RBF neural networks and cross validation-based signal reconstruction for nonlinear multi-functional sensor

For signal reconstruction in nonlinear multi-functional sensor, there may be some outliers caused by systematic errors or gross errors in observations. Therefore, it is worth while to get rid of the outliers from experimental data during the calculation to ensure the reliability and precision of the system. Based on the Radial Basis Function neural network and the law of cross validation, this paper presents an iterative regressing method in consideration of the existence of outliers. Cross validation is repeatedly used for random sampling the experimental data as the training data set, with which RBF neural network can complete the regressing. By repeating such procedure and updating the estimated parameters, the training data set and system function of multi-functional sensor can be optimized. Accordingly, the reconstruction of any signals can be accomplished with the selected model. The theoretic analysis and the experimental results show that the approach is effective, robust and practicable.