Abstract:

Aiming at the problem that the soft sensing precision of key variables deteriorates when unmodeled dynamics and uncertain disturbances exist in the complex industrial process, an on-line soft sensor based on stable Hammerstein model (H model) was presented. H model was composed of wavelet neural network with time-varying stable learning algorithm as nonlinear gain and ARX model with RLS (recursive least square) algorithm as linear part. The boundedness of identification error for H model was proved according to the Input-to-State Stability theory. Wavelet neural network could represent strong nonlinearity of the process, and the stable learning algorithm could restrain the influences of unmodeled dynamics and uncertain disturbances and improve prediction precision and self-adaptability. Simulations based on a nonlinear system and the wastewater treatment process showed that the soft sensing method presented in this paper possessed high prediction precision.

Key words: Hammerstein model, on-line modeling, soft sensor, prediction, stable learning, wastewater treatment process, stability

摘要：

针对复杂工业过程中由于存在未建模动态和不确定干扰,导致关键变量的软测量精度下降的问题,提出了一种基于稳定Hammerstein模型(H模型)的在线软测量建模方法。H模型的非线性增益采用带有时变稳定学习算法的小波神经网络模型,线性系统部分采用基于递推最小二乘的ARX模型,基于输入到状态稳定性理论证明了H模型辨识误差的有界性。其中小波神经网络具有表征强非线性的特性,稳定学习算法可抑制未建模动态和不确定干扰的影响,改善了模型的预测精度和自适应能力。以典型非线性系统和实际污水处理过程为例进行了仿真研究,结果表明,基于稳定H模型的软测量方法具有较高的在线软测量精度。

关键词: Hammerstein模型, 在线建模, 软测量, 预测, 稳定学习, 污水处理过程, 稳定性