Vehicle Simulationand Test System Data Fusion Based On Neural Network And Its Algorithm Research

The vehicle simulation and test system proposed by FAW-Volkswagen is an indoorsystem which can simulation the real vehicle driving environment. The system can savea lot of labor and resources which is wasted in the testing process. The emphasis of thevehicle simulation and test system is the data fusion of the information extracted fromcan bus.Artificial Neural Networks is a newly revived data processing technology from1980s. It is regarded correct in theory that the accurate fitting function between anynumber of inputs and outputs can be given by enough hidden layer nodes. NeuralNetworks is widely used in practical problems with a strong self-learning ability andgood fault tolerance.Considering the complex function relationship between inputs and outputs in thevehicle simulation and test model which cannot be described by a single mathematicalexpression, using the Artificial Neural Networks into data fusion of vehicle simulationand test system is proposed in this paper.Firstly, by analyzing the packet structure of the CAN bus, information isdownloaded into PC, and is drawing in MATLAB.Secondly according to the parameter which was cared by drivers while driving, weset up the vehicle simulation and test model. Studying the influencing factors of theparameter, the model’s input and output is fixed.Then, BPNN\cascade-NN\RBFNN is selected for they are widely used in functionfitting field recently. The basic principle, structure and learning algorithm of thenetworks are introduced in the paper. Among them, LM learning algorithm is aimproved algorithm of BPNN. NRBF and Classified-RBF is proposed against the poorgeneralization ability of ordinary RBFNN.At last, the networks is used in fitting the vehicle simulation and test model,achieving a good result, which shows that using the Artificial Neural Networks into datafusion of vehicle simulation and test system is feasible and effective. NRBF andClassified-RBF which is proposed in this paper can reduce the output error and improvethe generalization capability of the RBFNN. By comparing the fitting effect betweendifferent networks, we get the conclusion, that the requirements of training spend andaccuracy are contradictory. That means we should choose the network according to the vehicle simulation and test system’s needs. For example, to some off-line learningprocess, cascade-NN and BPNN is adopt for its high accuracy; and to the on-linevehicle security system, NRBF and Classified-RBF should be used because of their fasttraining spend.