- 92 views
- 87 downloads
Fault Diagnosis of Wind Turbine Gearboxes Based on Transfer Learning
-
- Author / Creator
- Wei, Dongdong
-
Operated under changing wind speed and harsh environment conditions, the rotating parts in wind turbine gearboxes, such as gears and bearings, will deteriorate and become faulty over time. By conducting real-time and accurate fault detection and diagnosis before significant failures occur, we can reduce the operation and maintenance costs of wind turbines. This is vital for the economic viability and stability of wind energy.
Vibration analysis based on deep learning technologies has emerged as a promising solution for fault diagnosis. Well-trained deep learning models can process large amounts of sensor data in real time and classify raw vibration signals into labels indicating different fault types, locations, and severity levels. However, these models typically require large amounts of labeled training data and may not generalize well to different working conditions or new fault modes. This limitation arises because these models are developed using the traditional supervised learning paradigm, which assumes a large and complete training dataset.
Other learning paradigms using the idea of transfer learning can be explored to address these limitations by leveraging knowledge gained from one diagnostic task or working environment to improve performance on another related task or environment. Models trained using transfer learning techniques show promise in 1) recognizing different fault classes under variable rotating speeds and load levels with high accuracy, 2) learning fault-discriminative knowledge with size-limited or incomplete dataset, and 3) providing improved interpretability and trustworthiness in the diagnosis process.
This thesis includes three topics. Topic #1 focuses on the learning paradigm of domain adaptation. A weighted domain adaptation network is proposed to adapt the diagnostic knowledge from multiple labeled datasets to an unlabeled dataset which is collected under a different working condition than those labeled datasets. Domain adversarial training and transfer learning using Maximum Mean Discrepancy are applied to align the learned features from different datasets. Topic #2 studies the open-set recognition (or open-set fault diagnosis) setting and proposes an evidential abstaining classifier that can classify both known faults that are seen in the training dataset and unknown faults which are not included in the training set. Synthetic auxiliary training samples are used to form better features and classification boundaries. Evidential learning is used to better quantify the prediction uncertainty of the model. Topic #3 explores the continual learning paradigm considering the accumulation of data and fault classes through time. A continual learning model is fine-tuned through a sequence of diagnostic tasks each features a different fault class and a different working condition. A task balanced sampling scheme is proposed to select training samples to represent previously learned tasks, and a multi-way domain adaption is conducted to adapt to different working conditions in different tasks.
The novelties explored in this research advance the development of intelligent diagnostic systems for various industrial applications, including wind turbines. The learning paradigms studied in this thesis are useful for building diagnostic systems across multiple life stages of a machine, from the early stages with only a few fault classes to the later stages with many faults to remember. Future research could explore other learning paradigms and advanced models, such as few-shot learning and the Transformer model.
-
- Graduation date
- Fall 2024
-
- Type of Item
- Thesis
-
- Degree
- Doctor of Philosophy
-
- License
- This thesis is made available by the University of Alberta Library with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.