Scholarlink

<p>Industrial Control Systems (ICS) are crucial for manufacturing operations but are vulnerable to failures, posing significant challenges to production continuity and operational reliability. Anomaly detection offers a proactive solution; however, most algorithms rely on labelled faulty data for training, which is scarce in industrial settings. Additionally, ICS data patterns are diverse, and existing approaches often misclassify novel healthy patterns as anomalies, leading to false positives. This research introduces a novel hybrid algorithm, NARXCNet, combined with an Artificial Intelligence-enabled Digital Twin to address these challenges. The proposed algorithm detects anomalies across diverse signal patterns without requiring labelled faulty signal data for training, while anomaly severity is quantified through classification into minor anomaly, severe anomaly, and complete fault using a very limited labelled dataset. Validation was conducted using real-world datasets recorded by sensors, encompassing healthy operations, minor anomalies, severe anomalies, and faulty scenarios. The performance of the proposed algorithm was compared with existing state-of-the-art machine learning, statistical and deep learning methods using the F1 score. The state-of-the-art methods achieved F1 scores of 0.78, whereas NARXCNet achieved an F1 score of 0.89 for dynamic operations, outperforming these methods. While existing algorithms performed well on data similar to that used during training, they struggled with novel patterns. In contrast, NARXCNet demonstrated exceptional performance on diverse datasets that were entirely different from those used for training. It successfully detected anomalies, quantified their severity through classification into minor anomaly, severe anomaly, and complete fault, and mitigated false positives in raw sensor datasets. This capability demonstrates the algorithm’s robust generalisation performance, underscoring its potential to significantly enhance manufacturing resilience and operational efficiency.</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

AI-driven digital twin for predictive maintenance in industrial cyber-physical systems: industrial control system case study

  • 期刊论文

  • 发表日期：

    2026年5月25日

查看全文
  • Mustafa Siddiqui,
  • Gayan Kahandawa,
  • Emanuele Principi

摘要

Industrial Control Systems (ICS) are crucial for manufacturing operations but are vulnerable to failures, posing significant challenges to production continuity and operational reliability. Anomaly detection offers a proactive solution; however, most algorithms rely on labelled faulty data for training, which is scarce in industrial settings. Additionally, ICS data patterns are diverse, and existing approaches often misclassify novel healthy patterns as anomalies, leading to false positives. This research introduces a novel hybrid algorithm, NARXCNet, combined with an Artificial Intelligence-enabled Digital Twin to address these challenges. The proposed algorithm detects anomalies across diverse signal patterns without requiring labelled faulty signal data for training, while anomaly severity is quantified through classification into minor anomaly, severe anomaly, and complete fault using a very limited labelled dataset. Validation was conducted using real-world datasets recorded by sensors, encompassing healthy operations, minor anomalies, severe anomalies, and faulty scenarios. The performance of the proposed algorithm was compared with existing state-of-the-art machine learning, statistical and deep learning methods using the F1 score. The state-of-the-art methods achieved F1 scores of 0.78, whereas NARXCNet achieved an F1 score of 0.89 for dynamic operations, outperforming these methods. While existing algorithms performed well on data similar to that used during training, they struggled with novel patterns. In contrast, NARXCNet demonstrated exceptional performance on diverse datasets that were entirely different from those used for training. It successfully detected anomalies, quantified their severity through classification into minor anomaly, severe anomaly, and complete fault, and mitigated false positives in raw sensor datasets. This capability demonstrates the algorithm’s robust generalisation performance, underscoring its potential to significantly enhance manufacturing resilience and operational efficiency.