Data from: A triboelectric airflow field sensor enables cross-scale multi-parameter sensing in meter-scale flapping-wing aircraft
Data files
Jun 24, 2026 version files 1.44 GB
-
Dataset_Splitter.py
6.18 KB
-
Flight_Test_Data.csv
36.51 KB
-
Large_Wind_Tunnel_AoS_Model.pth
1.77 MB
-
Large_Wind_Tunnel_Data_Prediction.py
32.40 KB
-
Large_Wind_Tunnel_Experimental_Data.zip
53.42 MB
-
Large_Wind_Tunnel_Transferred_AoA_Model.pth
47.98 MB
-
Large_Wind_Tunnel_Transferred_FF_Model.pth
558.71 MB
-
Large_Wind_Tunnel_Transferred_LWS_Model.pth
175.37 MB
-
README.md
4.49 KB
-
Small_Wind_Tunnel_Data_Prediction.py
17.78 KB
-
Small_Wind_Tunnel_Experimental_Data.zip
207.70 MB
-
Small_Wind_Tunnel_Pretrained_AoA_Model.pth
23.99 MB
-
Small_Wind_Tunnel_Pretrained_FF_Model.pth
279.35 MB
-
Small_Wind_Tunnel_Pretrained_LWS_Model.pth
87.69 MB
Abstract
Accurate in-situ sensing of unsteady aerodynamic parameters is essential yet challenging for closed-loop control in meter-scale flapping-wing aircraft. This task requires simultaneous monitoring of three coupled parameters: local wind speed (LWS), angle of attack (AoA), and flapping frequency (FF). Here, we report a conformal triboelectric airflow field sensor that converts airflow-induced aeroelastic vibrations into electrical signals. Its low-damping apertured cantilever design provides a broad sensing range and rich vibration features, enabling accurate aerodynamic parameter decoding with a deep-learning model. To address cross-scale discrepancies, we use transfer learning to adapt representations learned from low-cost small-scale wind tunnel data to data-limited full-scale conditions, achieving mean absolute errors of 0.04 m·s⁻¹ for LWS, 0.07° for AoA, and 0.06 Hz for FF. Outdoor flight tests demonstrate real-time trend reconstruction and maneuver-correlated monitoring of the decoded parameters. This study provides a scalable biomimetic airflow sensing solution for future closed-loop control of meter-scale flapping-wing aircraft.
Dataset DOI: 10.5061/dryad.sn02v6xm2
Description of the data and file structure
The data primarily consists of three parts: data collected from small wind tunnel experiments that could only accommodate partial wing models, data from large wind tunnel experiments that could accommodate full-scale wings, and data from outdoor flight tests of a flapping-wing vehicle. The wind tunnel data were used to train deep learning models that establish the relationship between sensor signals and key flight parameters (LWS, AoA, FF, AoS). The data from outdoor flight tests serves as an independent validation dataset.
Files and variables
File: Dataset_Splitter.py
Description: This code implements dataset splitting and windowing for single-channel or dual-channel raw data. It allows for adjustment of parameters such as dataset ratios, window length, sliding step size, and the number of windows.
File: Large_Wind_Tunnel_AoS_Model.pth
Description: This is a model trained on the large wind tunnel dataset for Angle of Sideslip (AoS) parameter identification.
File: Large_Wind_Tunnel_Data_Prediction.py
Description: Prediction code for the large wind tunnel models.
File: Large_Wind_Tunnel_Experimental_Data.zip
Description: This is the dataset from the large wind tunnel experiments. Each CSV file is named using the convention: LWS-AoA-FF-AoS.csv. Each CSV file contains columns:Time (s): Sampling time, Channel_A (V): Sensor output voltage, channel A, Channel_B (V): Sensor output voltage, channel B
File: Large_Wind_Tunnel_Transferred_AoA_Model.pth
Description: This is a large wind tunnel transfer model for Angle of Attack (AoA) identification, obtained by fine-tuning a pre-trained small wind tunnel model using the large wind tunnel data via transfer learning.
File: Small_Wind_Tunnel_Data_Prediction.py
Description: Prediction code for the small wind tunnel models.
File: Large_Wind_Tunnel_Transferred_FF_Model.pth
Description: This is a large wind tunnel transfer model for Flapping Frequency (FF) identification, obtained by fine-tuning a pre-trained small wind tunnel model using the large wind tunnel data via transfer learning.
File: Large_Wind_Tunnel_Transferred_LWS_Model.pth
Description: This is a large wind tunnel transfer model for Local Wind Speed (LWS) identification, obtained by fine-tuning a pre-trained small wind tunnel model using the large wind tunnel data via transfer learning.
File: Small_Wind_Tunnel_Experimental_Data.zip
Description: This is the dataset from the small wind tunnel experiments. Each CSV file is named using the convention: LWS-AoA-FF-AoS.csv. Each CSV file contains columns: Time (s): Sampling time, Channel_A (V): Sensor output voltage, Channel A
File: Small_Wind_Tunnel_Pretrained_AoA_Model.pth
Description: This is a pre-trained model for Angle of Attack (AoA) parameter identification, trained on the small wind tunnel dataset.
File: Small_Wind_Tunnel_Pretrained_FF_Model.pth
Description: This is a pre-trained model for Flapping Frequency (FF) parameter identification, trained on the small wind tunnel dataset.
File: Small_Wind_Tunnel_Pretrained_LWS_Model.pth
Description: This is a pre-trained model for Local Wind Speed (LWS) parameter identification, trained on the small wind tunnel dataset.
File: Flight_Test_Data.csv
Description: Independent outdoor flight test data. The time-series parameters (LWS, AoA, FF, AoS) were predicted by the deep learning models from sensor signals recorded during flight.
Variables
LWS: Local Wind Speed
AoA: Angle of Attack
FF: Flapping Frequency
AoS: Angle of Sideslip
Code/software
The involved Python code runs on the Ubuntu 22.04 system. The environment is configured with PyTorch 2.8.0, Python 3.12, and CUDA 12.8.
Workflow: First, use Dataset_Splitter.py to perform dataset splitting and windowing. Then, after correctly setting the model paths and dataset paths in Large_Wind_Tunnel_Data_Prediction.py and Small_Wind_Tunnel_Data_Prediction.py, the prediction results can be obtained.
Access information
Other publicly accessible locations of the data:
- None
Data was derived from the following sources:
- None