Abstract
Automated fault diagnosis can significantly reduce the workload of condition monitor- ing personnel. Deep learning based methods are one way to automate fault diagnosis. These methods can be used to infer machine faults from vibration data. However, deep learning methods based on standard supervised learning are very sensitive to changes in data distribution between training and testing. For fault diagnosis, they require fault data from all potential operating conditions. Because of this, the required amount of training data can become too extensive. Therefore, standard supervised learning based fault diagnosis models are rarely deployed in applications with variable speed and load conditions, such as ship thruster.
Few-shot learning can be used to overcome the data gathering challenges of standard supervised learning methods. Few-shot learning is a sub-field of deep learning, which is defined by tasks with very few labelled samples. Few-shot models are pretrained with domain adjacent data in such a way, that they can learn to classify the final classes from as little as one example per class. The advantage of few-shot learning is that it can be used to learn new classes with the use of existing similar data without having to gather a full new dataset.
This thesis expands previous research related to few-shot models for rotating machine fault diagnosis. A prototypical network based few-shot learning model was developed to diagnose gear faults from a powertrain. The generalization ability of the model was tested over changes in operating speed and sensor placement. In addition, model training was further optimized for rotating machine fault diagnosis with fast Fourier transform and a novel method of operating speed batching.
The results suggest that the developed few-shot learning model generalizes well to changes in operating speed and sensor placement, given just one sample from the new operating conditions. The results can be considered novel since previous research does not contain such large changes in operating conditions as this study and because of the combination of changes to the training algorithm. This study could have industrial relevance as it shows that few-shot models can significantly reduce the amount of data needed to perform deep learning fault diagnosis in changing operating conditions.
Few-shot learning can be used to overcome the data gathering challenges of standard supervised learning methods. Few-shot learning is a sub-field of deep learning, which is defined by tasks with very few labelled samples. Few-shot models are pretrained with domain adjacent data in such a way, that they can learn to classify the final classes from as little as one example per class. The advantage of few-shot learning is that it can be used to learn new classes with the use of existing similar data without having to gather a full new dataset.
This thesis expands previous research related to few-shot models for rotating machine fault diagnosis. A prototypical network based few-shot learning model was developed to diagnose gear faults from a powertrain. The generalization ability of the model was tested over changes in operating speed and sensor placement. In addition, model training was further optimized for rotating machine fault diagnosis with fast Fourier transform and a novel method of operating speed batching.
The results suggest that the developed few-shot learning model generalizes well to changes in operating speed and sensor placement, given just one sample from the new operating conditions. The results can be considered novel since previous research does not contain such large changes in operating conditions as this study and because of the combination of changes to the training algorithm. This study could have industrial relevance as it shows that few-shot models can significantly reduce the amount of data needed to perform deep learning fault diagnosis in changing operating conditions.
Original language | English |
---|---|
Qualification | Master's degree |
Awarding Institution |
|
Supervisors/Advisors |
|
Award date | 27 Feb 2023 |
Publisher | |
Publication status | Published - 20 Mar 2023 |
MoE publication type | G2 Master's thesis, polytechnic Master's thesis |
Keywords
- few-shot learning
- deep learning
- gear fault diagnosis
- vibration