aykutkabaoglu/uav-flight-anomaly-dataset

DATASET CARD for UAV-SEAD

Citation

BibTeX:

@article{kabaoglu2024uavsead,
  title={UAV-SEAD: State Estimation Anomaly Dataset for UAVs},
  author={Kabaoglu, Aykut and Sariel, Sanem},
  journal={arXiv preprint},
  year={2026}
}

DATASET INFORMATION

Dataset Description

UAV-SEAD (State Estimation Anomaly Dataset) is a comprehensive, large-scale, real-world flight dataset designed to advance research in Fault Detection and Identification (FDI) for Unmanned Aerial Vehicles. It contains 1,396 categorized real-world flight logs totaling over 52.4 hours of flight data. Unlike existing benchmarks, it avoids synthetic fault injection, providing naturally occurring anomalies across diverse environmental and hardware configurations.

• —Curated by: Aykut Kabaoglu
• —License: Creative Commons Attribution 4.0 International (CC BY 4.0)

Dataset Sources

• —Paper: "UAV-SEAD: State Estimation Anomaly Dataset for UAVs"

Toolbox for Dataset Preprocessing and Annotation

• —https://github.com/aykutkabaoglu/ulogannotationtool
• —Use readme in the ulog_annotation_tool repository to start using dataset. The first step is converting raw flight logs into csv files. Then the flight logs can be used for plotting and labeling.

DATASET CREATION AND USAGE

Dataset Structure

The dataset is structured as a collection of multivariate time-series sequences.

• —Point Anomalies: Spikes and outliers in sensor readings.
• —Contextual Anomalies: Readings that are only anomalous relative to the flight state (e.g., zero altitude during thrust).
• —Collective Anomalies: Sustained drifts or patterns indicating gradual system failure.

Use Cases

• —Benchmarking multivariate time-series anomaly detection algorithms.
• —Developing predictive maintenance and fault identification systems for UAVs.

Data Collection and Processing

Data was harvested from a fleet of custom-built, PX4-based UAVs.

• —Hardware: Multiple airframes with variations in motor sizes, weights, and sensor suites. Check paper for hardware details.
• —Environments: Flight missions conducted in indoor hangars, warehouses, and open outdoor spaces.

Annotation process

Annotation was performed via Physically-Constrained Expert Evaluation. Human experts identified state estimation divergences by comparing autopilot estimates with real-world observed outcomes. Labels were verified against physical flight dynamics to ensure the statistical anomalies correlate with actual physical events.

Bias, Risks, and Limitations of Data Collection

• —Platform Specificity: Limited to multi-rotor dynamics. Limited outdoor flights with GPS availability.

DATASET STATISTICS

Annotated Data Flight Source Statistics

Annotated Number of Flight: 1396

Annotated Data Class Statistics

RAW Data Statistics

Number of Flights: 3196

RAW Data, Data Fields

The dataset follows a Superset Schema of 1897 unique columns. Fields are grouped into 84 unique topics by PX4 UORB topic prefixes.

81 PX4 Topics out of 84 and their fields that has data at least a single flight log in the dataset are listed as below:

Dataset Features

Major PX4 topics that can mostly be found in flight logs and their descriptions: