Control‐Oriented Neural Network for Systems with Nonlinear Elements

This paper proposes a novel neural network (NN) architecture for the unified representation of typical nonlinear elements in control systems, including dead zones, input saturation, and hysteresis. The proposed architecture employs a compact structure composed of rectified linear units, enabling enhanced learning efficiency and reduced computational complexity in comparison with conventional general‐purpose recurrent neural networks. It is further demonstrated that the inverse models corresponding to these nonlinear elements can be formulated in a closed‐form expression without approximation, thereby facilitating precise nonlinear compensation even in data‐driven control frameworks. The effectiveness of the proposed NN is confirmed through several numerical examples.