Download 3681_0_DMPaper2002

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Document related concepts

Types of artificial neural networks wikipedia , lookup

Transcript
Neural Networks in Automotive Engine Tuning
Vong Chi-man, Wong Pak-kin, Li Yi-ping
Faculty of Science and Technology, University of Macau
ABSTRACT
Automotive engine performance is significantly affected with effective tuning. Current engine tuning relies on the
experience of the mechanics, and it is usually done by trial-and-error method. It may take days or even weeks, and may
even fail to tune the engine optimally because a formal performance model of the engine has not been determined yet.
With the aid of data mining methods, the engine tuning can be done automatically through the use of a computer. In
addition, a nearly optimal engine setting can be obtained which dramatically increases the engine performance. An
experiment has also been done to verify the usefulness of data mining methods in this application area.
1. INTRODUCTION
Automotive engine performance is significantly affected with effective tuning. In general, the original car manufacturers fix
the control parameters on most car engines with electronic fuel-injection systems. It is impossible to adjust the setting for
different engine conditions and driver practices. With the recent development of low-cost electronic components,
programmable electronic control unit (ECU) is widely adopted by racing and super production car engines. The unit
substitutes for the OEM ECU and allows the user to adjust the engine parameters via a PC.
Current technique of engine parameters tuning only relies on the experience of the mechanics, and it is usually done by
trial-and-error method. It may take days or even weeks, and may even fail to tune the engine optimally because a formal
performance model of the engine has not been determined yet. The difficulty of building such a performance model is due
to the set of parameters, which incurred highly nonlinear correlations.
With the emerging data mining techniques, the engine parameters can be tuned to achieve a sub-optimal performance
even its model is not known.
2. AUTOMOTIVE ENGINE PARAMETERS
Currently, one of major goals in modern engine tuning is the careful selection of air/fuel ratio "Ø" for different driving
conditions [5]. Actually the performance of the electronic fuel-injection engine mainly depends on air/fuel ratio control.
For achieving maximum engine performance, various air/fuel ratios should be produced for different engine speeds and
throttle positions. Some references [5,6] recommended the following air/fuel ratios for running four-stroke naturally
aspirated engines optimally.
Ø
6
Running condition
11.5
12.2
13.3
14.7
Rich burn limit (cold start and fast idle
speed)
Approximate rich best torque at wide-open
throttle (for acceleration)
Safe best power at wide-open throttle
Approximate lean best torque
(Stoichiometric ratio) maximum conversion
efficiency for catalytic converters
Lean cruise
Best fuel economy
Electronic fuel injection lean burn limit
15.5
16.5
22
* 2-6% variation of Ø is acceptable
Equations (1) & (2) show the relationship of adjustable engine parameters in the modern programmable ECU and air/fuel
ratio.
Ø = 0.147λ
  f ( p, s , a,
(1)
p
, b, r , h, i, j , c)
t
(2)
where
s:
p:
a:
b:
r:
h:
i:
j:
c:
t:
λ:
engine speed
throttle position/ manifold air pressure
acceleration enrichment
temperature compensation
rate of exhaust gas recirculation
altitude compensation
ignition spark advance
fuel injection time
fast idle time
time
% of stoichiometric
Adjusting the parameters in Equation (2) can produce different air-fuel ratios. However, the multivariable function in
Equation (2) is an unrecognized pattern – an unknown performance model. Currently, all tuning processes on equation (2)
are manually performed by trail-and-error method. The results totally depend on the mechanic’s experience. In addition,
different automotive engines have different performance models. That arises another obstacle for engine tuning.
3. DATA MINING
Data mining is an emerging technique that is also considered to be one of the ten most contributive scientific technologies
in the future 20 years. Nowadays, more and more companies are employing data mining techniques [1,2] to digest the
huge amount of data that they own, in order to find out valuable hidden knowledge. This knowledge is usually difficult or
even impossible to be noticed.
Data mining is a multidisciplinary field, drawing work from areas including database technology, machine learning, neural
networks, statistics, and data visualization. Because of this multidiscipline, data mining is a constitution of many methods,
addressing different kinds of knowledge discovery problems. Moreover, upon different kinds of data, different data mining
methods are employed.
In this paper, neural network (NN) – a method from the area of machine learning is employed to find out the sub-optimal
settings of selected parameters defined in equation (2).
3.1 NEURAL NETWORK
Neural network is a set of connected input/output units where each connection has a weight associated with it. Each of
these units is called a neuron. Neural network is a kind of supervised learning methods. The input of a neural network is
usually a large set of data. This set of data is called a training set. By passing this training set to the neural network, the
network learns by adjusting the associated weights of each neuron so as to be able to predict the correct class of the input
samples. As neural networks involve long training times, they are only suitable for batch-processing applications.
3.2 MULTILAYER FEED-FORWARD NEURAL NETWORK
A multilayer feed-forward neural network [4] contains three layers: input layer, hidden layer, and output layer. The inputs
are fed simultaneously into a layer of units making up the input layer. The weighted outputs of these units are, in turn, fed
simultaneously to a second layer of “neuronlike” units, known as a hidden layer. The weighted outputs of the hidden layer
are input to units making up the output layer, which emits the network’s prediction for given samples. The units in the
output layer are referred to as output units. Figure 1 shows an example of a feed-forward neural network.
Input
layer
Hidden
layer
Output
layer
.
.
.
.
.
.
x1
x2
.
.
.
xi
Oi
wij
Oj
wjk
Ok
Figure 1. A feed-forward neural network
3.3 MATHEMATICAL MODEL AND ALGORITHM
The algorithm used in neural networks is backpropagation [3]. Backpropagation learns by iteratively processing a set of
training samples, comparing the network’s prediction for each sample with the actual known class label. For each training
sample, the weights are modified so as to minimize the mean squared error between the network’s prediction and the
actual class. These modifications are made in the “backwards” direction, i.e., from the output layer back to the hidden
layer. In general, the weights will eventually converge, and the learning process stops. To compute the net input to a
certain unit in hidden layer and output layer, the formula is determined as:
I j   wij Oi  j
i
where wij is the weight of connection from unit i to unit j; Oi is the output of unit i; and θj is the bias of the unit.
For input layer, Oj = Ij.
And the algorithm is summarized as below:
Input: The training samples, samples; the learning rate, l; a multiplayer feed-forward network, network.
Output: A neural network trained to classify the samples.
Method:
(1)
(2)
(3)
(4)
Initialize all weights and biases in network;
While terminating condition is not satisfied {
for each training sample X in samples {
for each hidden or output layer unit j {
I j  i wij Oi   j ;
(5)
Oj 
(6)
(7)
Err j  O j (1  O j )(T j  O j );
for each unit j in the hidden layer
Err j  O j (1  O j )k Errk wkj ;
(10)
(11)
}
for each unit j in the output layer
(8)
(9)
1
1  e 1
for each weight wij in the network {
(12)
wij  l  Err j Oi ;
(13)
wij  wij  wij ;
(14)
for each bias θj in network {
(15)
 j  l  Err j ;
(16)
 j   j   j ;
(17)
}
}
}}
4. IMPLEMENTATION AND TESTING
As there are many existing data mining software in the market, it may even take longer time and cost for developing brand
new software for the above NN algorithm. The software package chosen for the implementation is SPSS Clementine 4.0,
which runs under Microsoft Windows XP. The engine used for testing is HONDA K20A i-VTEC (Figure 2) and the car
model is ACURA RSX DC5. K20A is one of the latest automotive engines in HONDA. In order to log necessary data from
the engine, a programmable ECU is used for substituting the OEM ECU. The programmable ECU chosen is E-Manage by
TRUST. A notebook computer is connected to the K20A’s ECU through E-Manage, as shown in Figure 3. A set of 50 logs
has been taken as training data for the NN. Each log just takes the parameters related to air-fuel ratio. An instance of the
logs is shown in Figure 4.
Figure 2. HONDA K20A i-VTEC engine
Figure 3. Connection of notebook computer to K20A’s ECU through E-manage
Figure 4. An instance of the logs recorded from E-manage
5. DISCUSSION OF RESULTS
After training of the NN, a set of weights corresponding to the adjustment of air-fuel ratio parameters is returned. These
parameters are then input to the E-manage as shown in Figure 5.
Figure 5. The parameters trained by the NN, which will be input to E-manage
The maximum road horsepower and torque of the car before tuning is 115 hp and 172 Nm respectively. After using the
new setting extracted from the NN, the maximum road horsepower and torque is increased to be 123 hp and 181 Nm
respectively. The horsepower shows an increase of 7%, while the torque demonstrates an increase of 5.2%. The
maximum horsepower and torque are measured on road using a G-force meter.
6. CONCLUSION
Automotive engine tuning is a very difficult task. For decades, nobody could tackle this problem effectively and only trialand-error method can work. This method is also highly dependent on the mechanic’s knowledge. However, this kind of
knowledge is hard to be formulated as mathematical models because there are too many variables for human to handle.
With the emerging data mining techniques, the problem of engine tuning can be augmented. Neural Network (NN) is one
of the data mining techniques. NN is employed in this research, as it is an optimization tool that perfectly fits the current
optimization problem. An experiment has been conducted to verify the improvement of engine performance after tuning
with the parameters supplied by NN. Although no optimal performance can be achieved so far, a near optimal
performance can be obtained at this stage.
REFERENCES
1.
2.
3.
4.
5.
6.
J. Han, M. Kamber, Data Mining: Concepts and Techniques. Morgan Kaufmann Publishers, 2000
T. Mitchell, Machine Learning. McGraw-hill companies, Inc., 1997
S. Russell, P. Norvig, Artificial Intelligence: A modern approach. Prentice Hall, Inc., 1995.
R. Johnson, D. Wichern, Applied multivariate statistical analysis, 3rd edition. Prentice Hall, Inc., 1992.
William H. Crouse and Donald L. Anglin, Automotive mechanics, 10th edition, McGraw-Hill, 1995
Jeff Hartman, Fuel injection: Installation, performance tuning, modifications, Motorbooks international powerpro series,
1993