Semantic and Engineering Knowledge based Embedding for Maintenance Work Order Clustering

In industrial maintenance, Maintenance Work Orders (MWOs) are the digital record for maintenance activities. These MWOs capture important information about an actual or impending failure event and the condition of the asset. These MWOs are therefore an essential data source for identifying different types of failures and informing maintenance strategies. Doing this failure identification involves a maintenance professional manually assigning a failure mode code (FMC) to the event described in the work order. Despite existing standard FMC lists, such as ISO 14224 for the oil and gas industry, many organizations rely on bespoke FMC lists developed internally by domain experts. These lists often suffer from having too many codes making them confusing and time-consuming to use, or have codes that are too general to be useful for analysis. As a result there is inconsistent coding within companies and across industry, limiting benchmarking, failure trend analysis and information exchange.
This recording presents work by Jadeyn Feng and the UWA NLP-TLP group to explore data-driven approaches to inform failure mode list development through the unsupervised clustering of MWOs. We test a hypothesis that sentence embeddings leveraging semantic and syntactic patterns in MWO texts and exploiting engineering knowledge of equipment function (based on the 16 inherent functions in IEC 81346-2) and equipment part relationships can inform development of a set of FMCs for a specific industrial domain. We use the MWO from the gold (3397 records) and silver (22,122 records) entity and relation annotated data sets available on the UWA NLP-TLP's open MaintIE data set. Available at github.com/nlp....
From these embeddings, we were able to perform clustering to group similar work orders together, and by analysing the resulting clusters, identify the good clusters that can be used as a generic category, each mapped to a failure mode.
We identified 9 generic categories from good clusters. Some example descriptors includes: guiding objects with material/structural failures, and generating/restricting/transforming objects having no charge, and driving/protecting object leaking.
From the bad clusters, we also gain important insights into other potential forms of similarity in maintenance work orders beyond inherent equipment function, and have narrowed down and grouped work orders for further analysis. We also show how semantic representation alone is insufficient to capture the engineering knowledge in MWOs.