An enhanced CLKAN-RF framework for robust anomaly detection in unmanned aerial vehicle sensor data

Autonomous flight and real-time control of unmanned aerial vehicles (UAVs) critically rely on onboard sensors, which are susceptible to mechanical and environmental disruptions. Sensor anomalies pose substantial risks to UAV safety, emphasizing the importance of anomaly detection (AD) methods. However, AD remains challenging due to the scarcity of real anomaly data and the intricate spatiotemporal dependencies in sensor readings, often obscured by random noise and interference. This paper presents an enhanced framework based on one-dimensional convolutional neural network (1D CNN), long short-term memory network (LSTM) and Kolmogorov-Arnold network (KAN) with residual filtering (CLKAN-RF), utilizing multivariate sensor data without labeled information. First, a correlation analysis is employed to avoid the negative impact of irrelevant parameters on model training. Second, a multiple regression model is designed to comprehensively extract spatial-temporal relationships using 1D CNN and LSTM, while KAN is incorporated to non-linearly process the complex patterns and optimize the learned features with high accuracy. To address the issue of random noise, a bi-directional adaptive exponentially weighted moving average (Bi-AEWMA) scheme is introduced to smooth residuals, complemented by an adaptive dynamic thresholding mechanism to further enhance detection performance. Finally, extensive experiments on real UAV sensor data highlight the superiority of the proposed CLKAN-RF framework, which improves the true positive rate and overall accuracy by an average of 6.43 % and 7.63 %, respectively, while reducing the false positive rate by an average of 11.96 % compared to existing methods, demonstrating its potential application in UAV prognostics and health management.