Scholarlink

<p>Industrial noise suppression constitutes a vital preprocessing step in machine condition monitoring, especially when dealing with noise stemming from heterogeneous environmental sources. However, many existing methods depend on access to clean reference signals, which constrains their applicability in unsupervised or real-world scenarios. To address this limitation, we propose the dual-domain multi-scale autoencoder (DDMS-AE)—an unsupervised, end-to-end framework that eliminates the need for clean training data. The proposed model operates concurrently in the time-frequency and waveform domains, employing multi-scale learning strategies to effectively separate noise from machine sounds without relying on predefined noise distribution assumptions. To empirically validate the effectiveness of the proposed method, we adopt the MIMII dataset and Ottawa bearing dataset, two widely recognized benchmarks in the development of automated facility maintenance systems. Experimental results demonstrate that DDMS-AE significantly improves anomaly detection accuracy by effectively suppressing complex real-world noise, thereby advancing the application of deep learning in industrial acoustic monitoring.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Deep learning-based noise suppression for industrial environments without clean data

  • 期刊论文

  • 发表日期：

    2025年11月23日

查看全文
  • Shuo Wang,
  • Jun Li,
  • Guangyu Hou,
  • Dezhi Yuan

摘要

Industrial noise suppression constitutes a vital preprocessing step in machine condition monitoring, especially when dealing with noise stemming from heterogeneous environmental sources. However, many existing methods depend on access to clean reference signals, which constrains their applicability in unsupervised or real-world scenarios. To address this limitation, we propose the dual-domain multi-scale autoencoder (DDMS-AE)—an unsupervised, end-to-end framework that eliminates the need for clean training data. The proposed model operates concurrently in the time-frequency and waveform domains, employing multi-scale learning strategies to effectively separate noise from machine sounds without relying on predefined noise distribution assumptions. To empirically validate the effectiveness of the proposed method, we adopt the MIMII dataset and Ottawa bearing dataset, two widely recognized benchmarks in the development of automated facility maintenance systems. Experimental results demonstrate that DDMS-AE significantly improves anomaly detection accuracy by effectively suppressing complex real-world noise, thereby advancing the application of deep learning in industrial acoustic monitoring.