Abstract
Flight data anomaly detection (AD) is essential for unmanned aerial vehicle (UAV) health management. Despite the current dominance of data-driven approaches, their effectiveness often requires sufficient data for model training. However, in practice, it is inevitable to face the situation of limited data, such as the high cost of data acquisition and the difficulty of collecting data in special scenarios, resulting in the performance degradation of the traditional data-driven methods with limited samples. This paper proposes an innovative data-driven approach leveraging transfer learning to detect and recover abnormal UAV flight data with limited samples through multi-source data fusion. First, a data-driven framework based on one-dimensional convolutional neural network and bi-directional long short-term memory (1D CNN-BiLSTM) with parameter selection and residual smoothing (1DCB-PSRS) is proposed. It employs the designed 1D CNN-BiLSTM prediction model for fully extracting spatiotemporal features of flight data, the maximum information coefficient (MIC) for parameter selection, and the exponentially weighted moving average (EWMA) for residual smoothing, thereby improving the AD and recovery performance. Second, multiple source domains with sufficient data are fused to pre-train the model to gain initialized parameters for the target domain. Then, the model is fine-tuned using limited training samples in the target domain through model-based transfer learning method and is evaluated using test data of the target domain. Finally, the effectiveness of the proposed method is verified on real UAV flight data.