Estimating Flow Velocity and Vehicle Angle-of-Attack from Non-invasive Piezoelectric Structural Measurements Using Deep Learning
Chandler B. Smith , S. Hales Swift, Andrew Steyer and Ihab El-Kady · 2026
Read Estimating Flow Velocity and Vehicle Angle-of-Attack from Non-invasive Piezoelectric Structural Measurements Using Deep Learning by Chandler B. Smith , S. Hales Swift, Andrew Steyer and Ihab El-Kady (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-23496.txt.
Introduction
Smith and three co-authors read speed and angle of attack from skin vibrations, not flow probes. A set of piezoelectric sensors lines the inside of an aeroshell. Turbulent pressure shakes the skin, and a neural network inverts those shakes into speed and angle. Tests ran in Sandia's hypersonic tunnel at Mach 5 and Mach 8. Sixteen runs covered steady and sweeping conditions. A median filter tames the output swings. Velocity error lands near a fifth of one percent, and angle error stays under half a degree.
Catalog record
| CREATOR | Chandler B. Smith , S. Hales Swift, Andrew Steyer and Ihab El-Kady |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 45733 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.23496
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Estimating Flow Velocity and Vehicle Angle-of-Attack from Non-invasive Piezoelectric Structural Measurements Using Deep Learning Chandler B. Smith , S. Hales Swift, Andrew Steyer and Ihab El-Kady Abstract Accurate estimation of aerodynamic state variables such as freestream velocity and angle of attack (AoA) is important for aerodynamic load prediction, flight control, and model validation. This work presents a non-intrusive method for estimating vehicle velocity and AoA from structural vibration measurements rather than direct flow instrumentation such as pitot tubes. A dense array of piezoelectric sensors mounted on the interior skin of an aeroshell capture vibrations induced by turbulent boundary layer pressure fluctuations, and a convolutional neural network (CNN) is trained to invert these structural responses to recover velocity and AoA. Proof-of-concept is demonstrated through controlled experiments in Sandia’s hypersonic wind tunnel spanning zero and nonzero AoA configurations, Mach 5 and Mach 8 conditions, and both constant and continuously varying tunnel operations. The CNN is trained and evaluated using data from 16 wind tunnel runs, with a temporally centered held-out interval within each run used to form training, validation, and test datasets and assess intra-run temporal generalization. Raw CNN predictions exhibit increased variance during continuously varying conditions; a short-window moving-median post-processing step suppresses this variance and improves robustness. After post-processing, the method achieves a mean velocity error relative to the low-pass filtered reference velocity below 2.27 m/s (0.21%) and a mean AoA error of 0.44 ∘ ( 8.25 % ) 0.44^{\circ}(8.25\%) on held-out test data from the same experimental campaign, demonstrating feasibility of vibration-based velocity and AoA estimation in a controlled laboratory environment. 1 Introduction Accurate estimates of aerodynamic state variables such as freestream velocity and angle of attack (AoA) are central to aerodynamic load prediction, flight control, and model validation. Conventional air data approaches (e.g., probes, vanes, or pressure port systems) can be difficult to integrate and calibrate on certain vehicle geometries and may disturb the local flow they seek to measure. These considerations motivate complementary sensing approaches that infer flow state indirectly from measurements taken within the vehicle structure. This paper investigates an internal sensing approach for estimating vehicle velocity and AoA using piezoelectric-based vibration measurements. The central premise is that the vibration signature of an aeroshell contains flow-dependent information through the coupling between turbulent boundary layer loading and structural dynamics. If successful, this approach could augment conventional air data instrumentation by providing additional, internal measurements that correlate with external flow state. Indirect flow sensing is motivated by decades of experimental and theoretical work showing that turbulent boundary layers generate stochastic surface-pressure fields that excite structural vibration through fluctuating normal stresses at the fluid-solid interface [ 1 , 2 , 3 , 4 , 5 , 6 , 7 ] . Building on this coupling, several non-intrusive approaches have used structural vibration measurements to infer flow parameters in settings where direct flow field sensing is undesirable or impractical. For example, pipe vibration-based flow measurements have demonstrated correlations between wall acceleration and flow rate via (i) direct flow-induced vibrations [ 8 ] , and (ii) flow-induced perturbations to mechanically excited plate waves [ 9 ] . Thompson et al. [ 10 ] characterized pipe wall vibration as a function of flow speed and pipe geometry for fully developed turbulent flow, establishing empirical relationships between flow parameters and structural response. More recently, machine learning models have been used to extract features from vibration signals that correlate with flow parameters (e.g., mass-flow rate) that may be difficult to measure directly [ 11 ] . In the aerospace sector, fiber Bragg grating (FBG) sensors have emerged as a particularly promising technology for indirect flow parameter estimation through structural response measurements, offering a dense field of structural strain measurements and minimal weight additions to the structure of interest. Nicolas et al. [ 12 ] used a dense network of FBG sensors distributed along a carbon-composite wing to determine both in-flight loads and deflected wing shapes during flight operations. Wada et al. [ 13 ] demonstrated a neural network approach wherein FBG strain data combined with the wing flap angles were used to estimate distributed wing loads and subsequently determine angle of attack, again without direct flow measurements. Furthermore, Kwon et al. [ 14 ] implemented an embedded FBG sensor network in a composite aircraft that successfully estimated wing loads during flight maneuvers. Collectively, these studies establish that internal strain/vibration measurements, when combined with data-driven inversion, can provide non-intrusive aerodynamic inference capabilities. While FBG-based approaches have demonstrated effectiveness for load and AoA inference, the present work explores an alternative sensing technique using piezoelectric sensors to estimate freestream velocity and AoA. Piezoelectric sensors are another type of low-weight, non-intrusive strain measurement that generate voltage signals when mechanically deformed due to the material’s inherent electromechanically coupled behavior. Our proposed method deploys a dense linear array of piezoelectric sensors, mounted to the interior skin of the aeroshell, to monitor vibration induced by turbulent boundary layer pressure fluctuations and thereby deduce the vehicle’s AoA and velocity. Because these interior-mounted sensors do not have direct access to the external flow, estimating freestream velocity and AoA becomes a challenging inverse problem. We address this inverse problem using a convolutional neural network (CNN) trained on labeled wind tunnel data. Specifically, this work establishes proof-of-concept through a three-phase methodology: (i) acquire piezoelectric response data under controlled wind tunnel conditions where velocity and AoA are measured using conventional instrumentation, (ii) train a CNN to map windowed piezoelectric vibration measurements to velocity and AoA labels using supervised learning, and (iii) evaluate the trained model on withheld test data to assess intra-run temporal generalization capability. Raw network predictions are optionally refined using a short-window sliding median filter to suppress high-frequency output fluctuations and improve robustness. Hypersonic wind tunnel experiments were used to exercise the methodology. We acquired tunnel freestream velocity estimates, AoA, and piezoelectric response data across sixteen runs spanning Mach 5 and Mach 8 conditions, multiple Reynolds-number states, and both constant and continuously varying tunnel operations. Piezoelectric signals were segmented into 16 ms windows and provided as input to a two-dimensional CNN that outputs estimates of velocity and AoA (via the inferred body-referenced velocity components). Performance is evaluated across the full set of runs, delineating between steady-state and time-varying tunnel conditions, to identify regimes where estimation is most reliable. The remainder of this paper is organized as follows. Section 2 details the proposed methodology and the neural network architecture. Section 3 describes the wind tunnel test setup, data acquisition, preprocessing procedures, and the range of flow conditions tested. Section 4 presents the zero AoA velocity estimation results (Section 4.1 ), then extends the analysis to the combined AoA and velocity estimation problem at nonzero AoA (Section 4.2 ). Finally, Sections 5 and 6 summarizes our findings and outlines directions for future work. 2 Method Overview This work estimates aerodynamic quantities from internally mounted piezoelectric vibration measurements using a supervised CNN. A dense array of piezoelectric sensors mounted to the interior of a sharp cone aeroshell measures structural response driven by turbulent boundary-layer pressure fluctuations. These spatiotemporal voltage measurements are segmented into fixed-duration windows and provided as input to a CNN that outputs estimates of freestream velocity (zero AoA study) or body-referenced in-plane velocity components (nonzero