Scholarlink

Purpose <p>Multistage compressors are vital industrial assets, and their continuous operation makes timely fault detection crucial to avoid unexpected downtime. Traditional vibration-based predictive maintenance methods largely depend on supervised learning techniques, which require extensive labeled fault data—data that are rarely available in real-world industrial environments. To address this limitation, this study presents a self-supervised learning (SSL) framework for vibration-based condition monitoring of multistage compressors, allowing fault-relevant features to be learned directly from unlabeled data.</p> Methods <p>The proposed method employs a contrastive Siamese convolutional architecture with domain-specific temporal augmentations, followed by a lightweight supervised classifier trained using only a small number of labeled samples. Furthermore, an Entropy Divergence Rate (EDR) metric is introduced to capture distributional shifts in the learned feature space, enabling early detection of anomalous behavior. The framework is validated using a six-month in-house multistage compressor dataset along with an adapted Case Western Reserve University bearing dataset.</p> Results <p>The results show an F1-score of 0.93 and a 28% improvement in early fault detection compared with FFT-based and fully supervised deep learning approaches.</p> Conclusion <p>Overall, the proposed framework substantially reduces reliance on labeled data while enhancing early fault recognition, making it well suited for scalable and practical industrial predictive maintenance applications.</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

Self-Supervised Vibration Analytics for Predictive Maintenance of Multistage Compressors

  • 期刊论文

  • 发表日期：

    2026年4月10日

查看全文
  • Rajagopal M,
  • Shashank R,
  • Shreeshanth R

摘要

Purpose

Multistage compressors are vital industrial assets, and their continuous operation makes timely fault detection crucial to avoid unexpected downtime. Traditional vibration-based predictive maintenance methods largely depend on supervised learning techniques, which require extensive labeled fault data—data that are rarely available in real-world industrial environments. To address this limitation, this study presents a self-supervised learning (SSL) framework for vibration-based condition monitoring of multistage compressors, allowing fault-relevant features to be learned directly from unlabeled data.

Methods

The proposed method employs a contrastive Siamese convolutional architecture with domain-specific temporal augmentations, followed by a lightweight supervised classifier trained using only a small number of labeled samples. Furthermore, an Entropy Divergence Rate (EDR) metric is introduced to capture distributional shifts in the learned feature space, enabling early detection of anomalous behavior. The framework is validated using a six-month in-house multistage compressor dataset along with an adapted Case Western Reserve University bearing dataset.

Results

The results show an F1-score of 0.93 and a 28% improvement in early fault detection compared with FFT-based and fully supervised deep learning approaches.

Conclusion

Overall, the proposed framework substantially reduces reliance on labeled data while enhancing early fault recognition, making it well suited for scalable and practical industrial predictive maintenance applications.