Fusion-Learning of Bayesian Network Models for Fault Diagnostics.

Area of Science:

• Engineering
• Computer Science
• Data Science

Background:

• Bayesian Network (BN) models are widely used for fault diagnosis, primarily trained on quantitative sensor data.
• Existing BN models often suffer from incomplete fault coverage due to sensor limitations.
• Maintenance logs contain valuable qualitative data, including unstructured natural language with technical terms, that is underutilized.

Purpose of the Study:

• To develop a fusion-learning method for Bayesian Networks (BNs) that integrates both quantitative and qualitative data sources.
• To improve the accuracy and fault coverage of BN models for enhanced equipment diagnostics.
• To incorporate a human-in-the-loop approach for expert knowledge elicitation using natural language data.

Main Methods:

• Proposed a fusion-learning approach for BNs, combining quantitative data (sensors, metrology) and qualitative data (maintenance logs, reports).
• Developed a method to fuse two separately learned BNs derived from different data types.
• Introduced a human-in-the-loop expert knowledge elicitation strategy, leveraging logged natural language data to aid BN structure definition.

Main Results:

• The proposed fused BN model demonstrated improved diagnostic capabilities compared to individual BNs.
• The method achieved wider fault coverage by incorporating diverse data sources.
• Real-world data from uninterruptible power supply (UPS) fault diagnostics validated the efficacy of the approach.

Conclusions:

• Fusion-learning of BNs by integrating quantitative and qualitative data significantly enhances fault diagnosis.
• The human-in-the-loop approach, aided by natural language processing of maintenance logs, improves BN model development.
• The developed method offers a more comprehensive and accurate fault diagnostic solution, increasing equipment uptime and customer service.