Model predictive control-based fault detection and reconstruction algorithm for longitudinal control of autonomous driving vehicle using multi-sliding mode observer

Abstract

This paper presents a model predictive control-based fault detection and reconstruction algorithm for longitudinal control of autonomous driving using a multi-sliding mode observer. In order to secure the safe longitudinal control of a vehicle, a numbers of factors must be ensured, such as the reliability of the longitudinal information, the data on the forward object from the environment sensor, and the acceleration of the ego vehicle. Thus, we propose a reasonable failure detection scheme for the acceleration signal of the host vehicle and the relative values of the front object of the radar. In order to identify the faults of the radar and the vehicle acceleration sensor related to the automated longitudinal control, the multiple sliding mode observer and prediction of model predictive control (MPC) algorithm are applied. The relative acceleration is reconstructed by applying a sliding mode observer (SMO) with clearance and relative speed measurements. The upper and lower limits of longitudinal acceleration were computed by analyzing human driving data under the preceding vehicle and reconstructed acceleration. A proper acceleration range can be defined precisely based on several reconstructed upper and lower bounds by using a multiple sliding mode observer with stored prediction data of relative values, making it possible to effectively identify the fault of the host vehicle's acceleration sensor. By applying MPC for this study, optimal control input and prediction of relative states can be obtained that are more reasonable than those using the linear prediction model. The proposed fault detection algorithm can identify the abnormal state of the environment sensors by using the accumulated past sensor data. By comparing the stored prediction of relative states with the stored data on current states for a given period, the signal faults of the longitudinal target information can be detected from environment sensors. With these fault indices of states, the final fault diagnoses of sensors can be determined by assessing confidence through statistical analysis of 27 sets of normal driving data. In order to obtain a reasonable performance evaluation, this study uses actual driving data and a 3D full vehicle model constructed in the MATLAB/Simulink environment. The test results reveal that the proposed algorithm can successfully detect the fault of the radar and acceleration sensor of the automated driving vehicle.