Abstract
Fault diagnosis plays a pivotal role in ensuring the health and stable operation of rotating machinery. The increasing computational resources and available data have paved the way for the application of deep learning techniques in intelligent fault diagnosis. Although deep learning-based fault diagnosis methods have shown great success, several issues remain unresolved in the fault diagnosis of rotating machinery. One concern is that the signal preprocessing methods for feature extraction rely on complicated computation and cause information loss. Furthermore, the signals used for fault diagnosis are frequently contaminated with noise, leading to significant performance degradation of existing deep learning models under noisy conditions. In addition, when dealing with multi-source domain generalization tasks in fault diagnosis, the interference between different source domains can adversely affect model convergence. The research aims to enhance the accuracy, robustness, and generalizability of fault diagnosis systems, ultimately contributing to the reliability and efficiency of rotating machinery maintenance. This thesis focuses on addressing the three major challenges: fault feature extraction with signal preprocessing, noise-robust feature learning for fault diagnosis, and multi-source domain generalization-based fault diagnosis.First, this thesis proposes Embedding Gramian Representation Network (EGR-Net) for addressing the challenge of fault feature extraction with signal preprocessing. In EGR-Net, a simple 1D-to-2D conversion method called EGR is proposed, where the converted representations have good separability. The signal and its corresponding EGR are inputted into a double-branch convolutional neural network to reduce information loss. By conducting experimental verification on gearbox and bearing datasets, the superiority of EGR-Net to other comparative fault diagnosis methods is demonstrated.
Second, this thesis proposes Multiscale Residual Attention CNN (MRA-CNN) for the challenge of noise-robust feature learning for addressing fault diagnosis. In MRA-CNN, the collected signals pass through the multiscale learning module and residual attention module successively. The fault-related locations and channels of the multiscale feature maps can be emphasized to realize noise reduction. The experimental results show the effectiveness of the proposed MRA-CNN model in fault diagnosis tasks against strong noise contamination.
Third, this thesis proposes the Gramian Time Frequency Enhancement Network (GTFE-Net) for addressing the challenge of noise-robust feature learning for fault diagnosis. In GTFE-Net, an efficient denoising algorithm called Gramian Noise Reduction is devised. Based on the periodic self-similarity of vibrational signals, the GNR transforms the noise reduction problem to the matrix rank reduction problem. The GNR is then integrated into a triple-branch CNN to form GTFE-Net, which can remove noisy information in the time and frequency domain significantly. The robustness of GTFE-Net to noise interference is demonstrated in various bearing fault diagnosis tasks.
Fourth, this thesis proposes Knowledge Distillation Domain Generalization (KDDG) for addressing the challenge of multi-source domain generalization-based fault diagnosis. In KDDG, a prototype aggregation loss is used to regularize the ensemble teacher model to learn domain-invariant features in the teacher pretraining stage, so that the student model can learn abundant knowledge from the ensemble teacher model in the transfer distillation. The out-of-domain generalization ability of KDDG is verified on the cross-machine bearing fault diagnosis task and cross-environment fan anomaly detection task.
| Date of Award | 27 Jun 2024 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Wai Shing Tommy CHOW (Supervisor) & Yixuan YUAN (Co-supervisor) |