Enhancing semantic search using ontologies: A hybrid information retrieval approach for industrial text

Abstract

Despite the increased focus on data in Industry 4.0, textual data has received little attention in the production and engineering management literature. Data sources such as maintenance records and machine documentation usually are not used to help maintenance decision-making. Available studies mainly focus on categorizing maintenance records or extracting meta-data, such as time of failure, maintenance cost, etc. One of the main reasons behind this underutilization is the complexity and unstructured nature of the industrial text. In this study, we propose a novel hybrid information retrieval approach for industrial text using multi-modal learning. Maintenance operators can use the proposed system to query maintenance records and find similar solutions to a given problem. The proposed system utilizes heterogeneous (multi-modal) data, a combination of maintenance records, and machine ontology to enhance semantic search results. We used the state-of-the-art Large Language Models (LLMs); BERT (Bidirectional Encoder Representations from Transformers) for textual similarity. For similarity among ontology labels, we used a modified version of Wu-Palmer’s similarity. A hybrid weighted similarity is proposed, incorporating text and ontology similarities to enhance semantic search results. The proposed approach was validated using an open-source dataset of real maintenance records from excavators collected over ten years from different mining sites. A retrieval comparison using only text and multi-modal data is performed to estimate the proposed system’s effectiveness. Quantitative and qualitative analysis of results indicates a performance improvement of 8% using the proposed hybrid similarity approach compared to only text-based retrieval. To the best of our knowledge, this is the first study to combine LLMs and machine ontology for semantic search in maintenance records.