Feature and Extrapolation Aware Uncertainty Quantification for AI-based State Estimation in Automated Driving

State-estimation is an integral method for automated driving as the need for more measurement data for vehicle control increases, despite them not always being directly measurable. In the field of state estimation, AI-based algorithms are increasingly attracting interest. However, an uncertainty measure is pivotal to use AI-based state estimation for safety-critical applications. This paper presents the implementation of a vehicle state estimator based on a recurrent neural network and a novel method for uncertainty quantification. The uncertainty quantification method comprises the sequential evaluation of four parts: feature importance algorithms to remove input features lacking informative value, novelty detection filtering data beyond the range of the training data, and prediction of an uncertainty measure and confidence interval with Monte Carlo dropout. The performance of the proposed approach is demonstrated using AI-based state estimation of the vehicle sideslip angle based on the simulation data from a nonlinear two-track model. The results achieved imply that the novel method can provide a reliable confidence interval and successfully identify cases where the estimation and uncertainty quantification are not trustworthy.

I. Introduction

Automated driving requires an increased number of measurements to guarantee safety and reliability, and to improve driving comfort and efficiency. Due to missing measurement principles, space restrictions, or costly sensors, some quantities cannot be measured directly. One of these crucial quantities is the vehicle sideslip angle, which is an essential vehicle dynamics quantity to assess the stability of the driving behavior. Missing measurements can be substituted by integrating state estimators, also called virtual sensors, into the control system of the vehicle (cf. [1]). State estimators are usually implemented based on a physical model. However, if the system cannot be modeled sufficiently accurately the results can be unsatisfactory. In this case artificial intelligence (AI)-based methods, particularly Recurrent Neural Networks (RNNs), can be utilized for modeling dynamic systems. A major challenge in machine learning applications is to provide an uncertainty measure (UM) of the predictions. Its use is especially important in automated driving to ensure safety and reliability when using AI-based techniques. Fig. 1 shows the used methods in this work.