APPLICATION OF BACK PROPAGATION NEURAL NETWORK IN PREDICTING HC EMISSION FROM I. C. ENGINES

One of the prominent emissions of I.C engines is the Hydro-carbons commonly known as HC which adversely affects our eco-system. In the present work, an attempt has been made towards the application of Back Propagation Neural Network (BPNN) for predicting the HC emission from a diesel engine for various engine settings such as compression ratio, injection timing and load. The data collected trains the Neural Network (NN) and these inputs were strategically combined to predict HC emission. It has been observed that by the right combination of input parameters to the NN, may effectively predict the level of HC emission with minimum Root Mean Squared (RMS) error of almost less than 7.4%.

variables. For pre-specified engine speeds and loads with LM algorithm, absolute percentage errors were found between 0.68% to 3.34% [2]. M.Ali Akcayol et. al (2005) attempted to improve cold start performance of catalytic converters for HC and CO emissions with the help of a burner heated catalyst tested in a four stroke spark ignition engine using back propagation learning algorithms of ANN for prediction of catalyst temperature, CO and HC emissions. Taking the training dataset from the experiment, it was found that the deviation coefficients for standard and heated catalyst temperature are less than 4.925%,and 1.602%, the same for standard and heated catalyst HC emissions are less than 4.798% and 4.926% and that for standard and heated catalyst CO emissions are less than 4.82% and 4.938% respectively [3]. Shivakumar et.al (2010) used blends of Hunge oil with diesel at various compression ratios as fuel to predict the performance and emission characteristics of a single cylinder, four stroke, and water cooled compression ignition engine using Artificial neural networks (ANN's). The ANN was trained with back propagation algorithm using compression ratio, blend percentage and percentage load as input variables whereas performance parameters together with engine exhaust emissions were used as output variables. ANN showed good convergence between predicted and experimental values for various performance parameters and emissions with mean squared error closed to 1 and mean relative error less 9% [4].

RESEARCH METHOD
The entire experiment was carried out at the I.C Engine laboratory in a computerized single cylinder, four stroke, multi-fuel, variable compression ratio (VCR) engine as shown in Fig.1. The fuel used for the experiment was diesel. The setup consists of single cylinder, four stroke, multi-fuel, research engine (specified in Table 1) connected to eddy current type dynamometer for loading. The operation mode of the engine can be changed from Diesel to Petrol or from Petrol to Diesel with some necessary changes. In both the modes, the compression ratio can be varied without stopping the engine and without altering the combustion chamber geometry.  Data on exhaust HC emission were collected by varying the controllable parameters of the engine among which are Compression Ratio (CR), Injection Timing (IT) and Load (W) on the engine are crucial. Also the parameters such as observed load (WOBS), water inlet and outlet temperature to and from the engine respectively (T1 & T2) engine exhaust temperature(T5) from calorimeter peak cylinder pressure(P.P),Crank angle corresponding to peak pressure(θpeak), indicated air pressure in mm of water column in the calorimeter(Air pr.) and rate of fuel into the cylinder(R.F.I) were recorded. In order to establish the domain of the experiment, we may consider each parameter or factor at several levels. =summing junction Information Processing Weighted sum (V) = W1.X1 + W2.X2 + W3.X3 -T, for i=1, 3, j = 1, 3 Now the neuron fires only when V≥0 and gives the output, generally using Sigmoid function (shown below); otherwise the output = 0.0 This type of network shown in Fig.3 is sometimes called multilayer perception (MLP) because of its similarity to perception networks with more than one layer.

Fig3. BPNN Architecture
The network consists of a number of layers called the input, hidden and output layers. The hidden and output layers contain a number of neurons or processing elements which are connected by links or connections to show the flow direction of signals and also to represent weight or strength of their respective connections. In an MLP of the back propagation type, the connections are first initialized by a set of uniformly distributed random numbers between 0 and 1. The calculations are made in feed forward manner until back propagation of errors is done. Following the processing in a single neuron (Fig. 2), outputs from the neurons of a certain layer (eq. 1) are given as inputs to the neurons of the next layer. Finally the output layer gives the calculated output (YK) from the BPNN and the back propagation begins on the basis of prediction error (eK). The flow chart shown in Fig. 3 (3) TR = training set, TS = testing set, NTR = no. of training set, NTS = no. of testing set, C = no. of output nodes. The iterations may be stopped for any of the following reasons: Either after a certain number of iterations Or after a desired precision level is achieved Or the RMS Testing error begins to increase (called Over learning/Over training) shown next [5].

(C) Strategic Analysis
The entire set of 63 data is divided into 42 nos. of training set and 21 nos. of testing set. The performance of the various input parameters for predicting the output (HC emission from the engine) are studied with the help of BPNN program. For this purpose, systemic analysis has been adopted by grouping the input parameters, which are being called as "strategies" listed in table 2 below.

HEURISTIC OPTMIZATION OF BPNN AND ITS PARAMETERS
(a) Firstly, the architecture of each strategy is optimized by changing the number of neurons in the hidden layer, keeping learning rate and momentum parameter [6,7] fixed respectively at 0.5 and 0.7 (the range being 0.1-20 and 0.7-5 for L.R and M.P respectively) to obtain a minimum mean squared error for the testing set of data or 25000 iterations, whichever occurs first (b) Secondly, the optimized architecture is further tested by varying the learning rate and momentum -parameter to further minimise the error for the testing set The optimized results are obtained by iterating the training and testing set with the program using back propagation algorithm. The

RESULT OF ANALYSIS OF HC
Best results compiled for HC emission are listed in Table 3. The learning curves and scatter diagrams of predicted and observed results for HC emission are depicted from Fig. 5 and 6.

INFERENCE
(i) It has been tested that the three basic controllable parameters compression ratio, observed load and injection timing as inputs gives bad result. The minimum difference between observed and predicted results of training and testing set of data for 25000 iterations amounts to 60.926ppm and 63.329ppm respectively.
(ii) The inclusion of sensor signal feature along with the controllable parameters leads to over learning after a short number of iterations but strategy III by far gives the best result although the conversion rate is not as fast as shown in Table 3 and Fig  5. The minimum difference between observed and predicted results of training and testing set of data for 25000 iterations amounts to 40.5178ppm and 50.152ppm for strategy III. Fig 5 indicates how close the set of predicted results are to the actual results obtained in practice (now in the experiment).
(iii) Strategy V and VI leads to poorer results where the minimum difference between observed and predicted results of training and testing set of data for 25000 iterations are 68.934ppm and 70.897, 66.546ppm and 68.926 respectively but with strategy VII there is marked improvement in prediction results where minimum difference between observed and predicted results of training and testing set of data for 25000 iterations amounts to 62.3296ppm and 53.4105ppm respectively.

CONCLUSION
As evident from the above analysis, strategy III gives the predicted value of emissions close to their observed values. Strategy IV comes closer to III. The temperature of water inlet and outlet to and from the engine respectively and engine exhaust temperature plays the most important role in emission prediction. It has been observed that the peak cylinder pressure and crank angle at peak pressure (θpeak) is a vital ingredient in NOX prediction whereas the inclusion of indicted air pressure (Air pr.) and rate of fuel input into the cylinder (R.F.I) as inputs to the BPNN lead to poorer results in emission prediction.