Visual detection of flap tears on abrasive flap wheels using deep learning
Citation Link: https://doi.org/10.15480/882.17996
Publikationstyp
Journal Article
Date Issued
2026-08-15
Sprache
English
TORE-DOI
Volume
146
Start Page
2557
End Page
2571
Citation
The international Journal of Advanced Manufacturing Technology 146: 2557–2571 (2026)
Publisher DOI
Scopus ID
Publisher
Springer
Peer Reviewed
true
During abrasive flap wheel grinding, flaps may fail and tear prematurely. Such flap tears increase local tool flexibility and can impair the uniformity of material removal in automated grinding processes. To ensure process reliability, this paper proposes a novel computer vision approach for the automated detection and quantification of flap tears using deep learning. The processing pipeline is based on images showing the radial tool structure, followed by a geometric normalization that transforms the flap arrangement into a linear coordinate system. A YOLOv8-based object detection model is employed to localize abrasive flaps. To overcome class imbalances, a hybrid approach is implemented where the neural network performs a robust one-class detection, while the final distinction between intact flaps and stubs is achieved through deterministic, scale-invariant post-processing filters. The system was validated on a dataset including various tool diameters, grit sizes, and wear conditions. For tools within the trained distribution, the system achieved 96.5% classification accuracy and a Mean Absolute Error of 0.14 and 0.12 for flap and stub counts. However, performance is limited when encountering significant geometric domain shifts, such as larger tool diameters. The results demonstrate that the combination of deep learning and geometric heuristics provides a robust basis for automated flap tear detection in industrial flap wheel grinding applications.
Subjects
Abrasive flap wheel
Flap tear
Deep learning
Object detection
DDC Class
670: Manufacturing
Funding Organisations
More Funding Information
The research leading to these results received funding from the Federal Ministry for Economic Affairs and Energy (BMWE) of Germany in the project LEGATO under the Grant Number 20Y2205C as part of the Federal Aeronautical Research Programme LuFo VI-3.
Publication version
publishedVersion