Thesis
Smart asset management: data-driven approaches for degradation prediction and maintenance policy
- Creator
- Rights statement
- Awarding institution
- University of Strathclyde
- Date of award
- 2025
- Thesis identifier
- T17391
- Person Identifier (Local)
- 202161377
- Qualification Level
- Qualification Name
- Department, School or Faculty
- Abstract
- With the use of advanced technologies and digital tools, asset management is advancing to a smarter stage with increasing availability of operational data. However, this also presents challenges in effectively utilising data to drive informed decisions. This thesis aims to address two critical tasks within the context of smart asset management through data-driven approaches: predicting degradation and developing maintenance policies. Three parts under different scenarios are involved in this thesis. To address the challenge of degradation prediction with limited observations, the first part of this thesis introduces a method that generates, selects, and reweights synthetic data to enhance prediction performance. Unlike existing methods that mix synthetic and real data without considering sample selection or weighting, this approach uses multiple data augmentation methods to generate time-series data, then applies an influence function to select effective synthetic samples, followed by reweighting via gradient descent. To further improve the performance of the deep learning algorithm, transfer learning is applied by pre-training the deep learning model and then fine-tuning it with real data. Numerical experiments demonstrate the frameworkâs effectiveness, especially for highly stochastic degradation data. The second part of the thesis explores a data-driven preventive maintenance problem where the true time-to-failure model is unknown, but past time-to-failure data and working conditions are observable. Traditional estimate-then-optimise methods separate estimation from optimisation, potentially propagating errors into the decision making process. To overcome this, an end-to-end framework has been proposed to directly determine the optimal preventive replacement time under specific working conditions without assuming a time-to-failure model. The end-to-end approach treats historical working conditions as features, mapping them to optimal maintenance decisions by minimising the maintenance cost rate. Supervised learning algorithms then train these features against the optimal decisions. The findings suggest that end-to-end learning can reduce error propagation and that a linear model, when aligned with the learning objective, may outperform more complex alternatives. Lastly, the third part of the thesis presents a condition-based maintenance policy considering component heterogeneity and dynamic working conditions. Bayesian Poisson and linear regressions are applied to analyse the shock occurrence and magnitude, updating parameters with new observations during online monitoring. The maintenance planning problem is framed as a Markov decision process. This approach establishes a tractable degradation model that accounts for heterogeneity and dynamic conditions and explores the structural properties of optimal maintenance policy. A heuristic algorithm based on the most likely distribution has been introduced to reduce computational complexity. Results reveal that maintenance thresholds, which serve as control limits, fluctuate with covariates and demonstrate the advantages of the proposed policy in handling varying working conditions and component heterogeneity. Collectively, this thesis tackles the challenges of degradation prediction and maintenance optimisation in smart asset management through data-driven approaches. The proposed methods offer improved solutions for degradation prediction with limited data, and maintenance policies with dynamic working conditions and heterogeneity, contributing to more effective and informed asset management strategies.
- Advisor / supervisor
- Liu, Bin
- Wall, Lesley
- Resource Type
- DOI
Relations
Items
| Thumbnail | Title | Date Uploaded | Visibility | Actions |
|---|---|---|---|---|
|
|
PDF of thesis T17391 | 2025-06-13 | Public | Download |