Closing the Loop Inside Neural Networks: Causality-Guided Layer Adaptation for Fault Recovery Control

Overview

As a Postdoctoral Scholar at Caltech, I developed adaptive neural controllers for nonlinear systems subject to actuator loss-of-effectiveness (LOE) faults.

Key Idea

We introduced a two-phase learning architecture:

1. Offline phase:
   Causal inference is used to identify the most causally relevant layers of a deep neural network-based fault compensator.
2. Online phase:
   Only the selected layer is updated using a Lyapunov-based gradient law to guarantee bounded tracking error.

Key Contributions

• First neural control framework with causality-guided internal adaptation
• Avoids unnecessary full-network retraining during fault recovery
• Preserves formal stability guarantees during online learning

Case Study

Fault-recovery control of a 3-axis spacecraft attitude control system under actuator faults.

This work demonstrates how learning can be embedded inside feedback loops without sacrificing safety.