Unmanned delivery aerial vehicles fault detection method based on enhanced spatiotemporal feature fusion framework and multi-head attention mechanism classifier

Large-scale multirotor unmanned delivery aerial vehicles (UDAVs) are increasingly used in low-altitude transport due to their high payload capacity and maneuverability. Fault detection is crucial for UDAV flight safety, as it identifies potential failures before they escalate, drawing widespread attention as a key technology. Despite its promising prospects, this technology faces the challenges of insufficient extraction of spatiotemporal features from flight data and difficulty in capturing fault information from the extracted features. To address these issues, an enhanced spatiotemporal feature extraction and fusion framework combined with a multi-head attention mechanism fault classifier (ESTEF-MH) is proposed for UDAV fault detection. First, the feature extraction and fusion framework introduces bidirectional long short-term memory and 2D convolutional neural network to extract features from the temporal and spatial dimensions. It then merges these features, enabling comprehensive capture of flight information. Second, a decoupled coordinate attention mechanism is used to enhance the spatiotemporal feature representation. Finally, a multi-head attention-based fault classifier is designed for efficient fault classification. The method is tested on four actuator faults and the normal conditions, with three ablation experiments and comparisons to classic methods. Furthermore, actual flight test was conducted with results showing 84.44% accuracy, validating the reliability and adaptability of our method in real-world operational environments. Compared to other methods, this approach shows superior performance in terms of key metrics, offering a reliable guarantee for flight safety.