Anomaly Detection for UAV Flight Data via Reconstruction–Prediction Co-Learning Attention Network

With the increasing deployment of unmanned aerial vehicles (UAVs) in critical applications, ensuring the accuracy and reliability of their onboard measurement systems is paramount. The anomaly detection (AD) in UAV flight control systems plays a crucial role in identifying deviations in the sensor data that may indicate system malfunctions. Traditional AD methods often rely on single-task models, such as prediction or reconstruction, which struggle to detect anomalies in complex flight data. While multitask approaches aim to combine these tasks, they often suffer from task conflicts or fail to properly aggregate losses, limiting their ability to prioritize anomaly-sensitive patterns. To overcome this limitation, a novel reconstruction–prediction co-learning attention network (RPCA-Net) is proposed. RPCA-Net integrates both prediction and reconstruction tasks into a unified framework, utilizing convolutional neural networks (CNNs) for feature extraction and long short-term memory (LSTM) networks to capture temporal dependencies in the data. Attention mechanisms are incorporated for both tasks. The temporal attention is designed for the prediction task, while the channel attention is applied to the reconstruction task, enhancing the model’s ability to focus on key features and improving its performance in AD. Furthermore, a custom loss function that combines prediction error and reconstruction error is designed to increase the model’s sensitivity to anomalies. Experimental results with the practical data demonstrate that RPCA-Net achieves superior performance in detecting a wide range of flight parameter faults, outperforming contrast methods in terms of detection accuracy and robustness.