2972
Energy & Fuels 2008, 22, 2972–2980
Combustion Faults Diagnosis in Internal Combustion Engines Using Angular Speed Measurements and Artificial Neural Networks Fernando Cruz-Peragon,*,† Francisco J. Jimenez-Espadafor,‡ Jose M. Palomar,† and M. Pilar Dorado§ Department of Mechanics and Mining Engineering, EPS de Jaen, UniVersidad de Jaen, Campus Las Lagunillas s/n, 23071 Jaen, Spain, Department of Energetic Engineering, ESI, UniVersidad de SeVilla, Camino de los Descubrimientos s/n, Isla de la Cartuja, 41092 SeVilla, Spain, and Department of Chemical Physics and Applied Thermodynamics, EPS, UniVersidad de Cordoba, Campus de Rabanales, Edificio Leonardo da Vinci, 14071 Cordoba, Spain ReceiVed March 3, 2008. ReVised Manuscript ReceiVed May 12, 2008
Recent studies have demonstrated that cylinder pressure profiles can be estimated along the operating cycle by means of artificial neural networks (ANNs), provided that the instantaneous angular speed is known. However, to be used in automotive applications, a higher level of confidence is required. Despite this restriction, ANNs can be considered a reliable tool for engines fault diagnosis. This paper goes in this direction. According to this, ANN has been applied to three different engines, that is a three-cylinder spark-ignition engine (SIE), a four-cylinder SIE, and a V-16 cylinder compression ignition engine (CIE). Results showed the suitability of the proposed methodology to diagnose faults in any internal combustion engine. To guarantee a correct identification, the characterization of the main parameters of the network was improved. Finally, a sensitivity analysis applied to the V-16 engine model concluded that the position of the faulty cylinder does not interfere in the suitability of the results.
Introduction Misfiring cylinder detection and faulty cylinder identification techniques using crankshaft angular velocity fluctuations have been under study for the last 20 years. Usual methods are summarized into four main groups:1 threshold criteria, pattern recognition, temporal filtering, and iterative techniques.2–5 Also, the complexity of the model and the analysis methods constitute other important features to be taken into consideration. However, the major obstacle to develop an effective technique is that a set of assumptions is required to match the unknown input data to the known measured data (constant coefficient driveline model, rigid crankshaft, constant load torque, and non-overlapping firing pulse assumptions). Among the previously mentioned methods, the iterative technique provides the only procedure that requires no assumptions. This technique can be applied to a complex system model, including a flexible crankshaft with a driveline model. The known torsional system response (torque and angular velocity) is related to the unknown system input data (individual cylinder contributions), until velocity waveform matches the measured data. The contribution of each cylinder is referred to * To whom correspondence should be addressed. Telephone: +34-953212367. Fax: +34-953-212870. E-mail:
[email protected]. † Universidad de Jaen. ‡ Universidad de Sevilla. § Universidad de Cordoba. (1) Williams, J. SAE Paper 960039, 1996. (2) Yang, J. G.; Pu, L. J.; Wang, Z. H.; Zhou, Y. C.; Yan, X. P. Mech. Syst. Signal Process. 2001, 15, 549–564. (3) Taraza, D. J. Eng. Gas Turbines Power 2003, 125, 797–803. (4) Geveci, M.; Osburn, A. W.; Franchek, M. A. Mech. Syst. Signal Process. 2005, 19, 1107–1134. (5) Moro, D.; Cavina, N.; Ponti, F. J. Eng. Gas Turbines Power 2002, 124, 220–225.
the indicated mean effective pressure, the instantaneous indicated torque, or the pressure profile. Moreover, state space deconvolution is valid for nonlinear and transient solutions using time or crank angle domain. Nevertheless, the indeterminate equation set does not converge to a unique solution and together with the unobservability at the top dead center (TDC) can suppose a serious problem.1 Currently, intelligent optimization techniques have been included in fault recognition methodologies, among them are contributions related to pressure reconstruction with very interesting results.6–10 In addition, when mathematical modeling is employed for pressure reconstruction as a substitute of waveform evaluation, some important parameters related to the combustion process (such as fuel consumption and injection or ignition timing) must be estimated, increasing engine behavior information. A literature review related to techniques to model and control combustion processes shows several applications conducted in the fields of chemical reaction mechanisms and emissions.11 This could also be applied to diesel engines analysis, provided that the fuel injection system is conveniently designed and modeled.12,13 (6) Lu, P. J.; Zhang, M. C.; Hsu, T. C.; Zhang, J. J. Eng. Gas Turbines Power 2001, 123, 340–346. (7) Johnsson, R. Mech. Syst. Signal Process. 2006, 20, 1923–1940. (8) Gu, F.; Jacob, P. J.; Ball, A. D. Proc. Inst. Mech. Eng., Part D 1999, 213, 135–143. (9) Jacob, P. J.; Gu, F.; Ball, A. D. Proc. Inst. Mech. Eng., Part D 1999, 213, 73–81. (10) Scaife, M. N.; Charlton, S. J.; Mobley, C. SAE Paper 930861, 1993. (11) Kalogirou, S. A. Prog. Energy Combust. Sci. 2003, 29, 515–566. (12) Kegl, B.; Mu¨ller, E. SAE Paper 981068, 1998. (13) Palomar, J. M.; Cruz, F.; Ortega, A.; Jimenez-Espadafor, F. J.; Martinez, G.; Dorado, M. P. Energy Fuels 2005, 19, 1526–1535.
10.1021/ef800159r CCC: $40.75 2008 American Chemical Society Published on Web 07/04/2008
Combustion Faults Diagnosis in Internal Combustion Engines
Energy & Fuels, Vol. 22, No. 5, 2008 2973
Table 1. General Characteristics of the Engines engine
three-cylinder SIE
four-cylinder SIE
V-16 power-plant CIE
number of cylinder stroke (mm) bore (mm) compression ratio maximum torque (N m) maximum power (kW)
3 72.0 68.5 9:1 61.7 (at 3000 rpm) 33 (at 6000 rpm)
4 75.5 74.0 9.5:1 102 (at 3500 rpm) 47 (at 6000 rpm)
V-16 (60°) 190 170 18:1 9550 and 1500 (at 1500 rpm) (nominal conditions)
Tocombinebothfeatures,Cruz-PeragonandJimenez-Espadafor14,15 applied two of the most extended techniques, such as genetic algorithms (GAs) and artificial neural networks (ANNs), to approach both the pressure waveform and associated combustion parameters. The proposal used the more complex system model without assumptions and a time domain space state deconvolution. Two validated combustion and dynamic submodels were applied to a single-cylinder compression ignition engine (CIE) and a three-cylinder spark-ignition engine (SIE). Furthermore, limitations of a high number of cylinders (where the overlapping firing pulse and flexible crankshaft behavior are critical) were evaluated theoretically through a V-16 power-plant CIE. Each one of them only used one encoder as a sensor, in front of conventional systems that also need a pressure transducer per cylinder. Most of the investigations over multicylinder engines show the same pressures in different chambers, except in one of the cylinders.5,8 An important contribution of these works is that all cylinders can show different pressure curves independently. Per contra, the main drawbacks are the presence of nonunique solutions. Although combustion parameters provide an acceptable pressure waveform, partial correlation indicates that some of them could be derived from others, thus leading to uncertainties near TDC.16 These problems are related to the characteristics of the analysis, system model, and no assumptions consideration, as previously mentioned. GA optimization methodology simulates each cylinder pressure better than other conventional methods, considering the derivatives of the objective functions. Thus, real pressure profile in cylinders can be approached with very low errors. Engine dynamics play an important role in the accuracy of the results. Because the different individual indicated torques in multicylinder engines are covered up, differences between modeled (optimized) and target (real) curves increase proportionally to the number of cylinders. This fact does not depend upon the type of engine (CIE or SIE). However, the main drawback of this technique is the high computation time required.14 The ANN technique can handle incomplete data, to deal with nonlinear problems, and once trained, can perform predictions and generalizations at high speed. Because the first fundamental modeling of neural nets was proposed in terms of a computational model,17 ANNs have shown their suitability in diverse fields as control, robotics, pattern recognition, forecasting, medicine, power systems, manufacturing, optimization, signal processing, and social and psychological sciences.18–20 The ANN technique needs a relatively little number of neurons, and it is (14) Cruz-Peragon, F.; Jimenez-Espadafor, F. J. Energy Fuels 2007, 21, 2600–2607. (15) Cruz-Peragon, F.; Jimenez-Espadafor, F. J. Energy Fuels 2007, 21, 2627–2636. (16) Cruz-Peragon, F.; Jimenez-Espadafor, F. J. An. Ing. Mec. 2004, 15, 1891–1896. (17) McCulloch, W. S.; Pitts, W. Bull. Math. Biophys. 1943, 5, 115– 133. (18) Casilari, E.; Jurado, A.; Pansard, G.; DiazEstrella, A.; Sandoval, F. Electron. Lett. 1996, 32, 363–365. (19) Williams, M. Talking Nets: An Oral History of Neural Networks; MIT Press: Cambridge, MA, 1998. (20) Yuste, A. J.; Dorado, M. P. Energy Fuels 2006, 20, 399–402.
able to present results with important time saving. The high number of required samples makes it necessary to train the network via simulations, which are model-dependent. The main advantages of this method are as follows: first, the net provides results very quickly; second, it results in a high level of confidence, thus estimating the pressure waveforms correctly; and third, the incorporation of some real data as a training pattern overcomes the initial misadjusting.15 This last contribution is based on radial basis function (RBF) networks.7–9 Despite the differences between contributions, all of them are able to reconstruct the pressure waveform with a high accuracy, estimating the maximum pressure in multicylinder engines with a deviation below 5% in most of the cases. Under less exigent applications, feasibility could approximate to 100%. However, because fault diagnosis in automotive and heavy-duty applications requires nearly 100% of feasibility, preferably using an online methodology, the preceding procedures are still far from this target. With the aim of increasing the reliability of the proposed methodologies in combustion fault detection, model implementation (in the sense of reducing the uncertainties related to system models), and network design (to improve accuracies) are required. When a multicylinder engine works in a large mechanism (such as marine or industrial purposes), the number of operative conditions drops drastically compared to automotive applications. Employees make use of diagnostic tools continuously to supervise the main parameters. Maintenance procedures are widely used, applying multiple sensors in different locations around the engines, generator, and so on. However, it may be more helpful to have indicators related to the health of the engine, such as pressure waveform, ignition timing, or mass of burnt fuel. In this way, some commercial applications with very acceptable results have been developed.21,22 These applications include a direct measurement procedure (using pressure or force transducers). Actually, sensors for continuous monitoring under rough conditions are very robust. Nevertheless, the high cost and physical complexity of direct methods constitute two important drawbacks, in front of the cheap procedures of indirect methods. According to this, the goal of the present work is to evaluate the potential of some designed networks to monitor engine performance and behavior, to improve their design. For this purpose, some systematic test sets will be simulated in engines by means of validated system models (results from combustion and dynamic submodels presenting a good agreement to experiments). Most of the investigations related to multicylinder engines show differences between inner pressures from different chambers. In the present study, a first systematic process will provide several faulty conditions to simulate, including differences in pressure waveforms in more than one cylinder, providing a more realistic scenario. Additionally, there are two critical characteristics of the designed networks that will be investigated through a sensitivity analysis: radius of the Gaussian function (σ) and number of (21) Kistler, 2008.
2974 Energy & Fuels, Vol. 22, No. 5, 2008
Cruz-Peragon et al.
Figure 1. System setup. Table 2. RBF Network Parameters in a z-Cylinder Engine learning patterns hidden layer neurons weight value on hidden layer bias weights output layer neurons weight value on output layer input layer vector length associate weights to hidden layer activation function of hidden layer associate weights to output layer activation function of output layer
XS1 S1 polynomial approximation S1 S2 ) zNp the same for any average speed R S1R Gauss (µ ) 0; σ ) 0.05) S2S1 linear
Table 3. Selected Values for Engine Networks from Table 2 engine
L-3 SIE15
L-4 SIE
V-16 CIE15
X R S1 S2
10 61 1000 3
10 85 1000 4
1 91 1500 16
learning patterns. First, the exponential decay of the Gaussian function ensures the local properties of the RBF model. When a larger σ is used, the output is less sensitive to a change in the Euclidean distance and a better extrapolation between centers can be achieved. Per contra, a smaller value enables the network to predict a slight change in the speed signature but gives poor interpolation between centers. Then, an appropriate value for this radius needs to be identified.9 Second, the simplest training pattern selection consists of randomly choosing a certain number of examples. They are used as radial basis centers in the hidden layer unit. A strict interpolation could make the problem appear over-determined, resulting in poor generalization to unseen inputs. In any case, it is desirable to construct a model with a few nonlinear terms. When a training pattern belongs to a systematic test procedure with healthy running conditions, the number of centers can be reduced and the internal interpolation for unknown conditions provides good results.7 Nevertheless, faulty conditions may introduce a high degree of uncertainty, resulting in poor approximations. Therefore, it is important to (22) ABB, 2008.
analyze how a designed RBF network behaves, provided that different numbers of learning patterns are applied to the training process. Materials and Methods Engines Characteristics: Model and Validations. The RBF network was initially applied to a three-cylinder SIE and a V-16 power-plant CIE.15 Additionally, a four-cylinder SIE with a validated model23 was analyzed. Table 1 shows the main characteristics of these engines. The training of the network needs both combustion and dynamic submodels to be run previously in one cycle. The combustion model is included in the first law of thermodynamics equation and is solved by first-order differential equations, resulting in a pressure curve for each cylinder. Pressure and mean angular speed data are the input data to the dynamic model. This provides the instantaneous angular speed curve for each condition, next to the load torque.15 The combustion submodel depends upon five parameters per cylinder in case a SIE is analyzed and seven parameters for a CIE.24 The two main parameters are fuel consumption and ignition/ injection timing, followed by the length of the different combustion phases. Other form parameters determine the shape of the rate of fuel heat released during the combustion process. The selection of parameters to be evaluated is critical for the effectiveness of the method, and it changes depending upon the number of cylinders. Although the model provides an acceptable pressure waveform through the combustion, the number of parameters must be reduced. In this sense, partial correlation indicates that some of them can be derived from others.16 Therefore, in SIE engines, combustion model parameters only depend upon the maximum pressure and torque. Because there are no experimental data to validate the system associated to the V-16 engine, the selected model is the simplest one.15 The dynamic submodel considers a flexible crankshaft and driveline components. Torsional vibrations have a high potential (23) Cruz-Peragon, F.; Palomar, J. M.; Mun˜oz, A.; Jimenez-Espadafor, F. An. Ing. Mec. 2000, 13, 1861–1866. (24) Cruz-Peragon, F. Análisis de metodologías de optimización inteligentes para la determinación de la presión en cámara de combustión de motores alternativos de combustión interna por métodos no intrusivos. Ph.D. Thesis, University of Seville, Seville, Spain, 2005, in Spanish.
Combustion Faults Diagnosis in Internal Combustion Engines
Energy & Fuels, Vol. 22, No. 5, 2008 2975
Table 4. Statistical Indicators of Deviation Errors (de) for Different Engines indicator error imepa deviation error in peak pressure (MPa) deviation error in peak pressure location (deg) maximum rate of pressure percent fuel consumption error ignition/injection timing error (deg) a
3-cylinder SIE15
4-cylinder SIE
µea ) 2.89% σea ) 8.6% (0.7 (1 µe ) 10.05% σe ) 52.5% µe ) 4.1 σe ) 14.5 µe ) -0.11 σe ) 2.75
µe ) 0.33% σe ) 14.6% (0.55 (1 µe ) 12.0% σe ) 47.2% µe ) 0.32 σe ) 13.5 µe ) 0.13 σe ) 3.2
4-cylinder CIE8
6-cylinder CIE7
dea ) (0.046 MPa (
