Integrating acoustic tapping with a UAV platform for tile condition classification

Technical Summary: Integrating Acoustic Tapping with a UAV Platform for Tile Condition Classification

Problem Statement
The structural integrity of building tiles is critical for public safety, yet traditional inspection methods face significant limitations. Visual inspections, whether manual or computer-vision-based, are often restricted to surface-level assessment, are susceptible to lighting and accessibility constraints, and struggle to detect subsurface defects like delamination. While Unmanned Aerial Vehicles (UAVs) offer a promising solution for rapid facade assessment, their integration with acoustic tap-testing—a reliable method for detecting internal defects—introduces a critical challenge: UAV-induced dynamic perturbations. The vibrations inherent in drone flight degrade the reliability of acoustic signals, and the extent of this degradation, along with methods to correct it, remains largely unquantified in current automated inspection methodologies.

Methodology
The study proposes a controlled experimental framework to isolate and quantify the effects of UAV-induced vibrations on acoustic tap-testing reliability, followed by the development of a signal correction method.

• Hardware and Setup: The research utilizes a custom-built tap-testing device featuring a crank-rocker mechanism driven by a DC motor to ensure consistent impact energy, direction, and frequency, eliminating human subjectivity. Acoustic data is captured wirelessly via a TASCAM DR-44WL recorder. To simulate UAV flight dynamics without the complexity of actual flight, a Stewart platform (hexapod) is employed to reproduce controlled oscillatory conditions (Roll, Pitch, Yaw) with varying amplitudes (1°, 3°, and 5°) derived from prior UAV flight characterization using Vicon motion capture systems.
• Specimens: Laboratory tests were conducted on tile specimens mounted on concrete blocks. Defects were simulated by applying adhesive to only 50% of the tile surface, creating "partially attached" (unhealthy) versus "fully attached" (healthy) conditions.
• Signal Processing and Classification:
  • Energy-Based Filtering: An energy-based signal processing workflow is introduced. Based on Parseval's theorem, the method segments acoustic time series, calculates Power Spectral Density (PSD), and applies band-limited energy thresholds. This approach filters out low-amplitude and high-amplitude artifacts associated with environmental vibrations while preserving the acoustic characteristics of the mechanical impact.
  • Dimensionality Reduction and Clustering: Principal Component Analysis (PCA) is applied to reduce the dimensionality of the wireless acoustic data, retaining components that explain at least 90% of the variance. K-means clustering is then used to group the transformed data.
  • Classification: A decision tree classifier is trained on the PCA-reduced data to distinguish between healthy and unhealthy tile conditions. Performance is evaluated using standard metrics: Precision, Negative Predictive Value (NPV), Recall (Sensitivity), Specificity, and Accuracy.

Key Results

• Impact of Vibration: Under ideal conditions (0° vibration), the classification model achieved perfect accuracy (1.00). However, as vibration amplitude increased, performance degraded significantly. At 3° and 5° amplitudes, accuracy dropped to 0.77 and 0.72, respectively, with a notable loss in sensitivity (Recall), indicating the model struggled to identify defective tiles amidst vibration noise.
• Efficacy of Correction: The proposed energy-based filtering method successfully mitigated the negative effects of vibration. By selectively preserving physically meaningful acoustic events and suppressing motion-related artifacts, the method restored classification accuracy to above 98% (specifically 0.99) even at the highest vibration level (5°).
• Feature Separation: PCA visualization demonstrated that while vibration caused data dispersion and shifted cluster boundaries, the energy-based method maintained clear separation between healthy and unhealthy clusters, preventing the overlap observed in uncorrected data.

Significance and Claims
The authors claim that this work establishes a necessary pre-field validation stage for robotic facade inspection systems. The primary contribution is the quantification of motion-induced acoustic degradation and the demonstration that an energy-based signal correction approach can systematically mitigate these effects. The study asserts that while UAV-induced vibrations significantly distort acoustic feature distributions, the proposed framework allows for robust damage detection in dynamic environments.

The paper positions this research as a bridge between laboratory acoustic testing and real-world UAV applications. It emphasizes that the method automates the traditional hammer test and provides a scalable approach to ecoacoustic monitoring. The authors modestly note that while the results are robust under controlled laboratory conditions with predefined trajectories, future work is required to validate these findings against real-world environmental factors such as wind gusts and thermal variations. The study concludes that integrating physical stabilization systems and extending the framework to higher-complexity models are logical next steps for field deployment.