Latent model-free adaptive control of nonlinear multivariable processes via virtual dynamic data modeling

Model free adaptive control has become an excellent method for complex processes with no model information available. However, the increasing scale of production in modern industries makes it difficult to model and control these processes. Therefore, a novel latent model-free adaptive control is proposed to deal with the high-dimension and collinearity problem of process variables in real-world industries. First, a nonlinear autoregressive moving average with exogenous input model is designed as a dynamical partial least squares inner relationship in the latent space to formulate the system input and output dynamics in a most common way. Then, a latent full-form dynamic linearization is developed to make the nonlinear model linearly parametric and a latent full-form dynamic linearization based virtual dynamical partial least squares data model is proposed consequently. An estimation algorithm is developed for identifying the unknown parameters of the virtual dynamical data model. By means of the virtual dynamical data model, the latent model-free adaptive control method is proposed by designing and optimizing a quadratic objective function. Theoretical analysis and simulation study confirm the efficiency of the latent model-free adaptive control.