LAES: A Local Adaptive Edge-enhanced Spectrogram Method for Unsupervised Anomalous Sound Detection

Abstract: Anomalous sound detection (ASD) plays a vital role in industrial monitoring due to its non-invasive and real-time nature. However, domain shifts caused by varying machine conditions, background noise, or sensor placement hinder generalization. Existing denoising and enhancement techniques often rely on global priors or fixed rules, failing to preserve subtle anomaly-relevant cues. This paper proposes a novel method, Local Adaptive Edge-enhanced Spectrogram (LAES), which suppresses redundant background components while emphasizing informative spectral structures. Inspired by edge detection, LAES employs a sliding dual-window mechanism with local adaptive thresholding and edge value preservation, overcoming the limitations of traditional global thresholding (Canny). Furthermore, a self-adjusting adaptive window eliminates the need for machine-specific tuning, enhancing cross-domain robustness. The resulting spectrograms are fed into a lightweight autoencoder under an unsupervised setting, enabling more accurate modeling of normal patterns. On the DCASE 2024 Task 2 dataset, LAES achieves a 0.134 AUC improvement over the baseline, while its adaptive version improves by 0.09 AUC with only 0.098M parameters. Additional evaluations on DCASE 2022 and 2025 confirm LAES generalization ability and suitability for deployment on low-power industrial edge devices.

Keywords: Anomalous sound detection; edge detection; dual window; autoencoder; log-Mel spectrogram

Recommended citation: J. Lu, W. Guan, M. Zhang, et al. LAES: A Local Adaptive Edge-enhanced Spectrogram Method for Unsupervised Anomalous Sound Detection[J]. Signal Processing, 2026: 110584.
Download Paper