Applying Natural Language Processing and Machine Learning to Support Underground Transmission Asset Management by Analysis of Maintenance Records

Work order records documenting the maintenance and defects underground components and systems constitute a valuable repository of information regarding the historical performance of these assets. However, the sheer volume and unstructured nature of such records significantly impede the efficient extraction and analysis of any valuable data.

To mitigate this challenge, researchers applied NLP and ML techniques and developed models capable of algorithmically processing historical UG defect and maintenance records from an electric utility and accurately classifying them according to a predefined taxonomy of category labels. This structured data then can be readily analyzed to enhance understanding of underground asset health and performance.

This study utilized a dataset of 17,605 work-order records on UG transmission assets. Results demonstrated strong model accuracy across all classification categories. Future work includes expanding training datasets with additional utilities’ data to improve model generalizability, refining classification hierarchies for more granular defect insights, and integrating these tools into broader asset management frameworks.