A Custom Word Embedding Model for Clustering of Maintenance Records
Abhijeet Sandeep Bhardwaj, Akash Deep, Dharmaraj Veeramani, Shiyu Zhou
Abstract
Maintenance records of industrial equipment contain rich descriptive information in free-text format, such as involved parts, failure mechanisms, operating conditions, etc. Our objective is to leverage this unstructured textual information to identify groups of similar maintenance jobs. In this article, we use a natural language based approach and propose a novel custom word embedding model, which utilizes two sources of information, first, maintenance records collected from in-field operations and second, industrial taxonomy, to effectively identify clusters. The advantages of our model include combined use of semantic and taxonomic sources of information for clustering, one step/simultaneous training, which enables knowledge sharing between the two information sources and reduces hyperparameters, and no dependence on third-party data. We demonstrate the efficacy of our model for cluster identification using a real-world dataset. The results show that simultaneous incorporation of semantic and taxonomic information enables accurate extraction of contextual insights for improving maintenance decision-making and equipment reliability.