Integrating Frequency Guidance into Multi-Source Domain Generalization for Acoustic-Based Fault Diagnosis in...

Area of Science:

• Engineering
• Signal Processing
• Machine Learning

Background:

• Acoustic-based fault diagnosis is crucial for industrial monitoring but suffers from performance degradation due to complex conditions and domain shifts.
• Unseen target domain data unavailability poses a significant challenge for existing models.

Purpose of the Study:

• To develop an advanced acoustic fault diagnosis network resilient to domain shifts.
• To enhance data diversity and feature discriminability for improved fault detection accuracy.

Main Methods:

• Propose an amplitude-phase collaborative augmentation network (AP-CANet) for adaptive feature alignment and label-consistent sample augmentation.
• Integrate a frequency-spatial interaction module to combine global spectral and local temporal information.
• Introduce a manifold triplet loss to improve intra-class compactness and inter-class separability for hard samples.

Main Results:

• AP-CANet demonstrates superior performance in domain-shift scenarios on the GPLA-12 and MIMII-DG datasets.
• The method effectively enriches data diversity while preserving semantic consistency.
• Improved feature discriminability leads to better detection of subtle fault distinctions.

Conclusions:

• AP-CANet offers a robust solution for acoustic fault diagnosis, particularly under domain-shift conditions.
• The proposed method shows potential for scalable and cost-effective industrial fault monitoring.
• The approach effectively addresses challenges posed by complex working environments and data variability.