Misfire Detection in IC Engine using Kstar Algorithm

Anish Bahri*, V.Sugumaran**, S. Babu Devasenapati***

*
SMBS, VIT University, Chennai Campus, Vandalur-kelambakam road, Chennai-600127,
anish.bahri@gmail.com
**
SMBS, VIT University, Chennai Campus, Vandalur-kelambakam road, Chennai-600127,
v_sugu@yahoo.com
***
Principal, Sri Subramanya College of Engineering and Technology, Palani,
babudeva@yahoo.com

ABSTRACT

Misfire in an IC Engine continues to be a problem leading to reduced fuel efficiency, increased

power loss and emissions containing heavy concentration of hydrocarbons. Misfiring creates a
unique vibration pattern attributed to a particular cylinder. Useful features can be extracted
from these patterns and can be analyzed to detect misfire. Statistical features from these
vibration signals were extracted. Out of these, useful features were identified using the J48
decision tree algorithm and selected features were used for classification using the Kstar
algorithm. In this paper performance analysis of Kstar algorithm is presented.

Keywords – Fault Diagnosis, IC Engine, Kstar algorithm, Misfire Detection, Statistical

Features.

1. INTRODUCTION
In case of misfire, uncombusted fuel accesses the exhaust system, where its combustion in the
catalytic converter can cause a dangerous increase in temperature, damaging the catalytic
converter due to thermal overload. It may also lead to increased emission of hydrocarbons.
Thus it becomes necessary to detect misfire to prevent pollution and wastage of fuel. Methods
using the torsional vibration signal of the crankshaft [1, 2] and the acceleration signal of the
engine head [3-5] were developed to detect misfire. Klenk et al. used crankshaft speed [6] for
misfire detection. Some works have already been carried out in this field using features such as
cylinder deviation torque [8], instantaneous angular velocity [9] and instantaneous exhaust
manifold pressure [10]. These methods can be costly and cumbersome. Machine learning
approach has been used for fault diagnosis. Sugumaran et al. have used decision tree [11] and
support vector machines [12] as classifiers. For fault diagnosis in bearings, Sugumaran et al.
used support vector machines [13]. Sakthivel et al. used histogram features with decision tree
[14]. Muralidharan et al. carried out fault diagnosis of monoblock centrifugal pump using
Naïve Bayes and Bayes Net classifier [15]. The use of engine vibration data requires minimum
instrumentation. Machine learning when used with this data gives appreciable results and the
system can be trained for changing conditions. Here the vibration signals from the engine head
were recorded using an accelerometer and stored. Statistical features were derived from these
signals and finally feature selection followed by feature classification was carried out. Here the
feature selection was done using J48 decision tree algorithm and feature classification was
carried out using Kstar algorithm.

2. EXPERIMENTAL SETUP

The experimental setup comprises mainly of the spark ignition IC engine test rig with
provisions made in order to manually cause misfire in a particular cylinder and the data
acquisition system. An accelerometer is attached on the engine which measures the vibration
signals. The acquired signal is then passed through an ADC to collect the required data from
which features are extracted. Fig.1 shows the basic flow of the steps involved in the whole
process.

2.1 Engine Test Rig

Table 1: Specifications of the engine used

FEATURES SPECIFICATION

Make Hindustan Motors

Number of cylinders/stroke Four cylinders/four stroke

Fuel Gasoline(Petrol)

Rated power 7.35 kW

Rated speed (alternator) 1500 rpm

Engine stroke length 73.02 mm

Engine bore diameter 88.9 mm

Cooling Water cooled

 The engine test rig consists of a 10 hp four stroke vertical four cylinder engine. It has
the provisions to simulate the misfire by cutting the electric supply to individual spark plugs.
The engine accelerator is firmly attached at the desired position via screw and nut mechanism.
The speed is measured and monitored by means of a tachometer. The specifications of the
engine are provided in Table1.

Engine Test Rig

Data acquisition and signal

conditioning

Feature extraction and

feature selection

Training data Testing data

NO Is Kstar algorithm
trained?

YES

Trained algorithm

Algorithm output

Misfire detection result

Fig. 1 Flowchart for Engine Misfire Detection

2.2 Data Acquisition System

Accelerometers were used to record the vibration signals. They have the ability to
detect both small and large vibrations. The accelerometer is placed in such a way that it can
measure vibration signals from all the four cylinders. The signal from the accelerometer is fed
to a DACTRON FFT analyser which converts the analogue signal to digital signal. This data is
transferred and stored in a computer for further processing. Fig 2 shows the experimental setup
used.

Fig 2 Experimental Setup

2.3 Experimental Procedure

Firstly the engine was started. This was done at no load by electrical cranking and
warmed for 15 minutes. Now the FFT analyser was switched on and the data is taken only after
the engine gets stabilised. All the data was collected for 1500 rpm, a sampling frequency of 24
kHz and a sampling length of 8192. For the present study five cases were considered i.e.
normal condition, misfire in cylinder one, two, three and four. All the events were simulated at
1500 rpm. Fig. 3, 4, 5, 6 and 7 show vibration signals for no misfire, misfire in cylinder one,
two, three and four respectively at 1500 rpm.
 .

Fig. 3 Vibration signal for no misfire Fig. 4 Vibration signal for misfire in
cylinder 1

Fig. 5 Vibration signal for misfire in Fig. 6 Vibration signal for misfire in
cylinder 2 cylinder 3

Fig. 7 Vibration signal for misfire in

cylinder 4

3. FEATURE EXTRACTION AND FEATURE SELECTION

The vibration signals acquired cannot be used directly. Useful information needs to be
extracted from them which will represent the signal. Here statistical features are extracted.
Descriptive statistics for a particular signal gives a wide range of parameters namely mean,
standard error, median, mode, standard deviation, sample variance, kurtosis, skewness, range,
minimum, maximum, sum and count. The description of these parameters is given in Table 2.
 Table 2: Description/Formulae of all statistical features
Name of the Description/Formulae
feature

Mean For a data set, mean is equal to the sum of all the data points divided by the number of

data points

Standard Error It refers to the error encountered when a statistic of sampling distribution varies from
its value.

Median Median is that value that separates the higher values from the lower values in a data
sample.

Mode It is that value that appears the most number of times in a data sample.

Standard It shows how much variation exists in the data from the average value.
deviation

Sample Variance It is a measure of how far a set of data is spread out.

Kurtosis It is used to describe the distribution of given data around the mean.

Skewness It describes asymmetry from the normal distribution in a set of statistical data.

Range In a set of data, range is the difference of highest and lowest values.

Maximum Value It refers to the maximum value in a given data set.

Minimum Value It refers to the minimum value in a given data set.

Sum It refers to the sum of all feature values of a given sample.

Count It refers to total number of data points.

 All the features extracted may not be useful for feature classification. Thus feature
selection is carried out. Feature selection is the process of selecting the features from the above
features which contribute highest to the classification accuracy. At the end of feature selection,
only useful features will be left. This reduces the computational effort and thus reduces the cost
and time. This is very important as the final aim is to convert this into an on-board diagnostic
system.

For the present study, J48 decision tree algorithm was used for feature selection. Fig. 8
shows the decision tree obtained. It can be observed from the tree that root node is sample
variance which is the feature contributing maximum to the classification accuracy closely
followed by kurtosis and standard error. Coming down the tree, the following features continue,
minimum, mean, standard deviation, skewness and range. Maximum, sum and count are not
visible indicating that they do not contribute significantly to the classification accuracy.

After feature selection, the effect of number of features on classification accuracy was
studied. Sample variance, standard error, kurtosis, minimum, mean, standard deviation,
skewness and range were identified as the useful features and were thus used. Other non-
relevant features were discarded.

Fig. 8 Decision tree

4. CLASSIFIER

Here Kstar algorithm is used as the classifier. The Kstar algorithm uses entropic
measure, based on probability of transforming instance into another by randomly choosing
between all possible transformations. Using entropy as appraise of distance has numerous
utility. Information theory helps in computing distance between instances. The complexity of a
transformation of one instance into another is actually the distance between instances. This is
achieved in two steps. First define a finite set of transformations they will map one instances to
another. Then transform one instance (a) to another (b) with the help of “program” in a finite
sequence of transformations starting at a and terminating at b.

4.1 Specification of Kstar

Given a set of infinite points and set of transformations predefined T. Let t be a value of set T.
This t will map t: I→I. To map instances with itself σ is used in T (σ (a) =a). σ terminates P, the
set of all prefix codes from T*. Members of T* and of P uniquely define a transformation on I.

Where t = t1...tn

P is a probability function on T*. It satisfies the following properties:

As a consequence, it satisfies the following:

The probability function P* is defined as the probability of all paths from instance a, to instance
b:

It is easily proven that P* satisfies the following properties.

The K* function is then defined as:

)=- . Please note that K* is not exactly a distance function.

5. RESULTS AND DISCUSSIONS

For misfire detection machine learning approach was used here. From a fairly large number of
features, sample variance, standard error, kurtosis, minimum, mean, standard deviation,
skewness and range were selected. The results are discussed below.

5.1 Effect of number of features on classification accuracy

All the features extracted may not be useful for feature classification. Thus the irrelevant
features need to be filtered out. This was done using feature selection explained in earlier
section. Table 3 shows the effect of number of features on classification accuracy.

Table 3: Classification accuracy of Kstar algorithm

No. of features Kstar algorithm accuracy (%)

1 72.8

2 72.8

3 82.2

4 80.2

5 78.2

6 79.2

7 81.4

8 82.6

9 81.4

10 79

11 80.4

12 80.4

Thus it is observed that maximum classification accuracy is achieved using only 8 features out
of 12 features. This reduces the computational effort and time.
5.2 Classification using Kstar algorithm

Table 4 Confusion matrix for Kstar algorithm

TESTING C1mis C2mis C3mis C4mis Normal
C1mis 83 0 4 13 0
C2mis 0 100 0 0 0
C3mis 3 0 55 42 0
C4mis 3 0 22 75 0
Normal 0 0 0 0 100

Here results obtained from the Kstar algorithm will be discussed. Table 4 shows the
confusion matrix obtained. In the confusion matrix, „C1mis‟ stands for misfire in cylinder 1 and
„C2mis‟, „C3mis‟,‟C4mis‟ stands for misfire in cylinders 2, 3 and 4 respectively. Normal
means the engine is in normal conditions with no misfire. The first location in the first row
shows the number of signals which are correctly classified as those from the first cylinder. The
second position in the first row shows the number of signals which are wrongly classified as the
signal from cylinder two. Similarly third and fourth positions show the number of signals
wrongly classified as those from cylinder three and cylinder four. The first position in the
second row shows the number of signals from cylinder two wrongly classified as those from
cylinder one. The second position in second row shows the number of signals correctly
classified as those from cylinder two. Thus the diagonal in the matrix shows the number of
correctly classified points. It can be seen from the matrix that no faulty signal is classified as a
normal signal which is very crucial. Thus Kstar algorithm can differentiate between good and
faulty signals with 100% accuracy. Because of its ability to distinguish between the faulty and
good signals with 100% accuracy no false alarm will be generated, thus saving time and
resources which may be spent to check if the engine misfired. However there are some
misclassifications among the faulty conditions as it is evident from Table 4. Kstar algorithm
shows an overall classification accuracy of 82.6% which is appreciable because no time needs
to be spent to pin point which cylinder misfired, and thus the problem can be directly rectified,
thus saving money and resources.

6. Conclusion
Considering the emissions and fuel wastage caused, Kstar algorithm was proposed as a possible
classifier for misfire detection using vibration signals from the engine block. Kstar performed
satisfactorily with a classification accuracy of 82.6%. Also the ability to distinguish between
faulty and good conditions with 100% accuracy makes it suitable for industry applications.
REFERENCES
[1] Rizzoni, G., Estimate of indicated torque from crankshaft fluctuations: A model for the
dynamics of the internal combustion engine, IEEE Transaction Vehicle Technology, 38(3),
1989, 168–179.
[2] Zhang, Y., Randall, R.B., The In-Cylinder Pressure Reconstruction and Indicated Torque
Estimation Based on Instantaneous Engine Speed and one Measured In-Cylinder Pressure,
Comadem Conference, Faro, Portugal, 2007.
[3] Citron, S.J., O‟Higgins, J.E. and Chen, L.Y., Cylinder by cylinder engine pressure and
pressure torque waveform determination utilizing speed fluctuations, SAE Transactions, SAE
890486, 1989.
[4] Shiao, Y. and Moskwa, J. J., Misfire detection and cylinder pressure reconstruction for SI
engines. SAE Paper No. 940144, 1994.
[5] Y. Ren, Detection of knocking combustion in diesel engines by inverse filtering of
structural vibration signals, PhD Dissertation, the University of New South Wales, Australia,
1999.
[6] Klenk, M., Moser, W., Mueller, W., and Wimmer, W., Misfire Detection by Evaluating
Crankshaft Speed – A Means to Comply with OBDII, SAE Paper 93099, 1993.
[7] Lee, D., and Rizzoni, G., Detection of Partial Misfire in IC Engines Using Measurement of
Crankshaft Angular Velocity, SAE Paper 951070, 1995.
[8] Yunsong Wang, Fulei Chu, Real-time misfire detection via sliding mode observer,
Mechanical Systems and Signal Processing, 19(4), 2005, 900–912.
[9] Zhixiong Li, Xinping Yan, Chengqing Yuan, Zhongxiao Peng, Intelligent fault diagnosis
method for marine diesel engines using instantaneous angular speed, Journal of Mechanical
Science and Technology, 26(8), 2012, 2413-2423.
[10] V. Maciána, J.M. Lujána, C. Guardiolaa, A. Perlesb, A comparison of different methods
for fuel delivery unevenness detection in Diesel engines, Mechanical Systems and Signal
Processing, 20 (8), 2006, 2219–2231.
[11] V.Sugumaran, K.I. Ramachandrana, S. Babu Devasenapatia, Misfire identification in a
four-stroke four-cylinder petrol engine using decision tree, Expert Systems with Applications,
37 (3), 2010, 2150–2160.
[12] V.Sugumaran, K.I. Ramachandrana, S. Babu Devasenapatia, Misfire Detection in a Spark
Ignition Engine using Support Vector Machines, International Journal of Computer
Applications, 5(6), 2010, 0975 – 8887.
[13] V.Sugumaran, K.I. Ramachandrana, Effect of number of features on classification of roller
bearing faults using SVM and PSVM, Expert Systems with Applications, 38, 2011, 4088–4096.
[14] N.R. Sakthivel, V. Indira, B.B. Nair, V. Sugumaran, Use of histogram features for decision
tree based fault diagnosis of monoblock centrifugal pump, Inter-national Journal of Granular
Computing, Rough Sets and Intelligent Systems (IJGCRSIS), 2011, 23-26.
[15] V. Muralidharan, V. Sugumaran, A comparative study of Naïve Bayes classifier and Bayes
net classifier for fault diagnosis of monoblock centrifugal pump using wavelet analysis, Journal
of Applied Soft Computing, 2012, 1-7.