1. Introduction
With the development of unmanned aerial vehicles (UAVs), their applications in civilian and military fields have expanded, including agriculture [
1], transportation [
2], and fire protection [
3]. However, as UAVs play an increasingly important role, their flight safety problems have become more prominent [
4]. Network attacks can lead to UAV failures, and physical component failures such as elevators and rudders can also affect UAV flight safety. For example, in June 2020, a US Air Force MQ-9 “Death” UAV crashed in Africa, causing a loss of USD 11.29 million [
5]. In February 2022, a DJI civilian UAV crashed out of control, resulting in a personal economic loss of up to 16,300 RMB [
6]. According to the Civil Aviation Administration of China, the number of registered UAVs in China alone has reached 8.3 million [
7]. Therefore, it is necessary to establish a UAV safety detection model to ensure the safety and reliability of UAV flights. Improving the flight safety of UAVs has become a major research topic in the field of UAVs. Currently, a common method to ensure UAV flight safety is to monitor UAV flight data for anomalies [
8]. Abnormal flight data indicates that the UAV may have hardware failure or misoperation, and timely identification of the cause of the failure can effectively prevent UAV flight accidents.
Figure 1 shows the main components of a typical UAV anomaly detection system.
UAV flight data is mainly extracted from attitude estimation data of different UAV sensors [
9,
10], which include the POS data and the system status (SS) data. These data enable the detection of UAV flight status. The POS data consists of a triple of values in the x, y, and z directions, while the SS data contains only a single value. Additionally, these data are closely related to UAV guidance, navigation, and control (GNC) [
11,
12]. The early UAV anomaly detection method was based on flight data rules; however, the rule-based anomaly detection method has a low detection performance [
13]. To better ensure the flight safety of UAVs, ML and deep learning methods have been introduced into the research field of UAV safety. The development of these methods has opened up new ideas for the research of UAV anomaly detection. However, the traditional anomaly detection method ignored the difference between POS data and SS data used to evaluate the flight status of UAVs in the frequency domain, resulting in the loss of some key feature information in-flight data. This limitation restricts the performance of UAV anomaly detection models. To address these problems, this paper proposes a method of extracting frequency domain information by setting timestamp slices and proposes a UAV anomaly detection model based on a multi-separable convolution neural network fusion method. It should be noted that this paper takes the time of UAV failure as the dividing point and does not consider the recovery process.
In the next part of this paper,
Section 2 describes the related research.
Section 3 introduces the processing method of the ALFA dataset [
14] and proposes the TS-MSCNN anomaly detection model.
Section 4 carries out experiments from various angles and analyzes the experimental results of binary and multi-class classification. The final section provides a summary and conclusion of this paper.
2. Related Works
This section provides a review of research related to UAV anomaly detection, covering rule-based algorithms and those based on ML and deep learning methods.
Regarding rule-based algorithms, Chen et al. [
15] investigated the impact of attackers’ behavior on the effectiveness of malware detection technology and proposed a specification-based intrusion detection system that showed effective detection with high probability and low false positives. Mitchell et al. [
16] considered seven threat models and proposed a specification-based intrusion detection system with specific adaptability and low runtime resource consumption. Sedjelmaci et al. [
17] studied four attacks—false information propagation, GPS deception, jamming, and black hole and gray hole attacks—and designed and implemented a new intrusion detection scheme with an efficient and lightweight response, which showed high detection rates, low false alarm rates, and low communication overhead. This scheme was also able to detect attacks well in situations involving many UAVs and attackers.
In terms of the UAV anomaly detection model based on traditional ML methods, Liu et al. [
18] proposed a real-time UAV anomaly detection method based on the KNN algorithm for the UAV flight sensor data stream in 2015, which has high efficiency and high accuracy. In 2016, Senouci et al. [
19] focused on the two main problems of intrusion detection and attacker pop-up in the UAV-assisted network. The Bayesian game model was used to balance the intrusion detection rate and intrusion detection resource consumption. This method achieved a high detection rate and a low false positive rate. In 2019, Keifour et al. [
20] released an initial version of the ALFA dataset [
13] and proposed a real-time UAV anomaly detection model using the least squares method. This method does not need to assume a specific aircraft model and can detect multiple types of faults and anomalies. In 2021, Shrestha et al. [
21] simulated a 5G network and UAV environment through the CSE-CIC-IDS-2018 network dataset, established a model for intrusion detection based on the ML algorithm, and also implemented the model based on ML into ground or satellite gateways. This research proves that the ML algorithm can be used to classify benign or malicious packets in UAV networks to enhance security.
However, some outliers can be difficult to detect using traditional machine learning (ML) techniques [
22]. To address this challenge, deep learning (DL) methods have been increasingly used to improve the detection accuracy of UAV anomalies, especially when processing high-dimensional UAV flight data. In 2021, Park et al. [
23] proposed a UAV anomaly detection model using a stacking autoencoder to address the limitations of the current rule-based model. This model mainly judges the normal and abnormal conditions of data through the loss of data reconstruction. The experimental results on different UAV data demonstrate the effectiveness of the proposed model. In 2022, Abu et al. [
24] proposed UAV intrusion detection models in homogeneous and heterogeneous UAV network environments based on a convolutional neural network (CNN) using three types of UAV WIFI data records. The final experimental results demonstrate the effectiveness of the proposed model. Dudukcu et al. [
25] utilized power consumption data and simple moving average data of the UAV battery sensor as the multivariate input of the time-domain convolution network to identify the anomaly of the instantaneous power consumption of the UAV battery. The simulation results show that the time-domain convolutional network can achieve good results in instantaneous power consumption prediction and anomaly detection when combining simple moving average data and UAV sensor data. In addition, some studies have explored the use of probability models, time series data, and data dimensions for anomaly detection, achieving effective results [
26,
27,
28], which have important implications for this study.
All of the previously mentioned methods have been successful in detecting anomalies, but they have not taken into account the differences between the POS data and SS data used to evaluate UAV flight status in the frequency domain. This has resulted in the loss of some key feature information in the flight data, which limits the improvement of anomaly detection model performance. The differences in the frequency domain can be seen in two aspects: first, the feature information amount of the POS data and the SS data in the frequency domain is inconsistent in the same time domain; second, the data structure is different. The feature of POS data in the frequency domain is triple, while SS data is a single value. When the amount of feature information is inconsistent, a feature vector with variable length is generated, which leads to the loss of key feature information in the model training process. Additionally, the difference in data structure causes POS data and SS data to lose some key information due to the confusion of feature information during the anomaly detection model’s feature extraction process.
To address the issues mentioned above, this paper proposes several solutions. Firstly, a specific timestamp size is set, and the frequency domain information of UAV data is divided and extracted to fuse key feature information, addressing the problem of inconsistency between POS data and SS data in the frequency domain. Secondly, POS and SS data are reconstructed into grayscale images. Lastly, the MSCNN is utilized to learn and fuse the key features of POS and SS data, overcoming the problem of key feature information loss caused by the structural differences between POS data and SS data. The following sections will provide a detailed description of these solutions.
5. Conclusions
UAV flight anomaly detection is a common safety measure to ensure the safety of UAV flights by identifying abnormal UAV flight data. However, the conventional anomaly detection model neglects the difference in POS data used to evaluate UAV flight status in the frequency domain, resulting in the loss of some crucial feature information that limits the improvement of the UAV anomaly detection model’s accuracy. Therefore, without considering the recoverable operation of UAV, this paper proposes a TS-MSCNN anomaly detection model based on timestamp slice and the MSCNN. Firstly, by setting a specific timestamp size, this paper extracts and fuses the frequency domain key features of POS data and SS data in the UAV flight log time domain. Then, the POS data and SS data are transformed into two-dimensional grayscale images to serve as the input data of the TS-SCNN model through data reconstruction. Finally, the TS-SCNN model accurately learns and fuses UAV grayscale image data features. The final experimental results demonstrate that the TS-SCNN model outperforms the comparative algorithm in the experimental results of binary classification and multi-classification, which validates the effectiveness of the TS-SCNN model proposed in this paper.
The deep learning model used in anomaly detection has a high time complexity, and UAVs typically have limited resources. Therefore, in future research, the authors of this paper will investigate a lightweight UAV anomaly detection model, taking into account both the timeliness of the anomaly detection model and the computational resources required by the model. The goal is to develop an anomaly detection model that can meet the resource constraints of UAV-embedded systems.