Hierarchical Self-Distillation with Attention for Class-Imbalanced Acoustic Event Classification in Elevators.

Area of Science:

• Engineering
• Artificial Intelligence
• Acoustics

Background:

• Acoustic anomaly detection is vital for elevator predictive maintenance and safety.
• Real-world challenges include acoustic interference and imbalanced data, hindering minority event detection.

Purpose of the Study:

• To propose a novel hierarchical self-distillation framework for robust acoustic anomaly detection in elevators.
• To address challenges of multi-source interference and data imbalance in elevator acoustic monitoring.

Main Methods:

• A hierarchical self-distillation framework with auxiliary classifiers for knowledge transfer.
• Integration of a self-attentive temporal pooling layer to manage temporal overlap.
• Utilization of focal loss in the teacher model to focus on minority class samples.

Main Results:

• The proposed model achieved over 90% accuracy and F1-score on benchmark and industrial datasets.
• Demonstrated substantial improvements in recognizing rare acoustic events.
• Validated robustness for critical elevator acoustic monitoring applications.

Conclusions:

• The hierarchical self-distillation framework effectively enhances acoustic anomaly detection in challenging elevator environments.
• The model shows significant promise for improving elevator safety and predictive maintenance through advanced AI techniques.