Sensitivity Study And Application Of B-spline Weight Function Neural Network

Professor Daiyuan Zhang proposed a new neural network model named spline weight functions learning algorithm neural networks algorithm in the monograph, named"The new neural networks theory and method". Spline weight functions neural networks breaks down the traditional shortcomings of the complex neural networks structure completely and simplify the networks structure. The theory and ways completely overcome the troubled years traditional algorithm academia questions such as local minima, slow convergence and difficult to obtain the global optimal point. B-spline weight function neural network owns the advantages of weight function neural network and B-spline curveWhen the trained neural networks are disturbed by noise, its weight will fluctuate and its output will change. Additionally, if the sample itself has input noise, it will make the output of the neural network changes. These changes and influence can be analyzed by the concept of sensitivity. The sensitivity analysis of neural networks has been studied by some scholars at home and abroad. Professor Daiyuan Zhang obtained statistical sensitivity formula of three-layer feed-forward neural network using recursive method. This paper firstly studied the B-spline weight function neural network and its training algorithm, and then deduced B-spline weight function neural sensitivity formula according to statistics sensitivity. Finally, simulation experiments show that B-spline weight function neural network has good approximation capability of generalization ability and the correct ness of theoretical sensitivity formal.This paper uses B-spline weight function neural network sensitivity for the modulation signal recognition based on the above analysis of B-spline weight function neural network sensitivity, according to the theory of digital modulation signal recognition and feature extraction of transient signals of five instantaneous parameters. From the simulation results we can see that B-spline weight function neural network classifier not only has the traditional advantages of neural network classifiers, but also simple structure and high recognition rate. The recognition rate will be higher after adding the sensitivity analysis.