The continual expansion of the range of applications for unmanned aerial vehicles (UAVs) is resulting in the development of more and more sophisticated systems. The greater the complexity of the UAV, the greater the likelihood that a component will fail. Due to the fact that drones often operate in close proximity to humans, the reliability of flying robots, which directly affects the level of safety, is becoming more important. This review article presents recent research works on fault detection on unmanned flying systems. They include papers published between January 2016 and August 2022. Web of Science and Google Scholar databases were used to search for articles. Terminology related to fault detection of unmanned aerial vehicles was used as keywords. The articles were analyzed, each paper was briefly summarized and the most important details concerning each of the described articles were summarized in the table.

In the modern technological advancements, Unmanned Aerial Vehicles (UAVs) have emerged across diverse applications. As UAVs evolve, fault diagnosis witnessed great advancements, with signal processing methodologies taking center stage. This paper presents an assessment of vibration-based signal processing techniques, focusing on Kalman filtering (KF) and Discrete Wavelet Transform (DWT) multiresolution analysis. Experimental evaluation of healthy and faulty states in a quadcopter, using an accelerometer, are presented. The determination of the 1024 Hz sampling frequency is facilitated through finite element analysis of 20 mode shapes. KF exhibits commendable performance, successfully segregating faulty and healthy peaks within an acceptable range. While the six-level multi-decomposition unveils good explanations for fluctuations eluding KF. Ultimately, both KF and DWT showcase high-performance capabilities in fault diagnosis. However, DWT shows superior assessment precision, uncovering intricate details and facilitating a holistic understanding of fault-related characteristics.

Unmanned aerial vehicles are being used increasingly in a variety of applications. They are more and more often operating in close proximity to people and equipment. This necessitates ensuring maximum stability and flight safety. A fundamental step to achieving this goal is timely and effective diagnosis of possible defects. Popular data-based methods require a large amount of data collected during flights in various conditions. This paper describes an open PADRE database of such measurements for the detection and classification of the most common faults - multirotor propeller failures. It presents the procedure of data acquisition, the structure of the repository and ways to use the various types of data contained therein. The repository enables research on drone fault detection to be undertaken without time-consuming preparation of measurement data. The database is available on GitHub at https://github.com/AeroLabPUT/UAV_measurement_data. The article also introduces new and universal quality indicators for evaluating classifiers with non-uniform parameters, are proposed. They allow comparison of methods tested for a variety of fault classes and with different processing times.

This article presents a new actuator fault detection and isolation method for multi rotor unmanned aerials (UAVs). The UAV community raises the need to develop a highly efficient classifier capable of early detection of failures and fully inde-pendent of other systems. This paper presents the entire process of preparing the method discussed and its implementation in the embedded system. The measurements of four accelerometers were digitally processed data, the main element of which was the frequency domain analysis. The feature vectors prepared in this way were used to train the artificial neural network. The network model has been implemented on a microcontroller. The tests were carried out using data collected during actual flight in various configurations of damaged propellers. The overall accuracy of the proposed method was 98.08 % without the presence of false alarms. The total processing time was also tested, demonstrating real-time classification capability onboard the autonomous flying robot. The efficiency results were compared with the random forest method.

This paper presents the robust Active Disturbance Rejection Control method in an aerodynamical object with two degrees of freedom. Full derivation of control algorithm equations and the settings values depending on the desired properties of the system, based on the dynamics of the second order oscillating object, have been shown. The equations of the control algorithm were discretized and its discrete implementation was proposed. The tests were carried out on the Inteco aerodynamical laboratory kit, whose mathematical model is strongly non-linear, has two degrees of freedom and crosscoupling. The quality of operation was assessed on the basis of time graphs as well as integral quality indices. The conducted research shows the advantage of the applied algorithm over PID control for the tested system. The presented control algorithm is also more portable – the settings are selected only on the basis of the object order and once selected are correct regardless of the mathematical model. The proposed discrete equations can be implemented using any programming language theoretically for any object, taking into account the order of its dynamics.

In the paper, a new machine learning technique to recognize movement patterns, is proposed. The efficient system designed for this purpose uses an artificial neural network (ANN) model implemented in a microcontroller to classify boxing punches. Artificial intelligence enables the processing of sophisticated and complex patterns, and the X-CUBE-AI package allows the use of these possibilities in portable microprocessor systems. The input data to the network are linear accelerations and angular velocities read from the sensor mounted on the boxer’s wrist. As a result, the categorization is done in real time. The extensive experiment was carried out for two groups with different levels of boxing skills. The developed model proved high efficiency in the identification of individual types of blows.