Drones have become an indispensable component of modern equipment systems, and as the number of components in these systems continues to increase, so does their complexity. Consequently, the requirements for system quality—particularly reliability, maintainability, and functionality—are also becoming increasingly stringent. Sensor data fusion, fault detection, fault-tolerant estimation, and fault-tolerant control are all important means of ensuring UAV safety. Among these methods, researchers tend to place particular emphasis on fault detection, which can be broadly divided into two categories: model-based approaches and data-driven approache.
1. Introduction
The traditional way to detect drone failures is through model-based approaches that use mathematical models to analyze UAV safety. These models can be further subdivided into qualitative or quantitative methods (Table 1). However, because model-based approaches often focus only on specific components, they may not always detect failures in other parts of the system. This means that, if an unmanned aerial system suddenly fails due to its complexity, resulting in higher costs, multiple model pairs may need to be deployed. Therefore, in this research, faults at the system level are mainly studied. Building a fault mode database using a large amount of fault data can facilitate the analysis of fault modes and fault propagation processes at the system level. The analysis process can help fault diagnosis personnel quickly locate the problem and save diagnostic time. At present, the research on failure mode databases lacks depth. Drone systems are complex systems with multiple tasks and configurable elements. There are both common and unique failure modes among different systems. The evaluation center where the authors of the current study work has a large body of failure data. However, due to the lack of organization and induction of these data, there has been no in-depth analysis of the mechanisms of common faults. How to use historical fault data to generate fault mode databases and utilize these databases to better achieve fault detection is a problem worth studying.
Table 1. Model-based approaches to Fault Detection.
| Authors |
Year |
Fault Location |
Used Methods |
| D’Amato et al. [8] |
2021 |
sensor |
sensor fusion algorithm based on particle filter |
| Fu et al. [9] |
2021 |
sensor, actuator |
adaptive observer to realize sensor and brake fault detection |
| Maqsood et al. [10] |
2021 |
sensor |
sensor fault detection and isolation method for quadrotor aircraft |
| Miao et al. [11] |
2021 |
sensor |
adaptive nonlinear proportional integral (PI) observer |
| Nejati et al. [12] |
2021 |
actuator |
three-level central difference Kalman filter (RThSCDKF) |
| Sun et al. [13] |
2021 |
pitot tube |
two identical synthetic air data systems |
| Cao et al. [14] |
2022 |
actuator |
interval observer and extended state observer |
| Gai et al. [15] |
2022 |
elevator, event-triggered
intervals |
dynamic event-triggered Hi/H∞ optimization method |
| Lin et al. [16] |
2022 |
sensor |
crash probability density (CPD) |
Another typical method used for drone fault detection is the data-based method. In recent years, the reduction in the difficulty of data collection and analysis has driven the development of the field of data analysis. With the support of various high-speed processors, the means of mining required information from data have become increasingly powerful. Similarly, data-based methods can be divided into two categories: quantitative and qualitative. Quantitative methods include those based on statistical theory and neural networks. Qualitative methods typically include expert systems, fuzzy logic, pattern recognition, and qualitative trend analysis. Table 2 lists some typical studies, including the location of fault detection and the methods used.
Table 2. Data-driven approaches to Fault Detection.
| Authors |
Year |
Fault Location |
Used Methods |
| Altinors et al. [17] |
2021 |
motor, propeller |
method combining statistical feature extraction and machine learning |
| Park et al. [18] |
2021 |
GPS spoofing, DoS attack, rudder, elevator, aileron, engine |
unsupervised learning |
| Souza et al. [19] |
2021 |
motor |
signal analysis technique based on chaos |
| Zheng et al. [20] |
2022 |
actuator, aero engine, equipment |
data assembly annotation method based on flight data and BIT records |
| Cabahug et al. [21] |
2022 |
actuator |
k-means clustering algorithm |
Working out how to optimize the original data or improve the relevant algorithms has also become a major research trend in recent years. Zhang et al. [
22] proposed a robust deformed denoising CNN (RDDCNN) to solve the offset pixels of feature maps from noisy images. This approach can effectively improve data quality and support data-based UAV fault diagnosis. Tian et al. [
23] proposed a multi-stage image denoising CNN with a wavelet transform (MWDCNN) via three stages. A dynamic convolution is used in a CNN to address the limitations in depth and width of lightweight CNNs to achieve good denoising performance.
The performance of data-based methods mainly depends on the quality of the original data. The fault data of drones is often multi-level linkage and cross-influence data, which is not conducive to analyzing complex types of faults. In addition, data-based methods are often offline and difficult to analyze in real time. This method cannot support the real-time safety monitoring of drones. It is more of a means of analysis after accidents, and cannot detect and handle abnormal situations during the operation of drones. Therefore, this method cannot guarantee the safety of drones during operation.
2. Failure Mode Databases
In recent years, failure mode databases have become a research hot spot in database system engineering, knowledge engineering, and other fields, and this approach has been applied in many fields. Failure mode databases are structured, easy-to-use, comprehensive, and organized failure clusters—a collection of interrelated patterns stored, organized, managed, and used in computer memory in a certain way in order to solve the needs of problems in some fields. These failure modes include theoretical knowledge related to the domain, factual data, and heuristic knowledge derived from expert experience. UAV failure mode databases are of great significance to the study of reliability trends. How to form a real scientific failure mode database is one of the key contents of the research.
A common solution to this problem is to extract UAV failure patterns from the records contained in a UAV accident database by cleaning, parsing, and normalizing the fault data. For example, by obtaining accident records from different countries and analyzing the causes of accidents. Suitable data sources include:
-
NASA and the FAA Aviation Safety Reporting System (ASRS) [
24];
-
U.S. Aviation Safety Communique (SAFECOM) data [
25];
-
Australian Transport Safety Bureau (ATSB) aviation safety investigations and reports [
26];
-
UK Air Accident Investigation Branch (AAIB) data [
27].
The attribute fields in the database should include the data source, accident year, month, location, flight phase, aircraft model, flight hour history, flight time of the mission prior to the accident, altitude, weather, mission being performed, root cause, and accident outcome. At present, the research on UAV accident databases mainly focuses on the collection and analysis of accident data. Wild et al. [
28] collected and analyzed 152 incident and accident records, but only 40 records deal with UAV accidents, and the data date back to 2006–2015. More recently, a total of 160 accident reports from 2015 to 2020 were collected and documented in an integrated database [
29].
3. Runtime Verification
Runtime verification techniques are formal verification techniques, which use logical formulas to describe properties and transform them into formal structures. Runtime verification tends to check that there is a poor path, that is, up to the current time, whether the system’s running path is within safety parameters. After being proposed, this method has received continuous attention from academia and industry. One of its characteristics is that it plays a role in the system software—it is in the real run environment used to monitor the system—so it can find potential defects that the traditional software testing method misses. It complements classical verification techniques (for example, theorem proving and model checking) and provides a more practical method for the verification of system running trajectories. At the cost of limited execution coverage, runtime validation provides accurate information about the runtime behavior of the monitored system for subsequent analysis. The object of action can be a software system, a hardware or information physical system, a sensor network, or any system where dynamic behavior can be observed. The following is a brief overview of some recent research on runtime validation.
Abbas et al. [
30] introduced private runtime verification. Liu et al. [
31] introduced the idea of incremental verification and proposed an incremental probabilistic model-checking method based on heuristics. Y-Rozier et al. [
32] explore the connection and difference between simulation and runtime verification. Stockmann et al. [
33] proposed architecture runtime verification. Teixeira et al. [
34] studied the properties of a runtime verification-specified API and wrote a minimalist specification language. Bicevskis et al. [
35] discuss data quality checking during the execution of a business process by using runtime verification. Lee et al. [
36] developed an effective model-checking method for IoT system operation verification. Legunsen et al. [
37] proposed an aware runtime verification technique. Geng et al. [
38] introduced a new smart tagging method to verify the completion of tasks involving one-to-many and many-to-one dependencies. Ring et al. [
39] significantly reduced the size of the state space of the verification process and reduced the complexity of the detection process. Ye et al. [
40] proposed an adaptive runtime verification method based on multi-agent systems. Tsigkanos et al. [
41] proposed a service-oriented software architecture and technical framework to support the runtime verification of decentralized edge-dense systems. Hu et al. [
42] proposed a runtime verification method based on the Robotic Operating System (ROS). Tracy et al. [
43] introduced an open-source framework to achieve efficient and high-performance runtime monitoring. Ye et al. [
44] developed a new approach based on approximate computation to achieve sufficiently fast and accurate repeated execution of security verification. Kong et al. [
45] proposed a method aimed at monitoring traces that reveal the runtime state of the software. Jung et al. [
46] proposed an automatic runtime prioritization method based on a classification tree. Miranda et al. [
47] proposed a method to automatically detect whether the errors reported by the monitor are real errors. The authors of [
48] proposed the concept of UAV safe operation monitoring and the operational limits to be monitored. A prominent example of such operational limitations is geofencing. Geofencing uses virtual fencing to prevent drones from entering restricted airspace. Felipe et al. [
49] proposed a solution based on stream runtime verification, which offers a high-level declarative language to describe sophisticated monitors together with guarantees on execution time and memory usage. They showed how monitors can be combined with temporal planning not only to monitor assumptions but also to support mitigation and remediation in UAV missions. Bonnah et al. [
50] presented a rewriting-based algorithm for runtime monitoring of safety requirements expressed in TWTL for specifying time-bounded serial tasks.
Runtime verification is widely used in academia and industry to ensure the reliability and security of the system, whether it is before deployment, testing, verifying, debugging, or after deployment. However, the setting of monitoring conditions still lacks a solid basis. Therefore, this paper intends to analyze the failure mechanism of UAVs through the UAV failure mode database to extract the monitoring conditions. According to the above monitoring conditions, the UAV is monitored in real time to improve its reliability.
This entry is adapted from the peer-reviewed paper 10.3390/drones7080486
Encyclopedia
