Name: LUCIANO HENRIQUE PEIXOTO DA SILVA
Publication date: 23/04/2024
Examining board:
Name |
Role |
|---|---|
| CLAUDINE SANTOS BADUE | Examinador Interno |
| JOÃO PAULO PAPA | Examinador Externo |
| THIAGO OLIVEIRA DOS SANTOS | Presidente |
Summary: Diagnosing faults in electrical submersible pumps using intelligent methods is a known challenging task. The complexity of the task further increases when employing deep learning techniques for directly extracting features from the vibration signals, rather than relying on predefined human-engineered features derived from the extensive expertise of specialists in this field.
A significant limitation of this approach is the absence of foundational models for machine fault diagnosis using vibration signals. This contrast becomes evident when compared to image classification tasks, where a plethora of pre-trained networks is readily available.
To address this limitation, our work proposes a novel method that leverages RGB images generated from time-domain signals. This method combines various 2D transformations to fine-tune pre-existing image classification networks from the literature to diagnose faults in electrical submersible pumps. The results obtained from our approach indicate its superiority over our previous deep-learning method, which was based on metric learning. Furthermore, the performance of this new method is shown to be comparable to a solution based on features defined by a human expert without the need of knowledge of an expert.
