Health Monitoring via Neural Networks

Monitoring of semi-autonomous systems increases situational awareness and is highly applicable to small remotely piloted aircraft. The complex dynamics of aircraft flight were abstracted by modeling the system with a neural network. The neural network accurately modeled the non-faulty dynamics for comparison to the actual system. Since, the system included unmodeled lag, the residual calculation needed to account for unmodeled lag. The method presented here decreased the effects of the lag on the residual. A ratio between the predicted and measured system outputs was used to understand the faults and understand how the future performance will be affected by the fault. A state machine abstraction offered a flexible fault-detection framework that can be adapted and designed to detect different faults without requiring a full dynamics model of each fault. Through using abstractions, the complex dynamics models required for fault detection were simplified and a state machine offered the flexibility to detect different suites of faults.