University of Twente Student Theses

Login

Detection of areas of interest in bending strain data for pipelines through 1D object detection

Couwenbergh, W. (2022) Detection of areas of interest in bending strain data for pipelines through 1D object detection.

[img] PDF
4MB
Abstract:Bending strain is a metric used to evaluate the build-up of stress within pipelines and subsequently their risk of breaking. Its analysis is a time-consuming and tedious process mostly involving manual evaluation which this paper aims to address. Bending strain data is 1D, similar to time series. However, in this domain detection and localization of areas of interest is still a relatively new field. U-Net has shown to work well on similar 1D data such as EEG and ECG data, but object detection algorithms have yet to be used. This paper will therefore explore the feasibility of applying a YOLO v4 based model to detect areas of interest within 1D bending strain data. The model’s performance will be evaluated using a 1D U-Net as a baseline, as it has already been established to work well on 1D data. Though, to allow for direct comparison with YOLO, the segmentation map of U-Net will be converted into a set of bounding boxes. The results show that U-Net outperforms YOLO in terms of detecting bends (with an average precision of 0.71 and 0.51 respectively), but that YOLO outperforms U-Net in detecting strain areas (with an average precision of 0.084 and 0.065 respectively). Moreover, while both models achieve comparable results (suggesting that YOLO performs on par with U-Net on 1D data), they are still found lacking in performance especially when detecting strain areas. In the best-case scenario, using an IoU threshold of 0.5, both models were able to attain an average precision of about 0.15, which is not sufficient to be used. Using the same threshold, however, both models were able to achieve an average precision of about 0.8 for bends which is a lot more promising. It was later found that some possible inconsistencies within the data and the labelling of said data might be the cause for this performance disparity. Future work using this data should therefore first aim to standardize the data and remove any inconsistencies. Thereafter, the focus of any future work should be on improving the detection performance of strain areas within bending strain data.
Item Type:Essay (Master)
Clients:
ROSEN Technology and Research Center GmbH, Enschede, Netherlands
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Subject:54 computer science
Programme:Computer Science MSc (60300)
Link to this item:https://purl.utwente.nl/essays/93957
Export this item as:BibTeX
EndNote
HTML Citation
Reference Manager

 

Repository Staff Only: item control page