Learning-enabled fault diagnosis based on abstract behaviors of the system

Fault diagnosis is essential for maintaining the safety and reliability of complex systems. An effective way to address this problem is to capture and analyze the abstract behavior of the system in the form of a sequence of events using “Discrete Event Systems” framework. However, such methods often suffer from high computational cost. In particular, the synthesis and continuous update of the fault diagnosis tools, named diagnosers, can become computationally intensive when applied to large-scale or dynamically evolving systems. To address these limitations, this paper proposes a learning-based fault diagnosis framework that combines discrete-event modeling with deep learning. The system behavior is first represented using deterministic finite automata, and a recurrent neural network is then trained to learn and reproduce the event-driven dynamics of the system. This hybrid approach reduces the complexity associated with redesigning the fault diagnosing system. The proposed framework is validated using a case study on a hybrid aircraft platform, demonstrating its effectiveness in accurately detecting and isolating faults with minimal computational cost. The results demonstrate that combining discrete-event modeling with machine learning can provide a scalable and computationally efficient framework for fault diagnosis in complex systems.