In the past, the component was automatically inspected for cracks using a camera. The detection of a crack by blob analysis under stable process conditions is over 99 percent. However, problems occur mainly due to pseudo faults. These pseudo faults from so-called "cloudiness" are caused by roughened surfaces and discontinuities in the oil film due to tool wear or chips. These were falsely identified as cracks by this method. The main objective of the project was to guarantee a sustainable and accurate detection of faulty press-fit elements as well as to avoid pseudo faults.
BMW AG
> PoC for Crack Detection Using Deep Learning Algorithms
> Machine Learning-Assisted Lightweight Robots
The BMW Group, including its brands BMW, MINI, Rolls-Royce and BMW Motorrad, is the world's leading premium manufacturer of automobiles and motorbikes. Furthermore, it offers premium financial and mobility services. The company employs around 125,000 people worldwide.
PoC for Crack Detection Using Deep Learning Algorithms
Initial Situation
Solution
Robotron has provided the RCV (Realtime Computer Vision) platform for crack detection.
Key success factors for the project were the identification of the cracks by means of object recognition and the creation of an approach using deep learning methods.This approach is robust enough to recognize that images with pseudo defects (e.g. caused by a chip or an oil drop) were not cracks.
To start off an object recognition model was trained, which was optimized for different defect classes. With this model it is possible to distinguish between cracks and particle contamination or oil drops on the press-fit part.
RCV platform communicates with the production robot and sends an IO or NIO signal to the robot after identification for the respective pressed part. In addition, the robot's control system was modified for the project in such a way that the component detected as faulty could be automatically removed from the process. Thus saving a further manual step.
Benefits
Using the RCV solution the rejection rate of faulty pressed parts could be reduced by a factor of ten from approx. two percent to 0.2 percent. This decrease contributes significantly to the reduction of manual re-inspection. The pseudo-defects that occure are reliably identified using the RCV solution. The complex models ensure that the pseudo-defects are ignored during the scrap determination process. In addition, the detection of pseudo errors supports the detection of anomalies in earlier process steps.