Data-Driven Probabilistic Fault Detection and Identification via Density Flow Matching
New AI method uses neural vector fields to spot actuator and sensor faults with lower false alarms than traditional filters.
A team from Caltech and NASA's Jet Propulsion Laboratory has published a novel AI framework for Fault Detection and Identification (FDI) in complex systems like spacecraft. The paper, "Data-Driven Probabilistic Fault Detection and Identification via Density Flow Matching," tackles the critical challenge of spotting simultaneous actuator and sensor faults in nonlinear control systems. Instead of traditional point estimates, their method models the entire evolution of a system's state as a probability density flow. They use the 2-Wasserstein metric—a measure of distance between probability distributions—to quantify the separation between normal and faulty operation, deriving theoretical guarantees for detectability based on system parameters.
The core of their data-driven approach is a conditional flow-matching scheme that trains neural networks to learn the vector fields that govern how these probability densities propagate under various fault conditions. To generalize across a continuous range of fault magnitudes, the team incorporated Gaussian bridge interpolation and Feature-wise Linear Modulation (FiLM) conditioning. When tested on a spacecraft attitude control system, this density-based method demonstrated superior performance against a standard augmented Extended Kalman Filter (EKF) baseline. The results showed improved discrimination between different fault scenarios and, crucially, a lower false alarm rate, which is essential for safety and operational reliability in high-stakes aerospace applications.
- Uses 2-Wasserstein metric and stochastic contraction analysis to provide quantifiable guarantees for fault detectability.
- Employs a conditional flow-matching scheme with neural vector fields to learn density propagation under different faults.
- Outperformed an augmented Extended Kalman Filter in spacecraft tests, offering better scenario discrimination and lower false alarms.
Why It Matters
Provides more reliable, probabilistic fault detection for critical infrastructure like spacecraft and industrial systems, reducing costly false alarms.