This study proposes a noise-robust unsupervised acoustic anomaly detection framework for early identification of abnormal operating conditions in rotating machinery under submarine-like environments with severe data scarcity. In such environments, underwater background noise and onboard interference sources significantly degrade signal quality, while limited computing resources constrain the deployment of high-complexity deep learning models. To address the lack of labeled fault data, the publicly available MIMII dataset was adopted as a proxy platform, and representative submarine interference sources were physically modeled, including colored background noise, structure-borne resonance, band-limited auxiliary noise, tonal components, and sensor noise. These components were combined and scaled to predefined SNR levels (−6 to 6 dB) to generate realistic noise-augmented data. Three unsupervised approaches were compared under edge deployment constraints: (i) Gaussian Mixture Model (GMM) with statistical MFCC features, (ii) statistical-feature-based Ensemble Autoencoder, and (iii) Conv1D-based Ensemble Autoencoder using 1-s log Mel-spectrogram segments. Performance was evaluated in terms of AUC, F1-score, and computational cost. Results show that GMM provides competitive detection performance with minimal computational burden, whereas Conv1D achieves superior accuracy when temporal fault patterns dominate, at the expense of higher complexity. The study provides practical design guidelines for acoustic anomaly detection under multi-noise and resource-constrained conditions.
Kim et al. (2026) studied this question.