ELECTRONICS DESIGN
should also be permanently expanded during the life cycle with significant examples from ongoing production.
The concept of the AI advisor is accordingly flexible. On the one hand, it is possible to start with an already predefined database after installation, which is retrained during operation with further image data from the corresponding user decisions. On the other hand, a model can also be trained without a predefined database only on the basis of classification decisions historically stored in the database of the inspection system. In this case, too, after the initial training, further image data from the corresponding user decisions, which are available in productive operation, are used
Figure 3: AI advisor and the AI backend in the context of GOEPEL software applications
production systems could cause enormous damage. Cloud solutions can therefore often not be realised and edge-based AI solutions are then the only way out.
For these reasons, the modern architecture of the AI backend of the AI advisor software covers all possible use cases. For electronics manufacturers who have only one production line, all AI software modules can be installed on the same PC as the verification station software.
Figure 4: Integration of the AI advisor into a single production line to retrain.
In productive operation, the AI software receives inspection data from the inspection system, assigns the inspection data to the respective AI model instances, performs the inference and transmits AI decisions to the verification software.
The AI software is designed in such a way that model instances are created on the basis of the error images in a rule-based and thus completely autonomous manner. To ensure that a valid training basis is created in this autonomous process and that it remains valid throughout the life cycle, various mechanisms were integrated. These range from the monitoring or exclusion of image data that are too similar to image data already present in the training set, to additional expert interviews (active learning) for individual new image data to be trained.
The developed AI software ecosystem also includes an AI framework for the distributed training and delivery of Deep Learning models and their instances. This tool, which has a web-based, system-independent interface and can thus be accessed conveniently via a web browser, makes it possible to manage the instances with the error images (samples)
Figure 5: Integration of the AI advisor into several production lines
and their assignment (labels) or, if necessary, to delete samples from the training set.
Integration into production infrastructure
AI solutions that are to be integrated into industrial production processes must first and foremost fit seamlessly into the corresponding IT infrastructure. In many cases, production processes are completely sealed off from the outside world for security reasons, as a loss of sensitive data or manipulation of the
If the production consists of several lines, the AI can run in the company network on a separate AI PC. For users who do not want to install additional computing power, however, a cloud-based solution can also be provided. In the field of automated optical and X-ray inspection, the use of AI contributes in particular to enabling inspection systems to make better inspection decisions or to simplify manual processes. But also the optimisation of production processes based on the collected inspection data is an important topic. The newly developed AI advisor software module provides an AI-based assistance function for the PILOT Verify verification software, which ensures that incorrect decisions are prevented during defect verification and that no detected defects are subsequently classified as “pass”. Through the interaction of humans and AI, the classification process is optimised and the human is relieved. In a further development stage, the artificial intelligence then becomes the decision-maker and the occurring defects are classified independently. Verification by the operator is only necessary in exceptional situations when the AI cannot provide a clear result. The automated classification of almost all errors reduces the workload of the operators at the verification stations even more.
GOEPEL Electronic
www.goepel.com/en
DECEMBER/JANUARY 2023 | ELECTRONICS TODAY 23
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36 |
Page 37 |
Page 38 |
Page 39 |
Page 40 |
Page 41 |
Page 42 |
Page 43 |
Page 44
