Explainability of anomalous sound detection (ASD) decisions under domain shifts

Develop methods to explain the decisions of anomalous sound detection systems, particularly when operating under domain-shifted conditions, so that the rationale behind normal versus anomalous classifications can be understood in settings where recording environments, sensors, or source properties change between training and inference.

Background

The paper reviews challenges in anomalous sound detection (ASD) arising from domain shifts, such as changes in recording environments or sensors, which can strongly affect signals and degrade performance if not properly handled. These shifts often cause domain mismatches between source and target domains, complicating robust detection.

While domain adaptation and domain generalization strategies are surveyed, the authors identify explainability as an unresolved need: understanding why an ASD system makes a particular decision is especially important when domain shifts occur, where the system must disentangle anomaly-related factors from domain-related variability. The authors explicitly state this remains an open problem and highlight it as a key future direction.