Fully automated traceabillity of QC results for ETO Production

The project had an ambitious goal of developing an automated quality control for small series production: the AutoQC system. The challenge is that there will always be a considerable variation in the process conditions when comparing different product series. Is it still possible to train a machine learning system to detect quality deviations based on sensor data from production? At Oshaug Metall AS, small series of NiAl bronze components are cast for the maritime industry, and a significant part of the production is delivered to Brunvoll AS to be part of their thruster systems. There are strict requirements for quality, and for components to be included in classified vessels there are standardized procedures for approval. The propeller blades supplied by Oshaug Metall must meet stringent requirements which, among other things, limit the presence of porosity on the surface of the component. For each blade, a penetrant test is performed. There are several reasons why porosity can occur in cast components. One mechanism that can lead to microscopic pores is that metals in solid form typically have higher densities than in liquid form. This gives rise to so-called "solidification shrinkage" which creates a negative pressure in the component as it undergoes the transition from melt to solid. To limit pore formation, it is important to provide good supply of liquid metal to fill the vacuum that occurs at the solidification front. In practice, this is achieved by designing a directional solidification process with favorable cooling rates. With theoretical calculations, it is possible to estimate whether there is a large or small risk of pore formation, for example through the so-called Niyama criterion. Results We developed a proof-of-concept machine learning system that was trained to compare simulated casting processes with sensor data from the production process to estimate the danger of propeller blade porosity. Using two simulated gradients and temperature data from an actual cast as input, the AutoQC system provides an estimate of the Niyama criterion for the actual cast. The machine learning system was trained only using simulated data, and at the end of the project period the system was tried using actual temperature measurements as input. Then we saw that the AutoQC system behaved promisingly: In a number of theoretical calculations, we obtained a range of values for the Niyama criterion, and the AutoQC system narrowed these estimates to porosity for the actual cast. We were not able to evaluate the accuracy of the system, but note that the AutoQC system had a certain suspicion about how much porosity occurred in the actual cast, and this with little sensitivity to which theoretical calculations the system received as input. As part of the AutoQC system, as we see it, it is relevant to later implement an automated feedback system. The penetrant testing performed on each leaf involves photography that documents the absence or occurrence of porosity over the entire blade surface. These images can be processed automatically, giving the machine learning system continuous feedback on which blades where porosity actually occurs. In the project we developed guidelines for imaging in connection with penetrant testing, so that the images are best adapted for automated imaging.

Vi utviklet et proof-of-concept maskinlæringssystem som ble trent opp til å sammenligne simulerte støpeprosesser med sensordata fra produksjonsprosessen. Faren for porøsitet i propellbladet blir estimert. Med to simulerte forløp og temperaturdata fra et faktisk støp som input gir AutoQC-systemet et estimat av kvalitet (Niyama-kriteret) for det faktiske støpet. Arbeidsflyt hos industribedriftene var forbedret i prosjektperioden.

Mange norske små og mellomstore bedrifter (SMB-er) har en stor andel engineer-to-order (ETO) produksjon. For slike skreddersydde produkter utgjør ofte kvalitetssikring (QA) en stor andel av produksjonskostnadene. QA kan forsinke leveranser og medføre store direkte kostnader. Manuell QA er tidkrevende og krever høy kompetanse. SMB-er er gjerne sårbare for tap av nøkkelpersonell med påkrevd QA kompetanse. I senere tid har norske underleverandører til maritim og offshore industri opplevd stadig større konkurranse og høyere krav til kortere leveringstid blant sine kunder. De er nødt til å levere kundetilpassede produkter på betydelig kortere tid enn i dag for å kunne konkurrere i et globalt marked. I storskala serieproduksjon brukes maskinlæring for QA rutinemessig. Med løpende sensordata fra produksjonslinjen og moderne maskinlæringsteknologi er det store besparelser ved at systemet automatisk kan avgjøre om eventuelle avvik vil påvirke kvaliteten. For norske SMB-er, som er spesialisert på små serier og ETO produksjon, er en slik automatisk QA ikke tilgjengelig: Seriene er for små til å rekke å trene opp maskinlæringssystemet før bedriften må videre til neste produkt. Vi skal støtte opp om norsk industri gjennom å gjøre maskinlæringsbasert QA tilgjengelig for små serier av skreddersydde produkter. Prosjektet vil bidra til å skape mer konkurransedyktige norske utstyrsleverandører, til maritim og offshore sektor, men også andre markeder med høye krav til kort leveringstid.