Explainable AI for post-hoc and pseudo-post-hoc predictive maintenance of governor valve actuators

Area of Science:

• Engineering
• Computer Science
• Artificial Intelligence

Background:

• Governor valve actuators (GVAs) are critical for steam turbine safety and efficiency.
• High-dimensional data and complex faults challenge traditional diagnostic algorithms.
• Existing methods lack real-time performance, reliability, and generalizability.

Purpose of the Study:

• To develop an advanced predictive maintenance framework for GVAs.
• To enhance fault diagnosis, localization, and early warning capabilities.
• To provide interpretable attribution analysis for improved maintenance workflows.

Main Methods:

• Proposed a post-hoc and pseudo-post-hoc predictive maintenance (PPPM) framework.
• Leveraged advanced machine learning and SHapley Additive exPlanations (XAI).
• Optimized fault diagnosis and early warning models using sensor network optimization and attribution analysis.

Main Results:

• PPPM optimized 50% of sensor network measurement points for fault diagnosis.
• Attribution analysis improved the real-time, generalizability, and reliability of diagnosis models.
• Optimized pseudo-supervised warning models achieved a 5.997% improvement in W-F1 score and 6.942% in AUC.

Conclusions:

• The PPPM framework effectively addresses challenges in GVA fault diagnosis and early warning.
• XAI integration provides interpretable insights for predictive maintenance.
• The proposed method significantly enhances the performance and reliability of GVA monitoring systems.