Cylinder Pressure Reconstruction Based On Complex Radial Basis Function Networks From Vibration and Speed Signals

ARTICLE IN PRESS
Mechanical Systems
and
Signal Processing
Mechanical Systems and Signal Processing 20 (2006) 1923–1940
www.elsevier.com/locate/jnlabr/ymssp
Cylinder pressure reconstruction based on complex radial basis

function networks from vibration and speed signals
Roger Johnsson
Division of Sound and Vibration, Luleå University of Technology, SE-971 87 Luleå, Sweden
Received 9 November 2004; received in revised form 8 September 2005; accepted 8 September 2005
Available online 2 November 2005
Abstract
Methods to measure and monitor the cylinder pressure in internal combustion engines can contribute to reduced fuel
consumption, noise and exhaust emissions. As direct measurements of the cylinder pressure are expensive and not suitable
for measurements in vehicles on the road indirect methods which measure cylinder pressure have great potential value. In
this paper, a non-linear model based on complex radial basis function (RBF) networks is proposed for the reconstruction
of in-cylinder pressure pulse waveforms. Input to the network is the Fourier transforms of both engine structure vibration
and crankshaft speed fluctuation. The primary reason for the use of Fourier transforms is that different frequency regions
of the signals are used for the reconstruction process. This approach also makes it easier to reduce the amount of
information that is used as input to the RBF network. The complex RBF network was applied to measurements from a 6-
cylinder ethanol powered diesel engine over a wide range of running conditions. Prediction accuracy was validated by
comparing a number of parameters between the measured and predicted cylinder pressure waveform such as maximum
pressure, maximum rate of pressure rise and indicated mean effective pressure. The performance of the network was also
evaluated for a number of untrained running conditions that differ both in speed and load from the trained ones. The
results for the validation set were comparable to the trained conditions.
r 2005 Elsevier Ltd. All rights reserved.
Keywords: RBF network; Cylinder pressure; Pressure reconstruction; Engine performance assessment
1. Introduction
Combustion pressure waveform measurement and analysis plays an important role in the improvement of
performance, emission control, noise control and condition monitoring in internal combustion engines. Direct
measurement of the pressure inside the cylinder with a pressure transducer is not suitable outside laboratories
due to a number of limitations. For direct measurements a high performance pressure transducer has to be
used and the harsh environment in the cylinder causes the transducer to have a limited lifetime; which makes
the method expensive. It is also hard to find a good place to mount pressure transducers and effort has to be
made to avoid errors due to calibration, wave propagation, and deposits on the transducer. As a consequence,
Tel.: +46 0 920491791; fax: +46 0 920491030.

E-mail address: roger.johnsson@ltu.se.
0888-3270/$ - see front matter r 2005 Elsevier Ltd. All rights reserved.
doi:10.1016/j.ymssp.2005.09.003
ARTICLE IN PRESS
1924 R. Johnsson / Mechanical Systems and Signal Processing 20 (2006) 1923–1940
Nomenclature
y output signal of the RBF network

w weight matrix
w0 bias
wk weight vector that connects the kth centre to the output of the network
f(u,tk) kth radial basis function
u input vector
d target matrix
tk kth centre vector
K number of centres
N number of training sets
Ah start point for the Hanning window
Ae end point for the exponential window
efforts have been made over the last 20 years to find a stable and reliable method to reconstruct the pressure
waveform from indirect measurements. Two different approaches have been investigated, vibration
measurement-based reconstruction and crankshaft angular speed measurement-based reconstruction.
It has been shown that both vibration- and angular speed signals contain information about the cylinder
pressure but mainly in different frequency regions [1,2]. Angular speed fluctuation comes from the low-
frequency content of cylinder pressure as angular velocity is not as sensitive as structural vibrations to
instantaneous pressure changes [2]. Vibration measurements, on the other hand, suffer from low signal-to-
noise ratio at low frequencies, making the reconstruction of the compression/decompression phase of the
combustion process uncertain. In both cases the relationships between cylinder pressure and vibration
response and crankshaft speed is non-linear and changes with the running condition.
The objective of this paper is to derive and evaluate a non-linear model for the cylinder pressure
reconstruction based on complex-valued radial basis function (RBF) network using both the vibration and
speed signal. Earlier studies using RBF networks for the pressure reconstruction have shown the potential of
RBF networks [3–5]. Input to the networks has either been the crankshaft speed signal (in time domain) [3,4]
or the power spectrum of the engine structure vibration [5]. In the study presented in this paper the input to the
network is the Fourier transforms of the vibration and speed signals. By using the Fourier transforms of the
measured signals the amount of data can easily be reduced and the interesting frequencies for each
measurement method be selected. The use of the complex Fourier transforms also simplifies the task for the
RBF network since for example a small change in the phase for one frequency component will only affect that
frequency component in the frequency domain while the shape of the signal can change significantly in the
time domain.
The proposed method is derived and tested on a 6-cylinder ethanol powered diesel engine by comparing a
few commonly used parameters of the pressure waveform was compared between the reconstructed and
measured pressure waveform. The parameters are maximum pressure (pmax) and its location, maximum
pressure rise and its location, and indicated mean effective pressure (IMEP).
2. Background
2.1. Engine vibration measurements
The quick pressure change in a cylinder during combustion gives rise to engine structure vibrations. These
vibrations contain information about the combustion process and can easily be measured; normally on the
cylinder head or the engine block. Unfortunately the vibration signal also contains non-combustion related
vibrations that decrease the signal-to-noise ratio. Examples of unwanted vibration sources are piston slap,
ARTICLE IN PRESS
R. Johnsson / Mechanical Systems and Signal Processing 20 (2006) 1923–1940 1925
valve impacts and gear transmissions. Two methods have been proposed for the pressure reconstruction from
vibration measurements; inverse filtering [1,2,6–11] and neural networks [5].
Inverse filtering is based on the assumption that the engine structure can be modelled as a linear time-
invariant (LTI) system and that the measured vibration signal is a non-disturbed response to the cylinder
pressure. One difficulty with this method is that the calculation of the inverse of the FRF in the frequency
domain is an ill-conditioned operation. To obtain a more robust solution, different methods have been
investigated such as cepstral analysis [1,2,7–10] and time domain smoothing [11]. Another problem is that the
FRF changes both with speed and load conditions, indicating that the system is non-linear. For this reason
Zurita used a matrix of transfer functions for different running conditions [10].
Neural networks are non-linear and are therefore suitable for modelling the non-linear relation between the
cylinder pressure and engine structure vibration. Du [5] used a RBF network to predict the cylinder pressure
from the power spectrum of the vibration signal.
2.2. Crankshaft angular speed measurements
When a cylinder fires, the combustion torque will become greater than the external torque which then causes
the crankshaft to accelerate. As the next cylinder goes into compression the torque will drop below the
external torque and the crankshaft speed will decrease. The result is a fluctuating waveform of engine speed
versus crank angle that contain information about the cylinder-by-cylinder combustion pressure that
produced it. Several different methods have been proposed and investigated for pressure reconstruction;
mathematical engine models [12–16], pattern recognition [17], neural networks [3,4,18] and inverse filtering
[1,19].
Mathematical engine models are based on solving the torque balance equation numerically [12–16].
Drawbacks with methods based on engine models are the simplifications and assumptions that have to be
done in order to solve the torque balance equation. One common simplification is to use models with constant
polar moment of inertia. However, Shiao [20] showed that the moment of inertia is highly angular dependent,
especially for engines with few cylinders. Moskwa et al. [21] introduced a method to compensate for the
varying moment of inertia and thereby achieve what they call ‘‘synthetic’’ speed. The ‘‘synthetic’’ speed is the
speed that would have been measured if the moment of inertia was constant. The relation between ‘‘synthetic’’
speed and cylinder pressure is linear, which simplifies the modelling. If the ‘‘synthetic’’ crankshaft speed is
calculated, the cylinder pressure can be reconstructed using a FRF between the ‘‘synthetic’’ speed and cylinder
pressure. This approach has been investigated by Moro [18] who was able to reconstruct the first 3 harmonics
over 1801 for a four cylinder engine. Johnsson [1] combined speed measurements for reconstruction of the low
frequencies of the cylinder pressure and cylinder head vibration measurements for the higher frequencies in
order to extend the frequency region of the reconstructed cylinder pressure.
Pattern recognition uses a knowledge base of real measurements for a number of different engine running
conditions. A measured speed signal is then compared with the signals in the knowledge base. Because not all
running conditions can be saved in the knowledge base some interpolation technique has to be introduced to
overcome this limitation. Brown [18] used pattern recognition together with linear interpolation to predict the
peak pressure for a given speed and load both, with and without under-fueling. Using an appropriate non-
linear interpolation technique could make it possible to use the method for a larger set of running conditions.
Johnsson [19] used a simple feedforward neural network to predict the maximum pressure and start of
combustion (SOC) from the crankshaft speed fluctuation for a 6-cylinder diesel engine. A more sophisticated
method using a RBF network was developed by Jacob et al. [3] where the pressure waveform was recovered.
Gu et al., from the same research group as Jacob, was able to reconstruct the maximum pressure with an error
of less than 5% [4].
3. Analysis of pressure, vibration and speed signals
Earlier in this paper it was noted that the vibration and speed measurements contain information about the
pressure in different frequency regions. The crankshaft speed contains the low-frequency information of the
cylinder pressure while the vibration signal contains the higher frequencies [1]. A coherence analysis was
ARTICLE IN PRESS
performed to show if this is the case for the engine used in this study. Fig. 1 shows the time signals for pressure,
vibration and speed for the investigated 6-cylinder ethanol powered diesel engine for two different running
conditions, for the experimental set-up see Section 5. In Fig. 2, the same signals are presented but as the
magnitude of their Fourier transforms. As can be seen, the frequency content of the cylinder pressure changes
with running conditions. As the load increases the high frequency part of the pressure pulse decreases and the
time signal becomes smoother. The coherence between cylinder pressure and vibration and speed for the
Fig. 1. Time signals for cylinder pressure, engine structure vibration and angular speed for 1500 rpm and (a) 25% load, (b) 75% load.
Fig. 2. Magnitude of cylinder pressure, engine structure vibration and angular speed for 1500 rpm and (a) 25% load, (b) 75% load.
ARTICLE IN PRESS
Fig. 3. Coherence between cylinder pressure and engine structure vibration (solid) and angular speed (dotted) for 1500 rpm and (a) 25%
load, (b) 75% load.
Fig. 4. Frequency response function for different running conditions (solid 1200 rpm and 25% load, dotted 1500 rpm and 50% load,
dashed 1800 rpm and 100% load), (a) vibration and (b) speed.
running conditions of Figs. 1 and 2 is shown in Fig. 3. In Fig. 4 frequency response functions for both the
vibration and speed approaches in different running conditions are presented.
From the examples shown in Figs. 3 and 4 some conclusions can be drawn. The engine structure vibration
and the crankshaft angular speed contain information about the cylinder pressure mainly in different
frequency regions. The crankshaft angular speed variation has the highest coherence with the cylinder pressure
for the lowest frequencies while the engine structure has the highest coherence for higher frequencies. It is also
obvious that the frequency response function both for the vibration and angular speed approach changes with
running condition proving that the relations are non-linear. By combining both engine structure vibration and
crankshaft angular speed measurements with a method capable of non-linear modelling the cylinder pressure
reconstruction should be possible over a wide range of running conditions. The approach in this paper is to
use a complex RBF network for the cylinder pressure reconstruction. By transforming the signals to the
frequency domain the frequency region selection for each signal is easy and at same time the data amount is
reduced.
4. Complex-valued radial basis function network
The RBF network is a subgroup of the multilayer feedforward neural networks. One advantage with RBF
networks is that they can be trained in a more straightforward manner than is the case for neural networks
trained with the back-propagation algorithm. However, the RBF network can become larger than the network
trained with back-propagation in order to achieve the same accuracy.
ARTICLE IN PRESS
Fig. 5. Schematic of RBF network, with a single output [23].
Ordinarily, the development of RBF networks assumes real-valued input and output data. In this section a
more general complex-valued RBF network [22] will be discussed, where both inputs and outputs are complex-
valued. The real-valued RBF network is a special case of this more general RBF network. The topology of the
complex RBF network is shown in Fig. 5. The centres (tk) and weights (wk) are complex while the output from
the hidden layer is real-valued.
The input–output mapping for the RBF network (in this case with a single output) can be described as, with
notations as depicted in Fig. 5
X
K
y¼ wk jðu; tk Þ þ w0 . (1)
k¼1
The term j(u;tk) is the kth RBF that computes the similarity between an input vector u and its own centre tk.
The output from the RBF-unit is a non-linear function of the distance between u and tk. w0 is a bias value
while w1 to wK are the weights (scaling factors) to the outputs from each RBF-unit. The output (real-valued)
from a hidden unit is calculated as

1
jðu; tk Þ ¼ exp 2 ðu tk ÞH ðu tk Þ , (2)
sk
where s is the width of the Gaussian function and both input vector u and centre tk are complex-valued. The
exponent H stands for the Hermitian operation and is the conjugate transpose of the matrix. The complex
output from the RBF network is calculated by
XK
1 H
y¼ wk exp 2 ðu tk Þ ðu tk Þ þ w0 . (3)
k¼1
sk
The width s (real-valued) of the centres for each hidden unit can, both in the real- and complex-valued case,
either be different for each unit or set to a fixed value. A fixed value however, will result in a simpler training
strategy.
The simplest training algorithm is to randomly choose a number K of training examples and use these as K
radial basis centres (tk) in the hidden layer units. The weights can then be calculated by the method of least
ARTICLE IN PRESS
squares. This approach assumes that the training data is distributed in a representative manner. For this
method it is common to set the width of the Gaussian function to a fixed value and the weights in the output
layer is computed with the method of least squares [23].
A shortcoming with this approach is that it does not guarantee an accurate performance. Furthermore, for a
given level of performance, this approach often results in an unnecessarily large network and, as a
consequence, causes numerical ill-conditioning [22]. Several more advanced methods have been suggested for
the selection of centres, such as:
Recursive hybrid learning [22,23]

Stochastic gradient approach [23]
Orthogonal least-squares learning [22,24].
In this study a version of the recursive hybrid learning is applied. It consists of k-means clustering to find the
centres and the use of regularisation in the process to determine the weights. The k-means clustering divides
the training set into k clusters based on the similarities in the training cases. The regularisation will increase the
smoothness of the solution and thereby give a more robust solution.
4.1. k-means clustering
The k-means clustering algorithm computes k centres and thereby partitions the input data into k clusters.
k-means clustering treats each observation as an object having a location in space. It finds clusters where the
objects within a cluster are as close to each other as possible while the clusters are as well separated as possible.
Each cluster is defined by its members and its centre. The centre in each cluster is the object to which the sum
of distances from all objects in that cluster is minimised.
k-means clustering is an iterative algorithm that minimises the sum of distances from each object to its
cluster centre by moving objects between the clusters until the sum cannot be decreased any further. The result
is k clusters that are as compact and separated from each other as possible. One disadvantage with the k-
means clustering algorithm is that it can reach a local minimum which means that the final result depends on
the initial choice of cluster centres.
4.2. Regularisation and least squares method
The risk that the network learns features that are specific to the training data (noise in the training data),
resulting in decreased generalisation, can be reduced by introducing a regularisation parameter, l. The idea is
to introduce a penalty term that increases the smoothness of the solution whereby the effect of outliers will be
reduced. One effect of the regularisation is that the effective number of degrees of freedom in the model is
reduced. This makes it less flexible and reduces the possibility of a rough output; the solution has become
smoother. The regularisation parameters li is added to the cost function, which is defined as the sum squared
error [25].
X
N X
K
C¼ ðd i yi Þ2 þ lj w2j . (4)
i¼1 j¼1
The optimal solution for the weight matrix that minimises the cost function is given by the equation
w ¼ ðUT U þ lIK Þ1 UT d. (5)
In order to determine the optimum value of the regularisation parameter l the generalised cross validation
(GCV) [25] is used.
NdT P2 d
s2GCV ¼ , (6)
ðtraceðPÞÞ2
ARTICLE IN PRESS
where the projection matrix is defined as
P ¼ I N UðUT þ lIK Þ1 UT . (7)
5. Experimental set-up
The measurements were carried out on a 9-litre, 6-cylinder, inline, four stroke, turbo charged and ethanol
powered diesel engine in the engine laboratory at Luleå University of Technology. In order to minimise the
effect of unwanted signals in the measured vibration signal it is important to find a measurement point that is
dominated by the forcing function (e.g. cylinder pressure). Based on the results from earlier studies on similar
engines, one of the cylinder head bolts was selected as the measurement point [10].
The cylinder pressure was measured with a pressure transducer (AVL QC33C) mounted in cylinder 1. One
accelerometer (PCB 353M15) was mounted firmly to one of the cylinder head bolts on cylinder 1 and used to
measure the engine structure vibrations. The crankshaft speed was measured by using an angular sensor (AVL
365C) with 1800 pulses/revolution. The sampling of vibration and pressure was controlled by the angular
pulses (e.g. sampling frequency was 1800 pulses/revolution meaning that the sampling frequency in Hz varies
with the engine speed). This means that all measurements have been done in the angle domain (with Dy ¼ 0:21)
rather than the common time domain. All data collection was controlled by LabVIEW from National
Instruments, Fig. 6.
An anti-aliasing filter set to 10 kHz was used to filter the vibration signal. The speed fluctuation was
measured by counting the number of time pulses between the angular pulses using a counter with a time base
of 80 MHz. The speed resolution with this measurement set-up can be calculated by Eq. (8) [26] and is shown
in Fig. 7. At 2000 rpm the speed resolution is 1.5 rpm which is considered acceptable for this purpose.
60fn
s¼ n, (8)
60f Nn
where s is the speed resolution; f the sampling frequency (Hz); n the speed (rpm); N the number of angle
pulses/revolution.
Measurements for the training process were taken at speeds between 800 and 2000 rpm and loads between
10% and 90% of full load for a total of 39 different running conditions as depicted in Fig. 8. Each
measurement consists of pressure, vibration and speed over an entire engine cycle, measured with an angle
resolution of 0.21. For each running condition 50 engine cycles were measured. Thirty of the measurements for
each running condition were used for the training of the RBF network and 20 used for the verification of the
training process. For the validation, an additional validation set with 8 running conditions that differ from the
training set was measured, see Fig. 8.
Fig. 6. Experimental set-up.

ARTICLE IN PRESS
Fig. 7. Speed resolution for the measurements with 1800 pulses per revolution and a clock rate of 80 MHz.
Fig. 8. Measured running conditions (*—used for training, J—used for validation).
6. Pre-processing
The input to the RBF network is the Fourier transform of both engine structure vibration and crankshaft
speed fluctuation. As already mentioned the reason is that the crankshaft speed fluctuation is dominated by
the low-frequency content of the cylinder pressure while the structure vibration is dominated by the high
frequency part of the cylinder pressure which is mainly related to the high pressure rise due to rapid
combustion.
6.1. Cylinder pressure
All measured cylinder pressures in the training set were limited to 7601 of the crank angle around top dead
centre (TDC). With an angle resolution of 0.21 this means that one pressure pulse has 600 samples. Since the
cylinder pressure waveform was to be reconstructed, it was not transformed to the frequency domain as
vibration and speed signals. Implied is that the input to the RBF-network is complex-valued while the output
is real-valued.
ARTICLE IN PRESS
6.2. Vibration signal
A weighting function proposed by Zurita [10] was used to filter out the wanted vibration signal around
TDC. The aim with the window is to reduce noise and leakage effects and to minimise the content of non-
combustion related vibrations. The window consists of zero padding, a half-Hanning window and an
exponential window, Fig. 9.
The window is described by
W G ðaÞ ¼ ½Zero_padding; W H ðaÞ; W E ðaÞ, (9)
where

1 2pa
W H ðaÞ ¼ 1 þ cos Ah pao0; (10)
2 2Ah þ 1
W E ðaÞ ¼ eðaa0 Þ=t a0 ¼ 0; 0paoAe . (11)

The crossover point between the Hanning and exponential window was found by Zurita [10] to be set to
TDC. The argument was that the zero-padding and Hanning window should handle the compression part of
the signal while the exponential window should take care of the combustion part of the signal. A coherence
analysis of measurements on the engine used in this paper shows that at least for this engine the transition
between the Hanning and exponential window should be set to 61 after TDC as illustrated in Fig. 10. An
explanation to the increased coherence is that the new position better captures the vibration signal related to
the quick pressure rise as illustrated in Fig. 11.
Fig. 9. The window proposed by Zurita [10].
Fig. 10. Coherence between cylinder pressure and engine structure vibration for 1200 rpm and 25% load and 1800 rpm and 25% load.
Dashed line—Zurita’s window with the original transition point between half-Hanning and exponential window at TDC. Solid line—the
transition point moved to 61 after TDC.
ARTICLE IN PRESS
Fig. 11. Comparison of the suggested (solid) and original (dashed) position of the Zurita window to the vibration signal.
For high speeds there is also another problem; the vibration pulse does not die out completely before the
next combustion starts. Due to the short time period between combustion pulses the vibration pulses will
overlap each other. This effect, is however, minimised by this window thanks to the zero-padding at the
beginning of the window.
The Fourier transforms of the vibration signal was calculated and frequency components from the 21 to 50
harmonic order over 1801 was used as one part of the input to the RBF-network. The harmonic order over
1801 is defined as the number of cycles over 1801 of the crank angle.
6.3. Crankshaft speed
Before the Fourier transform of the speed signal could be calculated the mean speed had to be removed. A
rectangular window with half-Hanning tapers in each end was applied to reduce the leakage effect. The
Fourier transforms of the speed signal was calculated and frequency components of the 0–20 harmonic order
over 1801 was used as the second part of the input to the RBF-network.
7. Results
Several runs were performed using the training data to train the RBF network. The number of clusters as
well as the width of the Gaussian function was varied during the training process. The training process is a
trade-off between reconstruction accuracy of the training data and the network’s abilities to generalise the
knowledge to the untrained data (validation set). By aiming for a high reconstruction accuracy with the
training set there is a higher risk that the network will overfit the trained data (e.g. fit the noise) and thereby
lose some of its ability to generalise its knowledge to unknown conditions. The accuracy of the training set can
often be improved by increasing the number of hidden units in the RBF-network. The network’s performance
for unknown conditions, on the other hand, may decrease due to the risk that the network is too closely fitting
noise to that in the training set. The best compromise between accuracy and the RBF network’s ability to
generalise its knowledge to unknown conditions was found with 39 centres and a width of 6000. A difficulty
with k-means clustering is that the clusters derived, to some extent, are dependent on the initial choice of
cluster centres.
Figs. 12 and 13 compare the measured and reconstructed pressure waveform for a selection of the training
set and validation set. It is clear that the network has learned the pressure waveforms in the training set well
and is able to extend this knowledge to the untrained conditions. A more detailed analysis will be performed in
the remainder of this section. Parameters that were investigated are:
ARTICLE IN PRESS
Fig. 12. Measured (solid line) and reconstructed (dotted line) cylinder pressure waveform for a few of the running conditions in the
verification set. Each graph shows five measured and five reconstructed pressure waveforms.
(a) Maximum pressure (pmax)

(b) Location of the maximum pressure
(c) Maximum rate of pressure rise
(d) Location of maximum pressure rise
(e) Indicated mean effective pressure (IMEP).
7.1. Maximum pressure and its location
Figs. 14 and 15 show the differences between measured and reconstructed maximum pressure (pmax)
expressed in bars and percentages for the verification and validation set. For the verification set, the vast
majority of the prediction errors in the pmax prediction were less than 5% (5 bars). Two of the running
conditions (1800 rpm at 50% and 90% of full load) in the training set were not well learned. The reason is
most likely due to the forming of the clusters. For the validation set the prediction error was slightly higher,
the network seems to most often underestimate the pmax for these running conditions. The rms error for the
verification set was 2.9% (3.2 bars) and for the validation set was 3.4% (3.5 bars).
ARTICLE IN PRESS
Fig. 13. Measured (solid line) and reconstructed (dotted line) cylinder pressure waveform for a few of the running conditions in the
validation set. Each graph shows five measured and five reconstructed pressure waveforms.
Fig. 14. Deviation between measured and reconstructed maximum pressure for the verification set.
The location of the maximum pressure for the reconstructed pressure waveform differs up to 61 from the
measured position of the validation set as shown in Fig. 16. A reason for the discrepancy is that the
reconstructed waveform was smoother than the measured cylinder pressure, Fig. 17a. This was caused by the
ARTICLE IN PRESS
Fig. 15. Deviation between measured and reconstructed maximum pressure for the validation set.
Fig. 16. Comparison of location of maximum pressure for verification set (upper graph) and validation set (lower graph).
limited number of centres in the RBF network but was also a result of the regularisation. A second
explanation is that the pressure waveform is almost constant over a wide range for some running conditions
(for high loads especially at higher speeds), Fig. 17b.
ARTICLE IN PRESS
Fig. 17. Typical pressure waveforms for (a) low loads and (b) high loads. Solid line—measured cylinder pressure. Dashed line—
reconstructed cylinder pressure.
Fig. 18. The maximum rate of pressure rise for the measured cylinder pressure, the error for the predicted maximum rate of pressure rise,
and the difference between the position of maximum rate of pressure rise for the measured and predicted cylinder pressure waveform.
7.2. Maximum rate of pressure rise and its location
A correct prediction of the rate of pressure rise during the combustion phase is an important factor for a
correct calculation of the heat release. The maximum rate of pressure rise is highly dependent on the running
condition. For the engine tested in this paper the factor between the highest and lowest maximum rate of
pressure rise was more than five. As the load increased the rate of maximum pressure rise tended to decrease.
For high loads it was not always possible to determine a distinct pressure rise as there were very flat pressure
waveforms as may be seen in Fig. 17b.
The difference between the measured and predicted maximum rate of pressure rise and its location is shown
in Fig. 18. The error for the validation set (rms error 2.6 bars/1) was slightly higher than for the verification set
(rms error 2.4 bars/1). There can be many reasons for the high errors; first of all, to calculate the differential of
a function is itself an ill-conditioned operation [27]. The determined pressure rate is also highly dependent on
noise in the measured signal and is affected by the regularisation.
ARTICLE IN PRESS
Fig. 19. Difference between IMEP calculated from measured and predicted cylinder waveform over an angle region 7601 around TDC.
Upper graph shows the verification set and the lower graph shows the validation set.
The ability to determine the location of maximum rate of pressure rise was more accurate than the
prediction of the location of maximum pressure. For the verification set all deviations were nearly within 711
of the crank angle and for the validation set 721 of the crank angle.
7.3. Indicated mean effective pressure
IMEP is the only parameter among the ones compared in this paper that was calculated from the pressure
waveform. IMEP is the constant pressure that exerted over a complete cycle would produce the same work as
the measured pressure waveform. In a plot of cylinder pressure against volume, IMEP is the area contained by
the compression and expansion stroke divided by the displacement volume [28].
Z
1 y2 dV ðyÞ
IMEP ¼ PðyÞ dy. (12)
V y1 dy
In this case the IMEP has been calculated over the angle region 7601 around TDC, because this is the region
where the pressure waveform has been reconstructed. The prediction error for the IMEP is shown in Fig. 19.
The error was for almost the entire verification set within 71 and 72 bars for the validation set.
8. Conclusions
The use of both engine structure vibration and crankshaft speed fluctuation measurements as input to a
RBF network for cylinder pressure reconstruction has shown promising results. The RBF network was not
only able to learn the training set but was also able to generalise this knowledge to unknown running
conditions. The network was able to reconstruct the correct pressure waveform even though the waveform
changed dramatically with variations in running conditions.
The accuracy of the predictions has been evaluated by determining a number of pressure waveform related
parameters. The accuracy can be summarised as follows:
ARTICLE IN PRESS
(a) The rms-error for predicted maximum pressure for the entire validation set was 3.5 bars.
(b) The rms-error for predicted location of maximum pressure for the entire validation set was 1.51
(c) The rms-error for the predicted maximum rate of pressure rise for the entire validation set was 2.6 bars/1
(d) The rms-error for the predicted location of maximum rate of pressure rise for the entire validation set was
0.91
(e) The rms-error for IMEP calculated from the predicted pressure waveform was 0.7 bars.
In order to increase the accuracy in reconstructed pressure waveforms a larger RBF network will be needed.
In the study reported in this paper a network with only 39 centres was derived. As a consequence, a much
larger training set will be needed for a higher level of accuracy. A recommendation is to have as many different
running conditions as possible and a more limited number of measurements at each running condition instead
of having few running conditions and many measurements at each one. It is also important to include all
extreme running (high and low speeds, and loads) conditions since neural networks generally are good at
interpolating between the learned examples but worse at extrapolating this knowledge.
For the pre-processing of the measured signals it was determined that at least for this engine the proposed
window by Zurita [10] should be changed in such a way that the transition point between the Hanning and
exponential window is moved to 61 after TDC. This is because the new placement yields a higher coherence for
all measured running conditions than does the original placement.
References
[1] R. Johnsson, Indirect measurements for control and diagnostics of IC engines, Ph.D. Thesis, Luleå University of Technology,
Sweden, 2004.
[2] Y. Ren, Detection of knocking combustion in diesel engines by inverse filtering of structural vibration signals, Ph.D. Thesis,
University of New South Wales, Australia, 1999.
[3] P.J. Jacob, F. Gu, A.D. Ball, Non-parametric models in the monitoring of engine performance and condition. Part 1: modeling of
non-linear engine processes, Proceedings of the Institution of Mechanical Engineers Part D-Journal of Automobile Engineering 213
(D1) (1999) 73–81.
[4] F. Gu, P.J. Jacob, A.D. Ball, Non-parametric models in the monitoring of engine performance and condition. Part 2: non-intrusive
estimation of diesel engine cylinder pressure and its use in fault detection, Proceedings of the Institution of Mechanical Engineers Part
D-Journal of Automobile Engineering 213 (D1) (1999) 135–143.
[5] H. Du, L. Zhang, X. Shi, Reconstructing cylinder pressure from vibration signals based on radial basis function networks,
Proceedings of the Institution of Mechanical Engineers Part D-Journal of Automobile Engineering 215 (D6) (2001) 761–767.
[6] P.M. Azzoni, G. Cantoni, G. Minelli, D. Moro, Indirect pressure measurement in a small diesel engine, Small Engine Technology
Conference, Pisa, September 1–3, 1993, pp. 391–401.
[7] G.D. Cassini, W. D’Ambrogio, A. Sestieri, Frequency domain vs. cepstrum technique for machinery diagnostic and input waveform
reconstruction, ISMA21, Leuven, 1996, pp. 835–846.
[8] R.B. Randall, Y. Ren, H. Ngu, Diesel engine cylinder pressure reconstruction, ISMA21, Leuven, 1996, pp. 847–856.
[9] P. Azzoni, A. Paniero, Cylinder pressure reconstruction by cepstrum analysis, 22nd International Symposium on Automotive
Technology and Automation, Florence, May 14–18, 1990, pp. 629–637.
[10] V.G. Zurita, Vibration based diagnostics for analysis of combustion properties and noise emissions of IC engines, Ph.D. Thesis, Luleå
University of Technology, Sweden, 2001.
[11] Y. Gao, R.B. Randall, Reconstruction of diesel engine cylinder pressure using a time domain smoothing technique, Mechanical
Systems and Signal Processing 13 (5) (1999) 709–722.
[12] S.J. Citron, J.E. O’Higgens, L.Y. Chen, Cylinder by cylinder engine pressure and pressure torque waveform determination utilizing
speed fluctuations, SAE Transactions, Section 3, SAE 890486, 1989, pp. 933–947.
[13] K. Iida, K. Akishino, K. Kido, IMEP estimation from instantaneous crankshaft torque variation, SAE Transactions 99 (SAE
900617) (1990) 1374–1385.
[14] G. Rizzoni, F.T. Connolly, Estimate of IC engine torque from measurement of crankshaft angular position, SAE Transactions,
Section 3, SAE 932410, 1993, pp. 1937–1947.
[15] S.K. Chen, S. Chen, Engine diagnostics by dynamic shaft measurement—A progress report, SAE Transactions, Section 3, SAE
932412, 1993, pp. 1964–1979.
[16] F.T. Connolly, A.E. Yaggle, Modeling and identification of the combustion pressure process in internal combustion engines,
Mechanical Systems and Signal Processing 8 (1) (1994) 1–19.
[17] D. Moro, N. Cavina, F. Ponti, In-cylinder pressure reconstruction based on instantaneous engine speed signal., Journal of
Engineering for Gas Turbines and Power 124 (2002) 220–225.
ARTICLE IN PRESS
[18] T.S. Brown, W.S. Neill, Determination of engine cylinder pressure from crankshaft speed fluctuations, SAE Paper 920463, 1992, pp.
771–779.
[19] R. Johnsson, A. Ågren, M. Klopotek, Prediction of points and tendencies of the pressure waveform from crankshaft speed
measurements, ISMA 25, Leuven, September 13–15, 2000, pp. 885–890.
[20] Y. Shiao, C.H. Pan, J.J. Moskwa, Advanced dynamic spark ignition engine modeling for diagnostics and control, International
Journal of Vehicle Design 15 (6) (1994) 578–596.
[21] J.J. Moskwa, D.J. Bucheger, A new methodology for use in engine diagnostics and control, utilizing ‘‘synthetic’’ engine variables:
theoretical and experimental results, Journal of Dynamic Systems, Measurement, and Control 123 (2001) 528–534.
[22] S. Chen, S. McLaughlin, B. Mulgrew, Complex-valued radial basis function network. Part 1: Network architecture and learning
algorithms, Signal Processing 36 (1994) 175–188.
[23] S. Hykin, Adaptive Filter Theory, third ed, Prentice-Hall, Englewood Cliffs, New Jersey, 1996.
[24] S. Chen, C.F.N. Cowan, P.M. Crant, Orthogonal least squares learning algorithm for radial basis function networks, IEEE
Transactions on Neural Networks 2 (1991) 302–309.
[25] M.J.L. Orr, Introduction to Radial Basis Function Networks, Centre for Cognitive Science, University of Edinburgh, 1996.
[26] R. Johnsson, Crankshaft speed measurements and analysis for control and diagnostics of diesel engines, Licentiate Thesis, Luleå
University of Technology, Sweden, 2001.
[27] H.W. Engl, M. Hanke, A. Neubauer, Regularization of Inverse Problems, Kluwer academic publishers, Dordrecht, 1996.
[28] D.R. Lancaster, R.B. Krieger, J.H. Lienesch, Measurement and analysis of engine pressure data, SAE preprints, SAE 750026, 1975,
pp. 155–172.

Cylinder Pressure Reconstruction Based On Complex Radial Basis Function Networks From Vibration and Speed Signals

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Cylinder Pressure Reconstruction Based On Complex Radial Basis Function Networks From Vibration and Speed Signals

Uploaded by

Copyright:

Available Formats

ARTICLE IN PRESS

Cylinder pressure reconstruction based on complex radial basis

Tel.: +46 0 920491791; fax: +46 0 920491030.

y output signal of the RBF network

2.1. Engine vibration measurements

2.2. Crankshaft angular speed measurements

3. Analysis of pressure, vibration and speed signals

4. Complex-valued radial basis function network

Fig. 5. Schematic of RBF network, with a single output [23].

Recursive hybrid learning [22,23]

4.1. k-means clustering

4.2. Regularisation and least squares method

where the projection matrix is deﬁned as

P ¼ I N UðUT þ lIK Þ1 UT . (7)

Fig. 6. Experimental set-up.

6.1. Cylinder pressure

6.2. Vibration signal

W E ðaÞ ¼ eðaa0 Þ=t a0 ¼ 0; 0paoAe . (11)

Fig. 9. The window proposed by Zurita [10].

6.3. Crankshaft speed

(a) Maximum pressure (pmax)

7.1. Maximum pressure and its location

7.2. Maximum rate of pressure rise and its location

7.3. Indicated mean effective pressure

You might also like