensive Nuclear-Test-Ban Treaty
(CTBT) International Monitoring System (IMS). Therefore charac-
teristic extraction and recognition of infrasound signal are very
important for disaster monitoring.
A great amount of theoretical research and engineering applica-
tions have been focused on both the detection and identification of
the different characteristics of infrasound sources. An exploratory
study of feature extraction, launched by Liszka and Holmstrom
[1], performed a scale spectrum of continuous wavelet transform
on infrasound signals. Schmitter formed a feature vector by com-
bining the energy of wavelet decomposition at different scales with
kurtosis and skewness of time series signal [2]. Ham et al. proposed
a set of feature vectors by use of Mel-frequency scale cepstral coef-
ficients and their associated derivatives for each infrasound signal
which was effective to classify the infrasound signals of volcanic
eruptions and avalanches [3–6]. In an effort to classify time series
⇑ Corresponding author.
E-mail address: liuxueyong@cugb.edu.cn (X. Liu).
and extracting features from infrasound data and compared three
feature extraction techniques [7,8]. In the research of Liszka, the
spectral entropy was used for extracting the feature of infrasonic
data. The research proved that the spectral entropy feature is one
of the useful distribution parameters of relevant variables [9].
On the basis of previous research results, Liu et al. presented a
new classification method based on Hilbert–Huang transform
(HHT) and a support vector machine (SVM) [10]. This method
was used to discriminate between three different natural events
and was shown to successfully improve the classification accuracy.
However, these feature extraction methods are complex, due to the
complexity of the algorithm and long running time, it was difficult
to implement using the FPGA and other hardware. These methods
are not suitable for real-time monitoring and analysis of infra-
sound signals.
In the classification theory of SVM, the support vectors of differ-
ent categories of data are most similar. Therefore, the entropy of
the support vectors is the largest by comparison since the entropy
is greater as the information is more confusing. The best classifica-
tion interface on the feature vectors extracted by information
entropy can therefore be found by SVM algorithm. The SVM algo-
rithm was used for the detection of infrasound signals from snow
avalanche and volcanic activity, which achieved a significant
reduction of false detections [11,12].
An attempt to combine the spectral entropy theory with the
SVM algorithm was undertaken in the field of face recognition by
Ye et al. [13]. In the medical field, Mehta et al. and Rud et al.

http://dx.doi.org/10.1016/j.apacoust.2016.06.019
0003-682X/Ó 2016 Elsevier Ltd. All rights reserved.
 M. Li et al. / Applied Acoustics 113 (2016) 116–120 117

applied an entropy-based algorithm for the detection of low fre- value and K represents the number of effective singular characteris-
quency heart murmur using support vector machine [14,15]. The tic values that are used to calculate the wavelet singular entropy.
numerical results indicated that the algorithm achieved a good
detection rate. The combination of approximate entropy and sup- 2.1.2. Wavelet Power Spectrum Entropy (WPE)
port vector machines showed strong generalization ability for the The distribution of the infrasound signal energy in the fre-
electroencephalogram signal classification [16,17]. quency domain is reflected in the power spectral entropy [10]. This
In this study, we proposed a feature extraction method based on indicates that the power spectral entropy is effective to extract the
information entropy and then applied it to infrasound signal clas- features from each type of infrasound signal.
sification. We then compared the proposed method with Liu’s Suppose x(t) is the original signal and the signal power spec-
method [10] on the effect of classification and the complexity of trum S(x) can be expressed as follows:
algorithm. The results indicate that, in comparison to other meth-
ods, the proposed method is a little bit less accurate; however, the 1
SðxÞ ¼ jXðxÞj2 ð2Þ
operating speed of the classification algorithm is greatly improved. 2pN
This makes the proposed method to be a feasible method for real- In Eq. (2), X(x) represents the discrete Fourier transform of the orig-
time monitoring and analysis of infrasound signals. inal signal x(t), N is the points number of Fourier transform.
Si represents the i-th value of power spectrum. S = {S1, S2, . . ., SK}
2. Materials and methods can be seen as a type of split on the original signal power division
[23]. In the frequency domain the power spectral entropy is
In this study, a classification framework combining information defined as:
spectrum entropy [18,19] with SVM method [20,21] is proposed to
X
K
extract effective feature vectors of infrasound signals generated by Wpe ¼  pi ln pi ð3Þ
earthquakes, tsunamis and volcanic events, followed by infrasound i¼1
signal classification. The integrated classification framework pro-
posed in this study is also shown in Fig. 1. In Eq. (3), pi ¼ PSKi , it represents a proportion of the i-th power
S
i¼1 i

value in the whole power spectrum, K represents the number of

2.1. Feature extraction based on information entropy power value.

Entropy was first introduced into the information theory by

2.1.3. Wavelet Energy Spectrum Entropy (WEE)
Claude Elwood Shannon and defined as the uncertainty of a dis-
The time domain and the frequency domain represent two
crete random variable. Information entropy is a quantitative
different dimensions of a signal. In order to obtain the whole infor-
assessment indicator for the uncertainty of a signal or system sta-
mation, we use the feature indicator WEE, in a joint time–fre-
tus. It is a better reflection of the inherent information of a system
quency domain, to analyze the signals. WEE can reflect the
and can be used to extract the characteristics of a signal [22]. On
uncertainty of the feature distribution of the signals in the time–
the basis of information entropy, the research of feature extraction
frequency domain.
on three types of infrasound events was performed. The informa-
A wavelet transform on the signal of finite energy is performed
tion entropy features were extracted in different transform space.
in order to obtain wavelet energy spectrum E, then the wavelet
These extracted feature vectors can be used for infrasound signal
energy entropy can be calculated using the following Eq. (4):
classification.
X
K
Wee ¼  pi ln pi ð4Þ
2.1.1. Wavelet Singular Spectrum Entropy (WSE)
i¼1
The information measure of WSE is a type of feature extraction
method that integrates the singular value decomposition of the In the above Eq. (4), pi ¼ PEKi represents a proportion of the energy
E
wavelet transform with feature extraction based on entropy. i¼1 i

WSE has the capability to reflect the uncertainty of the infrasound of the i-th scale in the energy of all scales, K represents the decom-
signal energy distribution in the time–frequency domain. Wavelet position number of scale.
Singular Spectrum Entropy is defined by the following formula (1):
2.2. Infrasound signal classification based on support vector machine
X
K
Wse ¼  pi  ln pi ð1Þ
i¼1
Support vector machine has many unique advantages for the
classification and regression prediction in solving small sample,
where Wse represents the extracted wavelet singular spectrum nonlinear and high dimension problems [24,25]. Due to the small
entropy. pi ¼ PkKi , represents a proportion of the i-th singular amount of simulation data and high dimension of feature vectors,
k
i¼1 i
we chose SVM as a classifier.
value in the singular value spectrum. ki represents the i-th singular
The support vector machine constructs a hyperplane in a high-
or infinite-dimensional space, which can be used for classification.
Wavelet Power f ðxÞ ¼ w  x þ b: ð5Þ
Spectrum Entropy

where x is the training data, w is the normal vector of the hyper-

Infrasound Wavelet Singular Feature SVM
Data Spectrum Entropy Vectors Training plane, and b is the offset.
The problem of maximizing the distance of data points of two
Wavelet Energy classes can be expressed as a quadratic programming optimization
Entropy Output of
Results
problem. The method introduces non-negative slack variables, ni,
SVM
Classification which measure the degree of misclassification of the data xi,
i = 1, 2, . . ., N, where N is the number of training data. The objective
Fig. 1. Classification model framework proposed in this paper. function is then increased by a function which penalizes non-zero
118 M. Li et al. / Applied Acoustics 113 (2016) 116–120

ni, and the optimization becomes a tradeoff between a large margin used in this study. The details of the infrasound data collected from
and a small error penalty. different areas are shown in Table 2. Each of the 132 infrasound
! signal recordings generated by three different event types comes
1 XN
min kwk2 þ C ni from nine different monitoring stations around the world. The
W;b 2 i¼1 three different infrasound events are volcano, tsunami and earth-
quake. The sampling frequency of all 132 signal recordings is
s:t: yi ðw  xi þ bÞ P 1  ni ; i ¼ 1; 2; . . . ; N ð6Þ 20 Hz.

The constant C (>0) is a penalty factor, which balance the gen-

eralization ability and misclassification. yi is equal to 1 or 1, 3.2. Data pre-processing
which represents the different classifying result respectively.
By introducing Lagrange multipliers a, the prime problem has The infrasound data used in this research were provided by the
its dual form. The kernel K(xi, xj) deals with nonlinear IMS of CTBT. The recording work from the IMS was primarily
classification. obtained using a microbarograph or micro-microphone. Currently,
" # the most widely used sensor is the model that includes the
XN
1X N X N MB2000, MB2005 and the MB3000, which was developed by a
max ai  ai aj yi yj Kðxi ; xj Þ French agency. These sensors also collect some interference signal.
a
i¼1
2 i¼1 j¼1
In order to eliminate the noise, a fourth order Butterworth band-
8 N pass filter with cutoff frequencies ranging from 0.01 Hz and
< Xa y ¼ 0
> 10 Hz was also used to reduce the overall impact of other
i i
s:t: ð7Þ frequencies.
> i¼1
:
ai 2 ½0; C
3.3. Simulations setup
Then the resulting optimal classification is as follows:
" #
XN The proposed classification framework was built according to
f ðxÞ ¼ sgn ai yi Kðxi ; xj Þ þ b ð8Þ Fig. 1. The details of data used in this study are shown in Table 1.
i¼1 The entropy features were extracted from the recorded infrasound
The choice of kernel function is very important and different signals using the above three types feature extraction methods by
kernel leads to different SVM algorithm. Commonly used kernel information entropy. The extracted infrasound feature vectors
functions include the linear kernel, the polynomial kernel, the from the three different sources, earthquakes, tsunamis, and volca-
radial basis function (RBF) and the sigmoid kernel function. The noes, are shown in Fig. 2.
RBF kernel is commonly used because it has less parameter and Each type of feature vector was randomly divided into two
applicable in many situations, no matter how small and high groups, the training group and the testing group. The proportion
dimensions samples. The RBF kernel is defined as of the two groups of data is 2:1. The infrasound classes used for
! the training and testing are shown in Table 2. The number of sig-
jjxi  xj jj2 nals used for training is 88, and the total number of signals used
Kðxi ; xj Þ ¼ exp  ð9Þ
d2 for model testing is 44.

where d is a free parameter.

In this study, a generic SVM software package, the lib-svm tool-
box, which is developed by Professor Chih-Jen Lin was used. It is
fast, effective and easy to use. It is an open source software pack-
age. Everyone can download it for free from the author’s personal
homepage http://www.csie.ntu.edu.tw/~cjlin/.
3. Results
The following steps are the overall process for training and test-
ing the classification framework.
3.4. Simulation results
The training group data were first used to train the SVM model
in order to obtain the optimal parameters of the SVM classifier.
After model training, the residual vectors were used as a test set
in order to evaluate the performance of the SVM classifier. The
results from the final simulation results are shown in Fig. 3. The
abscissa on the graph represents the feature vectors for each test
and the ordinate represents the category of each test feature
vector; ‘‘o” represents the actual category of test feature vector;
‘‘⁄” represents the prediction category of test feature vector.

(1) Three categories of infrasound events were studied in this

study. The data were collected from 8 different infrasound Table 1
sensor arrays with different locations. Infrasound data summarya.
(2) Feature vectors, extracted from time domain and frequency
Event type (No. of Data sourceb (No. of Sampling frequency
domain, are based on Spectral Entropy [15]. signals) signals) (Hz)
(3) For training, 88 feature vectors were used and for testing, 44
Volcano (45) I48TN(24) 20
feature vectors were used. I59US(21) 20
(4) Support vector machine was used for classification. Tsunami (42) I07AU(8) 20
(5) The performance of proposed classifier was computed I30JP(14) 20
(Classification accuracy). I44RU(11) 20
I52GB(9) 20
Earthquake (45) I14CL(15) 20
3.1. Infrasound data for training and testing the proposed model I22FR(17) 20
I30JP(13) 20
The Beijing National Data Center is a branch of the CTBT global a
Data from Comprehensive Nuclear-Test-Ban Treaty Beijing National Data
infrasound monitoring data center and provided the data that were Center.
b
used in this study. 132 groups of infrasound signals in total were The station codes of IMS.
 M. Li et al. / Applied Acoustics 113 (2016) 116–120 119

Feature Extraction
2.5
Earthquake
Volcano
Tsunami
2

1.5

Amplitude
1

0.5

0
1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 2.8 3
Feature Vectors

(a) Feature vector set of three kinds of infrasound signal

Earthquake Volcano Tsunami
2.5 2.5 2.5

2 2 2

Amplitude
Amplitude

1.5
Amplitude

1.5 1.5

1 1 1

0.5 0.5 0.5

0 0 0
1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 2.8 3 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 2.8 3 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 2.8 3
Feature vectors Feature Vectors Feature Vectors
(b) Feature vector set of earthquake (c) Feature vector set of volcano (d) Feature vector set of tsunami
Fig. 2. Feature vectors extracted by spectral entropy.

Table 2
Infrasound classes used for training and testing.

Class number 1 2 3
Event Earthquake Tsunami Volcano
No. of signals (total of 132) 45 42 45
No. of signals used for training (total of 88) 30 28 30
No. of signals used for testing (total of 44) 15 14 15

Simulation results indicate that the classification accuracy of the

SVM Test for Classification
proposed method is approximately 86.36%. 3
The effectiveness of the proposed method used in this study Category of Actual
2.8 Category of Predicted
was verified by conducting a comparative test with the method
The Category of Test Feature Vector

that was used in Liu’s study [10]. Fig. 4 shows the classification 2.6
result of the method based on DWT and SVM method. The compar-
2.4
ative results compared both the classification accuracy and the
speed of the two methods, shown in Table 3. We can find in the 2.2
table that the classification accuracy is a little lower; however,
2
the operating speed is greatly improved. The faster algorithm
makes it more applicable to the real-time monitoring of infrasound 1.8
signals.
1.6

4. Conclusions 1.4

1.2
This study combined the information entropy and the SVM clas-
sification algorithm and utilized this method in the classification of 1
0 5 10 15 20 25 30 35 40 45
infrasound signal. Numerous simulation tests were carried out in
Feature Vectors for Test
order to validate the effectiveness of the proposed method. Com-
parison test shows that the classification accuracy of the proposed Fig. 3. SVM test results of classification.
120 M. Li et al. / Applied Acoustics 113 (2016) 116–120

Feature Extraction
Earthquake Tsunami
SVM Test for Classification
1 1 3
Normalized Amp

Normalized Amp
Category of Actual
0.5 0.5 2.8 Category of Predicted

The Category of Test Feature Vector

0 0 2.6

-0.5 -0.5 2.4

2.2
-1 -1
0 5 10 0 5 10
Feature Vectors Feature Vectors 2

Volcano 1.8
1
Normalized Amp

1.6
0.5

1.4
0
1.2
-0.5
1
-1 0 5 10 15 20 25 30 35 40 45
0 5 10
Feature Vectors for Test
Feature Vectors
(a) Feather vectors extracted by DWT (b) Test results of classification

Fig. 4. SVM test results of classification based on DWT.

Table 3
The comparison results of infrasound classification system.
Classification scheme Classification accuracy
Entropy combined with SVM 86.36%
DWT combined with SVM 91.00%
method in this study is 86.36%, which is a little lower than the
other, but the operating speed of the proposed algorithm is greatly
reduced. The improved operating speed is important for the real-
time monitoring of infrasound signals in engineering applications.
It also saves the response time for us to deal with the emergency
events.
The focus of this study was on the practical applications of the
infrasound classification algorithm, where the slight loss in the
accuracy is outweighed by the greatly improved operating speed
of the constructed model. Subsequent studies will attempt to fur-
ther improve upon both the accuracy and operating speed of mod-
els and algorithms used in analyzing infrasound signals.
Acknowledgements
This work was financially supported by the National Natural
Science Foundation of China (Grant No. 41572347 and Grant No.
41374185) and the Fundamental Research Funds for the Central
Universities (Excellent Scholars Fund, Grant No. 2652013112).
Many thanks for the Comprehensive Nuclear-Test-Ban Treaty
Beijing National Data Center providing the data.
References
[1] Liszka L, Holmstrom M. Extraction of a deterministic component from POSAT
X-ray data using a wavelet transform and the principal component analysis.
Astron Astrophys 1999;140(1):125–34.
[2] Schmitter ED. Characterisation and classification of natural transients. Int J
Appl Sci, Eng Technol 2008;2(1):210–3.
[3] Ham FM, Rekab K, Acharyya R, Lee YC. Infrasound signal classification using
parallel RBF Neural Networks. Int J Signal Imaging Syst Eng 2008;1:155–67.
[4] Ham FM, Acharyya R, Lee YC, Garces M, Fee D, Whitten C, et al. Classification of
infrasound surf events using parallel neural network banks. In: Int Joint Conf
Neural Networks. p. 720–5.
[5] Park SJ, Ham FM, Lowrie CG. Discrimination of infrasound events using parallel
neural network classification banks. Nonlinear Anal 2005;63:859–65.
[6] Ham FM, Rekab K, Park Sj, Acharyya R, Lee YC. Classification of infrasound
events using radial basis function neural networks. In: Int Joint Conf Neural
Networks. p. 2649–54.
Run time (s) [7] Chilo J, Jabor A, Liszka L. Filtering and extracting features from infrasound data.
In: 14th IEEE-NPSS Real-time Conference. p. 451–5.
1.017
[8] Chilo J, Lindblad T, Olsson R, Hansen SE. Comparison of three feature extraction
4.707
techniques to distinguish between different infrasound signal. In: Progress in
pattern recognition. Springer; 2007. p. 75–82 [chapter 8].
[9] Liszka L. Swedish-finnish infrasound network-the research program. The
International Scientific Studies (ISS) Conference, Vienna, Austria, 2009.
[10] Liu XY, Li M, Tang W, Wang SC, Wu X. A new classification method of
infrasound events using Hilbert-Huang transform and support vector machine.
Math Problems Eng 2014;2014:1–6 Article ID 456818.
[11] Thuering T, Schoch M, Herwijnen A, Schweizer J. Robust snow avalanche
detection using supervised machine learning with infrasonic sensor arrays.
Cold Reg Sci Technol 2015;111:60–6.
[12] Cannata A, Montalto P, Aliotta M, Cassisi C, Pulvirenti A, Privitera E, et al.
Clustering and classification of infrasonic events at Mount Etna using pattern
recognition techniques. Geophys J Int 2011;185(1):253–64.
[13] Ye N, Wang D, Dou LJ. Relationship between information entropy and support
vector. J Guangxi Normal Univ: Nat Sci Ed 2006;24(4):127–30 [in Chinese].
[14] Mehta SS, Lingayat NS. Development of entropy based algorithm for cardiac
beat detection in 12-lead electrocardiogram. Signal Process 2007;87
(12):3190–201.
[15] Rud S, Yang JS. A support vector machine (SVM) classification approach to
heart murmur detection. In: 7th international symposium on neural networks,
Shanghai, PEOPLES R CHINA.
[16] Zhang Z, Zhou Y, Chen Z, Tian XH, Du SH, Huang RM. Approximate entropy and
support vector machines for electroencephalogram signal classification.
Neural Regen Res 2013;8(20):1844–52.
[17] Hsu WY. Assembling a multi-feature EEG classifier for left-right motor
imagery data using wavelet-based fuzzy approximate entropy for improved
accuracy. Int J Neural Syst 2015;25(8).
[18] Yang CC, Su FC, Guo LY. Comparison of neck movement smoothness between
patients with mechanical neck disorder and healthy volunteers using the
spectral entropy method. Eur Spine J 2014;23(8):1743–8.
[19] Sang YF, Wang Zg, Liu CM. Period identification in hydrologic time series using
empirical mode decomposition and maximum entropy spectral analysis. J
Hydrol 2012;424–425:154–64.
[20] Osuna E, Freund R, Girosi F. Training support vector machines: an application
to face detection. In: IEEE Comput Soc Conf Comput Vision Pattern Recognit. p.
130–6.
[21] Wang RH. AdaBoost for feature selection, classification and its relation with
SVM, a review. Phys Procedia 2012;25:800–7.
[22] Pan MZ, Pan HX, Ren HF. Fault feature extraction research of automatic
mechanism based on wavelet transform information entropy. J Launch Control
2012;4:74–8 [in Chinese].
[23] Liu T, Xu RL. Fault diagnosis of rolling acoustic emission signal based on
information entropy. Mach Manuf 2012;50(576):64–7.
[24] Cortes C, Vapnik V. Support—vector networks. Machine Learning 1995;20
(3):273–97.
[25] Li S, Zhou W, Yuan Q, Geng S, Cai D. Feature extraction and recognition of ictal
EEG using EMD and SVM. Comput Biol Med 2013;43(7):807–16.