Parity Vector Compensation for Multi-Sensor Fault Detection

A scheme based on neural network is presented to improve the fault detection and isolation (FDI) performance of a redundant strapdown inertial measurement unit. The neural network is trained to eliminate the effects of input axis misalignment, scale factor error and biases on parity vector. One advantage of the proposed technique over other compensation algorithm is that it does not require dynamic equation of error states and statistics of noise. When the neural network is trained on line with delayed samples, the compensated decision function is sensitive to slow-varying type fault. The simulation results show that the technique could significantly improve the FDI performance of the skewed inertial sensor sets.