稳定RBF神经网络的在线软测量建模方法  

Stable Soft Sensor Based on RBF Neural Network and Its Applications

作　　者：丛秋梅[1] 邓淑贤[1] 赵宇[1] 王艳[1] CONG Qiu-mei;DENG Shu-xian;ZHAO Yu;WANG Yan(School of Information and Control Engineering, Liaoning Shihua University, Fushun 113001, China)

机构地区：[1]辽宁石油化工大学信息与控制工程学院

出　　处：《控制工程》2018年第5期823-828,共6页Control Engineering of China

基　　金：国家自然科学基金项目(61673199,61573364);辽宁省教育厅一般项目(L2015297);辽宁石油化工大学国家级科研项目培育基金(2016PY-017)

摘　　要：针对RBF（Radial Basis Function）神经网络在存在未建模动态或不确定干扰时,采用梯度下降法建模出现不稳定、实时性和鲁棒性较差的问题,提出了带有稳定学习算法的RBF神经网络在线软测量建模方法。以隐含层径向基函数为Gaussian函数的RBF神经网络为例,通过分析ISS（Input-to-State Stability,输入到状态稳定性）-Lyapunov函数,得到网络权值和径向基函数参数的稳定学习算法,并证明RBF神经网络辨识误差的有界性。稳定学习算法可抑制过程未建模动态和不确定干扰的影响,使软测量模型具有较高的预测精度和自适应能力。以非线性对象和实际污水处理过程为例进行了仿真,结果表明,以稳定RBF神经网络建立的软测量模型具有较好的鲁棒性和在线软测量性能。RBF neural network（RBFNN） learned by the conventional gradient descent algorithm is likely to have problems such as instability, weak real-time capability and robustness when unmodeled dynamics and uncertain disturbances exist. An on-line soft sensor based on RBFNN with stable learning rate is presented. The stable learning algorithm of the weights and the parameters of radial basis function of the RBFNN are derived through ISS-Lyapunov function, and the boundedness of the identification error for RBFNN is proved. Stable learning algorithm could weaken the influences of unmodeled dynamics and uncertain disturbances and the precision and adaptive ability of RBFNN can be improved. Simulation experiments of a nonlinear process and the wastewater treatment process show that the soft sensor based on stable RBFNN possesses high robustness and online predictive performance.

关 键 词：径向基函数 软测量 建模 稳定学习算法 输入到状态稳定性 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]