AI has multiple appliances in different domains and it also comes to MOMs systems, in which we specializes. A few months ago we started a project AIM2ASSIST in which we were experimenting with AI for supporting technologists in a factory while defining technologies of metal assembling products designed in CAD. We have just finished validation of the solution and wanted to share our results.
The project was realized in consortium of 3 partners: MCH Polska - metal products manufacturer whose role was to provide data for the AI training and to validate the solution, PCSS being involved as DIH, providing us with the computing infrastructure and continuous advisory and training services, and MASTA being the solution provider based on its MOM system.
The AI component of the solution employed deep learning techniques to extract business value from data provided in a form of CAD drawings. We applied dedicated neural architecture based on the convolutional neural units allowing for creating predictive models accepting pictorial representation of the detail. This architecture is responsible for finding patterns in the data and correlating the visual information of the detail with respective target quantity, which in this case is a list of production tasks needed for a detail production (such as cutting, welding or bending) and duration of the specific tasks. In order to achieve satisfactory predictive capabilities, the critical amount of the data needed to be provided. To mitigate potential problems with the data availability we applied data augmentation techniques.
We trained the model with use of 4000 details in total, each detail was visualized by 3D CAD project. Then we analyzed technologies proposed by the AI mechanism with technologies created by human specialist. For 20 validation products, 7 of them were exactly the same regardless the entity preparing the technology. An average deviation between duration of times predicted by AI and a technologist was below 7%. A time required by the technologist to implement a complete technology to the system from scratch was significantly reduced in the variant with AI compared to the variant without AI. An average was reduced from 80 minutes to about 32 minutes.
Although the solution requires further development, training and validation, we see its revolutionizing potential on the SMEs manufacturers market. Defining products technologies is crucial for proper MOMs system application and is quite challenging and time consuming. This is one of the barriers which cannot be entered by many SMEs lacking employees and cash resources. By integrating our solution with MeMOM we expect to achieve easier customer onboarding which is desired by anyone taking part in the process - the manufacturer, its employees and software consultants as well.
The trained AI model was already published on the AI4EU Experiments Marketplace.
The project AIM2ASSIST has received funding within the FSTP mechanism of the DIH4AI project founded from the European Union Framework Programme for Research and Innovation Horizon 2020 under Grant Agreement n° 101017057.
