TORQUE VECTORING NEURAL NETWORK

I along with a partner designed and simulated an adaptive controller working on a car model to later be implemented as a module of a small scale intelligent vehicle (SSIV) platform. It will be used as a learning tool for a future controls course in the Mechanical Engineering Department. This was a proof of concept effort and thus wasn't implemented on actual hardware. The intent with this project is to model the nonlinear behavior of a car moving through a turn by utilizing neural networks for system identification and control. Below is the basic outline of the Neural Network implemented in Matlab to identify the plant and act as a controller (two parts are marked in the diagram).

The plant was first trained to mimic actual vehicle slip data (volvo v70) using a nonlinear autoregressive network with exogenous inputs (NARX network) designed with 5 hidden layers and 10 neurons per layer. The Levenberg-Marquardt backpropagation training algorithm, ‘trainlm’ and a linear activation function, ‘purelin’ were used. Training was run for 13 epochs until it matched data sufficiently.

The weights and biases were locked as the NARX network was copied as the plant into the bigger network encompassing the controller and plant shown below. With the first network tracking its training set well, we created an MRAC (control loop side) and trained it in closed-loop mode.