Abstract:
It was shown in an earlier study that it is possible to predict the spectral radiance of rocket combustion plumes directly from the propellant composition and motor parameters. Little is published in the open literature on this subject, but the current trend is to use determinative methods like computational fluid dynamics and statistical techniques to simulate wide band radiance based on blackbody temperature assumptions. A limitation of these methods is the fact that they are computationally expensive and rather complex to implement. An alternative modeling approach was used which did not rely on solving all the nonlinearities and complex relationships applicable to a fundamental model. A multilayer perceptron based Neural Network was used to develop a parametric functional mapping between the propellant chemical composition and the motor design and the resulting spectral irradiance measured in a section of the plume. This functional mapping effectively models the relationship between the rocket design and the plume spectral radiance. Two datasets were available for use in this study: Emission spectra from solid propellant rockets and flare emission spectra. In the case of the solid rocket propellants, the input to the network consisted of the chemical composition of the fuels and four motor parameters, with the output of the network consisting of 146 scaled emission spectra points in the waveband from 2-5 microns. The four motor parameters were derived from equations describing the mass flow characteristics of rocket motors. The mass flow through the rocket motor does have an effect on the shape of the plume of combustion gases, which in turn has an effect on the infrared signature of the plume. The characteristics of the mass flow through the nozzle of the rocket motor determine the thermodynamic properties of the combustion process. This then influences the kind of chemical species found in the plume and also at what temperature these species are radiating energy. The resultant function describing the plume signature is: "Plume signature equal f{p1, T1, At, [epsilon], fuel composition}" It was demonstrated that this approach yielded very useful results. Using only 18 basic variables, the spectra were predicted properly for variations in all these parameters. The model also predicted spectra that agree with the underlying physical situation when changing the composition as a whole. By decreasing the Potassium content for example, the model demonstrated the effect of a flame suppressant on the radiance in this wavelength band by increasing the predicted output. Lowering the temperature, this drives the process of molecular vibration and translation, resulted in the expected lower output across the spectral band. In general, it was shown that only a small section of the large space of 2 propellant classes had to be measured in order to successfully generate a model that could predict emission spectra for other designs in those classes. The same principal was then applied to predicting the infrared spectral emission of a burning flare. The brick type flare considered in this study will ignite and the solid fuel will burn on all surfaces. Since there are no physical parameters influencing the plume as in the case of the rocket nozzles it was required to search for parameters that could influence the flare plume. It was possible to calculate thermodynamic Properties for the flare combustion process. These parameters were then reduced to 4 parameters, namely: the oxidant-fuel ratio, equilibrium temperature, the molar mass and the maximum combustion temperature. The input variables for the flares thus consisted of the chemical composition and 4 thermodynamic parameters described above. The network proposed previously was improved and optimised for a minimum number of variables in the system. The optimised network marginally improved on the previous results (with the same data), but the training time involved was cut substantially. The same approach to the optimization of the network was again followed to determine the optimal network structure for predicting the flare emission spectra. The optimisation involved starting out with the simplest possible network construction and continuously increasing the variables in the system until the solution predicted by the network was satisfactory. Once the structure of the network was determined it was possible to optimise the training algorithms to further improve the solution. In the case of the solid rocket propellant emission data it was felt that it would be important to be able to predict the chemical composition of the fuel and the motor parameters using the infrared emission spectra as input. This was done by simply reversing the optimised network and exchanging the inputs with the outputs. The results obtained from the reversed network accurately predicted the chemical composition and motor parameters on two different test sets. The predicted spectra of some of the solid propellant rocket test sets and flare test sets did not compare well with the expected values. This was due to the fact that these test sets were in a sparsely populated area of the variable space. These outliers are normally removed from training data, but in this case there wasn't enough data to remove outliers. To obtain an indication of the strength of the correlation between the predicted and measured line spectra two parameters were used to test the correlation between two line spectra. The first parameter is the Pearson product moment of coefficient of correlation and gives an indication of how good the predicted line spectra followed the trend of the measured spectral lines. The second parameter measures the relative distance between a target and predicted spectral point. For both the solid propellants and the flares the correlation values was very close to 1, indicating a very good solution. Values for the two correlation parameters of a test set of the flares were 0.998 and 0.992. In order to verify the model it was necessary to prove that the solution yielded by the model is better than the average of the variable space. Three statistical tests were done consisting of the mean-squared-error test, T-test and Wilcoxon ranksum test. In all three cases the average of the variable space (static model) and the predicted values (Neural Network model) were compared to the measured values. For both the T-test and the Wilcoxon ranksum test the null hypothesis is rejected when t less than -t[alpha equal 1.645 and then the alternative hypothesis is accepted, which states that the error of the NN model will be smaller than that of the static model. The mean squared error for the static model was 0.102 compared to the 0.0167 of the neural net, for a solid propellant rocket test set. A t-test was done on the same test set, yielding a value of -2.71, which is smaller than - 1.645, indicating that the NN model outperforms the static model. The Z value for this test set is Z equal -11.9886, which is a much smaller than -1.645. The results from these statistical tests confirm that neural network is a valid conceptual model and the solutions yielded are unique.