Scholarlink

Purpose <p>Anomaly detection in vibration-based time-series data is essential for identifying irregularities in industrial systems, such as 3D printers, where early detection of mechanical faults can prevent operational failures. Traditional anomaly detection models often struggle with subtle or gradually evolving anomalies, particularly in complex and dynamic environments. This study aims to address these limitations by developing a robust deep learning framework tailored for vibration-based anomaly detection and future prediction.</p> Methods <p>We propose an innovative framework that integrates an Autoencoder with a Multi-Task Gated Recurrent Unit (GRU) network. The Autoencoder extracts a compact latent representation from high-dimensional vibration sensor data, which is then processed by the GRU network. The GRU simultaneously performs two complementary tasks: (1) regression-based prediction of future vibration trends and (2) binary classification for anomaly detection. To optimize performance, we employ a dual-loss function, integrating mean squared error (MSE) for predictive accuracy and binary cross-entropy for anomaly classification. This multi-task learning approach ensures that the model effectively captures temporal dependencies while improving detection accuracy.</p> Results <p>Experimental evaluations on a real-world dataset of vibration signals from a 3D printer demonstrate that the proposed model significantly outperforms conventional methods. The model achieves a ROC of 0.9995, an F1 score of 0.9975, and minimizes False Negative (FN) samples to only 9, highlighting its superior anomaly detection and predictive capabilities. Comparisons with traditional approaches, such as Isolation Forest and Naïve Bayes, as well as other deep learning architectures, confirm the effectiveness of the proposed method for vibration-based monitoring.</p> Conclusion <p>The proposed Autoencoder-GRU framework offers a powerful solution for anomaly detection in vibration-based condition monitoring and fault diagnosis. Its ability to accurately detect anomalies while maintaining robust predictive performance makes it highly suitable for real-world industrial applications, particularly in 3D printer health monitoring and predictive maintenance.</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

EEMTL-GRU: Encoder-Enhanced Multi-Task Learning Gated Recurrent Unit for Vibrational Data-Driven Anomaly Detection

  • 期刊论文

  • 发表日期：

    2025年8月29日

查看全文
  • Abdulfattah Ba Alawi,
  • Bilal Usanmaz,
  • Ferhat Bozkurt,
  • Murad A. Rassam

摘要

Purpose

Anomaly detection in vibration-based time-series data is essential for identifying irregularities in industrial systems, such as 3D printers, where early detection of mechanical faults can prevent operational failures. Traditional anomaly detection models often struggle with subtle or gradually evolving anomalies, particularly in complex and dynamic environments. This study aims to address these limitations by developing a robust deep learning framework tailored for vibration-based anomaly detection and future prediction.

Methods

We propose an innovative framework that integrates an Autoencoder with a Multi-Task Gated Recurrent Unit (GRU) network. The Autoencoder extracts a compact latent representation from high-dimensional vibration sensor data, which is then processed by the GRU network. The GRU simultaneously performs two complementary tasks: (1) regression-based prediction of future vibration trends and (2) binary classification for anomaly detection. To optimize performance, we employ a dual-loss function, integrating mean squared error (MSE) for predictive accuracy and binary cross-entropy for anomaly classification. This multi-task learning approach ensures that the model effectively captures temporal dependencies while improving detection accuracy.

Results

Experimental evaluations on a real-world dataset of vibration signals from a 3D printer demonstrate that the proposed model significantly outperforms conventional methods. The model achieves a ROC of 0.9995, an F1 score of 0.9975, and minimizes False Negative (FN) samples to only 9, highlighting its superior anomaly detection and predictive capabilities. Comparisons with traditional approaches, such as Isolation Forest and Naïve Bayes, as well as other deep learning architectures, confirm the effectiveness of the proposed method for vibration-based monitoring.

Conclusion

The proposed Autoencoder-GRU framework offers a powerful solution for anomaly detection in vibration-based condition monitoring and fault diagnosis. Its ability to accurately detect anomalies while maintaining robust predictive performance makes it highly suitable for real-world industrial applications, particularly in 3D printer health monitoring and predictive maintenance.