Evaluation of ML Algorithms for System Dynamics Identification of Aircraft Pressure Control System

Abstract

In this work we investigate the behaviour of aircraft air conditioning and internal pressure control system. This system is critical for aircraft operations, because detected faults lead to flight cancellation and maintenance, while in-flight faults can substantially worsen crew or passenger flight conditions. Our goal is to investigate the quality of fault detection and identification as well as computational complexity of novel data-driven algorithms for system identification. The main algorithm studied is Sparse Identification of Nonlinear Dynamics, which identifies nonlinear dynamical systems from data promoting sparsity in the solution. Developing a SINDy-based prototype of the health monitoring system is carried out as follows. We are using simulation data from an industrial ground facility, which is capable of high altitude flight simulation. We generate time series data of system's normal operation and different types of faults, such as wear of the most sensitive elements of the system (air gate, temperature and pressure gauges). Then, using the obtained data we build a Koopman operator approximation for each known mode of operation. During the testing phase prediction error is used as anomaly indicator. It is shown that online model update and its comparison with reference model can help in unknown fault isolation.