A neural network approach to the prediction of cetane number of diesel fuels using nuclear magnetic resonance (NMR) spectroscopy

In this work, quantitative relationships between the structural parameters of diesel fuels, as observed by H-1 NMR spectroscopy, have been established, with their ignition delay characteristics, using the artificial neural network (ANN) technique. Sixty commercial diesel samples were analyzed for this study. The cetane number (CN) of the samples was determined on an ignition quality tester (IQT). The H-1 NMR spectra of the samples were used as their structural characteristics, and relative intensities of various regions in the spectra were used as neural network inputs. The spectra in each case were divided into 18 regions, representing paraffins (normal and iso), cycloalkanes, olefins, and aromatics (different types). The development of the ANN model presented difficulties, because the data set consisted of only 60 samples for 18 input (NMR) parameters and 1 output (CN) parameter. Therefore, the data set was compressed to 8 input parameters by training a primary neural network in which inputs and outputs were the same. The hidden layer of the developed primary network, containing eight nodes, was then used as the inputs and CN was used as the output for the development of the final network. The primary network for data compression and the final network for CN prediction were then appended together. The pattern set was appropriately divided into subsets for development and validation of the final model. The developed model when tested on the unseen data set, gave a very high correlation between the actual and predicted CN values.