Resource‐Aware Contrastive Scattering Meta‐Learning for Efficient Few‐Shot Acoustic Anomaly Detection

Abstract

Motivated by the increasing complexity of cyber‐physical systems and growing interest in resource‐aware artificial intelligence, this paper addresses the challenges related to anomaly detection using limited data with online adaptation. Specifically, we address challenges related to concept drift and generalization in acoustic anomaly detection applications. Within this context, this paper introduces a novel resource‐aware Contrastive Scattering Meta‐Learning (CSML) framework with an application to the few‐shot acoustic anomaly detection problem. The proposed framework leverages the appealing properties of wavelet scattering networks and the power of contrastive learning within a meta‐learning framework. Extensive experiments demonstrate its robustness in domain‐shifted and noisy environments, yielding promising results against state‐of‐the‐art methods on benchmark datasets from DCASE challenges (2020–2022) while maintaining a lightweight for 50 K learnable parameters.