Abstract
In safety-critical robotics, the ability to detect and identify actuator failures in real-time is paramount. Traditional model-based Fault Detection and Identification (FDI) methods often struggle with model uncertainties and environmental perturbations. This paper presents a model-free FDI framework for quadrotors that utilizes Long Short-Term Memory (LSTM) networks. By leveraging the inherent dynamic symmetry of the quadrotor platform, the proposed method allows a network trained on a single motor's failure data to generalize across all actuators. Experimental results demonstrate over 90% prediction accuracy and significant robustness to physical parameter variations and distribution shifts in control strategies.
Introduction
Unmanned Aerial Vehicles (UAVs), particularly quadrotors, are increasingly deployed in complex environments where actuator faults—such as a motor losing power or becoming unresponsive—can lead to catastrophic failure. Maintaining operational safety requires a robust Fault Detection and Identification (FDI) system capable of determining if a fault has occurred, isolating the specific faulty motor, and estimating the magnitude of the power loss.
Historically, FDI has relied on model-based approaches that compare real-time behavior against a mathematical model. These systems flag "residuals" when the physical drone deviates from the model's predictions. However, these methods are highly sensitive to model mismatch; changes in payload, mechanical wear, or wind can trigger false positives. Conversely, while model-free neural network approaches offer potential robustness, they typically require exhaustive datasets covering every possible failure mode for every motor, leading to high computational costs and training complexity.
Methodology
LSTM Architecture
The researchers developed a model-free framework centered on Long Short-Term Memory (LSTM) networks. LSTMs are uniquely suited for this task due to their ability to process time-series data and maintain temporal dependencies. The network receives a sliding window of system outputs—specifically position and attitude—alongside commanded control inputs. From this data, the network predicts a "fault vector" ($\Theta$) representing the health of the actuators.
Exploiting Dynamic Symmetry
The core innovation of this research is the application of dynamic symmetry to the training process. Because a quadrotor is physically symmetric, a failure in one motor is dynamically equivalent to a failure in any other motor, provided the coordinate frame is appropriately rotated.
Instead of training the model on failures for all four motors, the researchers trained the LSTM exclusively on data from Motor #2. During deployment, the system processes incoming data through four distinct rotational transformations. The transformation that results in the lowest fault value effectively identifies which motor is failing. This approach reduces the required training data by 75% and encourages the model to learn the fundamental physics of failure rather than memorizing specific motor indices.
Training and Evaluation
The model was trained using 10,000 trajectories generated by a standard Linear-Quadratic Regulator (LQR) controller. These trajectories included varying degrees of "loss of effectiveness" faults, ranging from partial power loss (e.g., 50%) to complete failure (0%). To test the system's ability to handle "distribution shifts," the researchers evaluated the model using trajectories generated by a Control Barrier Function (CBF)-based safety controller—a control logic the network never encountered during training.
Results and Discussion
Accuracy and Architecture Comparison
The LSTM-based FDI achieved a prediction accuracy exceeding 90% across various fault scenarios. When compared to a standard Multi-Layer Perceptron (MLP), the LSTM showed superior performance. The MLP’s inability to accurately identify faults underscored the necessity of temporal memory when interpreting the dynamic sequences of a failing UAV.
Robustness to Model Mismatch
A critical test of the framework involved a "perturbed dataset" where the quadrotor's physical parameters—including mass, inertia, and thrust coefficients—were altered by over 40%. While traditional model-based residual methods saw a significant decline in reliability under these conditions, the model-free LSTM remained stable. This suggests that the network learns the underlying relationship between control inputs and resulting motion, making it highly resilient to the "dirty" environments of the real world.
Generalization Across Controllers
The framework demonstrated impressive generalization by maintaining high accuracy when the drone was operated using a CBF-based controller. This indicates that the FDI system is not overfit to a specific control law but has instead captured the general dynamics of the vehicle's response to actuator failure.
Conclusion and Future Work
This research demonstrates that combining deep learning with physical symmetries can produce FDI systems that are both efficient and robust. By training on a single motor and generalizing via rotation, the framework achieves high accuracy with minimal data requirements.
Future research will address the challenge of simultaneous multi-motor failures and expand the fault model to include more complex, erratic signals beyond simple loss of effectiveness. Additionally, the team plans to adapt this symmetry-based approach for multi-agent systems, where identifying faulty neighbors is essential for the safety of the entire swarm.
Summary updated