Leveraging Latent Temporal Features for Robust Fault Detection and Isolation in Hexarotor UAVs

Abstract

Unmanned Aerial Vehicles (UAVs), commonly known as drones, have revolutionized various sectors including surveillance, agriculture, and disaster management due to their versatility and maneuverability. Hexacopter UAVs, equipped with six rotors, offer enhanced stability and payload capacity compared to their quadcopter counterparts. However, their operational effectiveness is contingent upon reliable fault detection and isolation mechanisms. In the dynamic operational contexts of hexacopter UAVs, potential faults such as motor failures, sensor malfunctions, or communication disruptions can lead to catastrophic consequences including loss of control, collisions, or data loss. Detecting and isolating these faults accurately and swiftly is imperative to ensure safe and efficient UAV operations. Our objective is to improve the reliability and accuracy of fault detection and isolation for a single motor failure in hexacopter UAVs. Commencing with a foundational exposition on hexacopter UAV dynamics, prevalent fault conditions, and classical machine learning classifiers, the research subsequently introduces LSTM networks and conducts a review of pertinent literature in fault detection and isolation, laying the groundwork for the proposed methodology. The principal contribution of this thesis revolves around the formulation of a fault detection and isolation paradigm that combines LSTM networks for latent temporal feature extraction with ensemble classifiers, notably Random Forests, aimed at enhancing fault detection efficacy. By harnessing the temporal intricacies inherent in UAV data using LSTM networks, the proposed model exhibits robust performance under measurement noise in fault detection and isolation tasks. The evaluation of the proposed approach encompasses comprehensive analyses conducted on synthetic and real-world datasets, encompassing examinations of noise resilience, fault detection and isolation timing, and the deployment on resource-constrained platforms such as the Raspberry Pi. Comparative assessments are conducted with benchmark models from both classical and deep learning domains, wherein our proposed approach demonstrates superior performance with an accuracy of 96.8% for stimulated datasets and 83.2% for real-world datasets. Furthermore, noise analysis highlights the resilience of our proposed method across varying noise intensities, underscoring its adaptability and robustness. The analysis conducted on synthetic data serves as a crucial validation step, instilling confidence in the model’s efficacy for real-world applications where data collection is inherently challenging. Subsequent experiments investigate fault detection and inference times, yielding insights into the temporal efficiency of the proposed approach. On an onboard microcontroller like the Raspberry Pi, the average inference time for our model is measured at 6.512 milliseconds, with additional statistical analyses providing further insights into performance metrics such as minimum and maximum algorithm run time and delay in prediction, alongside standard deviation. In conclusion, the thesis summarizes key findings and contributions, emphasizing the efficacy of the proposed approach in enhancing fault detection and isolation in hexacopter UAVs. Furthermore, it outlines potential avenues for future research, thereby underscoring the practical applicability of advanced machine learning techniques within real-world UAV systems. In essence, this thesis presents a novel fault detection and isolation framework that integrates LSTM networks with ensemble classifiers, thereby advancing fault detection and isolation capabilities in hexacopter UAVs operating in real-world scenarios