Scholarlink

In recent days, predictive maintenance is a major component in modern industrial systems, which minimizes downtime and operational costs by enabling early detection of equipment anomalies. Incomplete or noisy, equipment behavior and failures are some issues faced in anomaly detection. So, an innovative approach was developed using a Temporal Convolutional Neural Network (TCN) with a self-attention mechanism for predictive maintenance of anomaly detection. Initially, from Numenta Anomaly detect Benchmark (NAB), the data is collected that contains Information Technology (IT) and industrial sensors metrics. Furthermore, using Z-score normalization, collected data is preprocessed, and temporal data is divided into seasonal, and residual components using Seasonal Trend decomposition (STL). Finally, data is fed to Deep Convolutional Autoencoder (DCAE) model for feature extraction, and TCN with self-attention (SA) is used to detect the anomalies. The results shows that the proposed TCN-SA achieved better results with precision (99.93%), recall (99.89%), and f1-score (99.79%), respectively.

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

Anomaly Detection for Predictive Maintenance Using Temporal Convolutional Neural Network with Self-attention Mechanism

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Venkata Ramana Kaneti,
  • K. V. Sheelavathy,
  • Debashis Rudra Sarma,
  • V. Prabhudeva,
  • G. U. Vasanthakumar

摘要

In recent days, predictive maintenance is a major component in modern industrial systems, which minimizes downtime and operational costs by enabling early detection of equipment anomalies. Incomplete or noisy, equipment behavior and failures are some issues faced in anomaly detection. So, an innovative approach was developed using a Temporal Convolutional Neural Network (TCN) with a self-attention mechanism for predictive maintenance of anomaly detection. Initially, from Numenta Anomaly detect Benchmark (NAB), the data is collected that contains Information Technology (IT) and industrial sensors metrics. Furthermore, using Z-score normalization, collected data is preprocessed, and temporal data is divided into seasonal, and residual components using Seasonal Trend decomposition (STL). Finally, data is fed to Deep Convolutional Autoencoder (DCAE) model for feature extraction, and TCN with self-attention (SA) is used to detect the anomalies. The results shows that the proposed TCN-SA achieved better results with precision (99.93%), recall (99.89%), and f1-score (99.79%), respectively.