The present invention relates in general to detecting misfires occurring
during normal in-use vehicle operation of internal combustion engines and more
specifically to identifying the occurrence of individual misfires by digitally
processing a pressure wave from the exhaust manifold of the engine.
Catalytic converters are used in automobiles to reduce the amount
of pollutants in the engine exhaust. When a cylinder misfires so that no combustion
or incomplete combustion occurs, uncombusted fuel is introduced into the exhaust
which then burns in the hot catalytic converter. The heat from fuel burning in
the catalytic converter destroys the catalyst. Thus, it becomes desirable to detect
and count engine misfires and signal the operator of the vehicle upon occurrence
of excessive misfires so that steps may be taken to protect the catalytic converter.
Some prior art techniques for detecting misfires have employed monitoring
of engine crankshaft accelerations, monitoring of electrical properties of the
ignition spark, and monitoring various properties of the exhaust gas, such as,
exhaust gas pressure and exhaust gas temperature. However, each prior art method
has been found to have disadvantages which have prevented the attainment of an
acceptable misfire detection system. Low signal-to-noise ratios and slow operating
speed have required averaging of many events in some previous misfire detection
systems. Such systems are only capable of detecting recurrent misfires of any
particular cylinder rather than individual misfires. Other systems may require
expensive custom sensors and components or may require disadvantageous sensor
locations. Furthermore, existing misfire detection systems all suffer from poor
accuracy which prevents any ability to identify very low misfire rates. For example,
in order to protect catalytic converters and prevent excessive emissions of pollutants,
a misfire rate of about one or two percent must be detected. In addition, the identity
of the misfiring cylinder associated with each individual misfire must be determined
and stored to facilitate later servicing of an engine to correct the condition
leading to the misfires. Typically, such diagnoctic strategies must have very
low false alarm rates in order to be deployed in large numbers of vehicles.
Furthermore in EP-A2-381855 a misfire detection is indicated using
exaust gas pressure and a pattern recognition with an electronically controlled
combustion engine is generally known from US-A-5041976.
The present invention uses exhaust gas pressure to detect misfire,
so that misfire can be detected over a wide range of speed and load conditions
in the engine. The invention employs pressure transducer means in communication
with the extaust manifold of an engine for generating pressure signals responsive
to exhaust pressure within the exhaust manifold. Position means are employed for
generating a plurality of position signals during each cycle of operation of the
engine indicative of predetermined rotational positions within each a cycle of
the internal combustion engine. An analog-to-digital converter is coupled to sample
the pressure signals from the pressure transducer means in response to the position
signals from said position means for sampling the pressure signal at the predetermined
rotational positions to generate digitized pressure data. A trained classifier
has a data input coupled to the analog-to-digital converter and has a set of predetermined
internal coefficients for processing the digitized pressure data to generate an
output signal indicative of the detection of individual misfires. In a preferred
embodiment, the pressure transducer means is shared with an exhaust gas recirculation
system which uses the pressure signal in controlling the introduction of exhaust
gas into the engine intake. Thus, the present invention achieves the advantages
of high accuracy in a real-time system allowing low misfire rates to be detected.
The invention further exhibits advantages of low cost and a low part count without
requiring any expensive or customized components.
The invention will now be described further, by way of example, with
reference to the accompanying drawings, in which:
Figure 1 illustrates an on-board misfire detection system according
to the present invention.
Figure 2 shows a typical pressure sensor signal from the sensor in
Figure 3 shows the sample rate for operating the analog-to-digital
converter of Figure 1.
Figure 4 illustrates the frame rate for formatting digital data obtained
in Figure 1.
Figure 5 illustrates a test system for collecting digitized pressure
data with known misfires.
Figure 6 illustrates a method for simplifying the representation
of the digital pressure data.
Figure 7 illustrates a method for presenting simplified data to a
trainable classifier to determine the internal coefficients used in the present
Figure 8 illustrates a typical exhaust gas recirculation system.
Figure 9 illustrates a pressure transducer which is shared between
a misfire detection system and an exhaust gas recirculation system.
Figure 10 shows the sensor of Figure 9 in greater detail.
In accordance with the present invention, a data classifier (i.e.,
a pattern recognition system), such as a neural network simulation programme, is
used in conjunction with a high speed data acquisition system to produce a misfire
detection system that is trained to recognise data signatures of individually misfiring
cylinders. During training of the classifier, an engine is operated in a service
bay with intentionally introduced misfires (i.e., bugs), each such bugged operating
trial being labeled according to the identity of the misfiring cylinder or cylinders.
The misfires can be introduced by inhibiting the ignition spark for an individual
firing or cutting off fuel to a cylinder for an individual firing, for example.
Data from a normal (i.e., nonmisfiring) engine is also included in the training.
The data from a sufficiently large number of trials is then presented to the data
classifier as training vectors in a training operation. During training, a set
of internal coefficients in the classifier is recursively readjusted until the
classifier produces the correct label (classified output) for each training vector.
Subsequently, a classifier with the same internal coefficients is attached to
an in-use engine substantially identical with the test engine to monitor misfires
in real time.
A misfire detection system for monitoring engine misfires onboard
a vehicle is shown in Figure 1. An internal combustion engine 10 includes a right
hand exhaust manifold 11 and a left hand exhaust manifold 12 joined to an exhaust
conduit 13. Exhaust gases from engine 10 flow through manifolds 11 and 12 and conduit
13 to a catalytic converter 14, a conduit 15, and a muffler 16. Engine 10 drives
an output shaft 17, such as a crankshaft or a camshaft.
The present invention collects exhaust pressure data at predetermined
sample times within each cycle of engine operation. Thus, a pressure transducer
20 is in communication with the exhaust manifolds, as shown. An analog pressure
signal from transducer 20 is coupled to the input of an analog-to-digital (AID)
converter 21 which provides digital samples to a formatting register 22. The resulting
formatted digitized pressure data is coupled through a transformation matrix 23
which simplifies the data representation to a trained classifier 24 containing
predetermined internal coefficients obtained in a separate training process. Trained
classifier 24 provides an output 25 indicative of the misfire or nonmisfire classification
of each engine cycle (i.e., identities of any misfiring cylinders during the engine
Predetermined rotational positions within a cycle of the engine 10
are determined using a position sensor 26 connected to position generator 27. Position
sensor 26 may include a fixed variable reluctance (VR) sensor located in proximity
to a multi-toothed rotating wheel connected to output shaft 17. Position generator
27 produces a reference signal once per engine cycle which defines a frame rate
for formatting the digitized pressure data in formatting register 22. Position
generator 27 also generates a plurality of position signals within each engine
cycle at predetermined rotational positions, thus providing a sample rate to a
A/D converter 21. In a preferred embodiment, the position signals indicate rotational
positions separated by about ten degrees, resulting in about 72 samples in each
frame relating to an engine cycle. Thus, formatting register 22 collects 72 samples
from AID converter 21 into a single frame which is provided to transformation matrix
23 for transformation into a simplified representation in order to reduce the
amount of computation required in trained classifier 24. The simplified data representation
from transformation matrix 23 includes a plurality of digital values which comprise
an input vector which is processed in a pattern matching space (defined by the
set of internal coefficients) within trained classifier 24 to produce a classification
output 25. The meaning of output 25 depends on the manner in which the internal
coefficients are derived during training of the classifier. In the preferred embodiment,
trained classifier 24 includes internal coefficients which classify an input vector
according to properly firing or misfiring cylinders.
Figure 2 illustrates a typical sensor signal from sensor 20 of Figure
1 which characterises the exhaust pressure (i.e., acoustic) waveform. The waveform
contains sufficient information to allow detection of misfiring and nonmisfiring
cylinders. However, prior art misfire detection systems based on exhaust pressure
have used deterministic algorithms based on certain expert derived models in order
to detect a misfire. The development of such a deterministic algorithm requires
an intense expert study of the system to understand precisely the system operation.
Such expert system development takes a large amount of time and resources and has
only been able to define a rough approximation of system operation. Therefore,
prior art systems have had limited accuracy.
The present invention bypasses the drawbacks of expert system development
by acquiring digital pressure data and formatting it into input vectors for application
to a trained classifier. Figure 3 shows the sampling rate at which the sensor
signal is sampled by an analog-to-digital converter, and Figure 4 shows the frame
rate for formatting the digital data according to a reference signal occurring
once per engine cycle (i.e., after two rotations of a four-cycle engine). The
position sensor may conveniently be located on the engine camshaft to facilitate
identification of each engine cycle. Alternatively, a sensor for generating the
sample rate may be located on the engine crankshaft and a separate cylinder identification
sensor may be located on the camshaft or other means may be used to detect each
Trained classifier 24 in Figure 1 preferably employs coefficients
determined in advance using a test system, thereby eliminating any need for the
capability of actual training within classifier 24. Formatting register 22, transformation
matrix 23, and trained classifier 24 are preferably implemented using a microcomputer
28. The required computing power of microcomputer 28 is reduced by not including
trainability for classifier 24.
The predetermined internal coefficients for the trained classifier
are obtained as shown in Figures 5-7. A bugged engine 30, substantially identical
to the engine and exhaust system to be utilised in production vehicles, is operated
under a variety of conditions to collect training vectors which are compiled into
a data base 31. Engine 30 is bugged by deliberately introducing misfires and combinations
of misfires in an engine cycle. Data is generated using various engine malfunctions
that could lead to misfire and under a variety of speed and load operating conditions.
Each training vector compiled in data base 13 is labeled with a bug identifier
to identify which, if any, cylinders were misfiring in the training vector.
After collection of sufficient vectors in data base 31 to adequately
represent all possible normal firing and misfiring conditions over the full range
of engine speed and load, a technique, such as principal component analysis, is
employed to reduce or simplify the representation of data in data base 31. The
training vectors in data base 31 are initially represented in an arbitrary coordinate
system. Using principal component analysis, an alternative coordinate system is
found that results in a more compact and simplified representation of the data.
Once the coordinate system is found that results in the most compact data representation,
a transformation matrix is determined that remaps data from the original arbitrary
coordinate system to the new simplified coordinate system. Thus, as shown in Figure
6, data base 31 is input to a principal component analysis 32 which yields transformation
After the transformation matrix is identified, a classifier is trained
as shown in Figure 7. Data base 31 provides data through transformation matrix
23 to a trainable classifier 33. The bug identifier labels from data base 31 are
provided directly to trainable classifier 33 for identifying the proper response
associated with each training vector. During training, the classifier recursively
adjusts its internal coefficients until it has learned the proper association between
training vectors and their identifiers. The final values for the internal coefficients
after full training are employed as the predetermined internal coefficients for
a misfire detector system as shown in Figure 1.
In order to reduce the number of components, the present invention
may share a pressure transducer with an exhaust gas recirculation (EGR) system.
In the typical EGR system of Figure 8, exhaust flow within an engine exhaust manifold
40 or other point in the exhaust system has an exhaust pressure that is monitored
by a sensor 41 through a conduit 42. The EGR system reintroduces exhaust gas through
an EGR valve 43 to an engine air intake manifold 44 in order to lower combustion
temperatures and reduce the formation of oxides of nitrogen. An exhaust pressure
signal from sensor 41 is provided to an engine control assembly (ECA) module 45
that produces a variable duty cycle output signal based on inputs of engine speed,
engine vacuum, exhaust pressure, coolant temperature, and throttle angle in order
to control the amount of exhaust gas reintroduced. The variable duty cycle signal
is connected to an electronic vacuum regulator (EVR) 46 that utilises intake vacuum
to control the position of EGR valve 43 in accordance with the duty cycle signal.
As shown further in Figure 9, an engine 50 has a left hand exhaust
manifold 51 connected to an exhaust pipe 52. A right hand exhaust manifold (not
shown) on the opposite side of engine 50 is connected to an exhaust pipe 53. The
exhaust pipes are joined at coupling 54 which is further connected to a catalytic
converter 55. Pressure sensor 41 communicates with left hand exhaust manifold through
conduit 42 and a conduit 56. Pressure sensor 41 communicates with the right hand
exhaust manifold through conduits 42 and 56 and exhaust pipes 52 and 53. In a typical
EGR system, sensor 41 includes lowpass filtering of the pressure signal prior
to sending the signal to ECA 45. For purposes of misfire detection according to
the present invention, the pressure signal must not be filtered in the manner used
in the EGR system. Thus, a separate unfiltered pressure signal is provided from
sensor 41 to a misfire detector 60, in accordance with the invention.
An advantage of the classifier used in the invention is that the
pressure transducer need not be located symmetrically with respect to separate
left and right hand exhaust manifolds. The internal coefficients of the classifier
adapt for any nonsymmetrical position. However, a symmetrical position, such as
at coupler 54 in Figure 9, could simplify the computation required in the classifier.
Figure 10 shows sensor 41 in greater detail. A transducer element
61, such as a pair of capacitive plates or a resistive strain gauge, is connected
to an amplifier 62. The output of amplifier 62 provides a direct output to misfire
detector 60 and is coupled to a lowpass filter 63 for providing signal averaging
prior to connection to ECA 45.