Multi-scale feature fusion network with temporal dynamic graphs for small-sample FW-UAV fault diagnosis

Abstract

With the extensive application of fixed-wing unmanned aerial vehicles, accurate fault diagnosis becomes crucial for flight safety and system reliability. Traditional fault diagnosis methods often require large datasets that are difficult to obtain in practice. To address this, we start with the spatio-temporal correlation characteristics and multi-dimensional heterogeneity of UAV flight data, and propose a multi-scale feature fusion with temporal dynamic graph network (MFFTD) that enables efficient fault diagnosis using limited UAV flight data. In the spatial dimension, a multi-scale residual convolutional design captures feature representations at various levels. Furthermore, the global temporal dynamic graph models the topological dependencies between the feature representations. In addition, we introduce long short-term memory networks to capture long-term dependencies in the temporal dimension. For cross-domain joint learning in both the temporal and spatial dimensions, we propose a multi-head feature fusion module based on mutual information to address the issue of heterogeneity imbalance between feature representations. Experiments on four public datasets demonstrated that MFFTD improves the detection accuracy by eight percentage points compared to the latest models under the 90 small-sample settings of multi-class tasks and significantly enhances the generalization capability, offering superior decision support for UAV fault diagnosis. The code and data will be available at https://github.com/17992/MFFTD.