Fault Diagnosis of UAVs via Riemannian Manifold Representation Model and Generalized Learning Riemannian Space Quantization

With the increasing application of UAVs in various fields, fault diagnosis has become critical to ensuring their safe operation. However, existing model-based methods depend heavily on extensive prior knowledge, whereas datadriven methods suffer from high computational complexity, excessive resource consumption, and insufficient real-time performance due to large-scale neural network architectures. In order to address these issues, we present a fault diagnosis approach that combines a Riemannian Manifold Representation Model with Generalized Learning Riemannian Space Quantization (GLRSQ). First, using information geometry, flight state data are mapped onto a symmetric positive definite (SPD) manifold to construct a Riemannian representation model that preserves the intrinsic nonlinear structure of the data. Next, we compute the Riemannian geometric mean per data category to build prototypes, then apply stochastic Riemannian gradient descent to update them efficiently. During the diagnostic phase, only the geodesic distances between a test sample and each prototype are computed, with the label of the closest prototype serving as the diagnostic result. Experimental results demonstrate that, while maintaining high accuracy, the proposed method significantly reduces memory and CPU usage as well as inference time, thereby providing a low-complexity, resource-efficient, and highly real-time solution for UAV fault diagnosis.