Regularized RBF Network-Based Inferential Sensor and Its Application in Product Quality Prediction

The risk of overfitting on noisy data is of major concern in neural network design. Regularization provides a stable solution to function approximation with a tradeoff between accuracy and smoothness of the solutions. k-means cluster algorithm is applied to determine the network centers at first and an approach based on L-curve is then proposed to estimate regularization parameter. These estimations are conbimed with forward selection to update network parameters in training. Simulation results show that RBFN with a suitable regularization parameter can get a good generalization.