A Neural Network-Based Direct Inverse Model Application to Adaptive Tracking Control of Electronic Throttle Systems 2014-01-0197
This paper presents another application  of using Artificial Neural Networks (ANN) in adaptive tracking control of an electronic throttle system. The ANN learns to model the experimental direct inverse dynamic of the throttle servo system using a multilayer perceptron neural network structure with the dynamic back-propagation algorithm. An off-line training process was used based on an historical set of experimental measurements that covered all operating conditions. This provided sufficient information on the dynamics of the open-loop inverse nonlinear plant model. The identified ANN Direct Inverse Model (ANNDIM) was used as a feed-forward controller combined with an adaptive feed-back gains (PID) controller scheduled  at different operating conditions to provide the robustness in tracking control to un-modeled dynamics of the throttle servo system. The un-modeled dynamics are mainly related to the strong nonlinearity functions that may excite the system with external un-measurable disturbances and noise effects. The feed-forward ANNDIM is first used to emulate the inverse dynamics of the DC servo system. However, the variations in nonlinear dynamics of the throttle body during actual operation cause some error in the prediction of the exact inverse dynamic obtained from ANNDIM. Therefore, by adding the feed-back PID adaptive controller term in the control loop will compensate the model mismatches for these un-modeled dynamic variations and improve the overall control performance. Practical implementation results using rapid prototype real-time system testing are provided to illustrate the performance and effectiveness of the proposed method in tracking controls of multiple set-point changes at different operating conditions.