Kalman filter and neural network fusion for fault detection and recovery in satellite attitude estimation

Most satellite missions have extremely stringent requirements for attitude reliability. However, the Inertial Measurement Unit (IMU) in the Attitude Determination System (ADS), is susceptible to performance degradation in the space environment and can lead to mission failure. The proposed fault tolerance scheme includes two-layer fault detection with isolation and two-layered recovery. An Adaptive Unscented Kalman Filter (AUKF), quaternion estimator (QUEST) algorithm, and residual generator constitute the first layer of fault detection. At the same time, Radial Basis Function (RBF) neural networks and an adaptive complementary filter (ACF) make up the second layer of fault detection. These two fault detection layers aim to isolate and identify faults while decreasing the rate of false alarms. The AUKF and Fault Detection, Isolation, and Reconstruction (FDIR) residual generator make up the two-layered attitude recovery system. Compared to traditional fault-tolerant systems, this scheme solves the outlier problem of sensors and has higher accuracy. When one of the IMU sensors fails, it will be detected, and the proposed scheme can maintain accurate attitude estimation by leveraging a trained neural network. In addition, the secondary fault detection and isolation layer can minimize the rate of false alarms, meaning more reliable ADS for satellites.