Unlocking Maintenance Knowledge: An Industrial Knowledge Graph-enabled Pathway

Abstract

Maintenance activities play a critical role in the industrial systems lifecycle, encompassing failure identification and prevention, reliability analysis, and the recommendation of maintenance actions. The investigation of maintenance knowledge relies heavily on failure data collected during operational processes. Recently, the Knowledge Graph, an advanced tool that semantically represents entities and their interrelationships, has emerged as a promising approach to encapsulate failure information and corresponding maintenance actions derived from failure data. This has led to the development of the Industrial Knowledge Graph (IKG) for engineering systems, which serves as a foundational tool for investigating maintenance knowledge. However, the rapid growth of failure-related data, particularly unstructured textual maintenance logs that contain failure descriptions, maintenance times, maintenance actions, etc., poses significant challenges to traditional manual methods of constructing IKG, necessitating more automated and intelligent approaches.

This thesis explores an IKG-enabled pathway powered by Natural Language Processing (NLP) to support the maintenance of industrial systems. Specifically, the thesis aims to encapsulate and present maintenance knowledge related to engineering system failures in a user-friendly IKG, enabling the classification of incoming failure descriptions within the IKG and providing recommendations in response to failures.

The thesis is structured around four studies, each addressing a specific problem related to the IKG development and application. The first two studies focus on IKG construction, including the initialization of the IKG with incomplete knowledge and its extension through knowledge transfer. The latter two studies focus on IKG application, addressing fault diagnosis for new incoming failure descriptions and uncertainty quantification to support maintenance action recommendations.

The first work focuses on a semi-supervised method for IKG initialization from maintenance logs that capitalizes on incomplete failure knowledge. To address the need for sufficient supervision information for knowledge graph construction, a semantic module based on the Bidirectional Encoder Representations from Transformers (BERT) model is proposed to extract hidden contextual information from maintenance records and identify corresponding failure modes. The semantic module is trained by unlabeled maintenance records with the assistance of the proposed hard pseudo-label acquisition by using key phrases and the self-training algorithm. Subsequently, a taxonomy induction module is presented to automatically extract failure items and their relationships to construct IKG.

The second work explores the extension of IKG to new systems lacking available information by knowledge transfer. A significant challenge is the variation between systems, which can negatively impact the performance of failure extraction models trained on data from source systems. To address this, the Bidirectional Encoder Representations from Transformers (BERT) and Conditional Random Field (CRF) are employed for failure extraction, enhanced by iterative learning to facilitate the transfer of failure data from source systems to target systems. Furthermore, this study incorporates a rule-based pseudo-labeling module and an innovative replacement-based pseudo-sample module to deal with the problems of label errors and data imbalance during the iterative learning process. Once failure events are extracted, the associated components and corresponding failure modes are identified, enabling the automatic construction of IKG for extension.

The third study investigates the hierarchical classification-based method for IKG-informed fault diagnosis of new incoming failure descriptions, enabling multi-level granularity in failure information mining and supporting quantitative reliability analysis. To address the challenge of uneven failure mode distribution, a data augmentation module based on prompt engineering and Large Language Models (LLMs) is introduced. This module first identifies minority failure modes within the label hierarchy of IKG and subsequently generates additional synthetic failure descriptions for these underrepresented failure modes. Furthermore, leveraging the structured information of the IKG, a novel hierarchy-informed prompt-tuning approach based on the BERT model is proposed for text classification, particularly in scenarios with limited labeled samples.

The fourth study investigates a unified hierarchical text classification approach for fault diagnosis with uncertainty quantification in IKG and maintenance recommendation. Considering the diverse prompt-tuning approaches employed with various pre-trained language models, a unified prompt-tuning method with a multi-head verbalizer is proposed, accommodating encoder-only, decoder-only, and encoder-decoder types of pre-trained language models. Hierarchical information is incorporated through feature sharing and probability propagation, while entropy dropout with masking is employed to quantify classification uncertainty. In instances of low uncertainty quantification results, historical records of maintenance actions of the same failure mode are gathered as a knowledge base to put into LLM for providing recommended maintenance actions.

The feasibility and effectiveness of the proposed IKG-enabled pathway for maintenance knowledge mining and management are validated by failure data collected from operating wind farms in mainland China. Overall, this thesis provides an intelligent IKG-enabled pathway of a set of methodologies to enhance the capability of the failure data analysis for maintenance of the engineering sector, requires fewer human interventions, and supports better use of engineering information under big data scenarios. The outcomes further contribute to the optimization of operation and maintenance (O&M) processes for industrial systems by improving lifespan performance and achieving cost savings.

Date of Award	20 Aug 2025
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Min XIE (Supervisor)