Minimal Neural Networks for Real-Time Online Nonlinear System Identification

Complex nonlinear dynamical systems which evolve over time are difficult to characterise with Linear Time Invariant techniques. In this paper, a minimal neural network is proposed for tracking the state of a plant and is optimised for real-time embedded system deployment. A comparative investigation found that the Rectified Linear Unit activation function had superior performance over other common activation functions in terms of both convergence and processing time. The proposed minimal neural network based system identification technique was validated through MATLAB/Simulink simulation. It was found that a hidden layer size of six neurons was capable of representing a highly nonlinear plant with an average error of as little as 0.01% after a period of simulated supervised learning. A computationally restricted multilayer network was also tested, which was not as accurate on average as the single-layer network containing six hidden neurons. Preliminary hardware implementation also yields promising results and justifies the continuation of research on this algorithm. The original motivation for this study was inspired by the test control systems used for complex full-scale airframe fatigue tests, and the need to maintain load accuracy despite the interactions between the test article and test rig. Further research should be conducted into applying this algorithm for the characterisation of other complex Multiple Input Multiple Output systems.