Avionic Main Fuel Pump Simulation and Fault-Diagnosis Benchmark
A high-fidelity co-simulation generates 10,000+ time-series samples for anomaly detection...
Researchers from Helmut Schmidt University have released a high-fidelity, physics-informed co-simulation of a common aircraft main fuel pump system, built in MATLAB/Simulink Simscape Fluids. The benchmark addresses the critical lack of training data for anomaly detection and diagnosis in cyber-physical systems, especially in aviation where data protection and partial observability limit real-world datasets. The simulation generates thousands of time-series samples with annotated health and fault modes, enabling supervised and unsupervised learning.
To demonstrate feasibility, the team applied an unsupervised Recurrent Variational Autoencoder (RNN-VAE) for anomaly detection and a SOM-VAE for operating mode discretization, both successfully separating healthy and faulty conditions. This work provides a standardized, reproducible testbed for developing and benchmarking AI-driven fault diagnosis systems for aircraft fuel systems, potentially improving safety and maintenance efficiency in aviation.
- High-fidelity physics-informed co-simulation of aircraft main fuel pump in MATLAB/Simulink Simscape Fluids
- Generates annotated time-series data with health and fault modes for anomaly detection training
- Validated with unsupervised RNN-VAE and SOM-VAE models achieving separation of healthy and faulty conditions
Why It Matters
Provides a much-needed benchmark for AI-based fault diagnosis in aviation, improving safety and reducing maintenance costs.