Anomaly Detection for Unmanned Aerial Vehicle Based on Sensor Fusion and Machine Learning: A Case Study

The proliferation of Unmanned Aerial Vehicles (UAVs) necessitates reliable anomaly detection systems to ensure flight safety. This study addresses the limitations of traditional threshold-based methods by proposing an innovative framework that integrates multi-sensor fusion with unsupervised machine learning via a custom-designed hardware board. A multidimensional dataset comprising gyroscope, accelerometer, GPS, battery, pressure, and velocity data was pre-processed using correlation analysis and RobustScaler. Four unsupervised algorithms - Isolation Forest (IF), Local Outlier Factor (LOF), One-Class Support Vector Machine (OCSVM), and Autoencoder (AE) - were comparatively evaluated. A contamination factor of 0,01 was utilized to define nominal flight behaviour across the models. Key evaluation criteria included reconstruction error-based analysis for the AE and statistical consensus across the integrated models. Results indicate that the Autoencoder provided the most stable and effective performance, especially in identifying complex anomalies through non-linear relationships. The proposed sensor fusion approach significantly improved detection accuracy compared to single-sensor analysis by compensating for individual sensor sensitivities. The findings demonstrate that this integrated method offers a more reliable solution for operational safety. The primary scientific contribution is the development of a novel sensor-fusion-driven unsupervised framework and the comprehensive comparative benchmarking of these algorithms using real-world data collected through dedicated hardware integration.