Development of back propagation neural network model to predict performance and emission


Published on

Published in: Technology
  • Be the first to comment

  • Be the first to like this

No Downloads
Total views
On SlideShare
From Embeds
Number of Embeds
Embeds 0
No embeds

No notes for slide

Development of back propagation neural network model to predict performance and emission

  1. 1. International Journal of Advanced Research in Engineering and Technology (IJARET), ISSN0976 – 6480(Print), ISSN 0976 – 6499(Online) Volume 4, Issue 3, April (2013), © IAEME85DEVELOPMENT OF BACK PROPAGATION NEURAL NETWORKMODEL TO PREDICT PERFORMANCE AND EMISSIONPARAMETERS OF A DIESEL ENGINEShailaja M,Assistant Professor, Department of Mechanical EngineeringJNTUH, College of Engineering,Jagtial,Karimnagar, Andhra Pradesh, INDIA.V.Vijaya Kumar,Assistant Professor, Department of MathematicsSree Chaitanya Institute of Technological SciencesKarimnagar, Andhrapradesh, INDIAChandragiri Radha Charan,Assistant Professor, Electrical and Electronics DepartmentJNTUH, College of Engineering,Jagtial, Karimnagar, Andhra Pradesh, INDIA,Dr.A V Sitarama Raju,Department of Mechanical Engineering JNTUH College of Engineering,Kukatpallyl, Hyderabad, Andhra Pradesh, INDIA,ABSTRACTIn this paper, development of Back Propagation Neural Network (BPNN) is proposed topredict performance and emission parameters of a diesel engine fuelled with sunflower oil and itsblends with petro-diesel. Short term tests were conducted at various loads and with various blends offuel to collect data. Various combinations of network parameters were investigated by varyingnumber of hidden nodes, learning rate (η), momentum factor (α), and training algorithm. Thedeveloped neural network is able to predict brake specific fuel consumption (bsfc), torque, brakethermal efficiency, Hydro Carbon (HC) and Carbon Monoxide (CO) emissions within acceptablerange of correlation coefficients, with load and percentage of blends as inputs.Keywords: Back Propagation Algorithm; Correlation coefficient; Diesel Engine; PerformanceParameters; Emission Parameters.INTERNATIONAL JOURNAL OF ADVANCED RESEARCH INENGINEERING AND TECHNOLOGY (IJARET)ISSN 0976 - 6480 (Print)ISSN 0976 - 6499 (Online)Volume 4, Issue 3, April 2013, pp. 85-92© IAEME: Impact Factor (2013): 5.8376 (Calculated by GISI)www.jifactor.comIJARET© I A E M E
  2. 2. International Journal of Advanced Research in Engineering and Technology (IJARET), ISSN0976 – 6480(Print), ISSN 0976 – 6499(Online) Volume 4, Issue 3, April (2013), © IAEME86I. INTRODUCTIONA. Suitability of Sunflower oil in Diesel EnginesStringent regulations on automobile emissions and rapid depletion of fossil fuels inspiredresearchers to think towards alternative fuels. One of the best substitutes for petro-diesel is vegetableoils and/or their methyl esters as vegetable oils as properties of vegetable oils are very much similar topetro-diesel and requires no or little engine modifications. In the present work, sunflower oil and itsblends with petro-diesel in various proportions are used as fuels. S Kalligeros et al.[1] investigated useof sunflower oil bio diesel in diesel engines and reported satisfactory results of emissions compared topetro-diesel. Abuhabaya A et al. [2] prepared bio diesel from sunflower oil and tested in diesel engineand testified decrease in emissions with increase in proportion of bio diesel in blend. The followingtable presents comparison of properties of diesel and sunflower oil.TABLE 1. Comparison of Properties of Diesel and Sunflower OilProperty Diesel Sunflower oilDensity (Kg/m3) 0.832 0.918Calorific Value (kJ/kg.) 44800 43500Flash Point (°°°°C) 62 274Fire Point (°°°°C) 175 371B. Artificial Neural Networks (ANNs)Artificial Neural Networks has been motivated right from their inception by the recognitionthat the brain computes in an entirely different way from the conventional digital computers.This interconnected group of artificial neurons that uses a mathematical model or computationalmodel for information processing based on a connectionist approach to computation. In most cases anANN is an adaptive system that changes its structure based on external or internal information thatflows through the network. In more practical terms neural networks are non-linear statistical datamodeling tools. They can be used to model complex relationships between inputs and outputs or tofine patterns in data.C. Applications of ANN in the Field of Internal Combustion EnginesANNs are widely used as a tool for prediction of performance parameters such as efficiency,specific fuel consumption etc. and fault diagnosis in internal combustion engines.Michael L. Traver et al. [3] developed a model to predict emissions of a diesel engine based on in-cylinder pressure derived variables. A.M.Frith et al. [4] investigated on adaptive control of gasolineengine air-fuel ratio using artificial neural networks and reported successful results. F.Gu et al. [5]attempted to develop a RBF neural network model for cylinder pressure reconstruction in internalcombustion engines and reported that RBFNN is best suited for the data. Robert.J.Howlett et al. [6]investigated on accurate measurement of air fuel ratio with back propagation algorithm and reportedthat the predicted values are in good agreement with experimental values. Radial basis function neuralnetwork is found to produce good results to predict toxicity and exhaust gas components as perinvestigations by Brzozowska et al. [7]. Adnan Parlak et al.[8] developed ANN to predict specific fuelconsumption and exhaust temperature with Levenberg-Marquardt algorithm and reported relativeerror for prediction in the range of 1% - 9%. Recurrent neural network was developed and validated topredict NOx (nitrogen oxides) emissions and error lower than 2% was reported by Arise et al. [9].Hidayat Oguz et al.[10] investigated on development of ANN for the prediction of power, torque, fuelconsumption with speed and type of fuel as inputs and reported that BPNN with Tangent-Sigmoidactivation function was best suited for data. M.Ghazikhani et al. [11] developed ANN to predict soot
  3. 3. International Journal of Advanced Research in Engineering and Technology (IJARET), ISSN0976 – 6480(Print), ISSN 0976 – 6499(Online) Volume 4, Issue 3, April (2013), © IAEME87emissions and reported 6.4% mean absolute percentage error. Sumita Deb Barma et al.[12] developedBPNN to predict performance parameters with Levenberg-Marquardt algorithm and reportedsatisfactory results. R.Manjunatha et al. [13] compared performance of radial basis function neuralnetworks and BPNN and concluded radial basis function neural network is best suited for the data topredict emissions.II. DISCRIPTION OF EXPERIMENTAL SETUPA single cylinder Direct Injection type four Stroke water cooled vertical diesel engine test rigdeveloping 3.5 kilo Watts at approximately 1500 RPM coupled to AC alternator with loading bank forexperimentation purpose. AC alternator is fixed to engine flywheel and the engine is mounted on amild steel channel frame and further mounted on anti-vibration mounts. Panel board is used to fix theburette with 3-way cock digital RPM indicator and u-tube manometer.Load is varied by varyingresistance. The fuel is supplied from the main fuel tank to the measuring burette. An air drum is fittedon the panel frame and connected to engine through an air base. The air drum facilitates a magnifiedorifice and pressure pick up points are connected to end u-tube manometer limbs. The difference inmanometer readings is taken at different loads.Emissions CO and HC are measured using a gas analyzer which measures with a good accuracy.III. EXPERIMENTAL WORKExperiments were carried with eight different proportions of diesel, sunflower oil combinations asfuels which are given below.1. 0% of sunflower oil and 100% of diesel2. 10% of sunflower oil and 90% of diesel3. 20% of sunflower oil and 80% of diesel4. 25% of sunflower oil and 75% of diesel5. 30% of sunflower oil and 70% of diesel6. 40% of sunflower oil and 60% of diesel7. 50% of sunflower oil and 50% of diesel8. 100% of sunflower oil and 0% of dieselWith each fuel, steady state short term tests were conducted at six load settings over entire range ofengine operation and observations were recorded for emissions and to calculate performanceparameters. Short term experiments were conducted to collect data. For accuracy each observation isrecorded three times and averaged. Using conventional formulae brake specific fuel consumption,torque and brake thermal efficiency were calculated. An exhaust gas analyzer is used to record COand HC emissions.Development of Back Propagation Neural A neural network is capable of approaching anonlinear function to significant desirable degree of accuracy. A feed forward network with backpropagation algorithm is chosen. Back propagation algorithm is most popular for supervised trainingof multilayer perceptron as it is simple to compute locally and it performs stochastic gradient descentin weight space [14].Two input parameters (load and blend percentage) and five output parameters (brake specificfuel consumption, torque, brake thermal efficiency, Carbon Monoxide and Hydro Carbon emissions)are chosen. Data were collected over entire range of engine operation comprising of 46 sets. Entiredata is divided into two sets namely training set and test set. Training set consists of 87% of data (40sets) and test set consists of 13% of data (6 sets).Before training, it is often useful to scale the given data so that they always fall within a specifiedrange which facilitates network for better training. Data is normalized in the range [-1 1] by usingfollowing formula.
  4. 4. International Journal of Advanced Research in Engineering and Technology (IJARET), ISSN0976 – 6480(Print), ISSN 0976 – 6499(Online) Volume 4, Issue 3, April (2013), © IAEME88minminminmaxminmax( )*( )y yy x x yx x−= − +−(1)Tangent-Sigmoid transfer function is chosen for hidden layer which is represented by followingequations.1 exp( )( )1 exp( )xb xxσσ− −=+ −(2)( ) [(1 ( ))(1 ( )]2b x b x b xσ= + − (3)Network performance is measured in terms of error function. Mean Squared Error (MSE) is one ofthe network error functions. It measures the networks performance according to the MSE.N1 2MSE= e(k)K=1N∑ (4); where e(k)=t(k)-a(k)t(k) is the experimental output, a(k)is the output predicted by ANN and N is number of elements inoutput vector. MSE is set as 0.001. Best BPNN is arrived after numerous trials by varying networkparameters such as number of nodes in hidden layer starting from 3 nodes-20 nodes, number ofhidden layers 1 -2, learning rate,α (0.0001-0.3), momentum factor,η (0.75-0.98) and trainingalgorithm (gradient descent back propagation, gradient descent with adaptive learning rate backpropagation, gradient descent with momentum back propagation, and Levenberg-Marquardt backpropagation). Application of algorithm for back propagation theorem is presented in the followingsteps [15].1. Initialize weights (from training algorithm).2. For each set of input perform steps 3 to 5.3. For i=1… n; set activation of input unit, xi.4. For j=1,……, p;nz =v + x voj i ij-inj i=1∑ (5)5. For k=1,……., mpy =w + z w-ink ok j jkj=1∑ (6)y(k)=f(y )-ink (7)z j= hidden unit jvoj= bias on hidden unitswok = bias on output unit kvij= weight connection between ithinput node andjthhidden nodew jk =weight connection between jthhidden node and kthoutput nodey (k) = output unit kVariation of correlation coefficient with number of nodes in hidden layer is shown in following Fig. 2.Maximum value of Regression, R (0.9953) may be observed at 7 nodes in hidden layer.Based on performance, supervised learning with feed forward network by Levenberg-Marquardt backpropagation neural network, one hidden layer with seven neurons, learning rate α=0.0001, momentumfactor, η=0.9 is reported as best suitable network. Schematic block diagram is shown in Fig.1.
  5. 5. International Journal of Advanced Research in Engineering and Technology (IJARET), ISSN0976 – 6480(Print), ISSN 0976 – 6499(Online) Volume 4, Issue 3, April (2013), © IAEME89TABLE II :TRAINING DATA SETS.No.BlendLoad(amps)SFC(kg/kwhr)Torque(N-m)Brake ThermalEfficiency(percentage)HCs(partsper million)CO(percentage)1 0 0 0 0 0 22 0.0572 0 2.4 0.566 4.297 14.2 22 0.0573 0 5.1 0.339 9.261 24.68 22 0.0574 0 7.8 0.262 14.524 30.44 22 0.0575 0 10.6 0.233 19.848 34.27 22 0.0576 0 14.6 0.195 25.102 41.02 22 0.0577 0.1 0 0 0 0 19 0.0548 0.1 5.1 0.347 9.243 23.45 19 0.0549 0.1 7.8 0.273 14.54 29.75 19 0.05410 0.1 10.6 0.248 19.875 32.78 19 0.05411 0.1 14.6 0.227 25.138 35.73 19 0.05412 0.2 0 0 0 0 15 0.04913 0.2 2.4 0.578 4.524 14.31 15 0.04914 0.2 4.9 0.353 9.49 23.43 15 0.04915 0.2 10.2 0.247 20.343 33.46 15 0.04916 0.2 12.8 0.225 25.485 33.66 15 0.04917 0.25 0 0 0 0 16 0.04618 0.25 2.4 0.618 4.545 13.52 16 0.04619 0.25 5 0.359 9.718 23.27 16 0.04620 0.25 10.9 0.235 21.779 35.46 16 0.04621 0.25 13 0.229 26.008 36.44 16 0.04622 0.3 0 0 0 0 15 0.04623 0.3 2.4 0.639 4.3 13.19 15 0.04624 0.3 4.9 0.359 9.03 23.48 15 0.04625 0.3 12.9 0.226 25.8 37.29 15 0.04626 0.4 0 0 0 0 15 0.04627 0.4 2.3 0.638 4.3 13.55 15 0.04628 0.4 4.9 0.381 9.4 23.16 15 0.04629 0.4 7.5 0.281 14.9 30.56 15 0.04630 0.4 12.9 0.236 25.7 36.4 15 0.04631 0.5 0 0 0 0 14 0.04532 0.5 2.3 0.663 4.31 16.35 14 0.04533 0.5 4.9 0.38 9.3 23.01 14 0.04534 0.5 7.6 0.296 14.9 29.04 14 0.04535 0.5 10.2 0.255 20.3 34.02 14 0.04536 1 0 0 0 0 15 0.04337 1 2.4 0.661 4.459 14.61 15 0.04338 1 4.9 0.399 9.379 24.19 15 0.04339 1 7.6 0.305 14.79 31.63 15 0.04340 1 10.2 0.271 20.108 35.6 15 0.043
  6. 6. International Journal of Advanced Research in Engineering and Technology (IJARET), ISSN0976 – 6480(Print), ISSN 0976 – 6499(Online) Volume 4, Issue 3, April (2013), © IAEME90TABLE III.TEST DATA SETS.No.BlendLoad(amps)SpecificFuel Consumption(kg/kWhr)Torque(N-m)Brake ThermalEfficiency(percentage)HCs(parts permillion)CO(percentage)1 0.1 2.4 0.584 4.267 13.91 19 0.0542 0.2 7.6 0.28 14.92 29.57 15 0.0493 0.25 7.6 0.293 14.99 28.52 16 0.0464 0.3 7.6 0.283 14.9 29.72 15 0.0465 0.3 10.2 0.254 20.4 33.1 15 0.0466 0.4 10.2 0.257 20.3 33.43 15 0.0467 0.5 12.9 0.237 25.61 36.79 14 0.0458 1 12.9 0.246 25.642 39.01 15 0.043IV. RESULTSThe economic structure that resulted in minimum error and maximum efficiency duringtraining as well as testing is selected as the final form of the BPNN model. BPNN with 2-7-5architecture is trained and validated .Performance plot is shown in Fig.3(a). Regression analysis plotsfor training, validation and testing are presented in Fig. 3 (b), Fig. 3(c) and Fig. 3(d) respectively.Comparison of experimental values with predicted from BPNN is presented in Fig.5 to Fig.10.Correlation coefficients, R for outputs of specific fuel consumption is 0.99409, Torque is 0.99526,Brake Thermal Efficiency is 0.9952, HC is 0.90193 and CO is 0.89137 which are satisfactory and inacceptable range(shown in Fig. 4 - Fig.8).Fig.1 Architecture of Back Propagation Neural Network (BPNN)Fig.3 (a) Performance plot of BPNN Fig. 3(b) Regression analysis ofTraining of BPNN
  7. 7. International Journal of Advanced Research in Engineering and Technology (IJARET), ISSN0976 – 6480(Print), ISSN 0976 – 6499(Online) Volume 4, Issue 3, April (2013), © IAEME91After training BPNN is further validated with unseen data during training and predictedvalues from BPNN are in very good agreement with experimental values.Fig.9 to Fig.13 showscomparison of experimental and predicted values by BPNN.Average error of specific fuel consumption, Torque, Brake Thermal Efficiency, Hydro Carbon,Carbon monoxide for prediction of various outputs is shown in Table IV.V. CONCLUSIONSShort-term experimental data with Levenberg Marquardt back propagation neuron model isused to predict performance and emission parameters of a diesel engine. The BPNN has beensimulated in MATLAB 7.0 ®. As mentioned in results the correlation coefficients (R) by usingadaptive learning has been obtained as R1= 0.99526, R2 =0.99409, R3 =0.9952, R4 =0.90193 and R5=0.89173 respectively for outputs which are in acceptable range. Further work may be extended toconduct medium term and long term tests on the diesel engine.ACKNOWLEDGMENTAuthors are grateful to authorities of JNTUH College of Engineering Jagtial for permitting tocarry out the experiments in Thermal Engineering laboratory.REFERENCES[1] Abuhabaya, A. Ali, J. ; Fieldhouse, J. ; Brown, R. and Andrijanto, E.” The optimisation of bio-diesel production from Sunflower oil using RSM and its effect on engine performance andemissions”, 17th International Conference on Automation and Computing (ICAC), 2011 pp. 310– 314 pp. 1-15.[2] S Kalligeros, F Zannikos, S Stournas, E Lois, G Anastopoulos, Ch Teas and F Sakellaropoulo,“An investigation of using biodiesel/marine diesel blends on the performance of a stationarydiesel engine”, ELSEVIER, Biomass and Bioenergy, Volume 24, Issue 2, 2003, pp. 141–149. I. S.Jacobs and C. P. Bean, “Fine particles, thin films and exchange anisotropy,” in Magnetism, vol.III, G. T. Rado and H. Suhl, Eds. New York: Academic, 1963, pp. 271–350.[3] Michael L, Traver, Richard J. Atkinson and Christopher M.Atkinson, “Neural network baseddiesel engine emissions prediction using in-cylinder combustion pressure. SAE, InternationalSpring Fuels & Lubricants Meeting & Exposition, 1999.[4] A.M. Frith, C.R. Gent, and A.J. Beaumont, “Adaptive control of gasoline engine air-fuel ratiousing artificial neural networks”, 4th International Conference on Artificial NeuralNetworks,1995,pp.274–278 .Y. Yorozu, M. Hirano, K. Oka, and Y. Tagawa, “Electronspectroscopy studies on magneto-optical media and plastic substrate interface,” IEEE Transl. J.Magn. Japan, vol. 2, pp. 740–741, August 1987 [Digests 9th Annual Conf. Magnetics Japan, p.301, 1982].[5] F. Gu, P.J. Jacob, and A.D. Ball, ” A RBF neural network model for cylinder pressurereconstruction in internal combustion engines”, IEE Colloquium on Modelling and SignalProcessing for Fault Diagnosis, IEE Digest / Volume 1996 / Issue 260. pp.4-15 .S.No. Prediction ParameterAverage Error(Percentage)1. Specific Fuel Consumption (kg/kW-hr) 1.8542. Torque (N-m) 3.4983. Brake Thermal Efficiency (percentage) 2.1874. Hydro Carbon 5.715. Carbon monoxide 2.837
  8. 8. International Journal of Advanced Research in Engineering and Technology (IJARET), ISSN0976 – 6480(Print), ISSN 0976 – 6499(Online) Volume 4, Issue 3, April (2013), © IAEME92[6] Robert.J.Howlett, Simon. D.Walters, Peters.A .Howson and Ian.A.Park, ” Air-fuel ratiomeasurement in an internal combustion engine using a neural network”, IEEE Conference onAdvances in Vehicle Control and Safety, (AVCS’98), July 1998.H. Goto, Y. Hasegawa, and M.Tanaka, “Efficient Scheduling Focusing on the Duality of MPL Representatives,” Proc. IEEESymp. Computational Intelligence in Scheduling (SCIS 07), IEEE Press, Dec. 2007, pp. 57-64,doi:10.1109/SCIS.2007.357670.[7] Brzozowska,L., Brzozowski, K.; Nowakowski,J. “An application of artificial neural network todiesel engine modeling”, IEEE Conference on Intelligent Data Acquistion and AdvancedComputing Systems; Sept 2005, pp. 142-146.[8] Adnan Parlak, Yasar Islamoglu, Halit Yasar and Aysun Egrisogut, “Application of artificialneural network to predict specific fuel consumption and exhaust temperature for a diesel engine”,ELSEVIER Applied Thermal Engineering, Vol 26, pp 824-828, 2006.[9] Arise, Ivan; Marra, Dario; Pianese, Cesare; and Sorrentito Marco “ Real time estimation ofengine NOx via recurrent neural networks”, 6thIFAC Symposium Advances in AutomotiveControl, 2010.[10] Hidayet Oguz, Ismail Saritas and Hakan Emre Baydan, “ Prediction of diesel engine performanceusing biofuels with artificial neural network”, ELSEVER, Expert systems with applications, vol37 (2010) pp. 6579-6586.[11] M.Ghazikhani and I.Mirzaii,” Soot emission prediction of a waste gated turbo-charged DI dieselengine using artificial neural network”, Springer ,Neural Computing & Applications, vol. 20.2011, pp. 303-308[12] Sumita Deb Barma, Biplab Das, Asis Giri, S.Majumder and P.K.Bose, “ Back propagationartificial neural network (BPANN) based performance analysis of diesel engine using biodiesel”.Journal Renewable and Sustainable Energy , Vol 3, Issue 1.2011.[13] Rajinder Kumar Soni And Pranat Pal Dubey, “Diesel Engine Air Swirl Mesurements Using AvlTest Rig” International Journal Of Mechanical Engineering & Technology (IJMET) Volume 4,Issue 1, 2013, pp. 79 - 91, ISSN PRINT : 0976 – 6340, ISSN ONLINE : 0976 – 6359[14] V.Narasiman , S.Jeyakumar , M. Mani ,K.Rajkumar, “Impact Of Combustion On Ignition DelayAnd Heat Release Curve Of A Single Cylinder Diesel Engine Using Sardine Oil As A MethylEster” International Journal Of Mechanical Engineering & Technology (IJMET) Volume 3, Issue3, 2012, pp. 150 - 157, ISSN PRINT : 0976 – 6340, ISSN ONLINE : 0976 – 6359