Research On Flatness Recognition Model Based On Distributed Radical Basis Function Neural Networks

Flatness recognition is the foundation of shape control system of the high precisewide strip steel cold mill. A group of on line detected stress values is mapped into severalcharacteristic parameters on a small dimension in the flatness procedure. The least squaremethod used in industry cannot meet the requirement accuracy because of the limitation ofthe number and the detected errors of the shape meters. Therefore, the artificial neuralnetwork for wide strip steel cold mill is chosen to be research object with theoretical andengineering sense and it has achieved new fruit.Firstly, flatness stress data has complex structure and the initial values of networks’parameters in flatness recognition model are hard to be optimized. An approach to build adistributed RBF (Radial Basis Function) neural networks recognition model based onSVM (Support Vector Machine) classifier is presented. To reduce the influence amongdifferent kinds of samples, flatness data with different features are inputted into foursub-networks respectively according to classification result made by a trained SVMclassifier. Parameters in sub-networks are optimized by equivalence of kernel-based SVMand neural networks with three layers. The weight vector and center vector of trainedSVM are set to be initial values of RBF neural networks.Secondly, too many training samples for the neural networks elongate the trainingtime, and the noise data make the network hard to converge. Therefore neural networkstraining algorithm based on PSO (Particle Swarm Optimization) is put forward to solvethis problem. The recognition error of neural networks is set to be fitness value of the PSO,and the PSO algorithm is applied to searches for the best parameter vector for neuralnetworks. The inertia weight vector of swarms is set to be dynamic value provided by theGauss function to improve the performance of the PSO algorithm.Finally, the distributed neural networks based on SVM and the training algorithmbased on PSO is applied to solve the flatness recognition problem. The constructionmethod and the training method of flatness recognition model based on distributed neural networks are given in this paper. At the same time the model has been applied on the1220two-stand temper mill for industrial experiment to evaluate its performance.The experiment with industrial data has shown that the proposed model can meet theindustrial requirement on both recognition accuracy and efficiency.