Sensor Fault Detection for UAVs Using Improved Self-Attention LSTM Network With Similarity Space Mapping

With the application of unmanned aerial vehicles (UAVs) in various scenarios, ensuring their flight safety has emerged as a major concern. Due to that sensors play an important role in flight control systems, etc., the detection of sensor faults is crucial for ensuring the flight safety of UAVs. Currently, fault detection (FD) methods using attention mechanisms generally rely on a single similarity calculation, thereby restricting the comprehensiveness and completeness of model information learning. This limitation can result in a decline in the effectiveness of FD. To address these issues, this article proposes a sensor FD method for UAVs using improved self-attention long short-term memory network (LSTM) with similarity space mapping (MISAL). Initially, the maximum information coefficient (MIC) is used to identify input variables related to the detected sensor, thereby mitigating the interference of irrelevant variables. Simultaneously, it is integrated with self-attention mechanism to achieve a similarity space mapping of the value matrix, enhancing the completeness of model information learning. Moreover, LSTM is introduced to further learn spatial and temporal correlations between data, enhancing the capability to detect faults. Finally, FD is implemented by comparing the absolute value of residual with the statistical threshold. Experimental validation using simulation data and real flight data demonstrates the excellent performance of the proposed method for the sensor FD of UAVs.