| Author: | Iqbal, Misbah |
| Title: | Intelligent fault diagnosis of rotating machines using transfer learning |
| Advisors: | Lee, K. M. Carman (ISE) Ren, J. Z. (ISE) |
| Degree: | Ph.D. |
| Year: | 2026 |
| Department: | Department of Industrial and Systems Engineering |
| Pages: | xxiv, 257 pages : color illustrations |
| Language: | English |
| Abstract: | Rotating machines are the backbone of many production systems and manufacturing processes due to their high efficiency and durability. However, many malfunctions— contribute to reduced efficiency and premature mechanical failure, resulting in unscheduled downtime and financial losses. This thesis seeks to establish transfer learning approaches for fault diagnostics in rotating machines, especially under dynamic operational conditions, by leveraging knowledge learned from one domain and applying it to another rather than generalizing in dynamic environment by traditional deep learning approach. This research explores innovative transfer learning methods for enhancing the accuracy and reliability of fault diagnosis systems in addressing complex diagnostic challenges. This thesis primarily conducts an empirical investigation using three widely utilized bearing fault datasets as the foundation for the experiments performed. It is found that under same-domain conditions, conventional data-driven approaches yield promising classification accuracy for fault diagnosis across diverse motor speed ranges; however, under cross-domain conditions, the performance significantly declined, emphasizing the limitations of conventional models in addressing distributional shifts. This study highlights the necessity for more robust and adaptable methods, such as transfer learning, which allows models to utilize knowledge from one domain and effectively adapt to other domains. To address the challenges of domain shifts caused by time-varying speed conditions and data scarcity, a parameter-based transfer learning strategy is applied to diagnose specific faults, such as unbalance and misalignments, which are often underexplored in existing research. This study proposes an intelligent, data-driven approach based on the Triplex Transfer Long Short-Term Memory (TTLSTM) network, which leverages transfer learning and fine-tuning strategies. The proposed methodology uses empirical mode decomposition to extract pertinent features from raw vibrational signals and utilizes Pearson correlation coefficients for feature selection. L2 regularization transfer learning is utilized to mitigate the overfitting problem caused by limited target labeled data. Compared with traditional transfer learning approaches, the proposed model excels in identifying machinery faults and generalizes well to target domain data under varying working conditions, proving its applicability in real-world applications. Domain shifts in data distributions prevent transfer learning models operating at constant speed conditions from detecting invariant features and reducing their generalization efficacy. Moreover, the current TL methods for varying speed scenarios mainly focus on aligning the marginal data distribution while neglecting the influence of class-specific feature alignment on the diagnostic accuracy. To address these limitations, a novel end-to-end semi-supervised marginal and conditional distribution alignment network (Semi-MCDA-Net) is developed, particularly for insufficient labeled target data. This method integrates multi-kernel Maximum Mean Discrepancy and Wasserstein distance to develop a unified domain alignment module, systematically applied across multiple convolutional layers of a shared 1D-CNN. Two case studies are carried out to verify the efficacy and generalizability of the Semi-MCDA-Net method. Most of the existing research assumes closed-set domain adaptation, wherein fault classes are identical across domains. However, in real-world industrial applications, fault diagnosis is an open task, which means an unknown fault can arise during the testing phase that was not seen during the training. Moreover, the presence of noisy fault labels in the target domain and the unpredictability of the machine operations make fault diagnosis difficult. In light of this, a novel semi-supervised open-set domain adaptation with noisy target labels (SS-OSDA-NTL) is developed, which integrates multi-kernel MMD loss, inter-class discriminability loss, and intra-class compactness loss functions to align the marginal and conditional distributions across the domains, therefore improving the transferability and discriminability. To reduce the impact of noisy labels in the target domain, a bootstrapped CE loss function is implemented. Samples accompanying an unknown target class are identified using a confidence-based boundary. Multiple experiments are conducted to verify that the diagnostic model (SS-OSDA-NTL) developed for open-set conditions not only recognizes known faults but also identifies the unknown faults with different noise rates. |
| 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/200/14370