University of Twente Student Theses
Multi-Label Classification for Sewage Pipe Anomalies
Berkhout, M.J. (2024) Multi-Label Classification for Sewage Pipe Anomalies.
|
PDF
3MB |
| Abstract: | The sewage infrastructure necessitates continuous inspections to detect anomalies, such as defects, to prevent environmental contamination and ensure public health. However, human inspectors may introduce labeling inconsistencies during this process, leading to untrustworthy maintenance advice. This thesis explores the development of data-driven automated analysis methods to enable predictive maintenance. Especially, the objective is to automate the labeling process to mitigate inconsistencies and data imbalance issues. Therefore, pre-trained image classification neural networks that improve recall and F2 scores are identified, alongside effective data pre-processing techniques for handling imbalanced data. In practice, the data consists of fish-eye camera images captured by a robotic system within sewage pipes which is first understood, prepared, and cleaned to fine-tune a multi-label classification model to detect anomalies. Through iterative experiments involving various pre-processing methods, hyperparameters, and models, the study highlights that specific image-processing algorithms can enhance model performance, particularly when using classic and residual neural networks. A primary limitation identified for future research is the need to manage fluctuations of the data sizes per observation class during resampling, increase data diversity, apply cross-validation, automate hyperparameter tuning, and implement multi-label classification networks or unwrapped tubular images to account for temporal mislabeling. |
| Item Type: | Essay (Bachelor) |
| Clients: | Rolsch Assetmanagement, Enschede, The Netherlands |
| Faculty: | EEMCS: Electrical Engineering, Mathematics and Computer Science |
| Subject: | 54 computer science |
| Programme: | Creative Technology BSc (50447) |
| Link to this item: | https://purl.utwente.nl/essays/102015 |
| Export this item as: | BibTeX EndNote HTML Citation Reference Manager |
Show download statistics for this publication
Show download statistics for this publicationRepository Staff Only: item control page