AI bending tool selection

Exploring Project, Quick Check

Initial situation

If complex technical bent parts are to be produced automatically from high-strength spring steels, it is necessary to pre-determine the forming tools and the necessary overbending of the material. The material is a spring steel and tends to return to its original shape after bending. It is important that there are only a few correction loops in relation to the forming angles and that the forming radius is set correctly.

Currently, a machine setter determines these parameters based on empirical values. The specialist uses the material data and geometry to roughly define the probable bending radius and the necessary overbending angle in order to achieve the desired target result. However, the machine setters evaluate this differently depending on their experience. This is due to the fact that there is no standardized database because the responsible input parameters are not sufficiently defined.

Solution idea

The idea of the project was therefore to create a data set that depicts as many different machine parameters as possible. A neural network was then to be trained on this basis to predict the quality parameters.

The neural network can then be used to easily determine the machine parameters for the desired default values for the quality parameters.
In order to minimize the number of experiments, special methods of Bayesian statistics were also used.

Benefit

The development of a functional tool selection, in which the expected forming angle can be determined, makes it possible to digitize this process in the first place. Tool selection transforms the existing but inadequate knowledge of the specialist staff into an AI that is constantly learning better. From this, the correct tool selection and predefinition of the machine parameters can be made with increasing accuracy in order to produce the bent parts. The set-up and forming process can then be truly automated based on this learning curve. This makes it independent of specialist personnel and enables low-disruption production 24/7, which in turn offers high quality and a delivery precision that makes delivery times of up to 24 hours feasible. This time saving makes companies more competitive and the overall process more economical.

Implementation of the AI application

The aim was to create a tabular, labeled data set to train a neural network to predict quality parameters via regression. Two strategies were chosen to minimize the number of trials: firstly, focusing on one type of material, and secondly, using Uncertainty-Based Active Exploration to maximize the learning potential per trial.