Analysis Of Octane Infrared Spectrum Data Based On Neural Network

In recent years,the requirements for environmental protection have been continuously improved,and the quality of gasoline products has been continuously improved.As an important index factor of gasoline product quality,the content of octane number directly affects the qualification of gasoline.Near infrared spectroscopy(NIR),combined with the technology of spectral measurement and neural network,is an effective method to measure various parameters of gasoline,so that the quantitative analysis of the octane number of gasoline can be carried out quickly and conveniently.BP(Back Propagation)neural network and RBF(Radial Basis Function)neural network in the practical application process hold different merits as well as its priorities.In order to select the most suitable neural network model,the author measured 60 sets of data,a total of 401 wavelength points by means of NIR.Based on the absorbance data and octane value data,the author adopted the BP neural network and the RBF neural network to model them respectively,by which the octane number of gasoline can be predicted well,at the same time,the merits and demerits of the BP neural network and the RBF neural network can be well revealed.The analysis is mainly from the following aspects:(1)Robustness analysis of BP neural network and RBF neural network.Each time a random noise is used to replace one of the data in the training sample,and then the test sample is used to test it.In order to ensure the accuracy of the test results,the sample type and number of each test will remain unchanged.The average value of the prediction accuracy is calculated after the simulation for many times,and the relationship between the error rate of the training sample and the accuracy of the prediction is drawn.In order to make the data more authentic and credible,we can increase the gradient of noise and the number of simulations,and compare and analyze the results.Finally,the conclusion is that the BP neural network model and the RBF neural network model are robust in a certain range,and the beyond range will lose the robustness.However,in general,the robustness of RBF network model is stronger than that of BP network model,so the RBF neural network model should be chosen first in the practical application.(2)Self learning ability analysis of BP neural network and RBF neural network.The sample is divided into training sample and test sample.One sample is added to the training sample each time,and the data and types of each test sample are kept the same.The number of training samples is changed once every time,and all the prediction accuracy before statistics is calculated and the average value is taken as the last data..Finally,the relationship between the number of samples added and the prediction accuracy is obtained.In order to make the result more accurate and reliable,the final result of the number of samples added each time was analyzed as a contrast test.The results are as follows: when the number of training samples is less,the self-learning ability of RBF neural network is stronger than that of the BP neural network.On the contrary,the self-learning ability of BP neural network is stronger than that of the RBF neural network when the number of training samples is large.(3)the analysis of BP neural network and RBF neural network's anti-counterfeiting ability.First of all,the training samples and test samples are determined.In the case of constant training samples,a single noise is used to change the data of a test sample each time,and the accuracy of multiple prediction is obtained.Then the most final result of the average accuracy is obtained,and the error rate and the prediction accuracy of the test sample are drawn.In order to get more accurate and reliable results,the number of samples added to each sample is changed,and the final result is analyzed as a comparative experiment.It is concluded that both BP network model and RBF network model have certain ability to refuse to fake,and the relative error of prediction is within acceptable range,but the relative error of BP network model prediction is less than that of RBF network model..Therefore,when the precision is high in practical application,the BP network model should be selected first.The neural network is very widely applicated,especially in optics,such as optical image processing,optical detection,optical communication and adaptive optics and so on,and achieved good results indeed.