Heterogeneous Ensemble Fault Detection for Autonomous Vehicle

Abstract

Fault detection is an important safety guarantee for the normal operation of autonomous vehicles. Common fault detection methods based on specific assumptions lead to the bias of detection results. Considering this problem, a heterogeneous ensemble unsupervised fault detection method for autonomous vehicle integrated navigation data is proposed in this paper. Firstly, the unsupervised fault detection method based on different assumptions is selected as the base learner to detect the fault of autonomous vehicles, and the fault probability is obtained. The voting method is used to fuse the detection results to further integrate the advantages of multiple base learners. The integrated navigation data are collected on the “Xinda” autonomous vehicle platform of Chang'an University for experimental verification. The results show that the heterogeneous ensemble method can well detect fault samples and reduce the proportion of misjudging normal samples as fault samples. In general, the ensemble method is superior to the single fault detection method without using the ensemble method, with a maximum increase of 12% in Precision score and 10% in F1 score. The proposed Heterogeneous ensemble autonomous vehicle fault detection method can effectively detect faults.