Neural Adaptive Fault Tolerant Control

This direction is merged into another direction: active fault-tolerant control. Its last version stays readable and citable here.

Version 1 of 128 papers38 conceptsfrozen 2026-10-04

Traditional control systems are often designed under the "nominal" assumption: that every motor, sensor, and actuator will function exactly as specified. In the real world, hardware degrades. Actuators lose effectiveness, sensors drift, and mechanical components fail abruptly. Neural Adaptive Fault-Tolerant Control (FTC) is a research direction that moves away from rigid models toward systems that can "learn" to compensate for their own broken parts in real-time.

From Diagnosis to Active Compensation

Historically, handling a hardware failure required a two-step process: first, a dedicated unit would detect and isolate the fault (Fault Detection and Isolation, or FDI), and then the controller would switch to a pre-planned backup strategy. However, this "hard-switching" can cause dangerous oscillations or "jitters" during the transition.

Recent work by Hongbing Xia suggests a more fluid approach: embedding fault compensation directly into the control law. Instead of waiting for a diagnostic report, the controller treats the fault as an unknown, time-varying disturbance. By using adaptive laws to counteract the energy of these unknown faults, systems can stabilize mechanical movements within milliseconds of a failure. This "fault-agnostic" architecture, as explored by Deqiang Cheng, allows for high-precision tracking even when the specific details of the fault are unknown.

Neural Networks as Universal Function Approximators

The core challenge of FTC is that a failure changes the underlying physics of the system. A drone with a damaged motor is no longer the same mathematical model it was at takeoff. Changyin Sun and Donald C. Wunsch II highlight how neural networks can serve as a "resilience layer" to bridge this gap.

Because neural networks are universal function approximators, they can be used to model the "missing" dynamics of a failed actuator on-the-fly. For instance, Lei Wang and colleagues utilize Radial Basis Function (RBF) neural networks to estimate the loss of effectiveness in wind turbine pitch actuators. By treating hardware degradation as a compensable variable—an "adaptive health factor"—the system can maintain constant power output even as its mechanical components wear down.

Performance Funnels and Stability Guarantees

A primary concern when introducing learning into control loops is safety. If a neural network is "learning" to compensate for a fault, how do we ensure it doesn't learn a behavior that leads to a crash?

Researchers like Fang Wang and Qun Zong address this using "performance funnels" or prescribed performance control. They define strict boundaries within which the system's error must stay. By combining adaptive back-stepping with error transformation techniques, they ensure that even during high-speed maneuvers in a hypersonic vehicle, the tracking error never exits these safe zones. Similarly, Yanjun Shen uses neural observers within a backstepping framework to ensure all system signals remain "semiglobally uniformly ultimately bounded," providing a mathematical guarantee of stability despite significant hardware failures.

Hybrid Architectures for Complex Environments

The most robust modern systems often combine high-level planning with low-level adaptive "shields." Jintasit Pravitra and Naira Hovakimyan demonstrated this by pairing Model Predictive Path Integral (MPPI) control with $L_1$ adaptive control. In this setup, the MPPI handles the complex task of racing a drone through obstacles, while the $L_1$ layer rapidly cancels out model-plant mismatches—such as a 40% loss in motor power.

This hybrid philosophy extends to extreme environments like fusion reactors and underwater exploration. Jaemin Seo showed that deep reinforcement learning can proactively avoid instabilities in plasma reactors, while Shangbin Jiao developed "soft-switching" systems for underwater vehicles that use Bayesian probability to transition smoothly between different fault scenarios.

By treating hardware failures not as catastrophic stops, but as dynamic uncertainties to be learned and canceled, these researchers are creating autonomous systems capable of "graceful degradation"—staying operational and safe even when the physical world takes its toll.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_neural_adaptive_fault_tolerant_control_a4258272,
  title = {Neural Adaptive Fault Tolerant Control},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/neural-adaptive-fault-tolerant-control/rdr_a4258272be2440899a9f8267f7ec1f61},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_neural_adaptive_fault_tolerant_control_a4258272_v1,
  title = {Neural Adaptive Fault Tolerant Control},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/neural-adaptive-fault-tolerant-control/rdr_a4258272be2440899a9f8267f7ec1f61/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}