Deep SVDD-based anomaly detection of machine sound signals using active signal refinement

Area of Science:

• Machine Learning
• Signal Processing
• Industrial Acoustics

Background:

• Detecting industrial audio anomalies is crucial for operational safety and efficiency.
• Irregular background noise and imbalanced datasets present significant challenges in anomaly detection.
• Existing methods like Deep SVDD and autoencoders have limitations in robustness against noise and data scarcity.

Purpose of the Study:

• To develop a robust Deep SVDD-based framework for industrial audio anomaly detection.
• To enhance detection performance by integrating novel preprocessing techniques.
• To address challenges posed by background noise and data imbalance.

Main Methods:

• Proposed a Deep SVDD-based framework with two novel preprocessing modules.
• Shifted Sequential Segmentation: Applied overlapping segmentation and time-shifting to Log-Mel Spectrograms for data augmentation and temporal invariance.
• Active Signal Refinement: Utilized the Sobel operator to enhance signal regions and enable effective noise subtraction.

Main Results:

• Achieved a superior Area Under the Curve (AUC) of 0.83 on the MIMII dataset.
• Outperformed existing Deep SVDD (AUC 0.70) and autoencoder (AUC 0.61) methods.
• Demonstrated enhanced robustness against irregular background noise and data imbalance.

Conclusions:

• The proposed Deep SVDD-based framework with novel preprocessing significantly improves industrial audio anomaly detection.
• The integration of Shifted Sequential Segmentation and Active Signal Refinement enhances robustness and accuracy.
• This approach offers a promising solution for real-world industrial monitoring applications.