Multi-Class Sensor Fault Isolation for Vehicle Dynamics with Decision Trees

A decision tree algorithm for fault detection and isolation of additive sensor faults in an inertial measurement unit (IMU) is presented. The algorithm is trained and validated using real measurements with an extended Kalman filter (EKF) implemented on hardware in a production vehicle. The measurement data is cleaned from faulty measurements using different clustering algorithms , i.e., K-means, hierarchical clustering and density based clustering for applications with noise (DBSCAN). Then, decision tree models are trained based on the output estimates of an EKF using the original IMU data. Two different classification problems are considered. First, the binary classification problem distinguishing between the cases nominal and faulty is considered. Second, the multi-class classification problem distinguishing between errors in all three acceleration and angular velocity signals is considered.