Research On The Test Method Of Fuel Consumption Based On BP Neural Network

With the increase of automobile population in recent years, considering the energy crisis and the ecological environment, all the countries have paid more and more attention to the fuel economy, and have invested enormous human and material resources from the limit on the automobile fuel consumption to the innovation of the engine technology. The existing fuel consumption test method is divided into direct measurement and indirect measurement. Direct measurement method is complicated, time-consuming, and would undermine the construction of the vehicles. The main indirect measurement is carbon balance method, of which the main principle is calculating the carbon content of exhaust and obtaining the oil consumption value on the basis of mass conservation law. Based on analyzing the domestic research, a soft measurement method based on BP neural network to detect oil consumption rapidly is proposed in this paper.BP neural network is a multi-layer feed-forward network, with a back propagation error, one of the most widely used neural network models. It can store the mapping between the input and the output through the training, without taking the complexity of the system itself into account. Restructuring the network structure can implement any nonlinear mapping, and BP network consists of input layer, hidden layer and output layer, including any number of hidden layers, which can be made up of any number of parallel computing simple neurons, neurons between layers in fully connected manner, while neurons in the same layer with no connections.The principle of the carbon balance method was analyzed in this paper first, the mathematical model of calculating fuel consumption based on air-fuel ratio was designed according to the principle of carbon balance method, and a test platform was built and a data acquisition system was developed using Visual c++. Experimental platform is mainly used for measuring the concentration of all the gases in the vehicle exhaust, the engine speed, the real fuel consumption and other experimental data required by BP network training and testing. The main functions achieved by data acquisition system are to filter the collected experimental data, and calculate the fuel consumption using the air-fuel ratio model based on the filtered data, in order to compare with BP neural network method.For this paper, the setting of BP network structure is a difficult point. There is no fixed formula for the number of nodes in the BP network hidden layer, which can only be determined by experience formula or cut-and-trial method. In this paper, through increasing the nodes, by comparing the effect map of the network training error curve when the number of nodes in the hidden layer is 5,9 and 13, 13 is selected as the number of nodes in the hidden layer finally, and also by comparing the error curve when the maximum training time is 500 , 1000 and 1500, 1000 is determined as the maximum training time. Finally the network topological structure is identified as 8-13-1.The three training algorithms of GD, GDA and LM are used to train the network in this paper, after comparing the training error curve and analyzing, the conclusion can be achieved that GD algorithm can not converge, and LM algorithm can converge, but due to its own characteristics, the network convergence velocity depends on the automatic initialization parameters, likely to cause the system instable and the situation of insufficient or non-convergence learning to appear. GDA algorithm is characterized by the automatic adjustment of learning rate, of which the error curve is gently, the system is relatively stable, and which is identified as the ultimate training algorithm.In the final stage of the paper, by comparing the output data of BP network testing with the fuel consumption data based on the air-fuel ratio model and the data measured by the fuel consumption instrument from horizontal and vertical respectively, the conclusion that the application of BP network in fuel consumption testing is superior to the method based on air-fuel ratio model is obtained, meanwhile the average relative error of BP network output values is in the range of 5%, meeting the actual needs fundamentally. Finally, through the application, the advantages and disadvantages of neural network technology are summarized, and the prospects and developments of BP network in the fuel consumption the measurement are discussed in this paper.