Assessing AI Dependability in Acoustic NDT: Deep Learning Robustness for Gold Purity Verification under Industrial Noise

Abstract

Acoustic resonance testing (ART) is a widely utilized nondestructive evaluation method for material characterization. However, its measurement reliability significantly degrades under harsh industrial noise, creating undefined operational risks. This study presents a systematic framework to quantify the robustness of deep learning-based acoustic sensing systems against environmental interference. A controlled synthetic noise injection protocol is employed, using white and pink (1/f)) noise as standardized proxies for broadband and low-frequency-dominant industrial interference. This protocol enables systematic analysis of measurement performance degradation across varying signal-to-noise ratios (SNRs) under simulated industrial acoustic conditions. The results define the operational boundaries for reliable measurement, identifying a critical SNR threshold of 20 dB where measurement uncertainty significantly increases. Furthermore, we demonstrate that spectral masking and the “Artificial Damping Effect” are the primary physical drivers of measurement error in noisy environments. Although the methodology is demonstrated on gold alloy benchmarks, the proposed reliability envelope and noise compensation approach are applicable to the acoustic evaluation of various homogeneous materials. These findings establish a quantitative robustness envelope and provide a reproducible benchmark for industrial deployment. Further validation under directly recorded field conditions is required before practical implementation.