Automatic knowledge extraction from industrial maintenance short text using deep learning

The Challenge

Maintenance is essential for efficient operation and preventing asset failures which can present safety and environmental hazards, risk regulatory non-compliance and lead to significant financial loss. Within this context, the unstructured data captured by operations and maintenance technicians and stored in historical maintenance work order records is of immense value. The inherent value of these unstructured texts proves compelling case for automated Information Extraction (IE) as the task of manually processing the data is fraught with challenges, such as large volume of generated content, unique linguistic characteristics and low quality of textual data, and complex technical nature of maintenance domain itself.

Key Findings

The thesis addresses these challenges by developing a novel annotation software, language resources, and methods tailored to applying Natural Language Processing (NLP) to short text descriptions of maintenance work order records. Two novel tools are presented:

• LexiClean: a multi-task annotation tool for lexical normalisation, enhancing text quality and preserving privacy; and
• QuickGraph: enabling collaborative, multi-task IE.

Leveraging of these tools, the development of MaintNorm, MaintIE and MaintNorm IE tools set the benchmarks for deep learning-based lexical normalisation and IE within the maintenance domain, specifically, entity recognition and relation extraction.

Benefits to WA

The research proposes a novel method for constructing maintenance property graphs that systemically capture both explicit information and implicit domain knowledge through a combination of automated extraction and expert-guided annotation. The thesis advances both the field of NLP within industrial maintenance and the practical extraction of actionable insights from maintenance records, paving the way for more informed asset management decision-making.

Report DOI

doi.org/10.26182/hhg8-ra38