TL;DR
This paper chronicles the 30-year evolution of the Process Diagnosis System (PDS), an expert system shell developed in the early 1980s for predictive maintenance of power generation equipment. PDS revolutionized diagnostics by mimicking human expert reasoning to detect subtle anomalies, preventing costly failures and downtime, and continues to be a cornerstone of modern power plant monitoring.
What Problem Does This Paper Address?
In the early 1980s, power generation equipment like steam turbines and generators were massive, complex machines with thousands of sensors. While control systems could detect critical threshold breaches, they struggled to identify subtle, gradual deviations that often preceded major malfunctions. Repairing these units could take weeks, leading to significant downtime and economic losses. Westinghouse sought a way to leverage Artificial Intelligence to model their engineers' expertise, enabling early detection and prediction of potential failures, thereby allowing for proactive maintenance and avoiding forced outages. This required a system capable of handling vast amounts of sensor data, diagnosing not only the equipment but also the sensors themselves, and adapting to different operating states.
Methods
The core of PDS is an expert system shell that combines elements of backward chaining (like MYCIN) with forward chaining production systems. Key innovations include:
- Hybrid Rule Architecture: PDS uses a network of sensor nodes and intermediate hypothesis nodes, with rules propagating certainty factors. This allows for both diagnostic reasoning and the ability to explain findings.
- Context-Aware Rules: Rules are associated with specific operating contexts (e.g., "start-up," "steady-state"), ensuring that only relevant logic is applied at any given time.
- Time Series Analysis Integration: PDS processes sensor data over time, not just instantaneous readings. "Storage nodes" and "reading-transform" rules were introduced to store and analyze successive sensor values, enabling the detection of subtle trends and degradation.
- Sensor Diagnosis and Robustness: Recognizing that sensors themselves can fail or degrade, PDS incorporates mechanisms to diagnose sensor health. The system is designed to continue functioning even with a subset of failed sensors, using "parametric-alteration" rules to adjust the influence of degraded sensor data.
- Knowledge Representation: The system utilized SRL (System Representation Language) for knowledge representation, allowing for inheritance of sensor and rule classes, facilitating modularity and scalability.
- Composite Sensors: This feature allowed the combination of readings from multiple, overlapping sensors into a single "virtual sensor" reading, using techniques like voting and auctioneering.
Datasets
Not explicitly discussed in the paper. The paper focuses on the development and application of the PDS system itself, using real-world sensor data from power plants for its diagnostic capabilities.
Key Insights and Ideas
- The "Subtle Anomaly" Problem: The most significant insight was that predictive maintenance hinges on detecting gradual deviations before they cross critical thresholds. This requires sophisticated time-series analysis and expert reasoning, not just simple alarm triggers.
- Sensor Health is Crucial: The realization that sensor failures are more common than equipment failures and can directly lead to incorrect diagnoses highlighted the necessity of building sensor diagnosis into the system.
- Modularity and Reusability: The development of subsystems and the ability to instantiate rule bases with different suffixes (e.g., for different boiler types) demonstrated a powerful approach to managing large, complex knowledge bases and adapting them to diverse equipment.
- Evolutionary Design: PDS wasn't a static product. It underwent multiple rewrites and improvements, driven by new technology (gas turbines), market changes (deregulation), and customer feedback, showcasing the adaptability of the core concept.
- The Power of Context: Explicitly defining operating contexts for rules proved vital for accurate diagnosis, as equipment behavior and failure modes differ significantly between start-up, steady-state, and shutdown.
Key Results
- Commercial Operation: PDS went into commercial operation in 1985, monitoring seven power plant generators from a centralized diagnostic center.
- Scalability: The current PCPDS supports Siemens Power Diagnostics Center with centralized rule-based monitoring of over 1200 gas turbines, steam turbines, and generators worldwide.
- Efficiency Gains: The system's ability to identify subtle issues led to improvements, such as a calculated 1% increase in hydrogen purity for generators, directly justifying the service cost.
- Reduced Downtime: By predicting malfunctions, PDS enabled scheduled maintenance, significantly reducing costly forced outages.
- Comprehensive Monitoring: The system now monitors over 1,200 turbines and generators, processing over 5.5 million sensor-hours daily.
Limitations and Future Work
- Processing Speed: While the system is effective, improving the processing speed of the expert system is an ongoing area of development, particularly for the rule development and testing phases.
- Transparency: Enhancing the transparency of the encoded knowledge base for monitoring engineers is a current focus, aiming to make the system's reasoning even more accessible.
- Data Quality and Customization: Adapting the rule base to new units presented challenges due to variations in data quality, instrumentation, and unit customization.
- Future Work: The paper hints at continued evolution, including the integration of empirical modeling tools like Support Vector Machines (SVMs) to complement the rule-based system, further enhancing diagnostic capabilities. The ongoing growth of the gas turbine fleet also drives continuous development.
Summary updated