Unsupervised Acoustic Anomaly Detection for Rotating Machinery Under Submarine-like Environments: Considering Data...

Area of Science:

• Mechanical Engineering
• Signal Processing
• Machine Learning

Background:

• Rotating machinery requires early fault detection for operational reliability.
• Submarine environments present unique challenges: severe noise, data scarcity, and limited computational resources.
• Traditional methods struggle with complex acoustic signatures and low signal-to-noise ratios (SNRs).

Purpose of the Study:

• To develop a noise-robust unsupervised acoustic anomaly detection framework for rotating machinery.
• To evaluate unsupervised methods under realistic submarine-like noise conditions and edge deployment constraints.
• To provide practical guidelines for acoustic anomaly detection in resource-constrained, multi-noise environments.

Main Methods:

• Physically modeled submarine interference sources (noise, resonance, interference) to create augmented data.
• Compared three unsupervised approaches: Gaussian Mixture Model (GMM) with MFCC features, statistical-feature-based Ensemble Autoencoder, and Conv1D-based Ensemble Autoencoder.
• Evaluated performance using AUC, F1-score, and computational cost on the MIMII dataset.

Main Results:

• Gaussian Mixture Model (GMM) demonstrated competitive detection performance with minimal computational cost.
• Conv1D-based Ensemble Autoencoder achieved superior accuracy for detecting temporal fault patterns, but with higher complexity.
• Noise augmentation effectively simulated realistic submarine interference across various SNR levels (-6 to 6 dB).

Conclusions:

• Unsupervised acoustic anomaly detection is feasible in challenging submarine environments.
• GMM is suitable for resource-constrained edge deployments prioritizing efficiency.
• Conv1D models offer higher accuracy when temporal dynamics are critical, balancing performance with computational demands.