FUTURE MANUFACTURING
In this application , the image data was generated manually in a production cell setup utilizing a sample per all workpiece variants . However , this makes image generation not only laborious but also expensive , so only small data sets are available . In addition , there is a high degree of flexibility in production , since the manufacturer produces many workpieces , and the operator regularly change and adjust the production settings . This is a complex AI classification problem , as the trained model must not only be able to reliably learn a high number of combinations of turning and clamping motions . It must also distinguish between the similar features of the workpiece variants .
However , due to the small input data sets , there is a risk of overfitting . To illustrate , the learned AI model has simply memorized the ( small ) data set . Thus , while it works optimally with the existing data , but no longer provides useful results for new image data with slight domain deviation . For example , if the input values in the data set changed only slightly , such as the light setting , the model could no longer provide reliable instructions .
Train AI models with higher reliability
To develop reliable AI methods that meet the above challenges , Siemens used special methods when training the robot model :
Use data perturbation methods and improve the robustness of the system . These methods systematically generate a large amount of synthetic data from the raw data and from changes e . g ., in lighting conditions , geometric rotations and shooting angles . These data represent the application domain in greater variance . Thus , it is possible to train the AI models with higher reliability .
In addition , the goal is to provide production cell operator with an estimate of how uncertain or certain the AI-based classification model is with respect to the decision made . If the uncertainty value is high , the manufacturing cell operator will have to take particular attention when
selecting the robot program . In this way , the interaction between humans and machines can be made trustworthy .
Possibilities for use
The company can use the developed AI solution in various ways . The following scenarios are conceivable :
● The operator is assisted by a suggestion from the system as to which robot program should be used . It is the operator ' s responsibility to accept this suggestion or to decide on a different robot program or parameterization . Based on the suggestion , the operator can back up the own decision . In standard cases , the operator will therefore be able to select the robot program more quickly and use the associated freedom to invest more effort in the selection in special cases . This makes the overall process more reliable . This also means that smaller batch sizes , where the production cell must be reconfigured more often , can also be manufactured more economically .
● The selection of the robot program , including parameterization , can be carried out without the intervention of operator , so that a reconfiguration of the production cell can be carried out fully automated . This increases efficiency and enhances the flexibility of the production cell .
Ultimately , it is the responsibility of the operator of the production cell to decide between these two scenarios for a sensible application option . l
Thomas Hahn Chief Expert Software Siemens AG , Technology
Example of another cell for efficient production of small batches .
The operator of the manufacturing cell controls the result of the AI program .
Martina Risteska Senior Data Scientist Siemens AG , Digital Industries , Motion Control