A Dual-Modality Fusion Framework with Adaptive Weighting for Intelligent Fault Diagnosis of UAVs
TL;DR
A temporal and time–frequency dual-modality weighted fusion network (TTF-DWFNet) for UAV fault diagnosis and achieves state-of-the-art accuracy, outperforming existing CNN-based and Transformer-based methods.
Abstract
The rapid expansion of the low-altitude economy places unprecedented demands on the safety and reliability of unmanned aerial vehicles (UAVs). To address the limitations of single-modal feature extraction in capturing heterogeneous complementary information, this paper proposes a temporal and time–frequency dual-modality weighted fusion network (TTF-DWFNet) for UAV fault diagnosis. The proposed framework employs a parallel dual-branch architecture: a temporal modality branch utilizes a one-dimensional Transformer with a broadcast self-attention mechanism to model global temporal dependencies from raw vibration data, while a time–frequency modality branch transforms the data into spectrograms and leverages a two-dimensional convolutional neural network (CNN) to extract local texture features. Furthermore, an adaptive weighted fusion module is designed to automatically learn the discriminative contribution weights of the heterogeneous features at the channel level, enabling effective feature integration. Extensive experiments on two public datasets show that TTF-DWFNet achieves state-of-the-art accuracy of 96.47% and 92.76%, outperforming existing CNN-based and Transformer-based methods. The proposed TTF-DWFNet architecture has great practical application potential.