Indirect Data-Driven Observer Design Using Neural Canonical Observer Structures

An indirect data-driven observer design approach for nonlinear discrete-time systems based on an input-output injection with neural canonical observer structures is proposed. An artificial neural network auto-encoder structure, trained with recorded state, input, and output data, is used for the identification of a system in a nonlinear Brunovsky observer form with output transformation. The neural approximations of the transformations and the input-output injection can be used to construct an observer with linear error dynamics using methods from linear control theory. The approach is demonstrated on two academic examples and on an industrially-motivated problem with a sampled continuous-time model.