A new method for spread value estimation in multi-spread PNN and its application in ship noise classification

The use of Probabilistic Neural Network (PNN) is very common in supervised pattern recognition applications. PNN is based on Bayes decision rule and it uses Gaussian Parzen windows for estimating the probability density functions (pdf) required in Bayes rule. The conventional PNN needs a single spread value for pdf estimation which is proportional to Gaussian window width. In this paper we will suggest the use of a multi-spread PNN structure whose spread values are estimated using the training data. In addition, we will introduce several new discriminating features of acoustic radiated noise which can be used for ship noise classification. These features will be used as discriminating features in the conventional and multi-spread PNN. Finally, the performance of the conventional PNN and the suggested multi-spread PNN in classifying real ship noise data will be compared. Results of this comparison show that the performance of the multi-spread PNN is better than the conventional PNN.