Advanced Sliding Mode Online Training for Neural Network Flight Control Applications

The transfer of higher order sliding mode theory to the online training of neural networks used in adaptive flight control systems is presented and evaluated for a small unmanned aircraft system (UAS). Derived from variable structure theory sliding mode online training allows for the dynamic and stable calculation of the learning rate of a neural network which is a major improvement in comparison to empirical determination. By considering the training of a neural network a control problem, sliding mode control theory provides specific switching functions and reaching conditions which force the network error onto a predefined trajectory within the state space of such a network. On this sliding surface the network error and its derivatives are steered to the point of origin of the state space which ensures a stable training process. To increase the order of this sliding motion not only the sliding function but also its first derivative is utilized in the switching law. This way the speed of convergence of the training algorithm can be further increased. In this work a nonlinear feedback linearization approach is augmented with neural networks in order to approximate and cancel inversion errors that may occur due to parametric uncertainties. First and second order sliding mode learning algorithms are compared to a standard gradient decent training in a simulation of a small fixed wing UAS. In addition to external disturbances like wind and gust, system degradation and constant inversion errors are simulated to analyze and prove the robustness of the presented approach.