基于小波分析的多RBF神经网络轧制力设定模型

带钢热连轧生产过程中,轧制力预设定时的轧制力信号影响因素多、关联复杂,难以建立精确的机理模型.为此,文中应用小波多分辨分析方法,将轧制力分解重构为对应于不同影响因素的子信号,并建立了一个多RBF神经网络模型.模型中每个子网络分别对一个子信号进行建模,最后将各子网络输出综合为轧制力设定信号.各个子信号的影响因素不同,每个子模型输入参数和输出参数亦不同,从而能真实地反映轧制力变化的内在机理,具有明确的物理意义.仿真实验表明,这种建模方法降低了系统维数,能有效提高网络学习能力,轧制力预设定误差率从BP神经网络的10%降低到了5%.

A Multiple-RBF Neural Network Model to Set Rolling Force Based on Wavelet Analysis

Aiming at the problem of rolling force setting, a multi-RBF neural network model based on the multi-resolution wavelet analysis is proposed. In a hot strip rolling mill, the number of system parameters and disturbance factors is large, the relationships between each of them are complicated, so it is difficult to obtain accurate mathematical models for the process control, including the one for rolling force setting. The multi- resolution wavelet analysis method is employed to decompose the rolling force signal and reconstruct it as a serial of sub-components. Each sub-component has its own affecting factors. Then a multi-RBF neural network with several sub-networks is established, each sub-network is for the modelling of each sub-component, and their outputs values are added up as the rolling force value. Different to those sub-networks in other multi-model modelling methods, the sub-networks in the proposed method have different input parameters and outputs, so they can truly reflect the variation mechanism of the rolling force. Simulation results show that this proposed model can reduce the system dimension, and improve the learning ability of the network. The error rate of rolling force setting is reduced from 10% by BP neural network model to 5% by the proposed model.