Feature Extraction and NN-Based Enhanced Test Maneuver Deployment for 2 DoF Vehicle Simulator

Abstract

This paper presents a deployment method of various test maneuver scenarios for 2 degree of freedom (2 DoF) vehicle simulator by using feature extraction and neural networks (NN). A prototype version has been set up for the 2 DoF vehicle simulator. Then, a hardware in the loop (HIL) model with 2 inputs (torque, τ 1 -τ 2 ) and 3 outputs (acceleration, a x -a y -a z ) is created. System identification is performed to obtain the training data of NNs to be used for the deployment of test maneuvers. In the system identification process, 2 arbitrary sinusoidal torque signals (τ 1 -τ 2 ) are generated by using the actuator specs of the 2 DoF vehicle simulator. By applying the generated torque signals to the actuators, acceleration (a x -a y -a z ) data are collected from the inertial measurement sensor (IMU) on the 2 DoF vehicle simulator. It is determined to create 3 different NN models for the obtained data. The 1st NN model is trained with 3 inputs (a x -a y -a z ) and 2 targets (τ 1 -τ 2 ) training data. The 2nd NN model is trained with 6 inputs (amplitudes and phases of a x -a y -a z ) and 2 targets (τ 1 -τ 2 ) training data. The input data features for the 2nd NN model is extracted by using the Fast Fourier Transform (FFT). The 3rd NN model is trained with 6 inputs (amplitudes and phases of a x -a y -a z ) and 4 targets (amplitudes and phases of τ 1 -τ 2 ) training data. For the 3rd NN model, the features of input and target data are extracted by using the FFT. The NN training process continues until acceptable performance criteria are reached. Then, 3 NN models are run and analyzed under various test scenarios such as Double Lane Change, Constant Radius, Increase Steer, Fish Hook, Sine with Dwell and Swept Sine. Only for the 3rd NN, the actuator signals (τ 1 -τ 2 ) are recomposed by applying an inverse FFT process to the 4 targets (amplitudes and phases of τ 1 -τ 2 ). Finally, the reference trajectory tracking performances are evaluated by comparing the NN models that are run under the test scenarios.