Taxi Vibration Test for UAV Dynamics

• TVT is a structural vibration evaluation method that measures UAV dynamics during ground taxiing using robust, synchronized DAQ systems.
• It employs low-cost single-board computers and MEMS IMUs to acquire multi-channel acceleration data with precise temporal synchronization.
• The methodology uses advanced signal processing, including PSD estimation and linear time-scaling, to reliably identify and analyze modal frequencies.

The Taxi Vibration Test (TVT) is a structural vibration evaluation procedure commonly employed in aerospace applications to assess and verify the dynamic response and integrity of airborne platforms, specifically during ground taxiing phases. The method involves acquiring multi-channel acceleration data under non-stationary broadband excitation produced while the aircraft (here, a small fixed-wing UAV) is towed along an uneven surface at low speeds. Critical requirements for TVT campaigns include robust multi-sensor data acquisition systems (DAQs) capable of stable sampling, precise timestamping, and reliable data streaming across physically distributed measurement nodes. An exemplar TVT campaign and associated DAQ architecture are detailed and experimentally validated in "Affordable Data Collection System for UAVs Taxi Vibration Testing" (Yang et al., 12 Jan 2026).

1. Data Acquisition System: Architecture and Implementation

A central focus of recent research is the development of low-cost, scalable DAQ platforms circumventing the prohibitive cost and complexity of commercial aerospace DAQs. The cited implementation utilizes three Orange Pi 3 LTS single-board computers (SBCs), each chosen for their integrated Linux support, built-in I²C, Wi-Fi networking, and adequate computational resources to achieve a per-node sampling rate near 200 Hz. Sensing is performed by eight Adafruit LSM6DS3TR-C MEMS IMUs (each providing three-axis accelerometer and gyroscope data, though only accelerometry is used for TVT), configured for ±4 g range and up to 6.66 kHz output data rate, with I²C connectivity and low power consumption.

To address address conflicts inherent in using multiple identical I²C devices, two Adafruit TCA9548A 8-way I²C multiplexers are deployed per slave SBC, each allowing up to eight IMUs per bus. Ancillary hardware includes regulated power supplies, pull-up resistors, and Wi-Fi dongles. The resulting hardware bill of materials totals approximately €540.

The DAQ's software stack follows a Python-based master/slave architecture. The master node coordinates overall configuration (sampling rate $f_s = 208\,\mathrm{Hz}$, range, acquisition duration) and synchronizes TVT initiation via broadcast TCP commands over Wi-Fi. Slave SBCs initialize IMUs and multiplexers, enter a read loop for sequential IMU channel polling via I²C, timestamp sampled acceleration vectors using the local system clock, and stream formatted CSV data to the master through non-blocking Python threads. The master aggregates those streams, reconstructs sensor-specific datafiles, and closes the runtime session post-test (Yang et al., 12 Jan 2026).

2. Sampling, Synchronization, and Data Integrity

A critical aspect of multi-node structural testing is the preservation of temporal synchrony among distributed acquisition channels. Each slave SBC aims for a nominal $f_s = 208\,\mathrm{Hz}$ by configuring the IMU's output data rate and regulating polling intervals using Python's time-keeping primitives. Empirical drift during 60 s TVT campaigns is constrained to <5 ms absolute divergence, with inter-sample interval regulation correcting low-frequency drift to below 0.1 Hz. Slaves' clocks are synchronized at test onset by a simultaneous TCP "start" signal, achieving initial timestamp spreads below 10 ms.

To account for network and clock-induced asynchrony, each sensor's time series is interpolated offline to a common 208 Hz grid; any residual drift (<0.02 Hz) is compensated by linear time-scaling. This pipeline ensures that multi-sensor datasets are temporally coherent to within the sampling system's inherent resolution, a precondition for accurate cross-channel spectral analysis.

3. Signal Processing: PSD Estimation and Statistical Protocols

After acquisition, acceleration time series undergo a detrending stage to remove the global mean and mitigate spurious low-frequency content. Power Spectral Density (PSD) estimation employs Welch's method, with the one-sided PSD approximated as

$$\hat S_{xx}(f) = \frac{1}{L}\sum_{l=1}^{L} \frac{1}{U}\,|\mathrm{DFT}\{w[n]x_l[n]\}|^2$$

where $L$ is the number of 4 s segments (832 samples/segment), $x_l[n]$ is the windowed data from segment $l$, $w[n]$ is the Hamming window of length $N$, and $U = \frac{1}{N} \sum_{n=0}^{N-1} w^2[n]$ ensures window power normalization. The analysis windows overlap by 50% (416 samples). The resulting PSDs are normalized so the area under the curve matches the signal variance. Statistical repeatability and uncertainty are addressed by constructing 95% confidence intervals using chi-square bounds with $2L$ degrees of freedom.

4. Experimental Procedure: UAV, Taxiing, and Instrumentation

The test platform is a Volantex RC Ranger 2400 UAV with 2.4 m wingspan, 1.23 m length, and 1.7 kg mass (dry). TVT runs involve the UAV being towed across unpaved asphalt and dirt at 2–5 m/s, ensuring broad-band and non-stationary dynamic excitation. Each TVT covers approximately 60 s, with six independent replicates per campaign. Six tri-axial IMUs (three/symmetrical per wing) are mounted using double-sided foam adhesive—ensuring direct coupling to the wing skin—at specific coordinates along each wing spar to resolve spatial mode shapes.

A comparison condition, the Ambient Vibration Test (AVT), exposes the stationary UAV (on wooden blocks) to idle wind and ground input, offering a low-excitation baseline for system sensitivity and noise floor estimation. Two AVT runs (20 min each) were performed, enabling the identification of modal signatures at minimal excitation levels (Yang et al., 12 Jan 2026).

5. Results: Dynamic Response and System Performance

TVT time series yield peak accelerations of ±0.5 g and RMS values near 0.06 g for sensor 1_1 (x-axis); AVT under the same band (1–50 Hz) shows peaks of only ±0.02 g and RMS ≈ 0.005 g. TVT analysis consistently reveals five dominant spectral peaks at frequencies ~1.2 Hz, 2.5 Hz, 4.0 Hz, 6.8 Hz, and 8.5 Hz. From 10–50 Hz, TVT spectra display a broadband, flat-top characteristic. AVT also resolves the same mode frequencies, but the final three appear with markedly diminished magnitude (below –20 dB). For TVT, the 95% confidence intervals for $f<10$ Hz are contained within ±1 dB, widening above due to reduced signal coherence. Amplitude variance at modal peaks remains <5% across all TVT runs, and identified modal frequencies are repeatable to within ±0.1 Hz. Data acquisition completed without dropouts or buffer overflows; all 18 channels streamed and recorded successfully throughout all tests.

6. Comparative Analysis: Cost, Capability, and Scalability

The DAQ’s total material cost is approximately €540, substantially lower than commercial aerospace DAQs with similar channel counts and per-channel bandwidth, which exceed €5,000. Hardware component costs are detailed below:

Component	Unit Price (€)	Quantity	Subtotal (€)
Orange Pi 3 LTS SBC	~50	3	150
LSM6DS3TR-C IMUs	~15	8	120
TCA9548A I²C Mux	~10	2	20
Wiring/Power/Networking/Ancillary	—	—	150
Enclosures & Mounting	—	—	100
Total			540

System-identified trade-offs include the dynamic range (configurable ±2 to ±16 g, vs. ±50 g in high-end DAQs), sensor resolution (16-bit MEMS; ~0.061 mg LSB, vs. professional 24-bit ADCs), and data bandwidth (Python/I²C limited to ~250 Hz vs. multi-kHz). Scalability is achieved by adding additional Orange Pi + multiplexer nodes, contrasting with the more constrained expansion and footprint of commercial units.

7. Limitations and Prospective Enhancement

The maximal stable sampling rate is constrained to approximately 250 Hz by the combined bottlenecks of Python software overhead and I²C bus saturation (operating at 400 kHz). Clock drift, while subcritical for the present 60 s tests (~0.1%), can accumulate non-trivially for extended campaigns in the absence of hardware timestamping. Gyroscope data, though available, is not currently utilized for TVT.

Future upgrades highlighted include migrating polling routines to C/C++ or employing real-time Linux for exceeding $f_s > 500$ Hz, exploiting on-IMU FIFO buffering to reduce I²C load and improve timestamp accuracy, and deploying higher-performance IMUs (e.g., ICM-42688-P) for enhanced noise performance. Real-time onboard processing (FFT computation) and system-wide sub-millisecond synchronization via Precision Time Protocol (PTP) or dedicated sync lines are advocated for next-generation implementations (Yang et al., 12 Jan 2026).

In summary, the TVT methodology, in conjunction with the described DAQ architecture, provides a cost-effective and reliable framework for capturing high-fidelity, multi-point vibration data on UAV platforms. The approach enables the resolution of principal structural modes and supports repeatable system identification, offering a scalable template for academic and resource-constrained experimental investigations in aerospace structural dynamics.