Decentralized Real-Time Anomaly Detection in Cyber-Physical Production Systems under Industry Constraints

Area of Science:

• Cyber-Physical Systems
• Industrial Automation
• Machine Learning

Background:

• Modern cyber-physical production systems require robust anomaly detection for security and efficiency.
• Existing centralized anomaly detection methods face limitations in industrial environments.
• Early anomaly detection prevents failure propagation in manufacturing.

Purpose of the Study:

• To introduce an unsupervised, decentralized, and real-time anomaly detection concept for cyber-physical production systems.
• To address the constraints of industrial setups, including communication and processing limitations.
• To provide an automated installation process requiring no expert knowledge.

Main Methods:

• Utilized several 1D convolutional autoencoders within a sliding window approach.
• Implemented a decentralized execution of anomaly detection across individual cyber-physical systems.
• Focused on unsupervised learning to handle data-driven limitations without expert intervention.

Main Results:

• Successfully evaluated the concept in a real industrial cyber-physical production system.
• Confirmed the ability to detect anomalies in all separate processes within each cyber-physical system.
• Demonstrated adequate prediction performance and fulfillment of real-time requirements.

Conclusions:

• The proposed decentralized anomaly detection concept is effective for cyber-physical production systems.
• The approach offers flexibility and meets typical industrial communication and processing constraints.
• This method shows promise for enhancing the security and reliability of industrial manufacturing through decentralized anomaly detection.