Researches Of Probabilistic Neural Networks Optimized By Fish Swarm Algorithm

In Probabilistic neural network (PNN) uses Bayes’ theorem and the minimumrisk-based Bayesian decision rule for classifying new samples of neural network. Itstraining time is short and it is easy to converge to a local maximum. But traditionalPNN’s identical smooth factor easily lead to low recognition rate andmisclassification problems. In addition smooth factor has a huge impact on theclassification and it is difficult to determine. And at last the number of pattern layerneurons is determined by the number of training samples. That is to say huge trainingsample set will lead complex network structure. To solve these problems in this papertwo improvements are made on PNN network:(1) The number of smoothing factor isbased on the number of categories of pattern, the smoothing coefficient from a singletransformed into a smoothing coefficient vector, so that the neurons in the hiddenlayer has a higher flexibility.(2) Select smoothing vector and optimize the networkstructure PNN with AFSA which is improved in this paper.Three improvements are made in this paper on AFSA:(1) use the rules of PSOwhich are used to update the speed to manual the step factor in AFSA and named thisas automatically adjust behavior.(2) merger the original three acts: see the foragingbehavior as the default behavior of following behavior and flocking behavior, whenthe following and flocking behavior cannot find a new location foraging behavior willbe simulated. Make each behavior fully take into account of individual informationand overall information. So after improving algorithms only have two basic behaviors.(3) To accelerate the convergence rate define a jumping behavior, performed aftereach iteration: after each iteration, sort all artificial fish according to the fitness, thefish ranked at the lower half copy the information of the ones at the top half. Throughthis action, increase the affection of the better artificial fish on the entire population ineach iterative process and avoid a negative impact of the poor artificial fish. Thus theconvergence rate can be speed up.Finally, the algorithm in this paper is applied in fault classification of industrial equipment and compared with basic PNN and other traditional classificationalgorithms, the algorithm in this paper achieved good classification results. Byanalyzing the deficiencies of traditional fault classification model, based on PNNdefine a "sub-health" fault classification model, and experiment shows that theclassification model advantages over the traditional fault classification model.