Research and analysis of the BP neural networkstructure and features. Find its shortcomingsand propose an improved method for the deficiencies, and establish the neural network softwarereliability of the new model.Through MATLAB simulation tools forexamples of simulation, confirmed the new model year with the traditional modelof high-precision, the characteristics of generalization stronger.

RBF neural network have advantages of training simple, fast efficiency of learning, easy to fall into local minima, etc..It is widely used to solve the problem in signal processing and pattern recognition. Although the common RBF network is relatively easy to build, but because of the structure is usually fixed or high complexity, resulting in learning time is too long or network resource waste. For these reasons, proposed using extended Kalman filter as the RBF learning algorithm, and using double radial function in the hidden layer. By approaching the basis of the results of the analysis clearly shows that the network model than the other categories have a stronger generalization.

Based on the power spectral density (PSD) function of stochastic irregularities of the standard grade road and by means of inverse fast Fouerier transform (IFFT) based on discretized PSD sampling, an equivalent sample of stochastic road surface model in time domain was built. A one-dimensional model of stochastic road was developed into a 2D model of stochastic road surface. Through computer simulation practice based on the MATlab, a 2D sample of stochastic road surface in time domain was regenerated. Furthermore, given the sample data, the PSD was estimated and then compared with the theoretical 2D PSD Equation deduced from the one-dimensional PSD expreesion so as to prove the effectiveness and accuracy of the time-domain model regeneration of 2D stochastic road surface by means of IFFT method. The 2D stochastic road surface model directly provided basic road excitation input data for virtual prototyping (VP) and virtual proving ground (VPG) technology.

By using time domain modeling obtained from PSD for stationary random road irregularities, concerning about non-uniform moving-vehicle conditions and non-stationary road characteristics, via model assembly and integration to generate non-stationary random road time domain modeling for inputting road excitation. On the assembly of the non-stationary random road time domain modeling, Time-Frequency was analyzed via wavelet analysis. From road data decomposed by the way of Wavelet, both time domain characteristics of signal components in different bands can be obtained and anomaly contained within the signal can also have accurately time-frequency localization. The signal separation can help achieving non-stationary sub-band analysis, and can further study the affection of road signal in different band focus on vibration response of the vehicle, and also providing a strong theoretical basis for road design, quality inspection and grading.

Abstract. The impact of artificial neural network model output precision technology widespread attention. Quality sample study of neural network output accuracy is not much affected, most of the research is the structure (number of layers and the number of nodes), the impact of this paper to analyze samples of artificial neural network output for the neural network, to improve the output of neural network accuracy is important.