Scholarlink

<p>Modern manufacturing typically achieves high production yields, yet even a small number of undetected defects can have disproportionate impacts, including costly product recalls and significant risks to brand reputation. Improving production efficiency and reducing resource waste have therefore become crucial priorities as the industry faces increasing environmental and economic pressures. Traditional anomaly detection methods often face limitations when anomaly data are scarce or poorly represented, which constrains their practical applicability in real-world scenarios. To address this challenge, this paper proposes an Anomaly Detection Mechanism for Multiscale-Frequency Domain Features (ADMF). The framework leverages raw high-frequency sensor data and applies Adaptive Continuous Wavelet Transform (ACWT) to extract localized transient patterns, while Mean Amplitude Feature (MAF) aggregates energy distributions across frequency bands to highlight defect-sensitive regions. A feature fusion step integrates ACWT and MAF representations, which are then used as input to a ResGANomaly detection model trained on normal samples. Experimental results on welding data demonstrate that ADMF improves Balanced Accuracy by 8.01% compared with baseline GAN-based methods, confirming its effectiveness in detecting defects with limited anomaly data. These findings suggest that ADMF provides a practical solution for continuous manufacturing environments, where enhanced anomaly detection helps reduce defective outputs, minimize production risks, and safeguard long-term reputation.</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

An Anomaly Detection Mechanism Based on Multiscale Frequency Features for Transient Signals

  • 期刊论文

  • 发表日期：

    2025年8月1日

查看全文
  • Hsiao-Yu Wang,
  • Cheng-Hui Chen

摘要

Modern manufacturing typically achieves high production yields, yet even a small number of undetected defects can have disproportionate impacts, including costly product recalls and significant risks to brand reputation. Improving production efficiency and reducing resource waste have therefore become crucial priorities as the industry faces increasing environmental and economic pressures. Traditional anomaly detection methods often face limitations when anomaly data are scarce or poorly represented, which constrains their practical applicability in real-world scenarios. To address this challenge, this paper proposes an Anomaly Detection Mechanism for Multiscale-Frequency Domain Features (ADMF). The framework leverages raw high-frequency sensor data and applies Adaptive Continuous Wavelet Transform (ACWT) to extract localized transient patterns, while Mean Amplitude Feature (MAF) aggregates energy distributions across frequency bands to highlight defect-sensitive regions. A feature fusion step integrates ACWT and MAF representations, which are then used as input to a ResGANomaly detection model trained on normal samples. Experimental results on welding data demonstrate that ADMF improves Balanced Accuracy by 8.01% compared with baseline GAN-based methods, confirming its effectiveness in detecting defects with limited anomaly data. These findings suggest that ADMF provides a practical solution for continuous manufacturing environments, where enhanced anomaly detection helps reduce defective outputs, minimize production risks, and safeguard long-term reputation.