基于自适应高效递推规范变量分析的多模过程软传感器建模
Soft sensor modeling for multimode process based on adaptive efficient recursive canonical variate analysis

由于多模过程中各模式间的均值和协方差发生了改变, 多变量单模高斯分布的基本假设不再成立. 基于递 推方法的多模过程软传感器建模存在两点问题: 其一, 递推建模方法不能及时的跟踪多模过程的改变; 其二, 递推 建模方法的在线计算负荷非常高. 为了解决上述问题, 本文提出了一种基于自适应高效递推规范变量分析的多模 过程软传感器建模方法. 首先, 采用指数权重滑动平均来更新过去观测矢量的协方差矩阵; 然后, 利用基于模型输 出误差范数的可变遗忘因子来跟踪多模过程的动态变化; 最后, 通过引入一阶干扰理论(first order perturbation, FOP) 来实现递推奇异值分解, 与常规奇异值分解相比递推奇异值算法的计算负荷显著降低. 将提出的方法用于田 纳西伊斯曼(tennessee eastman, TE)化工过程进行仿真验证, 其结果表明了该方法的可行性和精确性.
Because of the mean shift and covariance changes between process modes, the basic assumption of Gaussian distribution for the multivariate unimode processes does not hold for the complex multimode processes. Disadvantages of soft sensor based on recursive modeling approach for multimode processes include the following two points: first, recursive methods have difficulty in timely tracking with the changes of multimode processes; second, the online computational cost of recursive modeling becomes much more expensive. To solve these problems, we propose a soft sensor for multimode processes on the basis of the adaptive efficient recursive canonical variate analysis (AERCVA) modeling approach. First, the exponential weighted moving average is adopted to update the covariance of past observation vectors. Then, the variable forgetting factor based on the norm of output residuals is used to track the changes of the multimode process. Finally, the first order perturbation (FOP) is introduced to realize recursive singular value decomposition (RSVD), the computation cost of RSVD is significantly reduced in comparison with the conventional singular value decomposition. The proposed method is validated in the Tennessee Eastman benchmark process. Simulation results demonstrate the feasibility and accuracy of the proposed method.