Implementation and testing of an image-based quality inspection process for laminated safety glass

Defects in laminated safety glass can not only obscure the view, but can also impair the safety of people and objects behind them if they are too pronounced. For this reason, GuS intends to inspect every manufactured workpiece intensively and automatically for streaks, scratches and other defects. The inspection currently takes the form of a complex, manual visual inspection, which requires experienced, qualified personnel. In a preliminary study as part of it'sOWL, it was demonstrated that all types of defects in laminated safety glass with integrated heating lines can be detected using image processing based on deep learning. As part of this project, the previously developed inspection concept was automated as a prototype and set up in the immediate vicinity of production. In addition, the inspection area was scaled to complete glasses of different shapes and characteristics. In the preliminary study, only sections (30x30cm) were tested for better handling. The on-site setup enabled the quality assurance staff to test the automated inspection close to the process and familiarize themselves with the inspection procedure. The test was carried out step by step, so that deviations and error detections led to new training of the neural networks in coordination with quality assurance. This allowed the recognition quality and accuracy to be increased. The tests confirmed that automated testing is suitable for in-process testing. However, they also revealed further fields of action and optimization possibilities that are to be addressed in subsequent follow-up projects.

Defects in laminated safety glass can not only obscure the view, but can also impair the safety of people and objects behind them if they are too pronounced. For this reason, GuS intends to carry out intensive automated testing of every manufactured workpiece for streaks, scratches and other defects. For the finished glasses, the inspection currently takes the form of a complex, manual visual inspection, which requires experienced, qualified personnel. In a preliminary study as part of it'sOWL, the extent to which laminated safety glass with integrated heating lines can be inspected using image processing based on deep learning was examined. As part of the project, it was found that systems available on the market cannot reliably hide heating lines in laminated glass. The project, on the other hand, showed that this can be achieved using deep learning methods. In addition, it was possible to detect all the marked defects under suitable lighting and camera settings and under laboratory conditions. Due to the accuracy of the process, contamination and dust can lead to false detections. As these influencing factors cannot be ruled out in the harsh production environment, this must be checked on site. In addition, only sections (30x30cm) of glass laminates were tested in the preliminary study for handling reasons. The different shapes, sizes and characteristics (e.g. light transmittance) of the complete glasses represent a practicable challenge in the implementation of the test. Finally, the extent to which automated testing and the associated handling of heavy glass assemblies can be integrated into the production process as time-efficiently as possible should be examined.

The aim was to set up and test an automated deep learning-based quality inspection system for laminated safety glass in GuS production, consisting of light guidance, camera positioning and control, the necessary software and hardware and the mechanical positioning of the test samples. This setup was to be tested in the context of production in cooperation with the quality assurance department of GuS. On the basis of these tests, the practicable suitability of the process was to be checked and tolerances and defect delimitation fine-tuned. The aim was also to optimize the neural networks by retraining them. The test parameters were to be determined as part of the tests so that all product variants of GuS could be tested equally in terms of size, composite structure and shape. Following a requirements analysis, a room concept was drawn up with regard to the selection and positioning of the individual components required and procurement processes were started. At the same time, the software development for the automated implementation of the image processing steps began. Once production space became available, the system was set up and commissioned at GuS in Lübbecke. Individual aspects and functions were first tested on a modular basis. This was followed by integration and the creation of a graphical user interface. This was followed by the cooperative test phase in collaboration with GuS quality assurance, which was concluded with a presentation and discussion of the results.

The test of the automated prototype implementation in production confirmed the suitability of the process. Defects (e.g. scratches, streaks, etc.) could be detected in complete glass laminates and differentiated from heating wires. In addition, the depth of defects in the glass laminate could be determined by the stereo structure. Dust effects can be masked out using compressed air and several stacked images. However, thorough cleaning of the glass fronts is absolutely essential. In the context of the test, various challenges of everyday production and optimization possibilities were also revealed. Due to the different product dimensions and shapes, the lighting was aligned to the maximum dimensions. With a static setup, this leads to a visible brightness gradient in the images, where digital compensation is not helpful. Fine streaks and fluff can sometimes not be recognized in the dark areas. Glass assemblies with vertical heating wires cannot currently be tested reliably as the light sources have to be too far apart. In order to counter this problem, an automated movable exposure system is to be tested in future, which should enable more uniform exposure. Spot errors and small streaky areas are still prone to errors with regard to depth determination. By optimizing the automated stereo calibration, this should also be avoided in the future. These challenges are to be worked on in follow-up projects and future activities so that the inspection process can be integrated into the production process reliably and time-efficiently in the long term.