Intelligent fault diagnosis for multi-rotor unmanned aerial vehicles using extreme learning neuro-fuzzy systems

• Main Author: T Thanaraj
• Other Authors: Ng Bing Feng
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Engineering; Unmanned aerial vehicles; Intelligent fault diagnosis; Neuro-fuzzy systems; Extreme learning
• Online Access: https://hdl.handle.net/10356/175942
• Institution: Nanyang Technological University

Multi-rotor unmanned aerial vehicles (UAVs) have grown popular due to their versatility in applications such as videography, surveillance, and cargo transport. Despite their utility, these UAVs like many mechanical systems, are susceptible to component degradation. Particularly, faults in critical components, such as actuators, can jeopardise control, posing risks to property, people and aircraft in proximity. Aerial vehicles usually depend on hardware redundancies to manage or mitigate onboard faults. However, modern UAVs prioritise weight conservation, particularly small ones, which necessitates advanced fault diagnosis techniques using analytical means. Traditional fault diagnosis methods employ mathematical models, whose computational intensity can increase when greater nonlinearity is considered. Alternatively, intelligent fault diagnosis methods harness data-driven machine learning models, which study the observational data from UAV flights. Notably, neuro-fuzzy systems leverage both the superior learning capabilities of neural networks and the inferential strengths of fuzzy logic systems. However, they can be computationally intensive due to their gradient descent-based learning. Extreme learning neuro-fuzzy systems address this by offering computationally lighter and more efficient learning. Therefore, this thesis presents the development of an Intelligent Fault Diagnosis (IFD) model based on extreme learning for multi-rotor UAVs. Over the past decade, UAV fault diagnosis has attracted extensive research. This thesis offers a comprehensive review, organising these methods into traditional, intelligent and hybrid approaches. Following this, the thesis delves into traditional and extreme learning neuro-fuzzy systems, highlighting studies related to UAV fault diagnosis. Accurate, timely fault diagnosis is crucial for pilots and autonomous control systems to enact strategic mitigation actions, whether reconfiguring control for continued flight or executing a controlled landing. These decisions hinge on the multi-rotor UAV’s ability to remain controllable during actuator faults. Hence, a controllability analysis is presented to study the controllable scenarios for quad-rotor and hexa-rotor UAVs. Subsequently, this thesis assesses the proposed IFD model’s reliability using comprehensive numerical simulations. Initially, an ANFIS IFD model is developed as a binary classification model for actuator fault detection in a quad-rotor UAV. Upon applying it to a quad-rotor UAV, the model achieved high accuracy and F1 score of 92% and 95%, respectively. The IFD model was further extended into a multi-classification model for fault isolation capabilities. By integrating a model-based extended Kalman filter with the data-driven extreme learning neuro-fuzzy algorithms, a hybrid IFD model is generated. Three IFD models, R-EL-ANFIS, EL-ANFIS and Fuzzy-ELM, were developed, all of which demonstrated to train more than an order of magnitude faster than ANFIS. EL-ANFIS, while fast, underperformed in validation and sensitivity testing, likely due to overfitting. R-EL-ANFIS (EL-ANFIS with regularisation) achieves high performance, comparable to Fuzzy-ELM and ANFIS, albeit with slightly less computational efficiency. Fuzzy-ELM excelled with abrupt actuator faults, presenting the highest sensitivity with an average F1 score of 87.9% across faults with weaker magnitude and shorter duration. For incipient actuator faults, both Fuzzy-ELM and R-EL-ANFIS IFD models outperform the ANFIS IFD model in computational efficiency, being up to 5.56 times faster, while their fault isolation accuracy is marginally lower, at most 12.93% worse than that of ANFIS. These findings substantiate the reliability of extreme learning neuro-fuzzy-based IFD models for multi-rotor UAVs, underscoring their potential in real-world scenarios. The robustness of the IFD model is evaluated using hardware-in-the-loop (HITL) simulations with an autopilot controller. OS-Fuzzy-ELM (an online learning variant of Fuzzy-ELM) is employed for quad-rotor UAV actuator fault and GPS sensor fault isolation. The OS-Fuzzy-ELM IFD model presented improved accuracy and macro-averaged F1 score of 19.5% and 12.1%, respectively, as compared to SLFN. The Fuzzy-ELM IFD model is further adapted and refined to assess its robustness with a hexa-rotor UAV, for a multi-input-multi-output time-series classification problem. Results show that when compared to SVM and SLFN IFD models, SLFN exhibited superior post-fault isolation performance, which is observed to be due to overfitting. Fuzzy-ELM IFD model, however, excelled in generalisation, showcasing at least 23.1% and 34.4% lower weighted classification error rates than that of SLFN and SVM. Building on the robustness demonstrated by the Fuzzy-ELM IFD model from HITL simulations, the focus transitions to its practical applications in active fault tolerant control. The critical fault information that the IFD model generates can facilitate the active fault tolerant controller’s ability to dynamically reconfigure control following an actuator fault. This strategy utilises a gain-scheduled PID (GS-PID) controller for control reconfiguration, employed on a quad-rotor UAV. In a hover scenario, the lateral deviation and recovery time were reduced by up to 73.1% and 25.0% for the altitude controller, respectively, compared to a nominal controller. In a trajectory tracking scenario, the lateral deviation and altitude drop were reduced by up to 74.1% and 90.0%, respectively. The findings from these simulations substantiate the critical role an IFD model plays in an active fault tolerant controller, promptly identifying faults, suppressing large deviations and subsequently, facilitating quick recovery. In summary, this research demonstrates that the extreme learning neuro-fuzzy-based IFD model significantly enhances computational efficiency while preserving high actuator fault isolation performance in multi-rotor UAVs compared to gradient descent-based methods. Through numerical simulations, the Fuzzy-ELM model exhibited superior sensitivity and efficiency in diagnosing abrupt and incipient actuator faults. In hardware-in-the-loop simulations, Fuzzy-ELM outperformed conventional intelligent methods, such as SLFN and SVM, showcasing superior generalisation for temporal actuator fault isolation. Crucially, a Fuzzy-ELM-enabled active fault tolerant control displayed significantly faster recovery times and enhanced deviation suppression during hover and trajectory tracking scenarios. The incorporation of an IFD model into multi-rotor UAV systems, either onboard or offboard, enables analytical means of redundancy that can either complement or replace traditional hardware redundancy approaches. This improves the pilot or autonomous controller’s mitigation action towards actuator faults, thereby ensuring the safety and efficiency of multi-rotor flight operations.