Homogeneous ensemble model built from artificial neural networks for fault detection in navigation systems

Abstract

In this paper, authors present a modern approach to the detection of malfunctioning sensory systems. The proposed solution is based on artificial neural networks. The application example uses a navigation system based on a 9-axis IMU (Inertial Measurement Unit), the signal fusion data is converted into quaternions. The form of quaternions is then analyzed along with sensor samples by an artificial neural network. If the network detects data processing inadequate to the pattern then we obtain information about the malfunction of a specific sensing axis from the sensors.The results compare fault detection capabilities using an ensemble structure built from three types of artificial neural networks: fully-connected, recurrent and convolutional. We provide a comprehensive analysis of all models; the proposed measures include RMSE (Root-Mean-Square Error), NRMSE (Normalized RMSE), t-SNE (t-distributed stochastic neighbor embedding) visualization, ROC (Receiver Operating Characteristic) curve, precision vs. recall curve, AUC (Area Under Curve) and F-score.