Signal Processing - March 2017 - 77

...

and control [31]. In this method, the
Estimated Individual
FFT or DFT of the O2 signal is calcuCylinder Fuel Richness
lated first, and the amplitude of the first
∆φ1
harmonic is then used for fuel-air ratio
Array
Individual Cylinder
Raw O2
∆φ2
Linearization
Signal
Fuel Richness
imbalance decision making through a
Signal
Formation
Estimators
∆φL
comparison with a given threshold. Once
the imbalance is detected, it retrieves
fuel-air imbalance relations among cy- Figure 7. The system block for individual cylinder fuel richness estimation. (Figure adapted from [32].
linders through the study of the fre- Reprinted with permission from SAE International.)
quency characteristics of the cylinder
imbalance patterns. With this method, no synchronization sigWhen engine knocking occurs, pressure shock waves are
nal is needed, and thus it avoids the risks associated with posgenerated and excite resonance inside the combustion chamsible synchronization errors.
ber. The knock resonance frequency depends on the cylinder
The aforementioned methods are frequency-domain signal
bore diameter and combustion chamber gas temperature [40]
processing methods. A nonfrequency-domain-based signal
and typically ranges from 5 to 7 kHz. Two to four additional
processing method is presented in [32]. In the previously disfrequency peaks are also evident between the fundamencussed methods, the obtained fuel-air ratio information is a
tal frequency and 20 kHz. Because of the characteristics of
relative quantity, not the estimate of the actual fuel-air ratio
the knock signals, signal processing techniques are the most
or its deviation from the stoichiometric point. The method
effective means to enhance the SNR and reliably perform
shown in [32] uses a temporal array signal processing techknock detection.
nique to estimate the actual fuel-air ratio or equivalently its
Several types of sensors, such as the piezoelectric accelerdeviation to the stoichiometric point, through an array of data
ometer sensor, cylinder pressure sensor, and ion sensor have
samples from an oxygen sensor located in a confluence point
been investigated for knock detection. Due to reasons of cost
of runners fed into each cylinder's estimator, as shown in the
and technology maturity, however, the sensor most comblock diagram of Figure 7 [32]. For a switching O2 sensor, in
monly used in production is the piezoelectric accelerometer
knock sensor, which is usually mounted on the engine block
the method, the O2 signal sampled in the crankshaft angular
and generates vibration signals. Regardless of sensor type,
domain is first linearized, and then input into the signal array
the signals generated are inevitably contaminated by noise or
formation block to generate a signal array for the input of the
interference; thus, it is necessary to further process the raw
fuel richness estimator, which is used to estimate the absolute
signals to enhance the SNR to achieve reliable knock detecfuel-air ratio or its equivalent. Finally, the estimated fuel-air
tion. Additionally, the signals measured from knock-generated
ratio of each cylinder is used for OBD or fuel-air ratio balance
shock waves and vibrations are generally nonlinear and time
control. This method requires the installation of wide-range
varying. This makes knock detection more complicated and
sensors in each exhaust runner for the calibration of the fuel
challenging. Not only have automotive engineers been makrichness estimators.
ing significant efforts to develop practical and reliable knock
In summary, model-based methods typically originated from
detection systems, but many academic researchers have also
control theoretical concepts. They are relatively complicated in
been exploring advanced signal processing techniques for
design and calibration for mass production. Signal processing
knock detection. The result has been that knock detection has
methods (and the frequency filtering methods, in particular) are
become a very fruitful area for application of signal processing
more attractive to automotive manufacturers because of their
techniques in engine control and diagnostics.
implementation simplicity and relatively lower computational
Various knock detection methods that use signal processing
load and calibration efforts. Most filter parameters can be deterhave been intensively studied in recent decades. The methods
mined in the system development stage, instead of being calimainly use frequency-domain filtering, DFT [38], [39], power
brated in the vehicle release stage. This is a strong advantage for
spectrum analysis [40], time-frequency analysis [41]-[43], and
application of signal processing techniques.
wavelet transform [44]-[46]; refer to [35] for reviews of the
early
generation of knock detection algorithms used by the
Knock detection
automotive
industry.
Minimizing engine exhaust emissions and improving fuel
Knock detection based on frequency-domain filtering has
economy are two challenges in modern engine controls. One
been widely used in automotive production because of its simapproach to improving fuel economy is to maximize engine
plicity and low cost. Even today, it is still the main approach
efficiency by optimally controlling spark timing under given
used in mass production automotive applications. As shown
operating conditions [34]. Abnormal combustion, such as
in Figure 8, typically, filtering methods use bandpass filters
combustion that occurs too early, will result in a knocking
to retrieve the knock-based resonance signatures after preprophenomenon that not only limits engine power generation but
cessing of the signal from a knock sensor, such as antialiasing
also causes damage to engine pistons, rings, exhaust valves,
filtering and signal amplifying, use an envelope detector (or
and other engine components. Thus, knock detection is crurectifier) to obtain the knock intensity signal and then make a
cial for optimal spark timing and engine protection.
IEEE SIgnal ProcESSIng MagazInE

March 2017

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