DFF-Net: An intelligent recognition network with high precision for unbalanced flight regimes of unmanned aerial vehicles

Conducting condition monitoring for the whole flight process is of great significance to enhance the security and reliability of unmanned aerial vehicles (UAV). Existing fault detection approaches only focus on one or several specific flight phases instead of the entire flight process. Precise flight regime recognition is the prerequisite to realize cross-stage failure detection and implement status monitoring for the entire flight process. Nevertheless, UAV flight data has several specific characteristics such as significant sequence temporality, multivariate spatial associations, and sample imbalance of different flight phases, which pose a great challenge to the data analysis-based flight regime recognition. In this study, an end-to-end intelligent recognition approach named deep feature fusion network (DFF-Net) is developed. In DFF-Net, a series of specialized designs have been adopted to meet the unique characteristics of UAV flight data. First, a multivariate convolutional channel attention module is designed to model the spatial connections of different flight parameters. Subsequently, a cross-scale convolutional Transformer detector is developed to excavate comprehensive temporal information. This detector is in a pyramidal architecture and can recognize flight regimes with different durations by fusing multi-level features. In this detector, cross-scale temporal embedding and convolutional multi-head self-attention (CMSA) are specifically designed and combined to model the interaction of multi-scale local features and long-term temporal information. To address the problem of long-tailed sample distribution of different flight regimes, a novel dynamic hybrid class-balanced loss function is proposed to guide the model learning by simultaneously considering the marginal distribution and effective samples of different classes. Finally, in the inference stage, a priority-based interval detection and merging operation is designed to correct the short-term marking errors induced by data fluctuations to further improve the identification performance. Experimental results on the simulation and real flight data demonstrate that our approach can realize precise identification of UAV flight regimes.