abstract
- This study characterized shallow subsurface cavity growth using acoustic testing and the Hilbert-Huang transform (HHT). Controlled sandbox experiments were conducted by systematically developing shallow cavities and varying sand disposal amounts. Acoustic responses were collected at multiple locations to capture the spatial variability. Signals were processed using the HHT to extract intrinsic mode functions (IMFs) for nonlinear and nonstationary analyses. The dominant IMFs were used to reconstruct the denoised signals and estimate the pitch frequency to track the cavity-induced spectral shifts. IMF-based time¿frequency maps were converted to mel-spectrograms to estimate the distance to the cavity center via deep neural network (DNN) regression with different backbones, with raw-signal mel-spectrograms serving as a benchmark. The results show that empirical mode decomposition decomposes acoustic signals into oscillatory modes with distinct physical relevance: high-frequency IMF1 is dominated by noise, whereas mid-order IMFs (IMF2¿IMF4) capture cavity-sensitive oscillations. Statistical indicators (root mean square, kurtosis, and Kurtosis × RMS) identified these IMFs as the most informative and the reconstructed signals exhibited reduced noise and low-frequency drift. The pitch frequency increased gradually from approximately 110 Hz at the cavity center to 136 Hz near the cavity cone edge, followed by an abrupt increase to ~280 Hz outside the cavity, indicating a transition from cavity-dominated to intact soil behavior. Distance estimation using analytic IMF mel-spectrograms achieved strong performance across all DNN backbones (coefficient of determination ¿0.91), significantly outperforming the benchmark. By integrating physically informed signal decomposition with learning-based regression, the proposed framework provides an effective laboratory-scale approach for shallow subsurface cavity characterization and motivates further validation under more diverse conditions.