On the Diagnosis of Cyber-Physical Production Systems

Oliver Niggemann, Volker LohwegAAAI 2015 · DOI

Introduction

As manufacturing evolves into Cyber-Physical Production Systems (CPPSs)—highly distributed networks of sensors, actuators, and software—the task of diagnosing faults becomes a massive bottleneck. Traditional diagnosis relies on "first-principles" models, such as complex differential equations or manual rule sets created by engineers. However, modern plants are too dynamic and generate too much data for humans to model manually.

The Shift in Methodology: A Research Roadmap

The core challenge in CPPS diagnosis lies in three major gaps: the excessive effort required for manual modeling, the hybrid nature of signals (mixing discrete events with continuous values), and the mathematical explosion of root cause analysis in high-dimensional systems.

To address these, a transition is required from local, phenomenological approaches to system-wide, model-based approaches using learned generic physical models.

Comparison of state-of-the-art diagnostic approaches versus the proposed roadmap for CPPS.

This roadmap emphasizes moving away from manually-engineered component-granular models toward models learned automatically from system observations.

Automated Learning of Models

Instead of writing static rules, the proposed framework utilizes algorithms like RTI+ and BUTLA to observe normal system behavior and automatically generate "Timed Automata." These are state machines that capture not just the sequence of events, but the precise timing of transitions.

Learned Timed Automata models for a manufacturing plant demo.

In CPPSs, timing is a first-class citizen; a process that occurs correctly but 50ms too late is considered a fault. By learning these temporal signatures, the system can detect anomalies that traditional logic-based models would miss. Furthermore, by extending these to Hybrid Timed Automata, the system can learn "power signatures" of specific components, enabling energy-based anomaly detection.

Key Results and Insights

The effectiveness of these learned models has been demonstrated across various domains:

  • Timed Automata Performance: In process industry tests, Timed Automata achieved 98.5% accuracy, significantly outperforming traditional Neural Networks (41.2%) and Decision Trees (86.9%).
  • Energy Modeling: Hybrid models achieved over 80% accuracy in predicting energy consumption.
  • Wind Turbine Monitoring: A PCA-based approach reached a 90.13% balanced accuracy, proving effective for continuous system monitoring.
  • Component Disaggregation: The MoSDA algorithm successfully estimated individual drive power consumption from total system data with an average error of only 1.7% to 7.4%.

A critical insight is that the goal of AI in this context is often localization rather than full causal reasoning. If an algorithm can pinpoint an anomaly to a specific drive among thousands, human experts can typically identify the root cause quickly.

Conclusion and Future Directions

The future of CPPS diagnosis lies in unsupervised learning from "normal" data. While significant progress has been made in offline learning, future research must focus on online adaptation (learning "on the fly") and better integration of human-in-the-loop cognitive capabilities to support technicians in complex environments.

Summary updated

BibTeX
@inproceedings{niggemann2015diagnosis,
  author = {Oliver Niggemann and Volker Lohweg},
  title = {{On the Diagnosis of Cyber-Physical Production Systems}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2015},
  doi = {10.1609/aaai.v29i1.9762},
}
Figure