Unmanned Aerial Vehicle Flight Data Anomaly Detection Based on Multirate-Aware LSTM

Hui Lu, Zan Wang, Yuhui Shi

Abstract: With the widespread use of unmanned aerial vehicle (UAV), health status monitoring is critical for flight safety, making anomaly detection a focus of attention. In general, UAV flight data are multirate multivariate time series (MR-MTS) acquired by multiple sensors at different sampling rates, while the existing methods are mainly designed for multivariate time series (MTS). Moreover, traditional methods that tune MR-MTS to MTS may artificially change naturally occurring temporal dependencies. To address these challenges, a novel multirate-aware long short-term memory (LSTM), called MRA-LSTM, is proposed for UAV flight data anomaly detection. First, without manual tuning, a multirate-aware structure is proposed to directly model MR-MTS UAV flight data. Second, to avoid the higher rate variates “masking” the multiple temporal dependencies present in the lower rate variates, a multiscale update mechanism is proposed to trade off information entropy from different rates. Finally, a hierarchical state excitation (HSE) is proposed to adaptively control the importance of each inferred compressed representation to the output module. The experimental results demonstrate the effectiveness of the proposed method, as well as its superiority to other state-of-the-art peer competitors.