Researchers unveil MELA to auto-model complex cyber-physical systems
MELA cuts state-machine complexity by 49% while boosting accuracy by 42% in CPS modeling.
A team from the University of Ottawa and industry partner RabbitRun Technologies released MELA, a passive automata-learning framework that fuses statistical machine-learning with classical automata induction to extract behavioral models from numeric time-series data produced by cyber-physical systems (CPS).
MELA first applies information-theoretic variable selection and decision-tree-based range abstraction to convert continuous signals into a compact, interpretable set of symbols aligned with system states. Those symbols are then fed into an automata learner, yielding state machines that are on average 49.2 % smaller and 41.71 % more accurate than expertise-driven abstractions when evaluated on a commercial RabbitRun intrusion-detection appliance and an open aerospace autopilot benchmark. The learned models support system-level requirement verification and expose behaviors not explicitly documented in requirements.
- MELA is a passive automata-learning + ML hybrid that transforms raw CPS numeric traces into symbolic intervals aligned with states.
- Across RabbitRun’s intrusion-detection hardware and an aerospace autopilot benchmark, MELA reduced state-machine size by 49.2 % and boosted accuracy by 41.71 % versus expert-defined abstractions.
- Generated models enable faster formal verification and help engineers discover undocumented system behaviors.
Why It Matters
MELA accelerates formal verification of safety-critical CPS by automating model synthesis from live telemetry.