Reasearch On The Improvement Of RBF Neural Network Based On Optimization Algorithm And Its Applications

The hybrid algorithm of neural network and intelligent optimization algorithm is themain tool of intelligent information processing. Many hybrid models and and their improvedmodels are put forward and widely used in air quality forecast, economic forecast, sensor,radar, sonar, communication and other fields. The research of neural network mainly includethe topology of the network model, parameter selection, etc. The research of particle swarmalgorithm mainly include the population topology structure, the selection of inertia weight,PSO population diversity, etc. In intelligent information processing applications, based on theresearch of RBF neural network model mechanism and the intelligence PSO algorithm, thisarticle puts forward different models combined of radial basis neural network and improvedPSO algorithm.In order to better balance the global exploitation ability and the local exploration abilityof particle swarm, an exponential decreasing inertia weight PSO algorithm is put forward. Inearly iterative steps, the inertia weight has faster descent speed to drop, so the particle swarmquickly search the feasible solution region; in the late iterations, inertia weight has slowerdescent speed to drop, so the particle swarm make fine-tuning for the global optimal solutionin the feasible solution region. Using the modified EDIW-PSO algorithm to optimize the RBFneural network parameters, an EDIW-PSO-RBF model is estiblished. For comparing withother three kinds of RBF neural network models based on different inertia weight strategyPSO algorithms, the air quality forecast experiment is choosen. The experiment results provethat the exponential decreasing inertia weight strategy PSO algorithm is more effective fortraining RBF neural network than other algorithms and has higher prediction precision.In order to generalize the distribution shape of the radial basis function (RBF) and extendthe shape parameter selection of the standard Gauss function, the generalized gaussian function is adopted as the radial basis function, so that the generalized radial basisfunction(GRBF) neural network is built. GRBF neural network has another shape parametersneed to be adjusted, besides parameters such as centers, widths, and the connection weights.Combining the EDIW-PSO algorithm and GRBF neural network optimizes the parameters ofeach hidden neuron. Meanwhile, use the AdaBoost algorithm integrated learning ability forthe selection of GRBF neural network connection weights.We set each hidden neuron as aweak prediction machine, and set the whole GEBF neuron network as a strong predictionmachine. An EDIW-PSO-AdaBoost-GRBF neuron network model is established. A two foldsynthetic bi-dimensional application example and a time series prediction example for theShanghai securities composite index are experimented to verify the validity of the modelwhich has higher precision.The vector hydrophone array for detection of underwater target and direction of arrive(DOA) estimation is an important research direction in ocean exploration and military. On theone hand, the EDIW-PSO-GRNN neural network model and the EDIW-PSO–AdaBoost-GRBF neural network model are applied in MEMS vector hydrophone array DOA estimationreplacing the original MUSIC algorithm. First, a real-valued spectral MUSIC algorithm isproposed to reduce the calculation amount in the MUSIC algorithm for DOA estimation. Onthe other hand, the powerful search ability of the EDIW-PSO algorithm is used to optimizethe real-valued MUSIC algorithm to improve the precision of MUSIC algorithm for the DOAestimation. In the two neural network models, set the first line of the real-valued covariancematrix C as the input of neural network, and set the incidence angles of sources as the outputof the neural network. Before training the neural network, respectively take different incidentangles, according to the mapping relation of the input and vector array output model obtaindifferent real-valued covariance matrix, so that obtain a set of training sample. Take the firstline of the real-valued covariance matrix of data received from vector array as the test samples.Train the neural network, and then test the received data for the DOA estimation. A simulationexperiment for simple sound source、a simulation experiment for multi sound source and an experiment in an anechoic tank are experimented to compare the three proposed algorithmswith the original MUSIC algorithm. The three experiments verify that the proposedEDIW-PSO algorithm improves the performance of MUSIC algorithm, and also verify thatthe two kinds of neural network models is effective in goal orientation of MEMS vectorhydrophone array in engineering application.