Signal Processing - March 2017 - 78

[39] presented an approach for varying
the knock detection window length
Knock
Knock
Signal
Bandpass
Envelope
Knock
while using the single-point DFT detecSensor
Decision
Preprocessing
Filter
Detector
Decision
Signal
Output
tion method.
The aforementioned frequency filFigure 8. A block diagram of a typical frequency-domain-filtering-based knock detector.
tering and transform spectrum methods are relatively simple. However,
because
they
do
not
consider
the time-varying nature of resknock decision. The implementation of filtering and envelope
onance frequencies as well as nonlinear behavior of knock
detection approaches for knock intensity detection evolved in
waves due to the complicated changes of cylinder gas temperseveral stages from analog technologies, to mixed analog and
atures and other physical phenomena, the knock information
digital technologies, to today's fully digital technologies.
cannot be precisely detected or estimated for robust knock
As mentioned previously, the knock signals' resonance
control and optimal spark timing under some engine operatfrequency and its harmonics range from 5 to 20 kHz, and the
ing conditions. To more effectively retrieve time-frequency
signal sampling rate can be as large as 100 kHz. Existing comdependent information associated with engine knocking,
mon ECUs have difficulty performing real-time processing of
more advanced signal processing techniques such as timesignals with such a high sampling rate. Thus, knock detection
frequency analyses and wavelet transforms have been studied
is often implemented in a dedicated integrated circuit, and
recently [41]-[45].
then the knock intensity is transferred from a chip to an ECU
The advantage of using time-frequency signal representathrough a communication interface at a much slower data rate
tions is in their ability to recognize time-frequency-dependent
compared to the knock sampling rate. Today's knock detector
features, such as the frequency shifts that occur in the knock
devices have flexible configurations with programmable capasignal [41]. Both the Wigner-Ville distribution and the cross
bilities for gain, bandpass filter frequencies, and integrators.
Wigner-Ville distribution were studied for a V6 engine's
When an ECU receives knock intensity signals from a knock
knock signal analysis in [41], and the test results showed that
detector device, the knock decision is made by comparing the
the SNR improved significantly compared to the bandpass filknock intensity or its integration result to a preset threshold or
tering method. To overcome the drawback of the large number
the one calculated or obtained using look-up tables based on
of operations required to perform such transforms, a pseudoengine operating conditions, such as engine speed and load.
Wigner distribution was proposed for the time-frequency
To improve the detection performance and reduce calibration
analysis of the knock signals [42]. Figure 9 shows a typical
effort, some recent papers have proposed further processing
high-pass filtered knock signal and its Wigner-Ville spectrum
of knock intensity using statistical methods. These stochastic
generated from several knock signals, where the knock resoapproaches assume the knock intensity has a lognormal disnance frequencies are clearly displayed with changes corretribution and estimate a new metric, called the knock factor,
sponding to the engine crank angle [48].
that is related to the high and low percentiles for a lognormal
The wavelet transform is good for time-scale analysis of
distribution given the number of consecutive knock intensities
a signal. In recent years, many researchers, mainly from aca[37]. The knock factor then is used for knock control.
demic areas, studied the application of wavelet transforms
In addition to frequency-domain-filtering methods, methfor knock feature detection. Unlike the Fourier transform,
ods that perform spectrum analysis, such as power spectrum,
the wavelet transform provides the time-evolution of the sigFFT, and DFT, have advantages for knock detection [38]-[40].
nal at different scales. The discrete wavelet transform (DWT)
Among them, however, DFTs are the most viable solution for
is a computationally efficient implementation of the wavelet
real-time knock detection, at least until ECUs become comtransform and is more suitable for knock detection applicaputationally more powerful. A detailed strategy using DFTs
tions as reported by [44]. The paper [45] proposes a knock feais discussed in [38] for implementing the detection of knock
ture extraction method using wavelet packet transforms, and
signatures. The method uses multiple single-point DFTs to
through the examples they studied, they showed that the wavemonitor the fundamental frequency plus the vibrational modes
let packet transform improves the time-frequency resolution
of an engine. The DFT algorithm provides better frequency
and has advantages in extracting knock feature information,
discrimination than analog or digital filters. It is also less comknock intensity and the time of knock occurrences, even under
putationally intensive than an FFT when only a few frequency
light knock conditions, compared to the DWT method.
points are monitored, even though the FFT is more efficient
Instead of using accelerometer sensor signals for knock
in total time required when calculating across all allowable
analysis, some researchers have also studied wavelet transfrequency ranges. In real-time processing, DFT has another
forms using ion current signals [46].
advantage over FFT. That is, all samples for an FFT must be
In addition to the aforementioned mainstream knock detecstored in memory before its calculations, but the DFT can be
tion techniques, other signal processing methods were also
calculated one sample at a time because there is no linkage
explored for knock detection recently, such as neural networksbetween samples as there is with the FFT [38]. Similar to timebased knock feature extraction methods [47], model-based
based filtering, the DFT method uses a time window to proknock detection methods [48], as well as the Hilbert-Huang
cess the data points in the knock-possible period. The paper
78

IEEE SIgnal ProcESSIng MagazInE

March 2017

Table of Contents for the Digital Edition of Signal Processing - March 2017