Machine Anomalous Sound Detection Method Based on Lightweight Temporal Pyramid and ECA-MobileFaceNet

Area of Science:

• Machine Learning
• Signal Processing
• Industrial Acoustics

Background:

• Industrial anomalous sound detection faces challenges with limited anomaly samples and inadequate temporal feature modeling in lightweight models.
• Existing methods struggle with effective feature selection and capturing complex temporal dynamics.

Purpose of the Study:

• To propose a novel unsupervised framework for industrial anomalous sound detection.
• To enhance the modeling of multi-scale temporal dynamic features.
• To improve the feature selection capability of lightweight models.

Main Methods:

• Introduced a Lightweight Temporal Pyramid Module (LTPM) for multi-scale temporal modeling in TgramNet.
• Embedded the Efficient Channel Attention (ECA) mechanism into MobileFaceNet for adaptive channel recalibration.
• Implemented waveform-level data augmentation: noise perturbation, time shifting, and amplitude scaling.

Main Results:

• Achieved competitive performance on the DCASE 2020 Task 2 dataset across various machine types.
• Demonstrated optimal or highly competitive results compared to existing approaches.
• Verified model stability and generalization capability using minimum Area Under the Curve (mAUC) and ROC analysis.

Conclusions:

• The proposed lightweight, unsupervised method offers a promising solution for industrial anomalous sound detection.
• The approach effectively addresses challenges of scarce data and enhances temporal dependency modeling.
• This method is suitable for industrial condition monitoring applications.