Data-driven model for accommodation of faulty angle of attack sensor measurements in fixed winged aircraft

Abstract

On March 10, 2019, Ethiopian Airlines’ Boeing 737-8 MAX nose-dived and crashed shortly after takeoff in the south-east of Addis Ababa, near Ejere Town. The cause of the crash was a faulty angle of attack (AOA) sensor. In the past, many aircraft accidents were associated with faulty AOA sensors. The literature uses triplex and duplex AOA sensor fault detection, isolation, and accommodation (SFDIA). The triplex voting mechanism uses three AOA sensors. A critical problem with the triplex method is that the consolidated value inherently depends on the sensor measurements. This problem makes fault detection and isolation susceptible to simultaneous failure. The duplex fault detection and isolation uses two sensors, reducing cost and providing lower protection. We propose using two AOA sensors and a virtual AOA sensor for faulty AOA SFDIA. The proposed faulty AOA sensor detection and isolation algorithm is based on conventional residual analysis with a fixed threshold for faulty AOA sensor detection and isolation. The virtual sensor is a data-driven model based on a recurrent neural network (RNN) for AOA accommodation. We use a combination of simple RNN (sRNN) and gated recurrent units (GRU). The model aims to effectively use the encoder–decoder behavior of the GRU for better AOA accommodation in the case of faulty AOA measurement, faulty velocity measurement, and faulty pitch rate measurement. Test results show that the proposed method can detect, isolate, and accommodate faulty AOA sensors with a lower number of false alarms than a model that uses only long-term memory (LSTM).