An Auto-Encoder enabled Fault Detection and Isolation Scheme for enabling a Multi-Sensorial Distributed Pose Estimation

This article proposes a novel Auto-encoder based fault detection and isolation framework approach for supporting the operation of a novel multi-sensorial distributed pose estimations scheme. The proposed work detects weak and strong time-dependent anomalies in a decentralised fusion approach from the initial estimation layer. As it will be presented, at the end of the learning phase, the neural network-based auto-encoders provide synthetic actual position and orientation of the robotic system, based on the statistics of the learning data. As a result, the square error between the output and input signal of the auto-encoder can yield the actual outlier with reasonable success. On the other hand, an Extended Kalman Filter (EKF) based fault detection method has been introduced in this article, which consists of a set of judiciously designed EKF acts as filter assembly. Based on innovation obtained from each of the EKF an innovative detection logic is proposed to identify the outlier in sensor measurement autonomously at the appropriate time samples. Based on the degree of accuracy of detecting the anomaly, the estimated signal is accepted or rejected for each time sample in the second layer of the fusion architecture. Moreover, we will introduce two outlier detection methods for the demonstration purposes and outline a comparative study using experimental data from a micro aerial vehicle. An extensive analysis with supporting results demonstrate these two methods' effectiveness and accuracy.