A feature-based data-driven approach for controller design and tuning

Traditionally controller tuning is model based. In many practical applications, however, the process model cannot be obtained and model-free tuning is imperative. In industrial control the huge amount of data is available, but we lack effective controller tuning schemes that are data driven instead of model driven. To address this issue, in this paper we first introduce the concept of feature space that can capture the characteristics of a control process, either in the time domain, frequency domain, or others. (data space to feature space, dim reduction) Next we introduce the control basis function space and control parameter space. The features and parameters form a mapping relationship. The controller tuning process can thus be formulated into the inversion of the mapping that yields appropriate control parameters and minimizes the mismatching between reference features and actual features. When the inversion is not analytically solvable, the iterative learning tuning method can be used.