Crunch Analytics

Training algorithm for welding robot for Stas

Welding is one of the operations in manufacturing which has become gradually more automated over time. Todays welding robots are capable of creating welding seams on complex parts, but as is the case with all new technologies imperfections cannot be completely ruled out. 



We were contacted by a large manufacturer of moving floor trailers who observed that their welding robots where the cause of a substantial amount of re-work hours. This was mainly because the complexity of their parts is pushing the limits of what these welding robots can handle today. Moreover, the same complexity meant that the amount of manual rework to fix parts which have inadequate welding seams is very substantial.



As a solution to their manufacturing woes today was conceived as a system which would be capable of detecting when mistakes were being made, stopping the system if it did so. An operator could then validate if this was indeed a mistake, of simply a false positive and choose the right corrective action if required. 



Upon observing that various sensors were already present in these machines - and that the data from these sensors was relatively easy to capture - it was possible to build such an algorithm without the need for installing various extra sensors and infrared cameras. Training data was generated by purposefully re-creating the most frequently observed errors, and a predictive model based on this data proved to be capable of predicting errors both on intentionally triggered mistakes as well as on real life data. The end result of this project being a functional and cost effective system which can be rolled out to the complete production line at minimal investment cost.