BReAD: Boosting Relational Knowledge Distillation with Large Language Model for Acoustic Industrial Anomaly Detection

Acoustic Industrial Anomaly Detection (AIAD) supports audio surveillance, IIoT, and robotics. Self-supervised methods are appealing but typically require large datasets to learn generalizable representations, which conflicts with few-sample industrial regimes. Large language models (LLMs) are powerful but too computationally intensive for direct edge deployment. This work presents BReAD, a relational knowledge distillation framework for AIAD. A pretrained audio LLM serves as a teacher and induces a semantically structured contrastive representation space. A lightweight student is trained to mimic the teacher's contrastive relational knowledge by matching pairwise relations among audio clips rather than copying features. At inference, anomalies are scored using the Mahalanobis distance between teacher and student outputs. Experiments on the real-world MIMII dataset show that BReAD outperforms state-of-the-art AIAD methods.