Statistical Machine Learning for Automata-Based Modelling of Simulink Autopilot System

Abstract

Aircraft autopilot controllers are often developed as Simulink models. Verifying such controllers is challenging because their behaviour is mainly observed through numeric simulation traces, and there is no explicit behavioural model showing how the controller responds under different conditions. In this thesis, we use automata learning to derive state machines from simulation data to support analysis and verification of a Simulink-based aircraft autopilot. A key difficulty is that standard automata learning assumes a finite alphabet, whereas the model’s inputs and outputs are numeric signals. We address this with an ML-enhanced passive automata-learning approach (MELA) that combines machine learning with automata learning. Feature-importance analysis selects informative signals, and decision tree based range abstraction partitions their numeric ranges into intervals. These intervals are then used to abstract the time-series traces before applying passive automata learning. We apply this pipeline to a closed-loop Simulink aircraft autopilot and learn Moore machines that capture its behaviour. Our evaluation compares MELA with a Manual baseline that uses the same data generation and learning procedures but relies on manually chosen variables and numeric abstractions. Across four learning sets and six abstraction configurations, MELA reduces the number of states and transitions in the learned automata by an average of 11.6% and improves accuracy by an average of 18.5% compared to the Manual baseline. The learned state machines support verification and exploration: by expressing the autopilot's requirements as temporal queries and evaluating them on the models, we can check whether these requirements are satisfied and identify behaviours that were not known in advance.