A Labeled UAV Fault Diagnosis Dataset from Real Flight Experiments for AI-Based Anomaly Detection

This study introduces a real-world dataset designed to support fault diagnosis in unmanned aerial vehicles (UAVs) through artificial intelligence–based anomaly detection techniques. To construct the dataset, common failure types were identified through analysis of historical UAV accident reports, followed by the design and execution of flight tests simulating faults in structural, propulsion, and sensor subsystems. A total of 25 flight trials were conducted under both nominal and fault-induced conditions. During each test, multimodal flight data—including positioning, attitude, barometric pressure, control signals, and vibration—were recorded.The collected data were preprocessed and categorized into three distinct segments: original, normal, and error. This structured dataset supports both supervised and unsupervised learning models. To validate its effectiveness, two approaches were evaluated: a supervised model based on a long short-term memory (LSTM) network, and an unsupervised, transformer-based model for time-series anomaly detection. The transformer model achieved an F1-score of 98.12%, precision of 96.31%, recall of 100.00%, and accuracy of 99.18%, outperforming the supervised baseline.These results demonstrate the high reliability and applicability of the proposed dataset for real-time fault detection. The methodology introduced in this study can be extended to other cyber-physical systems requiring high-integrity operational diagnostics. The proposed dataset and labeling framework provide a practical foundation for future research on autonomous safety monitoring and predictive maintenance in UAVs and other intelligent systems.

本研究提出一款面向真实应用场景的数据集，旨在依托人工智能驱动的异常检测技术，支撑无人机（unmanned aerial vehicles, UAVs）的故障诊断工作。为构建该数据集，研究团队先通过分析无人机历史事故报告明确了常见故障类型，随后设计并开展飞行试验，模拟结构、推进与传感器子系统的各类故障工况。本次研究共计开展25次飞行试验，覆盖标称正常工况与故障诱发工况两类场景。每次试验过程中，均同步记录多模态飞行数据，涵盖定位、姿态、气压、控制信号及振动信号。所采集的数据经预处理后，被划分为原始、正常与错误三个独立类别。该结构化数据集可同时适配监督学习与无监督学习两类模型的训练需求。为验证数据集的有效性，研究团队评估了两种建模方案：一种是基于长短期记忆网络（long short-term memory, LSTM）的监督学习模型，另一种是用于时序异常检测的无监督Transformer模型。实验结果显示，该Transformer模型的F1分数达98.12%，精确率为96.31%，召回率为100.00%，准确率为99.18%，性能优于该监督学习基线模型。上述结果充分证明，本研究所提出的数据集在实时故障检测场景中具备极高的可靠性与适用性。本研究提出的方法学可拓展至其他需要高完整性运行诊断的信息物理系统。本次提出的数据集与标注框架，为无人机及其他智能系统的自主安全监测与预测性维护相关的未来研究提供了切实可行的基础支撑。