Estimation of Airflow Parameters for Tail-Sitter UAV through a 5-Hole Probe Based on an ANN

Xiaoda Li, Yongliang Wu, Xiaowen Shan, Haofan Zhang, Yang Chen2023 · DOI

TL;DR

This paper tackles the challenge of accurately estimating airflow parameters (angle of attack, sideslip angle, and incoming flow speed) for tail-sitter VTOL UAVs, especially at large angles of attack where traditional methods fail. They propose a novel approach combining a 5-hole probe with an Artificial Neural Network (ANN) to achieve high accuracy, paving the way for improved control and performance of these versatile aircraft.

What Problem Does This Paper Address?

Tail-sitter VTOL (Vertical Take-Off and Landing) UAVs are gaining traction due to their ability to combine the advantages of fixed-wing aircraft (speed, endurance) and rotary-wing aircraft (vertical takeoff/landing). However, their flight involves a transition phase where their attitude changes significantly, leading to large angles of attack and sideslip. Accurately measuring airflow parameters like angle of attack (AoA) and sideslip angle (AOS) is crucial for stable control during these maneuvers. Traditional air data sensors, such as pitot tubes and vanes, struggle with large angles of attack and low speeds, making them unsuitable for tail-sitters. While multi-hole probes offer a solution, existing estimation methods (look-up tables, interpolation, polynomial fitting) often have limitations in accuracy or applicability at large angles. This paper aims to bridge this gap by developing a robust estimation method for tail-sitter UAVs that can handle these challenging flight regimes.

Methods

The core of this research lies in a two-pronged approach:

  • 5-Hole Pressure Probe: This device, custom-made using 3D metal printing, measures pressure at five distinct points on its head. By analyzing the pressure differences between these holes and a static pressure port, information about the incoming airflow's speed and direction can be derived.
  • Artificial Neural Network (ANN): To translate the raw pressure data from the 5-hole probe into usable airflow parameters (AoA, AOS, and velocity), the researchers employed an ANN. Specifically, they utilized a Backpropagation (BP) neural network. This method involves:
  • Data Collection: Wind tunnel experiments were conducted to gather a comprehensive dataset of pressure readings under various controlled airflow conditions (different angles of attack, sideslip angles, and velocities).
  • Network Architecture: The ANN architecture was optimized through extensive experimentation. They determined that a four-hidden-layer structure with a specific neuron distribution (15-30-15-5) provided the best performance.
  • Training: The Levenberg-Marquardt algorithm was used for efficient training of the neural network. The network was trained to predict AoA, AOS, or velocity individually, based on the five pressure inputs.
  • Data Preprocessing: A Kalman filter was applied to the raw pressure data to reduce fluctuations caused by low airflow speeds and turbulence, which improved the training performance of the ANN.

Datasets

The primary dataset used in this study was generated through wind-tunnel experiments. The experiments were designed to cover a wide range of airflow conditions relevant to tail-sitter UAVs:

  • Angle of Attack (AoA): Ranged from -50° to 80°.
  • Sideslip Angle (AOS): Ranged from -30° to 30°.
  • Incoming Flow Velocities: 10 m/s, 15 m/s, and 20 m/s.

The paper does not mention any pre-existing public datasets for this specific problem. The experimental data collected is specific to the tail-sitter UAV and the 5-hole probe configuration used.

Key Insights and Ideas

  • ANNs Excel at Complex Relationships: The study highlights that ANNs are well-suited for modeling the complex, non-linear relationships between pressure measurements from a 5-hole probe and the desired airflow parameters, especially in challenging flight regimes.
  • Filtering is Crucial for Low Speeds: For low-speed airflow, pressure data can be noisy. Applying a simple Kalman filter significantly improved the ANN's training performance and estimation accuracy.
  • Large Angles are Achievable: The research demonstrates that with a carefully designed ANN and sufficient experimental data, accurate estimation of airflow parameters is possible even at large angles of attack (up to ±60° for AoA, ±50° for AOS), a significant improvement over traditional methods.
  • Velocity Estimation is Robust: The incoming flow velocity estimation proved to be particularly robust, likely due to the inherent and clear relationship between pressure and velocity, requiring less attention for optimization.
  • Asymmetry in Errors: An interesting observation was the asymmetry in estimation errors between AoA and AOS. The authors attribute this anomaly to differences in the experimental ranges and data distribution for AoA and AOS, suggesting a need for further investigation.

Key Results

The proposed 5-hole probe and ANN method achieved the following performance:

  • Angle of Attack (AoA) Estimation: Error within ±1° for AoA in the range of -60° to 60°. Even at angles up to 70°, the error remained within ±3° for angles less than 80°.
  • Sideslip Angle (AOS) Estimation: Error within ±1.5° for AOS in the range of ±50°.
  • Incoming Flow Velocity Estimation: Error remained relatively small across the entire measurement range.
  • Improved Performance with Filtering: The ANN trained with filtered data showed significantly better performance compared to training with raw data.

Limitations and Future Work

  • Onboard Experimentation: The authors acknowledge that the study was limited to wind-tunnel experiments and that onboard flight testing has not yet been completed. This is a crucial next step to validate the method in real-world flight conditions.
  • Data Distribution Anomaly: The observed asymmetry in AoA and AOS estimation errors warrants further investigation and validation with more experimental data.
  • Generalizability: While the method shows promise for tail-sitter UAVs, its generalizability to other aircraft types or different probe designs might require further study.
  • Real-time Implementation: The paper mentions importing the trained network into Simulink for verification and simulation, but the practical challenges and computational load of real-time implementation on an embedded system for a UAV would be a subject for future work.

Summary updated

BibTeX
@misc{li2023estimation,
  author = {Xiaoda Li and Yongliang Wu and Xiaowen Shan and Haofan Zhang and Yang Chen},
  title = {{Estimation of Airflow Parameters for Tail-Sitter UAV through a 5-Hole Probe Based on an ANN}},
  year = {2023},
  doi = {10.3390/s23010417},
}
Figure