Content uploaded by Amirhossein Moshrefi
Author content
All content in this area was uploaded by Amirhossein Moshrefi on Jul 12, 2018
Content may be subject to copyright.
Abstract— Knock is one of the undesirable phenomenon, which
leads to many defects in internal combustion engine. Usually this
effect creates vibrations in engine block, which can be detect by
knock sensor. However, because of high level of noise with the
knock signal, detection process is so difficult. Nowadays, wavelet
decomposition is used as a powerful method for noise reduction in
signal processing. In this study, by employment of wavelet and
adaptive filter denoising, detection accuracy of this effect has been
improved. This method is applied to real knock signal, which
shown superiority compare to previous works. In addition, due to
use of one sensor, it is so economical.
Index Terms— Knock Sensor, Wavelet Decomposition,
Adaptive Filter, Noise Reduction, Internal Combustion Engine.
I. INTRODUCTION
Auto ignition of air and fuel mixture in combustion chamber
due to the increased local temperature and pressure causes the
knock phenomenon [1]. The occurrence of knock has direct
effect on the output power, heat efficiency and gasoline engine
lifetime, because it limits the compression ratio and volumetric
efficiency and damages engine by creating intensive pressure
waves [2]. When spark plug ignites, pressure and temperature
will be increased in the end of chamber region, air and fuel
mixture will be ignited simultaneously and led to pressure
oscillation in chamber with high amplitude, in a short time and
non-uniform ignition. This event can be identified with several
methods which divide to direct and indirect categories. Direct
method, employs internal cylinder parameters and indirect
method is based on external vibrations. By employing of knock
sensor, oscillations will be measured. Knock sensor is cheaper
and one sensor can be used for all cylinders. However level of
noise polluted signal in measurement is higher than cylinder
pressure signal [3-4]. Engine noises contain white Gaussian
background noise that is generated by the totality of machine
events and color noises generated from combustion process,
valve opening, or closing events [5-6]. In references [7-8] are
applied some filtration approaches to enhance the knock
detection on the cylinder pressure sensor. However, these
methods have some drawbacks. The first, there is no constant
bandwidth for the engine noise and we cannot consider specific
frequency region for it, also when noise and signal blend
together, this approach cannot be efficient.
References [9-10] have proposed a probability function and
noise spectrum estimation. In addition, references [11-13] have
introduced some models for the knock event prediction.
However, these results for the cylinder pressure signal are
achieved which is not currently practical in car industries.
References [14-16] have employed Short Time Fourier
Transform (STFT) and Fast Fourier transform (FFT) for knock
event detection but these approaches for knock signal, which
does not have periodic and constant treatment, are not efficient.
References [17-18] for enhancement of knock detection, have
used empirical mode decomposition (EMD) method. However,
in this method, by frequency division to multilevel, we cannot
consider some levels as the pure noise and generalize it for all
decomposed signals, because by elimination of each level,
some parts of data will be removed and this issue occurs
especially for non-stationary signals such as knock signal.
In addition, in low frequencies, the signal with more levels and
more detail will be decomposed than that in high frequencies
and there is part of significant data in high frequencies in knock
signal [19], and decomposition levels cannot distinguish noise
and signal in high frequency explicitly.
References [19-20] have used Discrete Wavelet Transform
(DWT) for increasing the accuracy of the knock detection and
have shown maximum summation value of the first and second
levels of this decomposition have an acceptable correlation for
cylinder pressure data. In addition, references [21-22] also have
used the soft threshold in DWT to eliminate the knock signal
noise. In major of references, artificial knock signal has been
used.
For signal de-nosing by single sensor and by Wiener filter
method, references [23-25] have estimated the noise for audio
signals.
In knock calibration application, we must determine values as a
reference, which requires a powerful and appropriate method
for improvement of the knock detection accuracy.
In this paper, the improvement of knock sensing has been
studied by employment wavelet denoising and adaptive filter
methods. The proposed method has been evaluated for actual
knock signal. This method led to the improvement of the actual
signal compared to the previous state of the art methods.
Structure of this paper is organized as follows: Section I
introduces the knock and also mentions some methods for better
understanding of this phenomenon. In Section II, a brief
explanation of wavelet and adaptive filter methods for noise
reduction and the main idea in this paper have been presented.
In Section III, the presented method is evaluated with the actual
signal recorded from an engine. Section IV concludes the paper.
Amirhossein Moshrefi, Omid Shoaei
School of Electrical and Computer Engineering, University of Tehran, Tehran, Iran
a.moshrefi@ut.ac.ir
Improved Knock Detection Method Employing
Wavelet and Adaptive Filter Based on Vibration
Sensor Data
II. METHOD
A. Discrete Wavelet Transform (DWT)
Discrete wavelet is a filter bank in which signal passes
through a high pass filter sequentially and then it is down
sampled. Consequently, detail coefficients are derived.
Furthermore, initial signal passes through low pass filters and
is down sampled leading to approximated coefficients. The
decomposition performed by low pass and high pass filters
depends on the type of the mother wavelet function [19].
Wavelet denoising is based on thresholding, which is called
hard thresholding in case of linear elimination of the sub-
threshold data and soft thresholding in case of nonlinear
elimination of sub-threshold data.
B. Adaptive Filter
An adaptive filter determines the weights for accurate
estimation of the output signal by receiving feedback from its
output signal. A structure of the adaptive filter for noise
reduction is demonstrated in Fig. 1. Two inputs are used for
denoising. The first one is the primary signal which is actually
the noise contaminated signal should enter it and the second one
is a reference signal which has correlation with the noise. The
reference input passes through the filter and the result (i.e. the
estimated noise) is subtracted from the primary signal such that
the output of this system is the clean desired signal. The filter
coefficients are estimated adaptively from the denoised output
signal.
Considering s, as the clean signal, n as the noise,
0
ˆ
n
as the
estimated noise and
ˆ
y
as the denoised output. The following
equation can be obtained:
1
ˆ
ˆ
y s n n
(1)
The above equation can be formulated in the form of energy
(E):
1)
ˆ
ˆ
E( y s ) E ( n n
(2)
This means that, decreasing the energy of the difference
between the real and the estimated noise, results in more
effective extraction of the clean signal.
One of the most appropriate algorithms in the adaptive
filtration is Least Mean Square (LMS) method in which the
weights of the filter are determined based on calculation of the
root-mean-square error in (3) for which its relations are given
as follows.
0k 1 k k . k
ˆ
ˆ
W W 2.y .n
(3)
T0
k
1k
ˆˆ
n W n
(4)
k k k 1k
ˆ
ˆ
y s n n
(5)
Where Wk is the weight of the filter and μ is the step size that
controls the balance between the stability of the system and the
adaptation convergence rate [26].
Non-adaptive filters require information about the signal
and noise frequency bands. On the other hand, if the noise and
signal overlap in the frequency domain, the separation with
non-adaptive filters is not possible.
The reference input in the adaptive filter can be a signal
from a sensor located near the noise source where the pure
signal is weaker. However in practice, for knock detection, it is
not proper to apply another sensor; therefore, an estimation of
the noise from the single one sensor used in the system can be
applied to the adaptive filter.
C. Proposed Method
As mentioned earlier, despite the high efficiency feature of
an adaptive filter, its application for knock signal processing
requires additional sensor, which is not available in
conventional vehicles. Therefore, we propose to use the
Wavelet transformation for estimation of the reference input of
the adaptive filter. As shown in Fig. 2, in this method, first the
signal is decomposed into different levels by Wavelet block1
(WL1 block) and then by using a threshold on each level (in T
block) the values higher than thresholds are being eliminated
which
11
ˆ
y T .WL .( s n )
then the returned signal considered
as an estimation of the noise by the decomposition Wavelet
block1 (WL1-1 block) and applied to the adaptive filter block.
On the other hand, signals are decomposed to different levels
before entering the adaptive filter block (by WL2 block) which
causes elimination of the noise of each signal section in a
separate frequency band and the accuracy of the noise
elimination increases in this method by band division. Then the
output signal of the adaptive filter is returned (by WL2-1 block)
which is an estimation of the clean signal.
Adaptive Weights
+
-
Primary
Signal
Reference
Signal
s+n
n
^0
n
^1
y
^Denoised
Signal
Figure 2. Schematic of proposed method
WL1T WL1
WL2
WL2Adaptive Filter WL2
-1
-1
S+n
S
^
y(1)
1
y(2 )
1
i
y(1)
1
y(2 )
1
i
^
^
n0
n
^0n
^0
(1) (2 )
i
y(1)
2
y(2 )
2
i
y(1)
2
y(2 )
2
i
^
^
Figure 1. Adaptive filter for noise reduction
III. RESULTS
In this section, the results of evaluation of the proposed method
are presented and we used from the actual recorded signal from
the engine.
In this work, we shed light on the result of our method for
denoising by comparing the result with two hard and soft
thresholds on the actual signal. To determine suitable
thresholds, different Sure and Universal methods were studied
and the Universal method mentioned in [27] shows best results
and therefore was applied here.
Three mother wavelet functions include of daubechies, symlet
and coiflet were studied and symlet function for WL1 and
daubechies for WL2 which demonstrates better performance
have been used with 4 coefficients and 4 decomposition levels.
Adaptive filter rank was 10 and the value of step size was 0.1.
This method can be programmed by an electrical circuit and
microcontroller on Engine Control Unit (ECU) board which
references [28-32] have analysed the circuit challenges.
Here, we study results of applying proposed method on actual
data recorded from the engine. Table I shows main
characteristics of the engine used in this test.
In this study four cylinder pressure sensors from AVL GH12D,
were used and their signals were recorded simultaneously for
four engine cylinders. Fig. 3 shows a sample of the cylinder
pressure sensor signal recorded from four cylinders in
5800(rpm) speed and 125% load. The signals are sampled at
sampling frequency of 100 kHz from these sensors. The
cylinder pressure signal is filtered in interval of 4-40 kHz from
data between 10 to 70 crank angle degrees. Then by derivation
of its rectified signal and considering peak value, the knock
intensity from the cylinder pressure signal is achieved.
Now, by considering cylinder pressure sensor as a reference for
the knock intensity, the results after applying the mentioned
method on the knock sensor signal can be compared with this
reference intensity. For this comparison, the cross correlation
(CC) criterion is used according to (6).
value of that signal, the cross correlation was calculated for four
engine cylinders and the results of these calculations are shown
in Fig. 4 for three mentioned methods including our proposed
method for the fourth cylinder. The results of cross correlation
for four engine cylinders are indicated in Table II.
Unit
Value
Description
-
EF7 turbocharged
, gasoline
Engine Model
L
1.646
Displacement
mm
78.6*85
Bore * Stroke
-
9.5
Compression Ratio
kW
110@5500 rpm
Max. torque
-
4 (in-line)
No. of cylinders
Nm
215@2500 – 4500 (rpm)
Max. power
TABLE II. Results of cross correlation between cylinder pressure intensity and
knock sensor intensity for four cylinders and for three mentioned methods
Cylinder#
Method
1
2
3
4
Average
CC
Without
Denoising
0.6212
0.6483
0.6755
0.6224
0.6419
Hard
Threshold
0.6284
0.6522
0.6986
0.6363
0.6539
Soft
Threshold
0.6371
0.6657
0.7015
0.6519
0.6641
Proposed
Method
0.7123
0.7395
0.7468
0.7092
0.7269
N N N
i 1 i 1 i 1
N N N N
2 2 2 2 1/ 2
i 1 i 1 i 1 i 1
N x (i) y(i) x (i) y (i)
CC
[(N x (i) ( x (i) ) )* (N y (i) ( y (i) ) )]
(6)
Figure 4. Result of cross correlation between cylinder pressure intensity and
knock sensor intensity for fourth cylinder a) without denoising b) hard
threshold denoising c) soft threshold denoising d) proposed method denoising
Where x and y are knock intensity and N is cycle number which
in this study, the evaluation is done for 2000 cycles.
By applying the proposed method on the knock sensor signal
and determining the knock intensity based on the peak absolute
Figure 3. Pressure sensor signal for four cylinders in 5800(rpm) and 125% load
TABLE I. Engine characteristics
It can be observed that the proposed method has improved the
knock detection accuracy by about 13.2% (ratio of 0.7269 to
0.6419) compared to the original signal and 9.5% (0.7269 to
0.6641) compared to the soft thresholding method.
This introduced method can be used in the knock intensity
calibration for better tuning in the knock IC parameters or as an
algorithm in microcontroller in the ECU board.
IV. CONCLUSION
For improvement of knock signal, an effective method for noise
reduction in vibration sensor data has been presented. In this
method by applying wavelet denoising technique and adaptive
filtration, the noise can be reduced by making use of a single
sensor as a source, which is efficient yet economical. This
method was evaluated on the actual knock signal, which shown
superiority compare to previous works and led to improvement
of the knock detection accuracy, by up to 13.2% on this basis.
It can be used for enhancement of knock detection in ECU
board.
IV. REFERENCES
[1] J.B. Heywood, “Internal combustion engine fundamentals,” McGraw-
Hill Book Company, New York, 1988.
[2] M. Castagne, J. Dumas, S. Henriot, F. Lafossas, and T. Mazoyer, “New
knock localization methodology for SI engines,” SAE transactions
112.3: 1584-1594, 2003.
[3] X. Zhen, Y. Wang, S. Xu, Y. Zhu, C. Tao, T. Xu and M. Song, “The
engine knock analysis - An overview,” Applied Energy, 92:628-36,
2012.
[4] B. Iorio, V. Giglio, G. Police, and N. Rispoli, “Methods of pressure
cycle processing for engine control,” No. 2003-01-0352. SAE Technical
Paper, 2003.
[5] H.A. Noubari, and G.A. Dumont, “Towards an Improved Knock
Detection and Quantification using Wavelets and Entropy-based Noise
Compensation,” No. 2005-01-2269. SAE Technical Paper, 2005.
[6] C. Hudson, X. Gao, and R. Stone, “Knock measurement for fuel
evaluation in spark ignition engines,” Fuel 80.3: 395-407, 2001.
[7] E. Galloni, G. Fontana, and S. Staccone, “Numerical and experimental
characterization of knock occurrence in a turbo-charged spark-ignition
engine,” Energy Conversion and Management 85: 417-424, 2014.
[8] E. Galloni, “Dynamic knock detection and quantification in a spark
ignition engine by means of a pressure based method”. Energy
Conversion and Management 64 : 256-262, 2012.
[9] D. Linse, A. Kleemann, and C. Hasse, “Probability density function
approach coupled with detailed chemical kinetics for the prediction of
knock in turbocharged direct injection spark ignition engines”,
Combustion and Flame 161.4 : 997-1014, 2014.
[10] J. Spelina, J. C. Peyton Jones, and J. Frey, “Characterization of knock
intensity distributions Part 2: parametric models”, Proceedings of the
Institution of Mechanical Engineers, Part D: Journal of Automobile
Engineering : 0954407013496234, 2013.
[11] D. Siano, and F. Bozza, “Knock detection in a turbocharged SI engine
based on ARMA technique and chemical kinetics”, No. 2013-01-2510.
SAE Technical Paper, 2013.
[12] D. Kozarac, R. Tomic, I. Taritas, J. Chen and R. Dibble, “A Model for
Prediction of Knock in the Cycle Simulation by Detail Characterization
of Fuel and Temperature Stratification”, SAE International Journal of
Engines, 8.2015-01-1245: 1520-1534, 2015.
[13] Y. Sugure, S. Oho, S. Phatak and G. Saikalis, “Virtual Engine System
Prototyping with High-Resolution FFT for Digital Knock Detection
Using CPU Model-Based Hardware/Software Co-simulation”, SAE
International Journal of Passenger Cars-Electronic and Electrical
Systems 2.2009-01-0532: 177-185, 2009.
[14] K. Akimoto, H. Komatsu and A. Kurauchi, “Development of pattern
recognition knock detection system using short-time Fourier transform,”
IFAC Proceedings, 7 : 366-371, 2013.
[15] S. d’Ambrosio, A. Ferrari, and L. Galleani, “In-cylinder pressure-based
direct techniques and time frequency analysis for combustion
diagnostics in IC engines,” Energy Conversion and Management 99 :
299-312, 2015.
[16] S. Vulli, J. Dunne, R. Potenza, D. Richardson and P. King, “Time-
frequency analysis of single-point engine-block vibration measurements
for multiple excitation-event identification,” Journal of Sound and
Vibration 321.3 : 1129-1143, 2009.
[17] J. Zhang, C. Liu, F. Bi and Y. Pei, “Knock Threshold Detection in
Turbocharged Gasoline Engine Using EEMD and Bispectrum,” No.
2016-01-0643. SAE Technical Paper, 2016.
[18] F. Bi, T. Ma, and J. Zhang, “Knock Feature Extraction in Spark Ignition
Engines Using EEMD-Hilbert Transform,” No. 2016-01-0087. SAE
Technical Paper, 2016.
[19] D. Siano, F. Bozza, D. D’Agostino and M. Panza, “The Use of
Vibrational Signals for On-Board Knock Diagnostics Supported by In-
Cylinder Pressure Analyses,” No. 2014-32-0063. SAE Technical Paper,
2014.
[20] D. Siano, and D. D’Agostino, “Knock Detection in SI Engines by Using
the Discrete Wavelet Transform of the Engine Block Vibrational
Signals,” Energy Procedia 81: 673-688, 2015.
[21] D. Siano, M. Panza, and D. D'Agostino, “Knock detection based on
MAPO analysis, AR model and discrete wavelet transform applied to
the in-cylinder pressure data: results and comparison,” SAE
International Journal of Engines 8.2014-01-2547: 1-13, 2014.
[22] D. Siano, G. Valentino, F. Bozza, A. Lacobacci, and L. Marchitto, “A
Non-Linear Regression Technique to Estimate from Vibrational Engine
Data the Instantaneous In-Cylinder Pressure Peak and Related Angular
Position,” No. 2016-01-2178. SAE Technical Paper, 2016.
[23] L. Benaroya, and B. Frédéric, “Wiener based source separation with
HMM/GMM using a single sensor,” in Proc. Int. Conf. Ind. Compon.
Anal. Blind Source Separation (ICA’03), Nara, Japan, pp. 957-961,
2003.
[24] L. Benaroya, B. Frédéric, and R. Gribonval, “Audio source separation
with a single sensor,” IEEE Transactions on Audio, Speech, and
Language Processing 14.1: 191-199, 2006.
[25] B. Cornelis, M. Moonen, and J. Wouters, “Performance analysis of
multichannel Wiener filter-based noise reduction in hearing aids under
second order statistics estimation errors,” IEEE Transactions on Audio,
Speech, and Language Processing 19.5: 1368-1381, 2011.
[26] N.J. Bershad, E. Eweda, and J.C. Bermudez, “Stochastic analysis of the
LMS and NLMS algorithms for cyclostationary white Gaussian inputs,”
IEEE Transactions on Signal Processing 62.9: 2238-2249, 2014.
[27] M. Khan, R. Ashique, B. Liya, Md. Sajjad, Md. Rahman and M. Amin,
“New wavelet thresholding algorithm in dropping ambient noise from
underwater acoustic signals,” Journal of Electromagnetic Analysis and
Applications 7.03: 53, 2015.
[28] H. Aghababa, A. Khosropour, A. Afzali-Kusha, B. Forouzandeh, and M.
Pedram, "Statistical estimation of leakage power dissipation in nano-
scale complementary metal oxide semiconductor digital circuits using
generalised extreme value distribution," IET circuits, devices & systems
6.5 : 273-278, 2012.
[29] H. Aghababa, B. Forouzandeh and A. Afzali-Kusha, "High-performance
low-leakage regions of nano-scaled CMOS digital gates under variations
of threshold voltage and mobility," Journal of Zhejiang University
SCIENCE C 13.6 : 460-471, 2012.
[30] M. Ansari, M. Imani, H. Aghababa, and B. Forouzandeh, "Estimation of
joint probability density function of delay and leakage power with
variable skewness," 2013 International Conference on Electronics,
Computer and Computation (ICECCO), pp.251-254, 2013.
[31] H. Aghababa, N. Mozaffari, and B. Forouzandeh, "Analytical Modeling
of Parametric Yield Considering Variations in Leakage Power and
Performance of Nano-Scaled Integrated Circuits," The Seventh
Conference on Advances in System Simulation, pp.89-93, 2015.
[32] H. Aghababa and M. Kolahdouz, "A novel leakage power reduction
technique for nano-scaled CMOS digital integrated circuits," 2016 26th
International Workshop on Power and Timing Modeling, Optimization
and Simulation (PATMOS), pp.268-274, 2016.
