Applied Acoustics 143 (2019) 48–58

Contents lists available at ScienceDirect

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

Fault detection of injectors in diesel engines using vibration

time-frequency analysis
Ahmad Taghizadeh-Alisaraei a,⇑, Alireza Mahdavian b
a
Department of Biosystems Engineering, Gorgan University of Agricultural Sciences and Natural Resources, P. O. Box 386, Gorgan, Iran
b
Department of Biosystems Engineering, Tarbiat Modares University, P. O. Box 14115-336, Tehran, Iran

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

Article history: In CI engines, injector spraying into the combustion chamber has extreme importance in fuel atomization
Received 12 December 2017 and knocking control in the CI engines. Knocking and malfunction of engines due to faulty injectors can
Received in revised form 30 July 2018 lead to efficiency reduction, damages, and acoustic noise. Much research is developing methods of engine
Accepted 4 September 2018
knock detection. Hence, injector fault detection has not been addressed specifically, this research is
Available online 8 September 2018
focused on the subject and corresponding vibration amplitudes and frequencies, likely to cause the knock
phenomenon. Welch test, Short-Term Fourier Transform (STFT), Wigner-Ville Distribution (WVD), and
Keywords:
Choi-Williams Distribution (CWD) were employed for detailed scrutiny of vibrations generated by an
Vibration
Injector
under-load engine. For an ideal combustion, the acceleration peak values should be placed in the range
TFR analysis of 0–10 kHz in time-frequency (TFR) diagram. While a faulty injection unit can cause components at
Fault diagnosis higher frequency, between 10 and 25 kHz, in TFR diagram for each cylinder, and this can effects on the
Knocking engine performance. Regarding the results which are presented in this research it infers that, in real-
time performance monitoring of an engine, the STFT technique is more efficient for fault diagnosis of fuel
injection nozzles and knock detection. By comparing vibration response of healthy and faulty injectors,
the RMS and kurtosis of the faulty injectors showed an increase of 12.9% and 20.6% respectively.
Ó 2018 Elsevier Ltd. All rights reserved.

1. Introduction rotating machines, via mounted accelerometers on bearing pedes-

Knocking-combustion studies is important due to engine dura-
bility, fuel consumption, and power density, as well as noise and
emission performance [1]. Combustion pressure changing and
knock are the main source of engine vibration and noise. In spark
ignition (SI) engines, ignition of fuel mixtures farther from the
flame front causes knocking due to rapidly releasing an uncon-
trolled energy. This causes a rapid and irregular increase in gas
pressure, which develops knocking. In compression ignition (CI)
engines, by fuel injection into the combustion chamber at the
high-compression ratio, the fuel mixture ignites with a delay
[2,3]. Current SI engines suffer from conventional knock and
super-knock [1]. Compared with SI engines, knocking is more rig-
orous in CI engines when the ignition delaying increases [2,3]. In
this case, the knocking generates annoying noise and vibration,
burns the piston crown, and damages engine parts [4].
Vibration-based condition monitoring (VCM) has applied suc-
cessfully for detection and differentiation of faults generated in
⇑ Corresponding author.
E-mail addresses: ahmadtza@yahoo.com (A. Taghizadeh-Alisaraei), a.mahdavian@
modares.ac.ir (A. Mahdavian).
tals [5]. Vibration or acoustic signals derived from reciprocating
engines are cyclostationarity. A cyclostationary signal is one that
exhibits some hidden periodicity of its energy flow along with sta-
tionary random signals [6]. Impact excitations, time-varying trans-
fer properties and non-stationary random response are typical
characteristics of the reciprocating engine vibration. These charac-
teristics of the engine vibration make its dynamic analysis and sig-
nature extraction much more difficult than that of the rotational
machinery [7]. Induced mechanical vibrations of engines can inter-
fere with the knocking detection process. For example, valves clo-
sure impact often produces high-frequency vibrations that fall
inside the knocking detection window in the signal processing.
The valve closure impacts depend on their clearance, which control
their upraising speed over the camshaft. Thus, these impacts take
place at different intensities, and therefore, can disturb knocking
detection process [8].
There are several methods for capturing information on engine
knocking by instruments. Measuring the combustion pressure
inside a cylinder is one of the most reliable methods, since other
sources of mechanical vibrations do not affect this method. Never-
theless, due to its difficult implementation and high costs, mostly
accelerometers are used for knocking detection [8–10]. To do so,

https://doi.org/10.1016/j.apacoust.2018.09.002
0003-682X/Ó 2018 Elsevier Ltd. All rights reserved.
 A. Taghizadeh-Alisaraei, A. Mahdavian / Applied Acoustics 143 (2019) 48–58
signals were converted using various methods such as FT (Fourier
Transform), TFR transforms like STFT, WVD, CWD, ZAMD, WT
(wavelet transform), etc. [11–13]. TFR analyses are among the
novel and practical methods for structural health monitoring in
vibrating systems [14]. TFR approach can be used in analyzing
the behavior of some systems. This method has been used for anal-
ysis of friction-induced vibrations in engines. With any variation in
a system, a frequency change occurs in the TFR analysis caused by
vibration [15]. For example, STFT can be used to detect different
sources of vibration in an engine. Further, this method has been
used for identifying normal and abnormal combustion-related
knocking in a cylinder block. Using the FFT and FFT-based methods,
such as estimated power spectral density, is unsuitable for identi-
fying non-stationary events, including cross-terms and systems
with rapid changes in time and frequency. Nevertheless, these
methods are fast and capable of filtering noises [16]. Lee et al.
developed a method of knock detection using cylinder pressure,
block vibration and sound pressure signals from a SI engine. Knock
intensity greatly varies with engine operating conditions, knock
sensor locations and fuel characteristics [17]. Taghizadeh et al.
studied the combustion, vibration, and knocking in diesel engines
produced due to different levels of diesel-biodiesel fuel blends by
time-frequency methods. Each fault in injection units and incorrect
spraying causes the knocking phenomenon and high-frequency
vibrations between the 7 and 25 kHz [18]. Rizzoni and Chen
detected knock in an internal combustion engine using time-
frequency distributions. They presented an improved knock detec-
tion scheme capable of tracking variations in the knock resonance
frequencies. Experimental results obtained from a single-cylinder
engine were used to validate the proposed method [19]. Bares
et al. analyzed the frequency spectrum of the pressure signal of
an engine using FFT and a window function in two locations near
the maximum heat release and near the end of combustion, and
the result was compared with the classical maximum amplitude
pressure oscillation (MAPO) definition. Results showed that the
proposed knock index definition can avoid the strong knocking
events and reduces engine vibration [20]. Wang et al. used the
Wigner-Ville distribution (WVD) of vibration signals for different
states of valve train and displayed it as TFR diagram. Moreover,
probabilistic neural networks (PNNs) were employed to classify
the time-frequency diagrams after normalization. By these meth-
ods, the fault diagnosis of the valve train was associated to the
time-frequency (TFR) diagram [21]. Molinaro et al. carried out
the knocking recognition in engine vibration signal using the
wavelet transform [22]. For low signal-to-noise ratio at high engine
speed, signal parameters extracted using wavelet transform
improve knocking detection [22]. Antoni et al. developed a cyclo-
stationarity methodology for condition monitoring of internal
combustion engines based on angular sampling and cyclic signal
processing [23]. Cerdá et al. performed knock detection in diesel
engines based on time-frequency analysis of cylinder pressure sig-
nal using Choi-Williams Distribution (CWD) via [24]. Chauvin et al.
studied real-time combustion parameters of HCCI-diesel engine
obtained from the knock sensor [25]. Ettefagh et al. developed a
parametric model-based filter to measure the knock intensity of
a SI engine. The filter were designed based on advanced parametric
modeling method and decomposition of the non-stationary ran-
dom vibration signals [26]. In recent years, higher-order spectra
(HOS) methods like bispectrum and trispectrum have been suc-
cessfully applied for faults diagnosis in vehicles and rotating
machines [5,27,28]. The final targets of such techniques are to
achieve a significant reduction in the number of vibration trans-
ducers installed at each bearing pedestal, without losing valuable
information needed for the diagnosis [5,27]. The order bispectrum
analysis also eliminates the effects of rotating speed change on the
vibration signal [29].
49
Diesel engine’s injection process is necessary for optimal engine
operation in terms of efficiency, power, and torque and reduction
of emissions. In general, this process is highly depend on the con-
dition of injection pump and fuel injector. Combustion progress in
diesel engines is depend on the characteristics of injection process
like the number of injectors and their timing, fuel quantity, and
mean injection pressure. Therefore, any changes in characteristics
of injection process can effects on the engine block vibrations
[30]. It is difficult to extract these information through measuring
parameters such as needle position, vibration, or acoustic proper-
ties of direct air-borne measurements [31]. Elamin et al. done the
identifying of injector faults in a diesel engine using acoustic emis-
sion (AE) technique. The AE signals were recorded and processed in
the angular, frequency, and jointed angular-frequency domain. The
results of angular-frequency analysis was shown that AE can
clearly monitor the changes in the combustion process due to its
high signal to noise ratio compared to other methods [32]. Jianmin
et al. monitored a diesel engine fuel-injection system through
vibrations analyzing. The results revealed that vibration signal
can potentially be employed for fault diagnosis and obtain infor-
mation such as crash strength [33]. Albarbar et al. used adaptive
filtering to conditioning the injector impact excitations via
air-borne acoustic signals. Based on the results, Wigner-Ville dis-
tribution (WVD) were useful in analyzing the injector needle
impacts [31,34]. In another research, Albarbar et al. investigated
the characteristics of diesel engine air-borne acoustic signal by
time-frequency domain analysis, and the energy levels within the
injection process in the frequency band of 9–15 kHz [34]. Siano
and Agostino employed discrete wavelet transform (DWT) to ana-
lyze the cylinder block signals of a SI engine for knock detecting.
They used a Multi-Resolution Analysis (MRA) and conventional
index MAPO for pressure data. The research had very similar
results for the both methods [35]. Flett and Bone studied a fault
detection and diagnosis (FDD) system for a diesel internal-
combustion engine valve train with deformed valve spring faults
and abnormal valve clearance faults. Five classification methods
were implemented experimentally to monitor valve closing and
combustion impacts. The FDD approach using the Naïve-Bayes
classification method produced the best performance. The detec-
tion and classification accuracy values for multiple faults were
99.95% and 92.45%, respectively [36].
A literature review showed the lack of research focused on the
injection fault detection in CI engines using time-frequency analy-
sis. Most of works that studied engine knock detection and there-
fore, identification of injector faults has not been addressed
specifically. In this study of this vibration signal of a six-cylinder
diesel engine were evaluated using time-frequency analyses. The
experiments were conducted for detection of faulty injectors as
well as engine’s knocking. The proposed diagnostic method in this
paper can detect the faulty injectors through knocking phe-
nomenon. Finally, this research introduced and evaluated a fast
and reliable approach for injector fault detection in diesel engines.
2. Materials and methods
2.1. Experimental set up
A direct injection, four-stroke six-cylinder, diesel engine (Per-
kins 1006-6) was used for experiments. To apply the load on the
engine a dynamometer, model R5 made by NJ-FROMENT Co, was
employed (Fig. 1a) as experimental setup. In this research, the
engine was run with two series of healthy and faulty injectors.
Injectors which were worked for more than 6 years and did not
spray well, were chosen as faulty injectors. Although these faulty
injectors did not significantly effect on the engine performance. It
50 A. Taghizadeh-Alisaraei, A. Mahdavian / Applied Acoustics 143 (2019) 48–58
Fig. 1. Schematics diagram of experimental set up; (a) dynamometer attached to the engine for load control, (b) data acquisition set up.
should be mentioned that the whole injectors had the same fault
type. New set of injectors were used as healthy and standard.
A proximity sensor (AC 2-wire type model, Autonics, Korea) was
employed for measuring the crankshaft angle as well as for calcu-
lating engine speed. Based on the proximity-sensor-produced
pulses’ count, the engine rotational speed was calculated. So, the
angle of the crankshaft was recorded simultaneously during the
experiments. In this process, the proximity sensor was installed
at a 2 mm gap with the end of crankshaft. Eight knobs were circu-
larly placed at the end of the crankshaft to create the crankshaft
rotation pulses.
To acquire engine vibration signals, three accelerometers (made
by CTC -AC102-1A) were mounted on the engine body near the fly-
wheel. The sensor pulses and vibration signals were transferred to
a switchboard. This switchboard includes four interface-starting
circuit for the three accelerometers and the proximity sensor.
The output signals of switchboard were transferred to an analogue
to digital (A/D) converter (Advantech, USB-4711A). Finally, the out-
put data cable was connected to a computer port (Fig. 1b). In this
study, the sample rate was set on 50 kHz. According to the engine
specifications, the sampling frequency of data acquisition for
accelerometers was set on 50 kHz. In signal analysis, half of this
frequency (25 kHz) is represented. Before data acquisition, ana-
logue anti-aliasing filtering was applied. The accelerometers were
mounted in three orthogonal directions namely vertical, lateral,
and longitudinal (perpendicular to the other axes), respectively.
Data were recorded for one minute at each experiment sample.
Sampling time was triggered by an engine shaft encoder signal.
The crank angle resolution of the proximity sensor was set on
45° by a sampling frequency of 10 kHz. The experiments were con-
ducted in under-load engine conditions. Furthermore, the acceler-
ation data were collected for both faulty and healthy injection
units. The signal conditioning process was performed in time
domain, frequency domain, Welch test, and TFR analysis.
2.2. Theory of analytic methods
For spectral content of the signals, the Welch’s test was carried
out to detect the partial differences between the engine vibration
signals and total comparison between faulty and healthy injectors.
At the first step, original signal of x(t) was divided into the L over-
lapping parts with length of N. The second step, was calculating
Fast Fourier Transform (FFT) of the windowed parts prepared at
the previous step. FFT was calculated for each part according to
Eq. (1). At the third step mean squared discrete FFT calculated
for each section of the signal.
Z þ1
FFT ¼ xðtÞejxt dt; ð1Þ
1
where x(t) and x are time series and frequency center, respectively.
In Welch’s test, after dividing the signal into the overlapping
parts between zero and 50%, the PSD is calculated for each section
of the signal according to Eq. (2).
 A. Taghizadeh-Alisaraei, A. Mahdavian / Applied Acoustics 143 (2019) 48–58 51

2
jFFTðW t  X t ; f Þj 1 Choi and Williams introduced kernel in the form of an exponen-
PSD ¼  ð2Þ tial function, Eq. (9):
L Df
uðh; sÞ ¼ eh s =r :
2 2
where Xt is time series, L is the number of samples in time series, Wt ð9Þ
is window function, f is frequency center, and Df is frequency
The equation for the continuous Choi-Williams distribution of
spacing.
the input signal is given as
ZZ 
2.2.1. Short-term Fourier Transform (STFT) 1 s  s
CWDðt; xÞ ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi x l þ x l  eððltÞ =ð4s =rÞÞjxs dl ds;
2 2

For short-term Fourier transform, a signal is divided into small- 4ps2 =r 2 2

time frames. And then, the Fourier transform is performed for each ð10Þ
frame to develop a spectrum. The STFT is calculated using Eq. (3):
Z Where, r and x are positive-valued scaling factor and angular
þ1
1 velocity respectively. If r is large enough, the kernel will approach
F ðx; sÞ ¼ pffiffiffiffiffiffiffi xðt Þhðt  sÞejxt dt; ð3Þ
2p 1 to one and CWD will be close to WVD.
where, h(t) is a central windowing function at time t, that provides
the conditions of Eq. (4), 2.2.4. Implementation of the Welch, STFT, WVD, and CWD algorithms
Z in this research
þ1
Four mentioned signal transform methods were employed for
hðsÞds ¼ 1 ð4Þ
1 engine vibration signals at two cycles of engine working (four
crankshaft revolutions) with faulty and healthy injectors. Equa-
Spectrogram or energy density spectrum of the STFT is derived
tions of 1, 5, 6, and 9 were used for Welch’s test, STFT, WVD, and
from Eq. (5):
CWD analysis respectively. First, a discrete-time signal, x[n] with
 Z þ1 2
 1  the number, N, was selected. Then, the signal was divided into

STFT ðx; sÞ ¼ jF ðx; sÞj ¼  p 2
ffiffiffiffiffiffi
ffi xðtÞhðt  sÞejxt
dt : ð5Þ
2p 1 small sections with length m by window function of h (n). After
windowing, the Fourier transform was performed for each section
For the discrete-time analysis, the STFT was defined as Eq. (6), of x[n]*h[n-m] in time domain and its content were gained. Finally,
X
þ1 the calculated spectrogram values from each equation were drawn
STFT ðm; xÞ ¼ x½nh½n  mejxn : ð6Þ in a time-frequency diagram. For this purpose, two engine cycles
n¼1 were used after normalizing and then, with constantly changing
The STFT method can be used successfully for slowly-changing parameters, the best resolution for time-frequency diagram was
signals. The in time and frequency domain resolutions are largely obtained. These parameters were type and size of the window
depend on the windowing process. Although gained results from according to the sampling frequency (SF), number of data, time
the STFT is useful, it has disadvantages due to a constant specific interval (1/SF), and the overlap length. The number of data depend-
window size selected for all frequencies. However, this method sig- ing on engine rpm was from 6011 (2000 rpm) up to 7500
nificantly shows obvious advantages such as no cross-terms inter- (1600 rpm) with SF of 50 kHz (time interval, 0.02 ms). Further-
ference and overlaps in the spectral analysis, which are both more, hamming window with length of 27 was applied for individ-
disturbing in WVD, CWD and ZAMD representations [11,15]. ual TFR methods. To obtain an appropriate resolution and decrease
computational load, the number of frequency bins were chosen
2.2.2. Wigner-Ville Distribution (WVD) 256 (the length of overlap was 50% for each two adjacent hamming
This distribution was presented by Wigner in the field of quan- windows).
tum mechanics in 1932. It was employed for solving problems in
the communication theory by Ville in 1948 [11,15]. For a 3. Results and discussion
continuous-time signal x(t), the Wigner-Ville Distribution is
defined as Eq. (7), 3.1. Engine work timing
 Z
s   s
1
WVDðt; xÞ ¼ x t þ x t  ejxs ds; ð7Þ Vibration signals can indicate combustion pressure fluctuations,
1 2 2
misalignment, unbalancing, and inertia in engine parts. Fig. 2
where, x*(t) is complex conjugate of x(t). shows simulated compression and combustion pressure curves
WVD is more attractive, because this method ignores the for one cylinder corresponding to the engine body vibration versus
hypothesis of the short-term stationary in signal and overcomes crank angle degree. This diagram is based on a simulation of the
the resolution between time and frequency. This method has a dis- previous pressure data of the same engine in the similar condi-
advantage in displaying the spectrum called cross-terms, which tions. In four steps of the combustion process, including delaying
causes the frequency misidentification. in combustion, premixed combustion, controlled combustion, and
delayed combustion the vibration exists with different intensity.
2.2.3. Choi-Williams Distribution (CWD) Most of the engine vibration arises from the areas of 3 and 4. These
This distribution was introduced by Choi and Williams in 1989. areas are ‘‘controlled combustion” and ‘‘delayed combustion”
This distribution is derived from the generalized Cohen’s class of which may be containing the knocking. The maximum amplitude
distributions by Eq. (8), of vibration is for area 3, namely, controlled combustion due to
the maximum heat release and possibility of knocking is high at
1
C f ðt; x; uÞ ¼ this stage. In stage of 3 and 4, spraying of injectors play an impor-
2p
ZZZ  tant role in engine knocking control.
s   s
 uðh; sÞx l þ x l ejhtjxsjhx dl ds dh; ð8Þ For better understanding, the events and timing are discussed
2 2
in the time domain while 1600 rpm (Fig. 3). The occurring time
where, x(l) is the input signal, x*(l) is its complex conjugate, u(h,s) for the intake valve opening (IVO), exhaust valve closing (EVC),
is called kernel function, a representative of the spatial distribution intake valve closing (IVC), firing, and exhaust valve opening
function. When u(h,s) = 1, the equation (8) is WVD. (EVO) are depicted as follows:
52 A. Taghizadeh-Alisaraei, A. Mahdavian / Applied Acoustics 143 (2019) 48–58
Fig. 2. Simulated compression and combustion pressure curves corresponding to measured engine body vibration.
(a) IVO at 25° (39.57 ms) before top dead center (TDC).
(b) EVC at 28° (41.28 ms) after TDC.
(c) IVC at 60° (64.89 ms) after the bottom dead center (BDC).
(d) Combustion peak pressure with maximum vibration at 18°
(78.92 ms) after TDC (for cylinder 4).
(e) EVO at 120° (90.42 ms) after the TDC with dashed red line or
60° before BDC.
The engine firing order is 1-5-3-6-2-4, as well combustion-
induced impacts occur at respective intervals of 40.43, 53.05,
65.67, 74.29, 90.90, and 103.53 ms for each cylinder (start from
cylinder 3). In an ideal combustion, the valve’s opening-closure
impacts and fuel-injection nozzles spraying effect have fewer
intensities than the firing impact in the combustion chamber.
In abnormal combustion condition, the most of frequency accu-
mulation are observed in ranges of 0–10 kHz and 10–25 kHz
while in ideal and proper combustion; pressure pulses generate
vibrations with high amplitude in frequencies between 0 and
10 kHz.
3.2. Detection injector’s fault
At first, the experiments were performed with faulty fuel injec-
tors at the engine speeds of 1600, 1700, 1800, 1900, and 2000 rpm
and full load condition. The same tests were done for healthy ones
with new set of injectors. Fig. 4 shows some vibration signals at the
1800, 1900, and 2000 rpm for faulty and healthy samples. Pulses
from engine cylinders can be observed clearly in both cases. In
the engine with faulty injectors, combustion pulses come from
cylinders have irregularities compared with healthy injectors,
especially at 1800 and 1900 rpm. At 2000 rpm, the pulses are more
similar to each other and making distinction between the injectors
is difficult. As a result, lower speeds are more appropriate for such
fault diagnosis process.
For better analysis and judgment about the performance of
injectors, the Welch test was applied by Eq. (1) and (2), which is
difficult in time-domain signal.
Fig. 5 shows the Welch test diagram for engine with faulty and
healthy injectors. This method is appropriate for general comparison
 A. Taghizadeh-Alisaraei, A. Mahdavian / Applied Acoustics 143 (2019) 48–58 53

Start Start Start

compression injecon combuson
Intake Compression Power Exhaust

IVO IVC EVO IVO EVC

EVC

C I C I C I C I C I C I

TDC BDC
3 5 1 4 2 6 3

Fig. 3. Time domain of vibration signal on engine body with timing and events (1600 rpm).

Fig. 4. Time domain of vibration signal for engine with faulty (a) and healthy (b) for vertical direction.

between the test treatments. The signals of both faulty and healthy
injectors are shown in Fig. 5. On the left side of the diagram (large
dashed rectangle), the difference in peak points of both cases is
clear. In healthy injectors, the peak is single and flat for all three
speeds, however, the faulty mode is composed of several peaks.
On the other hand, the amplitude of acceleration for faulty mode
is more than the healthy mode in all three speeds. The main argu-
ment is in the right side of the diagram (small dashed rectangle) for
each case. The diagram (Fig. 5) is divided into two frequency bands
of 0–10 and 10–25 kHz. In the lower frequency range, the amount
of vibrations are high in both cases of faulty and healthy. This
vibration is related to fuel combustion and fluctuations of pressure
inside the cylinders. In the frequency range of 10 to 25 kHz, the
vibrations strongly are dropped for healthy injectors against the
faulty ones. It can be said that most of the fuel energy loses are
happened because of unwanted combustion, and this can generate
high-frequency vibration. Incomplete and poor combustion can
cause high-frequency vibrations in the range of 10 to 25 kHz and
makes knocking.
RMS and kurtosis parameters that are employed to study the
engine performance in this research are defined as follows.
RMS is defined by Eq. (11) with the number of points within a
time period, N, and acceleration data, x(tk):
" #1=2
1 XN
xRMS ¼ x2 ðtk Þ : ð11Þ
N k¼1
54 A. Taghizadeh-Alisaraei, A. Mahdavian / Applied Acoustics 143 (2019) 48–58

Fig. 5. Frequency domain of vibration signal for engine with faulty (a) and healthy (b) for vertical direction using Welch’s test.

The RMS is calculated according to the square of vibration data; engine performance in the first case. In the second case, the engine
thus, more weight is automatically given to sudden peaks (shocks). runs smoother and generates less noise and vibration, which can
To compare the results in different methods, the kurtosis result the less engine part damaging. The results revealed that,
method was also used for evaluation of engine’s performance using RMS and kurtosis of the engine vibration increased by 12.9% and
the vibration signal (Eq. (12)), [37–39], 20.6% due to faulty injectors. Approximately, the similar results
h P i were obtained for all engine speeds.
N
1
N
ðt k Þ
k¼1 x
4
Welch test presents an overall comparison of vibration for
Kurt ¼ h i2 : ð12Þ
PN 2 faulty and healthy injectors. To study each cylinder individually
k¼1 x ðt k Þ
1
N in details, the mentioned STFT time-frequency method was applied
for each case using Eq. (6) and procedure presented in Section 2.2.3.
Kurtosis is known as a noise sensitive feature, there for in prac-
Fig. 7 shows this analysis for under load engine at rotational speeds
tice, preliminary treatments of the signal (band-pass filtering,
of 1800, 1900, 2000 rpm.
envelope detection) are essential in order to benefit fully from this
For faulty injectors, the dominant observed vibration frequen-
indicator [37–39].
cies were between both 0–10 kHz and with fewer intensities in
Fig. 6 shows the RMS and kurtosis values calculated by Eqs. (11)
range of 10–25 kHz (Fig. 7a). For healthy injectors, there were
and (12), respectively, for all three directions of vertical, lateral and
vibrations with very lower amplitude within the frequencies of
longitudinal at the all speeds, the RMS and kurtosis are higher for
10–25 kHz (Fig. 7b). This shows the engine runs smoother and
faulty injectors than the healthy mode. Higher value of kurtosis
combustion interval is short. Comparing the results from this sec-
indicates sharpness in time domain presentation and existence of
tion and earlier sections, we can conclude that no knock and poor
numerous shocks in signal (Fig. 6-b). This shows very erratic
performance were observed for engine cylinders. This proves that
precise timing and proper injection play an important role in pre-
Faulty Healthy venting the engine knocking. Carlucci et al. has explained that
(a)
injection pressure and injected quantities affect vibration signals.
100 Injection timing affects the engine block vibration in a less obvious
way [15]. It seems an untimely fuel injection and fuel dripping are
RMS

80 the main reasons for high-frequency vibrations (between the 10 to

25 kHz) which produce resonance. This can be detected to needle-
60 spring malfunctioning in fuel-injection nozzles and opening dilata-
1500 1600 1700 1800 1900 2000 2100
tion. The dripping phenomenon adds an extra amount of fuel to the
hot gases inside a combustion chamber, and a sudden undesirable
(b) 11
firing produces knocking. In other words, two ignitions with differ-
ent power continually (at short time intervals) and the collision of
9
Kurtosis

two flames’ front occur in the cylinder chamber that leads to

7 unbalance and non-smooth engine parameters. These problems
are sometimes observed in a diesel engine and are the main pref-
5 erence in design and development of diesel engines, which can
1500 1600 1700 1800 1900 2000 2100
Engine Speed (rpm) be solved by more accurate fuel-injection nozzles. In Fig. 7 a, c,
and e for faulty injectors, the strongest combustion impacts
Fig. 6. The RMS and kurtosis value for faulty and healthy injectors. associated with faults in the fuel-injection nozzles are generated
 A. Taghizadeh-Alisaraei, A. Mahdavian / Applied Acoustics 143 (2019) 48–58 55

(a) (b)
Healthy-1800
Acceleraon (m/s ) Faulty-1800 rpm
2

Time (ms) Time (ms)

(c) (d) Healthy-1900 rpm

Faulty-1900 rpm
Acceleraon (m/s )
2

Time (ms) Time (ms)

(e) (f) Healthy-2000 rpm
Faulty-2000 rpm
Acceleraon (m/s )
2

Time (ms) Time (ms)

Fig. 7. The STFT analysis for engine’ body vibration signal with faulty (a, c, e) and healthy (b, d, f) injectors in 1800, 1900, and 2000 rpm.

in several cylinders. It can be concluded that primary source of
vibrations is knocking. Fig. 7b, d, and f shows the TFR diagram
for healthy injectors at1800, 1900, and 2000 rpm, respectively.
Here, the vibrations with very lower amplitude are found within
the frequency range of 10 and 25 kHz in diagrams, and engine
works without any problems. However, in this case the combus-
tion interval is more visible in the diagram due to accurately
injected fuel and more powerful ignitions in each cylinder in Fig. 7.
Other similar studies have been conducted on the engine vibra-
tion. For example, in a same study, the energy of 7–15 kHz fre-
quency bands was varied according to the injection pressure in a
research by Albarbar et al. [31]. Jianmin et al. studied the fuel-
injection system of a diesel engine by analyzing the cylinder head
vibrations which obtained almost the same results [33]. Siano and
Agostino applied discrete wavelet transform (DWT) to detection of
knocking on the vibration signals acquired by an accelerometer in
a spark ignition (SI) engine. The results showed that the proposed
knock detection approach is capable of monitor the engine status
and intensity of knock [35]. Condition monitoring using air-borne
acoustic signal is a useful approach due to this has rich informa-
tion. However, the obstacles in this technic are the background
noise contamination, interferences and the number of internal-
combustion engine vibro-acoustic sources [34].
3.3. WVD and CWD analysis
WVD and CWD analysis were performed using Eq. (7) and Eq.
(10), respectively, and procedure presented in 2.2.3 section. Fig. 8
displays the TFR diagram using WVD method. This diagram has a
good resolution in time axis, but it is impossible to mark off the
56 A. Taghizadeh-Alisaraei, A. Mahdavian / Applied Acoustics 143 (2019) 48–58

Faulty-1800 rpm Healthy-1800 rpm

(a) (b)
Acceleraon

Time Time (ms)

(c) Faulty-1900 rpm (d) Healthy-1900 rpm

Acceleraon (m/s )
2

Time Time (ms)

(e) Faulty-2000 rpm (f) Healthy-2000 rpm

Acceleraon (m/s )
2

Time Time (ms)

Fig. 8. The WVD analysis for engine’s body vibration signal with faulty (a, c, e) and healthy (b, d, f) injectors in 1800, 1900, and 2000 rpm.

frequency axis. Therefore, this technique is not capable of
identifying faults and knock clearly. Fig. 9 shows the TFR analysis
using CWD method. Compared to WVD methods, this method
presents a better resolution in the time axis as well in the
frequency axis, relatively. For example, cylinders with strong
combustion are visible in Fig. 9. This approach gives more weight
to the areas with higher energy compared to the WVD method.
CWD analysis has the resolution weakness compared with the STFT
in frequency axes.
3.4. Comparing the used analysis methods
Comparing the three methods, namely STFT, WVD, and CWD for
fault detection showed that the STFT method is more reliable for
measuring the vibration characteristics in cylinders. Overall, STFT
is the fastest method, and CWD is a time-consuming method for
the TFR analysis. The average computing times in STFT, WVD,
and CWD methods are 0.29, 0.54, and 3.25 s, respectively, for
two cycles of engine working vibration signal. Using a simpler
 A. Taghizadeh-Alisaraei, A. Mahdavian / Applied Acoustics 143 (2019) 48–58 57

Faulty-1800 rpm Healthy-1800 rpm

(a)
Acceleraon (b)

Time (ms) Time (ms)

Faulty-1900 rpm Healthy-1900 rpm
(c) (d)
Acceleraon

Time (ms) Time (ms)

Faulty-2000 rpm Healthy-2000 rpm
(e) (f)
Acceleraon

Time (ms) Time (ms)

Fig. 9. The CWD analysis for engine’s body vibration signal with faulty (a, c, e) and healthy (b, d, f) injectors in 1800, 1900, and 2000 rpm.

algorithm, STFT is a suitable approach for online and off-line knock
detection.
4. Conclusions
This study presents fault detection process of diesel engine’s
injectors using time-frequency analysis of vibration signal. The
results showed that the Short-Term Fourier Transform (STFT) is
the most efficient method for such fault detection process. In
healthy mode, the combustion and injectors work properly, cause
the high-amplitude vibrations in the range of frequencies from 0
to 10 kHz. Any fault in fuel-injection nozzles produces high-
frequency components, higher that 10 kHz. For a healthy and accu-
rate fuel injection unit, the vibrations higher than 10 kHz were
eliminated and. In total, STFT is a convenient method for identify-
ing knocking and faults in fuel injector nozzles for offline and on-
line applications. With a fast algorithm and a reduced resolution,
58 A. Taghizadeh-Alisaraei, A. Mahdavian / Applied Acoustics 143 (2019) 48–58
STFT method can be used for on-line knock detection with reason-
able accuracy. It can be concluded that for real-time engine perfor-
mance monitoring, the STFT technique is more efficient and faster
than the other techniques presented in this research. By replacing
the faulty injectors with healthy ones, increased the RMS and kur-
tosis of the engine vibration by 12.9% and 20.6%, which indicates
more irregularity in engine.
Acknowledgment
This study was supported by Gorgan-Golestan Universities,
Iran. Therefore, we would like to thank all the staff and professors
of the mentioned university who provided the necessities of the
research.
Appendix A. Supplementary data
Supplementary data associated with this article can be found, in
the online version, at https://doi.org/10.1016/j.apacoust.2018.09.
002.
References
[1] Wang Z, Liu H, Reitz RD. Knocking combustion in spark-ignition engines. Prog
Energy Combust Sci 2017;61:78–112. https://doi.org/10.1016/j.
pecs.2017.03.004.
[2] Saikaly K, Le Corre O, Rahmouni C, Truffet L. Preventive knock protection
technique for stationary SI engines fuelled by natural gas. Fuel Process Technol
2010;91:641–52. https://doi.org/10.1016/j.fuproc.2010.01.011.
[3] Goering CE, Hansen AC. Engine and tractor power. ASAE 2006 (4th ed.).
[4] Ettefagh MM, Sadeghi MH, Pirouzpanah V, Arjmandi Tash H. Knock detection
in spark ignition engines by vibration analysis of cylinder block: A parametric
modeling approach. Mech Syst Signal Process 2008;22:1495–514. https://doi.
org/10.1016/j.ymssp.2007.11.027.
[5] Yunusa-Kaltungo A, Sinha JK, Elbhbah K. HOS analysis of measured vibration
data on rotating machines with different simulated faults. In: Lect. Notes
Mech. Eng. Berlin, Heidelberg: Springer; 2014. p. 81–9. https://doi.org/
10.1007/978-3-642-39348-8_6.
[6] Antoni J. Cyclostationarity by examples. Mech Syst Signal Process
2009;23:987–1036. https://doi.org/10.1016/j.ymssp.2008.10.010.
[7] Geng Z, Chen J, Hull JB, Barry Hull J. Analysis of engine vibration and design of
an applicable diagnosing approach. Int J Mech Sci 2003;45:1391–410. https://
doi.org/10.1016/j.ijmecsci.2003.09.012.
[8] Ortmann S, Rychetsky M, Glesner M, Groppo R. Engine knock detection using
multi-feature classification by means of non-linear mapping. Proc ISATA
1997;97:607–13.
[9] Pruvost L, Leclère Q, Parizet E, Pruvost L, Lecle Q, Leclère Q, et al. Diesel engine
combustion and mechanical noise separation using an improved spectrofilter.
Mech Syst Signal Process 2009;23:2072–87. https://doi.org/10.1016/j.
ymssp.2009.04.001.
[10] Vulli S, Dunne JF, Potenza R, Richardson D, King P. Time-frequency analysis of
single-point engine-block vibration measurements for multiple excitation-
event identification. J Sound Vib 2009;321:1129–43. https://doi.org/10.1016/j.
jsv.2008.10.011.
[11] Debbal SM, Bereksi-Reguig F. Time-frequency analysis of the first and the
second heartbeat sounds. Appl Math Comput 2007;184:1041–52. https://doi.
org/10.1016/j.amc.2006.07.005.
[12] Peng ZK, Chu FL, Tse PW. Detection of the rubbing-caused impacts for rotor-
stator fault diagnosis using reassigned scalogram. Mech Syst Signal Process
2005;19:391–409. https://doi.org/10.1016/j.ymssp.2003.09.007.
[13] Franz Hlawatsch FA. Time-Frequency Analysis: Concepts and Methods. First
Edit. ISTE Ltd and John Wiley & Sons, Inc; 2008. doi:10.1002/9780470611203.
[14] Mertins A. Signal analysis: wavelets, filter banks, time-frequency transforms
and applications. John Wiley & Sons Ltd 1999.
[15] Cohen L. Time-Frequency Analysis 1995:299. doi:10.1002/9780470611203.
[16] Al-Badour F, Sunar M, Cheded L. Vibration analysis of rotating machinery using
time-frequency analysis and wavelet techniques. Mech Syst Signal Process
2011;25:2083–101. https://doi.org/10.1016/j.ymssp.2011.01.017.
[17] Lee J, Hwang S, Lim J, Jeon D, et al. A new knock-detection method using
cylinder pressure, block vibration and sound pressure signals from a SI Engine.
ASE 1998:981436.
[18] Taghizadeh-Alisaraei A, Ghobadian B, Tavakoli-hashjin T, Mohtasebi SS,
Rezaei-asl A, Azadbakht M. Characterization of engine’s combustion-
vibration using diesel and biodiesel fuel blends by time-frequency methods:
a case study. Renew Energy 2016;95:422–32. https://doi.org/10.1016/j.
renene.2016.04.054.
[19] Rizzoni G, Chen XC. Detection of internal combustion engine knock using
time-frequency distributions. In: Circuits Syst 1993 Proc 36th Midwest Symp.
p. 360–3. https://doi.org/10.1109/MWSCAS.1993.343035.
[20] Bares P, Selmanaj D, Guardiola C, Onder C. A new knock event definition for
knock detection and control optimization. Appl Therm Eng 2018;131:80–8.
https://doi.org/10.1016/j.applthermaleng.2017.11.138.
[21] Wang C, Zhang Y, Zhong Z. Fault diagnosis for diesel valve trains based on
time-frequency images. Mech Syst Signal Process 2008;22:1981–93. https://
doi.org/10.1016/j.ymssp.2008.01.016.
[22] Molinaro F, Castanie F, Denjean A. Knocking recognition in engine vibration
signal using the wavelet transform. In: Time-Frequency Time-Scale Anal 1992,
Proc IEEE-SP Int Symp. p. 353–6. https://doi.org/10.1109/TFTSA.1992.274166.
[23] Antoni J, Daniere J, Guillet F. Effective vibration analysis of IC engines using
cyclostationarity. Part 1: a methodology for condition monitoring. J Sound Vib
2002;257:815–37. https://doi.org/10.1006/jsvi.5062.
[24] Cerdá S, Romero J, Giménez A, Zurita G. Detection of the knock in diesel
engines by using the Choi-Williams distribution. Acta Acust United with Acust
2002;88:998–1004.
[25] Chauvin J, Grondin O, Nguyen E, Guillemin F. Real-time combustion
parameters estimation for HCCI-diesel engine based on knock sensor
measurement. IFAC 2008;vol:41. https://doi.org/10.3182/20080706-5-KR-
1001.01437.
[26] Ettefagh MM, Sadeghi MH, Rezaee M, Khoshbakhti R, Akbarpour R. Application
of a new parametric model-based filter to knock intensity measurement.
Measurement 2010;43:353–62. https://doi.org/10.1016/j.measurement.2009.
11.008.
[27] Yunusa-Kaltungo A, Sinha JK. Combined bispectrum and trispectrum for faults
diagnosis in rotating machines. Proc Inst Mech Eng Part O J Risk Reliab
2014;228:419–28. https://doi.org/10.1177/1748006X14524547.
[28] Czech P. Diagnose car engine exhaust system damage using bispectral analysis
and radial basic function. In: Proc. Int. Conf. Comput. Networks Commun. Eng.
(ICCNCE 2013), Paris, France. : Atlantis Press; 2013. https://doi.org/10.2991/
iccnce.2013.78.
[29] Kocur D, Stanko R. Order bispectrum: a new tool for reciprocated machine
condition monitoring. Mech Syst Signal Process 2000;14:871–90. https://doi.
org/10.1006/mssp.2000.1307.
[30] Carlucci APÃ, Chiara FF, Laforgia D. Analysis of the relation between injection
parameter variation and block vibration of an internal combustion diesel
engine. J Sound Vib 2006;295:141–64. https://doi.org/10.1016/j.
jsv.2005.12.054.
[31] Albarbar A, Gu F, Ball AD, Starr A. Acoustic monitoring of engine fuel injection
based on adaptive filtering techniques. Appl Acoust 2010;71:1132–41. https://
doi.org/10.1016/j.apacoust.2010.07.001.
[32] Elamin F, Gu F, Ball A. Diesel engine injector faults detection using acoustic
emissions technique. Mod Appl Sci 2010;4:3–13. https://doi.org/10.5539/mas.
v4n9p3.
[33] Jianmin L, Yupeng S, Xiaoming Z, Shiyong X, Lijun D. Fuel injection system fault
diagnosis based on cylinder head vibration signal. Procedia Eng
2011;16:218–23. https://doi.org/10.1016/j.proeng.2011.08.1075.
[34] Albarbar A, Gu F, Ball AD. Diesel engine fuel injection monitoring using
acoustic measurements and independent component analysis. Measurement
2010;43:1376–86. https://doi.org/10.1016/j.measurement.2010.08.003.
[35] Siano D, Agostino DD, D’Agostino D, Agostino DD, D’Agostino D. Knock
detection in SI engines by using the discrete wavelet transform of the engine
block vibrational signals. Energy Procedia 2015;81:673–88. https://doi.org/
10.1016/j.egypro.2015.12.052.
[36] Flett J, Bone GM. Fault detection and diagnosis of diesel engine valve trains.
Mech Syst Signal Process 2015:1–12. https://doi.org/10.1016/j.
ymssp.2015.10.024.
[37] Pachaud C, Salvetat R, Fray C. Crest factor and kurtosis contributions to identify
defects inducing periodical impulsive forces. Mech Syst Signal Process
1997;11:903–16. https://doi.org/10.1006/mssp.1997.0115.
[38] Wang Y, Xiang J, Markert R, Liang M. Spectral kurtosis for fault detection,
diagnosis and prognostics of rotating machines: a review with applications.
Mech Syst Signal Process 2016;66–67:679–98. https://doi.org/10.1016/j.
ymssp.2015.04.039.
[39] De Lorenzo F, Kurtosis Calabro M. A statistical approach to identify defect in
roller bearings. In: 2nd Int Conf Mar Res Transp. p. 17–24.