INITIAL SITUATION

Erosion and cavitation detection is currently performed by experts that evaluate image and video data. The goal of the project is to semi-automize the labelling and evaluation of images and videos by implementing and adapting a state-of-the-art computer vision model for erosion and cavitation detection. Further objectives besides cost savings are the reduction of subjective influences and improvements in the quality of the evaluation.
While the focus was on the detection of erosion on images of turbine surfaces and detecting cavitation in videos of turbines inside a laboratory, the solution should also general enough to be applied to other similar use-cases.

SOLUTION

A state-of-the-art Computer Vision technology was adapted and applied to erosion and cavitation detection on water turbines, leading to significant improvements in the quality of the results while reducing the costs by semi-automating the evaluation of the images and videos.

BUSINESS VALUE

• Reduction of costs for erosion and cavitation detection
  – ~70% precision & recall after just 50 training samples, potential for full automatization
• Reduction of subjective influences in image classification
• Improvements in the quality of the results
• Active learning workflow
  – Significant speed-up for labelling data
  – Semi-automized workflow to continuously improve the computer vision model
  – Can be quickly adapted for other use-cases
• Showing the potential of state-of-the-art machine learning technology and improving the company’s competitiveness

FRAMEWORK & TOOLS

• Object detection problem, manually done by experts
• (Semi-) automation of the process using an active learning workflow
• State-of-the-art computer vision technology used
• Flexible workflow that can be applied to similar object detection problems within the company