Acoustic-Sensing-Based Attribute-Driven Imbalanced Compensation for Anomalous Sound Detection without Machine...

Area of Science:

• Acoustic Sensing
• Machine Learning
• Condition Monitoring

Background:

• Anomalous Sound Detection (ASD) is vital for condition monitoring.
• Unsupervised training data (normal sounds only) presents challenges for robust ASD systems.
• Existing discriminative models rely on strong prior knowledge (e.g., machine ID), limiting real-world applicability.

Purpose of the Study:

• To develop a robust ASD system using weak attribute labels from acoustic sensors.
• To overcome limitations of models requiring strong prior knowledge.
• To enhance ASD performance in unsupervised and data-scarce scenarios.

Main Methods:

• Utilizing imbalanced and inconsistent acoustic sensor attributes (e.g., speed, microphone model) as weak priors.
• Training an attribute classifier with an imbalanced compensation strategy for model trainability.
• Employing a score fusion method to improve anomaly detection robustness.

Main Results:

• The proposed algorithm achieved a sixth-place ranking in the DCASE2023 Challenge Task 2.
• Demonstrated effective ASD performance by exploiting weak attribute knowledge.
• Showcased the framework's capability in scenarios lacking strong priors.

Conclusions:

• Exploiting acoustic sensor data attributes as weak priors offers an effective ASD framework.
• The approach provides a viable solution for robust condition monitoring when strong priors are absent.
• The developed method enhances ASD system performance in practical, real-world applications.