| Author: | Xia, Liqiao |
| Title: | A systematic graph-based methodology for cognitive predictive maintenance of complex engineering equipment |
| Advisors: | Zheng, Pai (ISE) Lee, Carman (ISE) |
| Degree: | Ph.D. |
| Year: | 2024 |
| Subject: | Machinery -- Maintenance and repair Computer software -- Development -- Graphic methods Hong Kong Polytechnic University -- Dissertations |
| Department: | Department of Industrial and Systems Engineering |
| Pages: | xxiv, 201 pages : color illustrations |
| Language: | English |
| Abstract: | Maintenance is a vital aspect of ensuring the reliability, availability, and safety of machinery and systems. Traditional maintenance approaches, such as corrective and preventive measures, depend on scheduled inspections and repairs to circumvent equipment failure. However, these methods can be expensive, time-consuming, and frequently lead to unnecessary maintenance actions. In recent years, an alternative strategy has garnered significant attention: predictive maintenance (PdM). This innovative approach employs data-driven techniques to monitor equipment conditions and anticipate when maintenance is needed. By identifying potential failures before they transpire, predictive maintenance can minimize downtime, prolong equipment longevity, and enhance overall operational efficiency. Although numerous data-driven models have demonstrated promising results in PdM, they still fall short of stakeholders’ cognition-oriented expectations. Specifically, these data-driven decision processes lack physical information, diminishing the confidence level and hindering the model’s robustness. Alternatively, complex engineering equipment is characterized by a relatively long lifespan, intricate interrelationships among components, and associated failure modes. The main challenge is to utilize this information effectively to offer maintenance-oriented services, such as failure explanation, causal relationship exploration, and failure association warnings. To address these issues, graph-based approaches (GbA) with cognitive intelligence are proposed, as they excel in semantic causal inference, heterogeneous association, and visualized explanation in graph form. This study presents various GbA for the PdM process: anomaly isolation, fault diagnosis, prognosis, and maintenance decision-making. First, a graph embedding-based Bayesian network is proposed to identify the component responsible for defective behavior symptoms, a prerequisite for predictive maintenance. Initially, a fault graph (a type of knowledge graph) is constructed by combining mechanical structure knowledge and maintenance experience. Subsequently, a multi-field hyperbolic embedding method is applied to vectorize the nodes and edges, optimally preserving the logical rules within the fault graph. A Bayesian network is then integrated with the graph, considering the trained embeddings, to predict the faulty component based on available evidence. To demonstrate the effectiveness of this method, a case study was conducted on oil drilling equipment, wherein the faulty component was successfully located using the provided evidence. This case study witnesses the embedding properties and inference performance of proposed method by comparing it with other cutting-edge methods and traditional scenarios. Second, a Hypergraph Convolution Network (HGCN) based approach is proposed for fault diagnosis and prognosis, both of which pertain to sensor-based data-driven tasks. This method represents a hybrid model-based and data-driven integrated approach, which seamlessly embeds equipment’s structure and operational mechanisms as a hypergraph form within a data-driven model, accounting for interactions among equipment components. A generic model-based hypergraph construction framework is first introduced, representing the synergetic mechanism of complex equipment. Subsequently, a multi-sensory data-driven Residual-HGCN approach, combining residual block and HGCN, is presented for fault diagnosis based on a pre-defined hypergraph. Similarly, HGCN is integrated with a Gated Recurrent Unit (time series model) to address the prognosis issue. Finally, two case studies of a oil drilling equipment are conducted and compared with other typical methods to highlight the superiority of the proposed approach. Third, a maintenance-oriented knowledge graph is proposed for decision-making following the diagnosis or prognosis results. Initially, a maintenance-oriented knowledge graph is established based on a well-defined domain-specific ontology schema and accumulated maintenance data. Subsequently, an Attention-Based Compressed Relational Graph Convolutional Network is introduced to predict potential solutions and elucidate faults in maintenance tasks. Finally, a maintenance case involving oil drilling equipment is executed, wherein the proposed model is compared with other state-of-the-art models to demonstrate its effectiveness in link prediction. Lastly, the scalability and quality of graphs are critical issues for GbA in PdM, yet many stakeholders lack their own KG, let alone access to graph-based algorithms. the emergence of Large Language Models (LLM) can supplement the typically scarce empirical knowledge by generating domain-specific knowledge, but the quality of generative knowledge remains uncertain. Integrating LLM and multiple stakeholders’ wisdom promises to elevate the knowledge quality and scale, yet this synthesis inadvertently raises privacy concerns for stakeholders. To overcome this challenge, Federated Learning (FL) is employed to refine KG quality using encrypted supervision from other industrial KGs that share common entities. Initially, a multi-field hyperbolic embedding method is implemented to vectorize entities, thus enhancing low-frequency entities and capturing embedded logical rules. Subsequently, the FL framework exposes and fuses common entities while concealing the remaining entities using encryption. Next, ground truth entity alignment ensures that each KG’s embeddings occupy the same space. Finally, the KG complement method aligns and disambiguates triplets to improve overall KG quality. A case study evaluating the proposed method across various industrial KGs demonstrates its effectiveness as a practical solution for KG collaborative creation without compromising data security. |
| Rights: | All rights reserved |
| Access: | open access |
Copyright Undertaking
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item:
https://theses.lib.polyu.edu.hk/handle/200/12827