Weighted Probabilistic Neural Network Based on the Sensitivity Analysis

Due to low training complexity, high stability, quick convergence and simple construction, the probabilistic neural network (PNN) has got extensive application in many fields. Because of the lack of additional weighting factors inside PNN model structure, the PNN using the Sensitivity Analysis (SA) has been improved in this paper. The weight coefficients and compensating factors are introduced into the network and put between pattern layer and summation layer to create the weighted probabilistic neural network (WPNN). The weights are derived using the sensitivity analysis procedure when the radial kernels are used as the output of the pattern layer. At the same time, compensating factors compensate the impact of the SA among the patterns. The performance of the WPNN is examined in contradistinctive experiments. Meanwhile, WPNN is used in fault diagnosis of the aircraft wing skin to prove feasibility of WPNN. The results show that the WPNN is feasible and has better performance in prediction accuracy.