| 摘要 |
A method for data processing and optimisation of a neural network for application in the determination of physical property data of hydrocarbon products from measured (N.)I.R. spectral absorbances, characterized by the steps of: a) measuring the (N.)I.R. spectra of a large set of hydrocarbon product samples from a wide range of sources; b) selecting an overtone (harmonic) region of the (near-)infrared spectra, thus obtained; c) selecting a number of discrete wavelengths in each (N.)I.R. spectrum, converting a number of the said wavelengths to absorption data and using said absorption data as an input to a neural network; d) training the neural network on the entire data set by repeated presentation of inputs and known outputs i.e. the near-infrared data for the hydrocarbon product and its relevant physical property data, to learn the relationship between the two, and monitoring the performance of its predicitions against the actual physical property data as measured by standard methods for the training data, thus correlating the absorbance values with said relevant physical property; e) generating a set of values of the interconnection weights and biases of the network as adjusted after the learning period of step d); and f) applying these adjusted values, utilizing the network algorithm to (near-)infrared spectra, taken under the same conditions, for hydrocarbon products of unknown physical property data. |
