an approach to full-range fault diagnosis of spark ... approach to full-range fault diagnosis of...

Int. J. Vehicle Systems Modelling and Testing, Vol. 6, No. 1, 2011 21

Copyright © 2011 Inderscience Enterprises Ltd.

An approach to full-range fault diagnosis of spark ignition engines’ intake system using normalised residual and neural network classifiers

Amir H. Shamekhi Mechanical Eng. Department, K.N. Toosi University of Technology, Tehran, Iran E-mail: [email protected] *Corresponding author

Mohammad H. Behroozi* School of Mechanical Eng., University of Birmingham, Birmingham, B152TT, UK E-mail: [email protected]

Reza Chini Department of Eng. & Applied Science, Memorial University, P.O. Box 70, Canada E-mail: [email protected]

Abstract: One essential part of automated diagnosis systems for spark ignition (SI) engines is due to elements of air path system. The faults that occur in this subsystem can result in deviation in the air-fuel ratio, which causes increased emissions, misfire and especially loss of power and drivability problems. In this article, a model-based diagnosis system for the air-path of an SI engine is developed. In addition, a non-linear four-state dynamic model of an SI engine is used, and then the diagnosis system is designed in the framework of an Artificial Neural Network (ANN) classifier. Simulation results show that the constructed diagnosis system for seven fault modes considering all three kinds of common fault, including the manifold air temperature (MAT) sensor fault, which has been comparatively less evaluated than other elements, is applied successfully. As another remarkable aspect of this work, all classes of faults are diagnosed in their full possible over-reading (positive) and under-reading (negative) ranges.

Keywords: full-range fault diagnosis; mean value engine modelling; MVEM; neural network classifier; normalised residuals; spark ignition; SI.

Reference to this paper should be made as follows: Shamekhi, A.H., Behroozi, M.H. and Chini, R. (2011) ‘An approach to full-range fault diagnosis of spark ignition engines’ intake system using normalised residual and neural network classifiers’, Int. J. Vehicle Systems Modelling and Testing, Vol. 6, No. 1, pp.21–55.

22 A.H. Shamekhi et al.

Biographical notes: Amir H. Shamekhi is an Assistant Professor in the Mechanical Engineering Department at K.N. Toosi University of Technology, Tehran, Iran. He teaches courses in engine dynamics and control. He received his MS and PhD in Mechanical Engineering from K.N. Toosi University of Technology, Tehran, Iran, in 1996 and 2004, respectively. His research interests are automotive control systems, FDI methods, and internal combustion engines.

Mohammad H. Behroozi received his BSc in Mechanical Engineering from the K.N. Toosi University of Technology, Tehran, Iran, in 2005. He received his MSc in Mechanical Engineering with a focus on vehicle dynamics at Iran University of Science and Technology (IUST) in 2008. He is presently pursuing his PhD in Mechanical Engineering in the School of Mechanical Engineering, The University of Birmingham, Birmingham, UK. He has been a member of Iranian Society of Mechanical Engineers since 2006 and a student member of IMechE since 2010.

Reza Chini received his BSc in Fluid Mechanical Engineering from K.N. Toosi University of Technology, Tehran, Iran, in 2005. He also studied Automotive Engineering as an MSc student at K.N. Toosi University of Technology, Tehran, Iran. He is currently a PhD candidate at Memorial University, Canada. He has been a member of the Iranian Society of Mechanical Engineers since 2006.

1 Introduction

On-board diagnostics (OBD) is a generic term referring to a vehicle’s self-diagnostic and reporting capability. OBD systems monitor the state of vehicle’s health to help driver and repairmen for various vehicle components. The number of available vehicle subsystem faults reported by OBD has widely increased since the first demonstration in the early 1980s (Birnbaum, and Truglia, 2000). Modern OBD implementation uses a standardised fast digital communications port to provide real-time data in addition to a standardised series of diagnostic trouble codes, which allow one to rapidly identify and remedy malfunctions within the vehicle.

In 1987, The California Air Resources Board announced that all vehicles sold in California and manufactured after 1988 had to have some basic OBD capability. These requirements are so called these days as OBD-I. In 1996, all the vehicles sold in USA were made obligatory to be equipped by OBD-II. The European Union made OBD-II mandatory for all petrol vehicles sold in the European Union after 2001 (SAE International, 2003). An ISO standard was introduced in USA at 2008 to car manufacturers as minimum necessary requirements of OBD options on the vehicle (ISO Standards, 2010). Latest developments include mandatory rules for heavy duty OBD (HDOBD) for specific heavy duty vehicles sold in USA (ISO Standards, 2010).

By increasing the number of possible faults in vehicle and covering more vehicle components in diagnosis system, the number of OBD codes will be drastically increased. It makes the diagnostic duty much more difficult in terms of implementation and adding new functions to engine on-board diagnosis system. Therefore, introduction of more flexible diagnosis approaches seems necessary in terms of implementation in future OBD development. Also, the health of engine system or other sub-systems can be monitored as

An approach to full-range fault diagnosis of spark ignition engines’ intake 23

a percentage of change instead of on-off light facilities in common OBD systems in case of using more flexible methods in diagnosis. Therefore, the sensitivity of the current OBD systems should be definitely improved by more flexible techniques.

Also, OBD II usually remains offline until the problem has occurred on two separate drive cycles before it turns on the diagnostic lamp. This is to reduce the number of ‘false warnings’ that might otherwise occur if the system turned on the lamp every time it saw something amiss. So if the lamp is on, it means the problem has occurred before and has occurred again. It is not just a temporary glitch but something that needs to be diagnosed and corrected.

Due to the complexity of spark ignition (SI) engines, the number of potential failures increases, which makes the role of fault diagnosis quite challenging. When faults frequently appear in industrial machinery, the component quality and flexibility of the system will be decreased and consequently the price of manufacturing will be adversely affected. Realising the characteristics, locations and effects of a particular fault on a system, the task of detection and isolation of failures becomes much easier.

Accordingly, a mathematical model is necessary to analyse the performance of a system and bases a reliable source of comparison for fault diagnosis and isolation (FDI) approaches. One of the methods applied to evaluate sophisticated and multi-input multi-output (MIMO) dynamic systems is mean value modelling. Simplicity and widespread utility of mean value models in engine operation points are the privilege of using this type of modelling methods. Hendricks and Sorenson (1990) developed this method for the first time. Mean value modelling is developed and more through the time by many researchers: Hendricks and Vesterholm (1992), Muller et al. (1998), Fons et al. (1999), Chevalier et al. (2000), Shamekhi and Ghaffari (2004) and Mostofi et al. (2006a).

First serious researches about fault diagnosis of the dynamic systems were done in 1970s. Beard (1971) and Jones (1973) initiated fault diagnosis based on observers in linear systems. Willsky (1976) provided a summary of their work. Clark (1978) introduced sensor fault diagnosis for the first time. Bakiotis (1979) and Filbert and Metzger (1982) introduced and developed the parameter estimation techniques. Patton and Chen (1991), Gertler et al. (1991) and Hofling (1993) used space relations in diagnosis systems. Nyberg (2003), Nyberg and Stutte (2004), Naidu et al. (2005) and Mostofi et al. (2006b) accomplished recent researches on automotive fault diagnosis.

Developing the artificial intelligence methods, research in FDI approaches has been introduced to other levels. Rumelhart and McClelland (1986) suggested application of neural networks in design of fault diagnostic systems for the first time. Frank and Seliger (1997) used neural network for estimation of physical parameters of system which is one of the characteristics of neural network. In this method, after estimation, the diversion of the parameters from normal mode indicated the occurrence of a fault. More recent works on applications of neural networks in fault diagnosis could be mentioned as Gen-Ting and Guang-Fu (2004) and Shen et al. (2006).

In this article, different fault types simulated in inlet system of a modified 1275 cc British Leyland SI engine have been diagnosed by a neural network-based diagnostic system. The mathematical model which is used to study the diagnostic scheme is presented in Section 2. This model is a modified mean value engine modelling (MVEM) of an SI engine. In Section 3, the characteristics and implementation procedure of faults into the model are demonstrated. The theoretical background of Artificial Neural Network (ANN) and its advantages in FDI approaches is discussed in Section 4. In


Section 5, the construction of a diagnostic system based on ANN is introduced, and in the end, the performance quality of proposed diagnosis system is analysed.

2 Mathematical model

2.1 Modelling

A mathematical model is necessary to analyse the performance of a model-based FDI. Here, the mathematical model which is used to study the diagnosis scheme is presented. As a result, a modified MVEM is applied to an SI engine. This model is based on basic and important principles described in Mostofi et al. (2006a). MVEMs are simplified and physically obtained. This simplicity makes them suitable for engine control and diagnosis applications on account of the fact that these can be expressed by a state-space model.

The isothermal assumption employed in most articles is insufficient over several cycles in transient situations. Therefore, it is important to consider both mass and energy balance to correctly describe the air temperature dynamics in intake manifold. Consequently, this led to a two-state manifold model instead of simple state model of isothermal assumption. Eventually, a non-linear four-state dynamic model of an SI engine is constructed. The model is simulated in SIMULINK/MATLAB toolbox and the results are validated with aforementioned real engine. Here, only the final differential state equations are shown. Detailed information about this model can be found in Mostofi et al. (2006a).

Note that notations and descriptions about the parameters and constants used in here are listed in nomenclature at the end of paper.

( )20 1 2 3 42

60

602

fc HV fi

t

Q mn a a n a n a a n p

I nπ

ηπ

⎧ ⎫⎪ ⎪⎡ ⎤= − + + + +⎨ ⎬⎣ ⎦⎪ ⎪⎩ ⎭ (1)

( )1 (1 )ff ff fiff

m m X Y mτ

= − + − (2)

( )

( )

( )( ) ( ) ( )

0 1 1 2

1 0

2 1 1

2 21

( )120

( ) 1 cos

2,1

1 2 ,1 1

d v ii at at i a

i

r r ri a

ia i

V npRp m m p TV R

p p pp p

p signp p

otherwise

κκ κ

κ κ

κ

ηβ α β

β α α α

κβ

κκ κ

+ −

−

⎧ ⎫⎡ ⎤= × + −⎨ ⎬⎣ ⎦⎩ ⎭= − −

⎧⎛ ⎞⎪ − ≥ ⎜ ⎟⎪ +⎝ ⎠⎛ ⎞ ⎪= − ×⎨⎜ ⎟

⎝ ⎠ ⎪ − ⎛ ⎞⎪ ⎜ ⎟⎪ + +⎝ ⎠⎩

(3)

( ) ( ) ( )0 1 1 2( ) 1120

i d v ii at at i a i

i i

RT V npT m m p T T

p V Rη

β α β κ κ⎧ ⎫⎡ ⎤= × + − − −⎨ ⎬⎣ ⎦⎩ ⎭ (4)

where a0, a1,…,a2 are model constants and α0 is the throttle angle in its closed position.


2.2 Model validation

The adiabatic MVEM model that presented above is simulated and compared by experimental results. The real engine is a modified 1,275 cc British Leyland SI engine without exhaust gas recirculation (EGR) (Chevalier et al., 2000). The engine specification is shown in Table 1. Differences between simulation and measurements can be seen in the Figures 1–4. Table 1 Leyland engine specifications

Specification Value

Capacity 1,275 cc Bore 70.6 mm Stroke 81.28 mm Maximum power 70 bhp @ 6,000 rpm Maximum torque 74 lb.ft @ 3,250 rpm Application 1969–1974: Austin 1,300 GT Number of cylinder 4

Figure 1 shows the throttle angle transients as an input to the system. Figures 2–4 show the results of above throttle manoeuvre on engine speed, manifold pressure and temperature respectively. In these figures, solid lines indicate simulation results of equations (1) and (3–4) while dashed lines represent experimental ones. As it could be seen in Figures 2–4, the maximum error is around 5%, 3%, and 8% respectively. It has to be mentioned that the major outcome of the model is its capability in modelling the dynamics of the manifold air temperature (MAT) which is a crucial factor in diagnosing the MAT sensor faults.

Figure 1 Throttle angle behaviour versus time (see online version for colours)

0 1 2 3 4 5 618

19

20

21

22

23

24

25

Time (sec)

Thro

ttle

(deg

)

SimulatedExperiment


Figure 2 Engine speed behaviour versus time (see online version for colours)

0 1 2 3 4 5 62050

2100

2150

2200

2250

2300

2350

2400

2450

2500

Time (sec)

Rev

olut

ion

(rpm

)

SimulatedExperiment

Figure 3 Manifold air pressure behaviour versus time (see online version for colours)

0 1 2 3 4 5 635

40

45

50

55

60

65

Time (sec)

Pre

ssur

e (k

Pa)

SimulatedExperiment


Figure 4 Manifold air temperature behaviour versus time (see online version for colours)

0 1 2 3 4 5 6275

280

285

290

295

300

305

310

315

320

325

Time (sec)

Tem

pera

ture

(K)

SimulatedExperiment

3 Simulation of faults

For the sake of generality, all three types of faults are diagnosed in this paper, and these are listed in Table 2. This list of operational modes was chosen based on the frequency of their occurrence and their effect on SI engines. Hence, seven operational modes are only studied here and other faults, despite their importance on system performance, are not considered. It has to be also noted that the throttle manoeuvre has been modified to provide wider ranges of below mentioned faults in order to completely clarify prediction and generalisation ability of trained ANN to diagnose defined faults (see Figures 9 and 14). Finally, each operational mode is described and compared with ‘no fault’ (NF) mode. Table 2 Abbreviation and description of diagnosed faults

Faults Description Fault type

NF No fault ---------------- IL Intake manifold leakage Component fault FAG Fuel injector actuator gain-fault Actuator fault MPSG Manifold pressure sensor gain-fault Sensor fault THAG Throttle actuator gain-fault Actuator fault MTSG Manifold temperature sensor gain-fault Sensor fault RSG RPM sensor gain-fault Sensor fault

It would be worth to mention that the sensors are virtually used to monitor vehicle health and performance in this study not to use specifically in the close loop fuel control systems. It follows that the main functionality of the sensors is to provide raw data to


equip the FDI with proper evidences of faults occurred in the engine. In the intervening time, the possible faults in sensors are of the interest to investigate because of covering possible faults in engine in order to ensure FDI reliability.

3.1 No fault

In this operational mode, the sensors and actuators are fault free and there is no leakage in the intake system. As a result, the measured sensor values are equal to physical quantities and the performance of actuators is set by the controllers due to the reference values.

3.2 Intake manifold leakage

For the engine that is used in this work, the intake manifold sub-model pressure during normal operation is always lower than ambient pressure. Hence, a leak in this part is always results in an air flow into the air tube and there would be no air flow in the direction out from it at any working conditions. As a result, faults in this part can not have negative values. The model for intake manifold leakage (IL) is obtained by taking the model for NF operational mode, but replacing it with equation (5) and putting a gain-fault gIL with maximum value of 15%. The constraint on the percentage parameter gIL, limiting the performance ranges of modelled SI engine, is gIL ∈ [0, 15]. Apparently, gIL = 0 represents the intake manifold mass flow in NF mode.

, , ,at IL at NF IL at NFm m g m= + (5)

Figures 5–9, as a case in point, show the effect of gIL = 7% on SI engine model simulation. As it could be seen, the injected fuel flow, revolution and manifold pressure are increased due to the higher air mass flow passing through the throttle while the manifold temperature is slightly affected by the fault. This behaviour of the state values is explained by equations (1) and (3–4).

3.3 Fuel injector actuator gain-fault

In fuel injector actuator gain-fault (FAG) operational mode, the fault is generated in the fuel injector actuator by adding a scalar parameter gFAG. This gain-fault is added to the physical quantity of injected fuel flow which is set by the amount of fuel mass flow needed to keep the charge stoichiometric; see equation (6). The restriction on gFAG is: gFAG ∈ [–8, 15].

, , ,f FAG f NF FAG f NFm m g m= + (6)

As it could be seen in Figures 5–9 for gFAG = 10% and in Figures 10–14 for gFAG = –5%, this fault causes the injected fuel flow to be increased. Thus, revolution is raised significantly. Leaping the engine speed, more air is pumping into the cylinders which would be led in manifold pressure drop in view of the fact that the throttle still remains stable at NF position. In fact, the manifold temperature is remained unaffected by the fault. The negative gain faults affected the model with diverse trend.


3.4 Manifold pressure sensor gain-fault

The model corresponding to this operational mode is related to a fault caused by a gain-fault added to the physical value of the manifold pressure sensor. For implementing the manifold pressure sensor gain-fault (MPSG), the gain factor of the sensor gMPSG is added to equation (7). The related constraint on gMPSG is: gMPSG ∈ [–15, 15].

, , ,i MPSG i NF MPSG i NFP P g P= + (7)

The revolution signal is rather congested by letting gMPSG = 7%, (see Figures 5–9). As it could be seen, the maximum and minimum absolute values of the revolution curve are decreased. The injected fuel flow remains unaffected while the manifold pressure decreases due to the fault. There are also no changes in manifold temperature diagram. This temperament of the state values is demonstrated by equations (1) and (3–4). The negative gain fault gMPSG = –8% affected the model with diverse trend (see Figures 10–14).

3.5 Throttle actuator gain-fault

This operational mode is representing a fault in the throttle actuator. The model of throttle actuator gain-fault (THAG) fault has an added parameter gTHAG which is multiplied to the positions of throttle in NF mode [equation (8)]. This scalar parameter can have both positive and negative values. Positive one means the actuator fault is escalating the NF-related values and the negative one affects it diversely. The gTHAG ranges, like other parameters, are limited by the operation ranges of the SI engine model gTHAG ∈ [–10, 10].

THAG NF THAG NFgα α α= + (8)

As it could be seen in Figures 5–14 for gTHAG = 10% and gTHAG = –7%, the positive faults would result in a distinct rise in the injected fuel flow, revolution and manifold pressure. The manifold temperature is remained unchanged. The negative gain faults affected the model with diverse trend.

3.6 Manifold temperature sensor gain-fault

The manifold temperature sensor gain-fault (MTSG) operational mode is modelled as an added gain-fault parameter gTSGF to the outcome of MAT sensor. In this case study, fault parameter gMTSG is set to gMTSG ∈ [–5, 15] percentage of the manifold temperature values in ‘NF’ working mode. Manifold temperature remains comparatively constant during the engine operation except for some spikes over transient positions of the throttle plate [see equation (9)].

, , ,i MTSG i NF MTSG i NFT T g T= + (9)

Figures 5–9 show that this operational mode (gMTSG = 7%) results in a significant surge in injected fuel flow. Increasing the fuel flow, the revolution and pressure are increased and decreased respectively. It has to be considered that the manifold temperature is decreasing due to the MTSG fault occurrence. This behaviour of the state parameters of the engine model is stated by equations (1) and (3–4) which were presented in


Section 2. The negative gain fault (gMTSG = –4%) affected the model with diverse trend (see Figures 10–14).

3.7 RPM sensor gain-fault

Normally, RPM sensor, which is also called revolution sensor, is very reliable. However, in this article the probability of occurring faults in this part is also taken into account. RPM sensor gain-fault (RSG) fault is implemented by a gain-fault gRSG which is added to the physical quantity of the RPM sensor [see equation (10)]. The ranges of gRSG are rather wider than other scalar parameters, and it is mainly because of the higher limitations of SI engine model operation ranges regarding to this change. Here, gRSG is constrained by gRSG ∈ [–19, 50]. The lower range is fixed but the higher one can be variable up to 140% (gRSG = 140) due to lower and upper operational ranges of the SI engine model (nSI,engine ∈ [1,000, 6,000] RPM). Although, it is theoretically possible to assume gRSG > 50, it results in applying an unrealistic magnitude of a fault in RPM sensor.

RSG NF RSG NFn n g n= + (10)

By implementation of this fault for instance gRSG = 25% to the engine model, (see Figures 5–9), the revolution curve is rather congested. In addition, the injected fuel rate and manifold temperature are unaffected by the RSG fault while the manifold pressure is decreased in this operational mode. This behaviour of manifold temperature is demonstrated by equations (1) and (3–4). Negative gain faults affected the model with diverse trend which is shown in Figures 10–14 for gRSG = –11%.

Figure 5 Engine speed in response to positive fault modelling – fault percentage: IL: 7%, FAG: 10%, MPSG: 7%, THAG: 10%, MTSG: 7%, RSG: 25% (see online version for colours)

0 5 10 15 20 25 30 35 40 45 502000

2200

2400

2600

2800

3000

3200

3400

3600

3800

Time (sec)

Rev

olut

ion

(rpm

)

NormalILFAGMPSGTHAGMTSGRSG


Figure 6 Manifold temperature in response to positive fault modelling – fault percentage: IL: 7%, FAG: 10%, MPSG: 7%, THAG: 10%, MTSG: 7%, RSG: 25% (see online version for colours)

0 5 10 15 20 25 30 35 40 45 50250

260

270

280

290

300

310

Time (sec)

Tem

pera

ture

(K)


Figure 7 Manifold pressure in response to positive fault modelling – fault percentage: IL: 7%, FAG: 10%, MPSG: 7%, THAG: 10%, MTSG: 7%, RSG: 25% (see online version for colours)

0 5 10 15 20 25 30 35 40 45 5038

40

42

44

46

48

50

52

54

56

58

Time (sec)

Pre

ssur

e (k

Pa)



Figure 8 Fuel injection in response to positive fault modelling – fault percentage: IL: 7%, FAG: 10%, MPSG: 7%, THAG: 10%, MTSG: 7%, RSG: 25% (see online version for colours)

0 5 10 15 20 25 30 35 40 45 506.5

7

7.5

8

8.5

9

9.5

10

10.5

11x 10-4

Time (sec)

Inje

cted

fuel

(Kg)


Figure 9 Throttle plate manoeuvre in response to positive fault modelling – fault percentage: IL: 7%, FAG: 10%, MPSG: 7%, THAG: 10%, MTSG: 7%, RSG: 25% (see online version for colours)

0 5 10 15 20 25 30 35 40 45 5020.5

21

21.5

22

22.5

23

23.5

24

24.5

25

Time (sec)

Thro

ttle

posi

tion

(deg

)



Figure 10 Engine speed in response to negative fault modelling – fault percentage: FAG: –5%, MPSG: –8%, THAG: –7%, MTSG: –4%, RSG: –11% (see online version for colours)

0 5 10 15 20 25 30 35 40 45 502000

2100

2200

2300

2400

2500

2600

2700

2800

Time (sec)

Rev

olut

ion

(rpm

)

NormalFAGMPSGTHAGMTSGRSG

Figure 11 Manifold temperature in response to negative fault modelling – fault percentage: FAG: –5%, MPSG: –8%, THAG: –7%, MTSG: –4%, RSG: –11% (see online version for colours)

0 5 10 15 20 25 30 35 40 45 50285

290

295

300

305

310

315

320

325

330

Time (sec)

Tem

pera

ture

(K)



Figure 12 Manifold pressure in response to negative fault modelling – fault percentage: FAG: –5%, MPSG: –8%, THAG: –7%, MTSG: –4%, RSG: –11% (see online version for colours)

0 5 10 15 20 25 30 35 40 45 5038

40

42

44

46

48

50

52

54

56

58

Time (sec)

Pre

ssur

e (k

Pa)


Figure 13 Fuel injection in response to negative fault modelling – fault percentage: FAG: –5%, MPSG: –8%, THAG: –7%, MTSG: –4%, RSG: –11% (see online version for colours)

0 5 10 15 20 25 30 35 40 45 505.5

6

6.5

7

7.5

8

8.5

9x 10-4

Time (sec)

Inje

cted

fuel

(Kg)



Figure 14 Throttle plate manoeuvre in response to negative fault modelling – fault percentage: FAG: –5%, MPSG: –8%, THAG: –7%, MTSG: –4%, RSG: –11% (see online version for colours)

0 5 10 15 20 25 30 35 40 45 5019

19.5

20

20.5

21

21.5

22

22.5

23

Time (sec)

Thro

ttle

(deg

)


4 Artificial neural network and its advantages

Several advantages can be attributed to ANNs, rendering them suitable to applications such as considered here. Firstly, an ANN learns the behaviour of a database population by self-tuning its parameters in such a way that the trained ANN matches the employed data accurately (Haykin, 1999). Secondly, if the data used are sufficiently descriptive, the ANN provides a rapid and confident prediction as soon as a new case, which has not been seen by the model during the training phase, is applied. Possibly, the most important aspect of ANNs is their ability to discover patterns in data that are so obscure as to be imperceptible to normal observation and standard statistical methods. This is particularly the case for data exhibiting significantly unpredictable non-linearities. Traditional correlations are based on simple models, which often have to be stretched by adding terms and constants in order for them to become flexible enough to fit experimental data, whereas neural networks are self-adaptable. Using a sufficiently large database for training, ANNs allow property values to be accurately predicted over a very wide range of’ input data. Moreover, ANNs are fast-responding systems. Once the model has been trained, generalising and predictions about unknown inputs are obtained with direct and rapid calculations without the need for tuning or iterative computations. Neural network methodology used in a more broad sense can help engineers to predict the indecent fault modes of engines.

Another major functionality of neural networks is classification, which makes them quite consistence for being used in FDI problems. Low calculation cost is another remarkable benefit of ANNs during real-time applications, which distinguish them from


classical methods like residual generation. Non-sensitivity to abnormalities of the various fault signals reasons the ANNs to have proper ability of generalisation and prediction for trained faults. That is, an ANNs trained to diagnose a certain amount of fault percentages, would be also capable of identifying related fault signals which are rather deformed. Meanwhile, by training an ANNs, it would be tried to regulate the network weights until proper outputs from ANNs are achieved. The neural network toolbox of MATLAB software was indeed able to learn how to find the complex relationships between inputs and outputs of available data. Capabilities for a neural network mentioned above, can be obtained using easy-to-use software. The capability of this toolbox was illustrated by generalising and predicting the outputs for those faults rather deviated from train faults. Deviation of real data from ones implemented by the ANN is reported as network errors. These errors are calculated based on mean square error (MSE) formula given below:

( ), ,

1 1

P Np i p i

real predictedp i

MSE Y Y= =

= −∑∑ (11)

where i is number of nodes in output layer, p is the number of samples and Ypredicted is network outputs and Yred is certified data of SI engine state variables, which extracted before from SI model test procedures.

5 Strategy and simulation

In this case of study, two distinctive networks are designed and then trained to detect and also isolate pre-defined faults. The first network diagnosis positive ranges of faults and the other one is dedicated to negative counterparts. In this case of study, network inputs are arranged as:

1 engine speed

2 intake manifold temperature

3 intake manifold pressure

4 injected fuel rate

5 throttle position.

It has to be mentioned that in this article, developed MVEM SI engine serves as both the real engine and the model. Faults are implemented in the model, and then five above variables are extracted and constituted inputs of diagnose system. In the same time, this model also generates NF mode conditions to provide a base for comparison.

Here, we take advantages of a principle to diagnosis faults based on model validity which in turn relies on a comparison between process variables and their predictions. Process variables originate from operational conditions of engine and the prediction ones result from developed engine’s model. To reduce sensitivity to noise and unordered disturbances, it is beneficial to normalise process variables obtained from working conditions of SI engine model in response to above mentioned faults (see Table 2). Normalisation eliminates the tendency of domination in training process by avoiding each of input variables to have much greater numerical value than others. It has to be


noted that output signals y(t) and related estimated signals ( )y t are extracted in each time step of engine performance (0.1 sec).

( ) ( ) ( )y t y t y tΔ = − (12)

To reduce sensitivity to noise and unordered disturbances, it is beneficial to normalise above established errors between ranges of [–1, 1]. As it is clarified in equation (13), each normalised quantity of Δy(t) is evaluated over the course of 0.1 second (the time step for solving the SI engine equations).

( ) ( )

y tnormalised process variablesy t

= (13)

Constraints on each fault were discussed in Section 3. In order to train two distinct neural networks, which are capable of diagnosing through entire extension of each fault, certain faults are chosen to perform as train data. These faults are determined in Table 3 which their process signals have shown previously through Figures 5–14 in Section 3 as illustrations. An important issue in selecting these particular faults is that we tried to pick them in the middle of the positive or negative ranges of their related fault modes. In this way, it is possible to make the most of generalisation and prediction ability of neural networks in order to isolate all fault modes in their whole ranges. Normalised residuals of above mentioned train data (see Table 3) are shown in Figures 15–19 for positive and Figures 20–24 for negative faults. Table 3 Faults used as trained data

Networks IL FAG MPSG THAG MTSG RSG

Positive 7% 10% 7% 10% 7% 25% Negative - –5% –8% –7% –4% –11%

Networks are then applied to MATLAB and trained by Quick-Propagation algorithm. Networks are then trained surprisingly well with MSE value of 0.0105 for positive network and 0.0145 for negative one. Numbers of samples are set to 501 for each input variables. Therefore, each fault mode has five sets of 501 samples which would be 5 * 3,006 samples for positive case while considering total six classes of fault modes. Having five classes of fault modes in negative case makes total number of samples 5 * 2,505. Consequently, the input matrix dimension would be 5 * 3,006 and 5 * 2,505 for positive and negative networks respectively as if these are arranged consecutively in MATLAB. The reason that the negative network has five sets of 501 samples less than positive one is that the IL fault is omitted in negative case due to its non-negative characteristics.

The outputs of the networks are arranged differently. As it could be seen in Tables 4 and 5 for sorting each class of fault modes including ‘NF’ mode, a vector with the arrangement of one or zero is selected. If the system is working on its fault-free mode, the first vector which has a one in the first element (representing the NF fault mode) and value of zero for other members is selected and in case of other fault modes, related vectors are applicable. As a result of this way of arranging and with consideration of the fact that the output samples have to be the same size of the input samples, the output matrix will be 7 * 3,507 and 6 * 3,006 in dimension.


Figure 15 Engine speed normalised residuals resulted from positive fault modelling (see online version for colours)

0 5 10 15 20 25 30 35 40 45 50-1

-0.8

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

0.8

1

Time (sec)

Rev

olut

ion

norm

aliz

ed re

sidu

als


Figure 16 Manifold air temperature normalised residuals resulted from positive fault modelling (see online version for colours)

0 5 10 15 20 25 30 35 40 45 50-1

-0.8

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

0.8

1

Time (sec)

Tem

pera

ture

nor

mal

ized

resi

dual

s



Figure 17 Manifold air pressure normalised residuals resulted from positive fault modelling (see online version for colours)

0 5 10 15 20 25 30 35 40 45 50-1

-0.8

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

0.8

1

Time (sec)

Pre

ssur

e no

rmal

ized

resi

dual

s


Figure 18 Injected fuel normalised residuals resulted from positive fault modelling (see online version for colours)

0 5 10 15 20 25 30 35 40 45 50-1

-0.8

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

0.8

1

Time (sec)

Fuel

inje

ctio

n no

rmal

ized

resi

dual

s



Figure 19 Throttle position normalised residuals resulted from positive fault modelling (see online version for colours)

0 5 10 15 20 25 30 35 40 45 50-1

-0.8

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

0.8

1

Time (sec)

Thro

ttle

posi

tion

norm

aliz

ed re

sidu

als


Figure 20 Engine speed normalised residuals resulted from negative fault modelling (see online version for colours)

0 5 10 15 20 25 30 35 40 45 50-1

-0.8

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

0.8

1

Time (sec)

Rev

olut

ion

norm

aliz

ed re

sidu

als



Figure 21 Manifold air temperature normalised residuals resulted from negative fault modelling (see online version for colours)

0 5 10 15 20 25 30 35 40 45 50-1

-0.8

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

0.8

1

Time (sec)

Tem

pera

ture

nor

mal

ized

resi

dual

s


Figure 22 Manifold air pressure normalised residuals resulted from negative fault modelling (see online version for colours)

0 5 10 15 20 25 30 35 40 45 50-1

-0.8

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

0.8

1

Time (sec)

Pre

ssur

e no

rmal

ized

resi

dual

s



Figure 23 Injected fuel normalised residuals resulted from negative fault modelling (see online version for colours)

0 5 10 15 20 25 30 35 40 45 50-1

-0.8

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

0.8

1

Time (sec)

Inje

cted

fuel

nor

mal

ized

resi

dual

s


Figure 24 Throttle position normalised residuals resulted from negative fault modelling (see online version for colours)

0 5 10 15 20 25 30 35 40 45 50-1

-0.8

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

0.8

1

Time (sec)

Thro

ttle

posi

tion

norm

aliz

ed re

sidu

als



Table 4 Classification vectors for positive fault classes

NF IL FAG MPSG THAG MTSG RSG

1000000

⎡ ⎤⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎣ ⎦

0100000

⎡ ⎤⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎣ ⎦

0010000

⎡ ⎤⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎣ ⎦

0001000

⎡ ⎤⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎣ ⎦

0000100

⎡ ⎤⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎣ ⎦

0000010

⎡ ⎤⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎣ ⎦

0000001

⎡ ⎤⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎣ ⎦

Table 5 Classification vectors for negative fault classes

NF FAG MPSG THAG MTSG RSG

100000

⎡ ⎤⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎣ ⎦

010000

⎡ ⎤⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎣ ⎦

001000

⎡ ⎤⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎣ ⎦

000100

⎡ ⎤⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎣ ⎦

000010

⎡ ⎤⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎣ ⎦

000001

⎡ ⎤⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎣ ⎦

For the purpose of reducing calculation cost and time, the numbers of layers and nodes should be as small as possible in trained network. It has to be taken into account that the numbers of hidden layers and related nodes are determined by trial and error. Hence, in this case study, both neural networks are designed with two hidden layers including seven nodes in the first layer and ten nodes in the second one. Although both networks have a five nodes input layer, positive network has seven nodes in its output layer while the negative one has six. Networks are then applied to MATLAB and trained by quick-propagation (‘trainrp’ function in MATLAB) method.

Figures 25 and 26 demonstrate precision of positive and negative trained networks. As it could be seen, networks are trained surprisingly well with MSE value of 0.0035 for positive network and 0.0029 for negative one. In order to test the ability of networks for diagnosing the exact trained faults, these faults are fed as inputs to networks that the results are shown in Figures 27–28. It is clear and predictable that both networks do extremely well in diagnosing these faults. In both figures, first 7 * 501 and 6 * 501 sets of outputs portray response of trained positive and negative networks to NF mode correspondingly while other series of 7 * 501 and 6 * 501 are dedicated to remaining fault modes according to Table 3. As it could be seen, networks do exceedingly fine while the related fault modes’ elements in output vectors are one in almost every 501 samples and the remaining fault modes’ elements are zero. Some deviations from one and zero are observed, but the performance of the diagnostic system is not affected.


Figure 25 Convergent diagram of positive network (see online version for colours)

Figure 26 Convergent diagram of negative network (see online version for colours)


Figure 27 Outputs of the positive trained neural network to its trained faults (see online version for colours)

500 1000 1500 2000 2500 3000 35000

0.5

1

IL

500 1000 1500 2000 2500 3000 35000

0.5

1

FAG

500 1000 1500 2000 2500 3000 35000

0.5

1

MP

SG

500 1000 1500 2000 2500 3000 35000

0.5

1

THA

G

500 1000 1500 2000 2500 3000 35000

0.5

1

MTS

G

500 1000 1500 2000 2500 3000 35000

0.5

1

RS

G

500 1000 1500 2000 2500 3000 35000

0.5

1

NF

Figure 28 Outputs of the negative trained neural network to its trained faults (see online version for colours)

500 1000 1500 2000 2500 30000

0.5

1

FAG

500 1000 1500 2000 2500 30000

0.5

1

MP

SG

500 1000 1500 2000 2500 30000

0.5

1

THA

G

500 1000 1500 2000 2500 30000

0.5

1

MTS

G

500 1000 1500 2000 2500 30000

0.5

1

RS

G

500 1000 1500 2000 2500 30000

0.5

1

NF


Figure 29 Normalised residuals resulted from RSG fault mode’s positive extreme boundaries (see online version for colours)

0 5 10 15 20 25 30 35 40 45 50-1

-0.8

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

0.8

1

Time (sec)

Nor

mal

ized

resi

dual

s

Injected fuelRevolutionPressureTemperatureThrottle position

Figure 30 Normalised residuals resulted from RSG fault mode’s negative extreme boundaries (see online version for colours)

0 10 20 30 40 50-1

-0.8

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

0.8

1

Time (sec)

Nor

mal

ized

resi

dual

s

Injected fuelRevolutionPressureTemperatureThrottle position

To test the generalisation and prediction capabilities of networks, different percentages of RSG fault mode are chosen as inputs. Hence, two extreme positive and negative extents of this fault mode were executed on SI engine model and results were fed into networks. As it is shown in Figure 29, solid lines indicate normalised residuals of gRSG = 3% while


dashed-dotted lines represent gRSG = 50%. Likewise, Figure 30 presents normalised residuals of gRSG = –3% through dashed-dotted lines and gRSG = –19% by solid lines which are negative extreme boundaries of RSG fault mode. MSE results of positive and negative networks related to above mentioned RSG test percentages are listed in Table 6. As could be understood from this table, the MSE values in different RSG fault percentages go up when fault percentages deviate from train data values. This behaviour of MSE values can be demonstrated from the ability of ANNs to identify faults which are rather deviated from their train data values. Table 6 MSE results of positive and negative networks related to various RSG test

percentages

Test errors of RSG in positive network Test errors of RSG in negative network

Fault percentages MSE

Fault percentages MSE

3% 0.0762 –3% 0.0908 15% 0.0179 –5% 0.0739 25% 0.0155 –7% 0.0558 35% 0.0344 –11% 0.0045 50% 0.0803 –15% 0.0387 –19% 0.1415

Uncertainties, noises and unordered disturbances are main causes of false alarm during the engine’s fault free mode. For the purpose of increasing immunity of diagnosis system to small faults and consequently false alarms, the ranges of each fault parameters described in Section 3 are extended to (–3, 3) for their NF mode in stead of a single point {0}. In this way, a scalar gain-fault parameter considers as a fault if its significance is higher than three or in another word if the fault is greater than 3%.

The positive network does well in diagnosis RSG faults’ utmost boundaries. According to Figures 31–34, it is obvious that NF, IL, THAG, MTSG fault modes are not sensitive to positive extensions of this fault mode while they are entirely zero in all of their outputs samples. This temperament could be also concluded from Figure 27 while the positive network has no output other than zero in response to RSG trained fault in four above mentioned fault mode rows. Consequently, no fluctuations are exhibited in the last 501 samples from 3,006 to 3,507 in these fault modes. In contrast, some deviations from zero are seen in the final 501 samples of FAG and MPSG fault modes of which illustrate sensibility to RSG. Moreover, in the ending row of Figure 27, which represents RSG, fault mode, there are some oscillations in the fourth part of the samples from 1503 to 2004. These dispositions of samples represent RSG fault mode dependency to MPSG. This behaviour somehow degraded the isolation performance of positive network in response to RSG and MPSG fault modes. As it could be seen in Figure 31, performance of positive network in gRSG = 3% case is devalued by this dependency of RSG and MPSG fault modes. As a result, some samples are not zero in MPSG row, which is not desirable in isolation RSG fault mode. It is also understood that FAG fault mode is not sensible to upper positive ranges of RSG. In gRSG = 3% case, the network performance reveals dependency between RSG and FAG fault modes. As it is clear in Figure 32, there are some FAG samples, which are not zero in response to RSG lower positive ranges. It is also mentionable that MPSG fault mode is not sensitive to this boundary of RSG.


In spite of the fact that the sensitivity of FAG and MPSG fault modes to RSG will worsen the ability of positive and negative networks to isolate entire extensions of RSG fault mode, networks are still completely able to fulfil the diagnostic task. In order to certify this fact Tables 7 and 8 are provided. In these tables, the responses of positive and negative networks to all ranges of each fault mode are presented. Table 7 represents the positive network responds and Table 8 is dedicated to negative one. Networks outcomes are evaluated by integrating the values of each fault mode’s samples generated by network [see equations (14) and (15)].

501

,1

[ ]ij F n

n

P O=

= ∑ (14)

501

,1

[ ]ij F n

n

N O=

= ∑ (15)

where P and N indicate the positive and negative networks’ performances, i and j represent applied fault modes and their percentages respectively. O stands for output samples resulted from network for each applied fault modes – which are between 0 and 1 – and F represents all fault modes that network is capable of diagnose. Finally, n is the amount of samples used for representing each fault mode in networks’ outputs. Performance indicators P and N are scalar vectors with seven and six members respectively.

Figure 31 Outputs of the positive neural network to gRSG = 50% (see online version for colours)

50 100 150 200 250 300 350 400 450 5000

0.5

1

NF

50 100 150 200 250 300 350 400 450 5000

0.5

1

IL

50 100 150 200 250 300 350 400 450 5000

0.5

1

FAG

50 100 150 200 250 300 350 400 450 5000

0.5

1

MP

SG

50 100 150 200 250 300 350 400 450 5000

0.5

1

THA

G

50 100 150 200 250 300 350 400 450 5000

0.5

1

MTS

G

50 100 150 200 250 300 350 400 450 5000

0.5

1

RS

G


Figure 32 Outputs of the positive neural network to gRSG = 3% (see online version for colours)

50 100 150 200 250 300 350 400 450 5000

0.5

1

NF

50 100 150 200 250 300 350 400 450 5000

0.5

1

IL

50 100 150 200 250 300 350 400 450 5000

0.5

1

FAG

50 100 150 200 250 300 350 400 450 5000

0.5

1

MP

SG

50 100 150 200 250 300 350 400 450 5000

0.5

1

THA

G

50 100 150 200 250 300 350 400 450 5000

0.5

1

MTS

G

50 100 150 200 250 300 350 400 450 5000

0.5

1

RS

G

Figure 33 Outputs of the negative neural network to gRSG = –3% (see online version for colours)

50 100 150 200 250 300 350 400 450 5000

0.5

1

NF

50 100 150 200 250 300 350 400 450 5000

0.5

1

FAG

50 100 150 200 250 300 350 400 450 5000

0.5

1

MP

SG

50 100 150 200 250 300 350 400 450 5000

0.5

1

THA

G

50 100 150 200 250 300 350 400 450 5000

0.5

1

MTS

G

50 100 150 200 250 300 350 400 450 5000

0.5

1

RS

G


Figure 34 Outputs of the negative neural network to gRSG = –19% (see online version for colours)

50 100 150 200 250 300 350 400 450 5000

0.5

1

NF

50 100 150 200 250 300 350 400 450 5000

0.5

1

FAG

50 100 150 200 250 300 350 400 450 5000

0.5

1

MP

SG

50 100 150 200 250 300 350 400 450 5000

0.5

1

THA

G

50 100 150 200 250 300 350 400 450 5000

0.5

1

MTS

G

50 100 150 200 250 300 350 400 450 5000

0.5

1

RS

G

Regarding the above definitions, when there is no dependency among various fault modes, the fault mode’s related member in these vectors has to become 501. In order to present complete depopulation among various fault modes, other members are required to take value of zero. In above case, it is necessary that each network output sample is just sensitive to their trained faults, or in other words, all faults are entirely decoupled. Such a separation among various fault modes is the goal of every FDI approaches. In RSG, fault mode instance, MPSG, FAG and RSG fault modes are not completely decoupled from each other. The performance results from networks brought in Tables 7 and 8 also put premium on this issue. In Table 7, the third and fourth members, representing FAG and MPSG fault modes, display substantial reaction to lower and upper ranges of the RSG fault mode respectively. Likewise, in Table 8, MPSG member is not zero.

Some noticeable characteristics among other fault modes are also observable in Tables 7 and 8. The NF mode is totally isolated in both networks with performance parameter of 501, located in the first element of its output vector. IL is likewise decoupled from other fault modes although THAG shows some sensitivity to its utmost upper positive boundary. Similarly, a complete isolation is seen in FAG fault mode in spite of RSG response to its lower negative extent. As it is described in above paragraphs, RSG and MPSG are somehow dependent to each other, which are apparent in both tables. In addition, FAG is sensitive to lower positive extreme of RSG. THAG is sensitive to IL in its lower positive and negative ranges to some extent. Here, performance dependency of networks to a selection of faults as train data is observable. Due to this issue, the negative network reveals better performance in THAG faults’ isolation rather than positive one. It is mainly because of that the THAG quantity used for training negative network (–7%) stands nearer to the middle of THAG negative extension ([0, –10]) in


comparison to its positive counterpart (10%). As a result, negative network is able to make the most of generalisation and prediction ability of neural networks rather than positive one, which utilised upper positive utmost range. Finally, MTSG fault mode is completely isolated. Apart from these dependencies among fault modes, the performance values of incident fault mode in both networks are higher than others in all cases. This issue leads to a safe isolation of right fault mode in this diagnosis system while the higher performance factor can correctly isolate the implemented fault mode. Table 7 Responses of the positive network to all ranges of each fault mode

Fault percentages NF IL FAG MPSG THAG MTSG RSG

NF 0% 501 0 1 1 0 0 0 IL 3% 0 490 0 0 0 6 0 5% 0 493 0 0 0 1.5 0 7% 0 496 0 0 0 0 0 10% 0 304 4 0 0 0 0 13% 0 193 9 0 6 0 0 15% 0 141 10 0 103 0 0 FAG 3% 0 4 474 0 0 11 6 5% 0 3 476 0 0 9 6 7% 0 2 479 0 0 5 6 10% 0 0 486 0 0 0 6 13% 0 0 488 0 0 0 6 15% 0 0 489 0 0 0 6 MPSG 3% 0 0 0 290 0 0 154 5% 0 0 0 447 0 0 37 7% 0 0 0 489 0 0 8 10% 0 7 0 491 1 0 8 13% 0 40 0 492 1 0 7 15% 0 66 0 492 1 0 6 THAG 3% 0 142 2 17 319 0 0 5% 0 33 1 1 356 0 0 7% 0 0 2 0 429 0 0 10% 0 0 0 0 501 0 0 MTSG 3% 0 0 0 0 0 483 0 5% 0 0 0 0 0 487 0 7% 0 0 0 0 0 499 0 10% 0 7 0 0 0 498 0 13% 0 9 0 0 0 491 0 15% 0 10 1 0 0 488 0 RSG 3% 0 0 241 26 0 0 465 15% 0 0 18 23 0 0 468 25% 0 0 0 24 0 0 469 35% 0 0 0 66 0 0 431 50% 0 0 0 144 0 0 350


Table 8 Responses of the negative network to all ranges of each fault mode

Fault percentages NF IL FAG MPSG THAG MTSG RSG

NF 0% 501 0 0 0 0 0 0% IL –3% 0 492 9 0 2 1 –3% –5% 0 494 8 0 0 1 –5% –8% 3 457 6 6 0 144 –8% –3% 0 0 257 0 0 234 –3% –5% 0 0 258 0 0 231 –5% –7% 0 0 435 0 0 53 –7% FAG –8% 5 0 475 0 0 16 –8% –10% 12 30 434 0 0 50 –10% –13% 17 58 430 0 4 56 –13% –15% 0 59 378 0 6 53 –15% –3% 0 97 23 401 4 0 –3% –5% 0 71 1 431 2 0 –5% MPSG –7% 0 0 0 500 0 0 –7% –10% 1 0 8 499 0 0 –10% –3% 0 6 0 0 490 0 –3% –4% 0 0 0 0 498 0 –4% –5% 0 25 0 4 498 0 –5% –3% 0 0 151 0 0 350 –3% THAG –5% 0 0 125 0 0 375 –5% –7% 0 0 98 0 0 400 –7% –11% 0 0 20 0 0 479 –11% –15% 0 1 99 0 0 410 –15% MTSG –19% 9 24 164 0 0 314 –19% 0% 501 0 0 0 0 0 0% –3% 0 492 9 0 2 1 –3% –5% 0 494 8 0 0 1 –5% –8% 3 457 6 6 0 144 –8% –3% 0 0 257 0 0 234 –3% RSG –5% 0 0 258 0 0 231 –5% –7% 0 0 435 0 0 53 –7% –8% 5 0 475 0 0 16 –8% –10% 12 30 434 0 0 50 –10% –13% 17 58 430 0 4 56 –13%

In robustness discussion, presented method is fed by normalised residuals instead of process variables, which have an enormous role in minimising effects of noises, disturbances and unordered uncertainties to high degrees. In addition, extending constraints on NF operating mode by ignoring small gain-fault parameters was immune the diagnosis system to false alarms. Moreover, as it is discussed above, this method has


the ability to diagnose a whole range of faults related to operating conditions of engine. This characteristic empowers this diagnosis system to show robust performance in any working conditions. This method works such an extent that we apply this method for real-world engines’ fault diagnostic applications.

6 Conclusions

In this paper, a diagnosis method based on residual generation and neural network classifier was developed on intake manifold of an SI engine. The main goal of this work was to present and apply a diagnostic method which is fast and accurate, and also has low computational cost. For this reason, the diagnosis system was designed by neural network which was shown to be a promising way of diagnosing faults occurred in the intake manifold of the SI engine. This method was capable of diagnosing not only the predefined trained faults, but also entire ranges of these faults, which are dependent on working condition of engine. Finally, in spite of the method simplicity, it was completely capable of diagnosing both positive and negative faults with substantially good accuracy and robustness.

The results of this study can be employed in new generation of fault diagnosis systems since the traditional ways are capable of warning the driver when the fault occurred. However, this method has the potentiality to introduce a gradual illustration of engine health even before the fault occurs in engine. Also, the accuracy will increase owing to sensitivity of diagnosis system to lower changes in faults in comparison with conventional approaches which use threshold criteria to address diagnosis problem.

References Bakiotis, C. (1979) ‘Parameter and discriminate analysis for jet engine mechanical state diagnosis’,

Proceeding of the IEEE Conference on Decision and Control, Piscataway, NJ, pp.1–11. Beard, R. (1971) Failure Accommodation in Linear Systems though Self-Reorganization, Dept.

MVT – 71 – 1, Man Vehicle Laboratory, Cambridge, MA. Birnbaum, R. and Truglia, J. (2000) Getting to Know OBD II, ISBN 0-9706711-0-5, New York. Chevalier, A., Muller, M. and Hendricks, E. (2000) On the Validity of Mean Value Engine Models

during Transient Operation, SAE 2000-01-1261. Clark, R.N. (1978) A Simplified Instrument Failure Detection Scheme, IEEE Transactions

Aerospace Electronic System, Vol. 14, pp.558–563. Filbert, D. and Metzger, K. (1982) ‘Quality test of systems by parameter estimation’, 9th IMEKO

Congress, Berlin. Fons, M., Muller, M., Chevalier, A. and Hendricks, E. (1999) Mean Value Engine Modeling of an

SI Engine with EGR, SAE 1999-01-0909. Frank, P.M. and Seliger, B.K. (1997) ‘Fuzzy logic and neural network applications to fault

diagnosis’, International Journal of Approximate Reasoning, pp.67–88. Gen-Ting, Y. and Guang-Fu, M. (2004) ‘Fault diagnosis of diesel engine combustion system based

on neural network’, Proceedings of the Third International Conference on Machine Learning and Cybernetics, Shanghai.

Gertler, J., Costin, M., Luo, Q., Fang, X.W., Hira, R. and Kowalczuk, Z. (1991) On-Board Fault Detection and Isolation for Automotive Engines Using Orthogonal Parity Equation, Invited Paper, Preprints of IFAC Conference on Fault Detection, Supervision and Safety, Baden-Baden, Germany, Vol. 2, pp.241–246.


Haykin, S. (1999) Neural Networks, ‘A Comprehensive Foundation’, 2nd ed., Prentice Hall Inc. Hendricks, E. and Sorenson, S.C. (1990) Mean Value Modeling of Spark Ignition Engines, SAE

900616. Hendricks, E. and Vesterholm, T. (1992) The Analysis of Mean Value SI Engine Models, SAE

920682. Hofling, T. (1993) ‘Detection of parameter variations by continuous time parity equations’, IFAC

World Congress, Sydney, Australia, pp.513–518. ISO Standards (2010) ‘Road vehicles – diagnostic communication over controller area network

(DoCAN) – Part 4: requirements for emissions-related systems’, ISO/FDIS 15765-4.2. Jones, H.L. (1973) ‘Failure detection in linear systems’, PhD thesis, MIT, Cambridge, MA. Mostofi, M., Shamekhi, A.H. and Gorji-Bandpy, M. (2006a) ‘Modified mean value SI engine

modeling (EGR included)’, Proceedings of the International Conference on Modeling and Simulation, Paper No. A076, Konya, Turkey.

Mostofi, M., Shamekhi, A.H. and Ziabasharhagh, M. (2006b) ‘Developing an algorithm for SI engine diagnosis using parity relations’, Proceedings of IMECE2006, IMECE2006-13463, Chicago, Illinois, USA.

Muller, M., Hendricks, E. and Sorenson, S.C. (1998) Mean Value Modeling of Turbocharged Spark Ignition Engines, SAE 980784.

Naidu, M., Shoepf, T.J. and Gopalakrishnan, S. (2005) Arc Fault Detection Schemes for an Automotive 42 V Wire Harness, SAE 2005-01-1742.

Nyberg, M. (2003) ‘Using hypothesis test theory to evaluate principles for leakage diagnosis of automotive engines’, Control Engineering Practice, Vol. 11, pp.1263–1272.

Nyberg, M. and Stutte, T. (2004) ‘Model based diagnosis of the air path of an automotive diesel engine’, Control Engineering Practice, Vol. 12, pp.513–525.

Patton, R.J. and Chen, J. (1991) ‘Robust fault detection using Eigen structure assignment, a tutorial consideration and some new results’, Proceeding of the 30th IEEE Conference on Decision and Control, Brighton, pp.2242–2247.

Rumelhart, D.E. and McClelland, J.L. (1986) Parallel Distributed Processing, MIT Press, Cambridge, Massachusetts.

SAE International (2003) On-Board Diagnostics for Light and Medium Duty Vehicles Standards Manual, Pennsylvania, ISBN 0-7680-1145-0.

Shamekhi, A.H. and Ghaffari, A. (2004) An Improved Model for SI Engines, ASME, ICEF-2004-818.

Shen, Y., Cao, L., Wang, Z., Zhou, S. and Gou, B. (2006) ‘Fault diagnosis of diesel fuel ejection system based on improved WNN’, Proceedings of the 6th World Congress on Intelligent Control and Automation, Dalian, China.

Willsky, A. (1976) ‘A survey of design methods for failure detection in dynamic system’, Automatica, No. 12, pp.601–611.

Nomenclature

I Moment of inertia (m4) m Mass flow rate (kg/s) n Engine speed (rpm) P Pressure (kPa) Q Heating value (kJ) R Universal gas constant (mole/kg.K)


Nomenclature (continued)

T Temperature (K) X Fuel split parameter v Volume (m3) FDI Fault diagnosis and isolation OBD On-board diagnosis MVEM Mean value engine modelling NF No fault IL Intake manifold leakage FAG Fuel injector actuator gain-fault MPSG Manifold pressure sensor gain-fault THAG Throttle actuator gain-fault MTSG Manifold temperature sensor gain-fault RSG RPM sensor gain-fault ANN Artificial Neural Network MAT Manifold air temperature SI Spark ignition engine HDOBD Heavy duty OBD MIMO Multi-input multi-output MSE Mean square error EGR Exhaust gas recirculatioin

Greek letters

α Throttle angle (deg)

η Efficiency

κ Gas atomicity coefficient

τ Time constant (sec)

Subscripts

a Ambient at past over the throttle plate

d Displacement

f Fuel

fc Fuel conversion

ff Fuel film

i Intake manifold

m Mean

r Manifold to ambient ratio

t Total

v Volumetric

an approach to full-range fault diagnosis of spark ... approach to full-range fault diagnosis of...

Documents