Physics-Based Residual Learning for Actuator Fault Identification and Health Monitoring in Quadrotors

The operational safety of multi-rotor Unmanned Aerial Vehicles (UAVs) is affected by actuator performance degradation. While sensor redundancy is standard practice in UAV architectures, incorporating physically redundant actuators, such as additional motors or drives, is largely impractical due to strict weight and cost constraints. Most existing studies approach actuator fault identification merely as a binary classification problem. However, evaluating the Loss of Effectiveness (LOE) to estimate the fault magnitude is essential to ensure operational continuity. To address this gap, this paper introduces a Physics-Based Residual Learning (PBRL) framework designed to isolate and quantify actuator anomalies. The proposed architecture pairs algebraic dynamic inversion with a Deep Neural Network (DNN). By learning the residual errors left unexplained by the physical model, the DNN effectively compensates for unmodeled aerodynamic effects, manufacturing-induced motor biases, and battery voltage depletion. The framework is validated using real flight data from a Parrot Minidrone and rigorously compared against two standalone baselines: a pure physical model and purely data-driven DNNs. The experimental results reveal distinct limitations of these standalone approaches: the pure physical model exhibits battery-induced drift and biases caused by unmodeled structural asymmetries, which can trigger false-positive alarms, whereas the pure DNN exhibits greater high-frequency fluctuations and, as a black-box model, provides limited physical interpretability. By combining the physical model with residual learning, the proposed PBRL framework reduces systematic modeling errors and high-frequency estimation variations, enabling accurate and stable LOE estimation.