Suspension Controls and Parameter Estimation Using Accelerometer Based Intelligent Tires

Abstract

This thesis aims at estimating vital vehicle states and developing control algorithms for automotive suspensions and vehicle stability. A parametric model of an automotive monotube damper is developed and several control algorithms for semi-active suspensions have been developed. An extensive comparison of different control algorithms has been done. Skyhook, Groundhook, Hybrid, Acceleration-driven, Power-driven, Groundhook-linear, Linear Quadratic Regulator (LQR) optimal, Genetic algorithm optimized Linear Quadratic Regulator optimal, Model-reference adaptive, H∞ robust, µ-synthesis, fuzzy-logic based, and Deep Reinforcement learning based control algorithms have been developed and simulated. A shock dyno is instrumented and skyhook and groundhook control algorithms have been implemented as well. In addition to this, a semi-active suspension switching based control algorithm is developed for reducing the effort of a direct moment yaw rate controller, and improve stability of a vehicle when turning. Accelerometer based intelligent tires have been used to estimate vehicle states like vertical load on tire, velocity of the vehicle, unsprung mass acceleration, and forces on a tire. All these estimations would be helpful in observing various parameters of a vehicle using data from only a tri-axis accelerometer inside the tire. Data was collected in an instrumented Volkswagen Jetta and a Trailer setup as well. The test vehicle was instrumented with a tri-axis accelerometer inside the tire, encoder, Inertial Measurement Unit (IMU), and VBOX Racelogic Global Positioning System (GPS) based velocity measurement unit. For payload estimation, the data collected by the in-tire accelerometer was converted into frequency domain using Welch's method of averaging, followed by feature extraction. The extracted features were fed to a trained bagged trees model. Root mean squared error of 11% was observed on the test dataset. For velocity estimation, the data collected by the accelerometer was fed to a variational mode decomposition process. The extracted mode was converted to time-frequency domain using Hilbert transform and features for machine learning were extracted. A root mean squared error of 1.02kmph was observed on the trained dataset. A Gaussian process model was trained for this application. For unsprung mass acceleration estimation, the test vehicle was instrumented with an accelerometer near the wheel spindle as well. For this estimation problem, Convolutional neural networks (CNN) were used. The time-frequency spectrogram of x, y, and z axis data of the in-tire accelerometer were considered as the three color channels of an image. With this, an image of 224 x 224 x 3 dimensions was generated, which represented the time and frequency variation of data. These images were used for training the CNN and a 96.8% coefficient of correlation was obtained for this regression task. For the last wheel force estimation problem, the concept of training the images generated by overlapping time-frequency matrices was used and an accuracy of 90.1% was achieved. With these estimation of vehicle states, better control algorithms can be developed and deployed for better handling, safety and comfort of vehicles using data from only tri-axis accelerometer in the tire.