From neural state-space description of marine powerplants to reduced-order Volterra models

The problem of reduced nonlinear input-output models for marine propulsion powerplants is visited by starting from the neural state-space description of the system, the neural networks of which can be trained by using steady-state data of the powerplant collected either from measurements or derived by use of conventional thermodynamic models. The analysis proposed is based on the Taylor expansion of the logistic sigmoid function, appearing in the neural torque approximators of the neural plant description, yielding a finite polynomial Volterra expression, approximating the dynamic behavior of the plant with desired accuracy. The reduced nonlinear model obtains finally the form of a sum of homogenous Volterra operators, and can be used for frequency domain characterization of the system and the design of nonlinear feedforward controllers.