Prediction Model For The Property Of Diesel Fuels Based On NIR-SVM

Diesel fuel is an indispensable fuel in our daily lives, and it is a necessary fuel for the heavy-duty vehicles for transportation, vessels, warships, tanks and conventional submarines as well. Diesel fuel for different purposes has different specifications, even though diesel fuel for the same use has different performance. So in production of diesel fuel, every performance specification must be measured accurately in order to satisfy different uses. Because the standard measurement method for the diesel fuel's cetane and viscosity is very complicated and time-consuming, the prediction models for its cetane and viscosity are established in this paper, based on the near infrared spectroscopy (NIR) data and the Support Vector Machines (SVM). Different data processing methods are introduced in the modeling.Firstly, the Euclidean Distance is taken to choose the data samples in order to remove the abnormal spectrum samples. Then, the Data Smoothing Method and the 1st Derivative Method are taken to overcome noises influences, such as high-frequent disturbance, temperature and color of samples. At last the sample dimension is reduced from 401 to7 or 18 by introducing the Principle Component Analysis Method, which greatly simplifies the test model, improve the modeling efficiency and predicting accuracy. The best parameters for cetane and viscosity model are gained after analyzing and comparing the model simulation results for several times.The simulation results show that the predictive accuracy and reliability of SVM is better than that of the BP network by comparing the SEC and SEP of SVM with BP , and there is over-fitting phenomenon in the modeling process for BP. In general, the diesel fuel's natural predictive model based on SVM is more valid than that based on BP.