In injection molding, process set-up, especially during start-up, still relies largely on operator experience. The machine-learning-based decision support systems proposed so far mostly address aesthetic defects. For warpage the problem is more complex: the deformation depends non-linearly on process, material and part geometry, and measuring it in line requires precise and fast instruments, often expensive or inflexible. This thesis develops a low-cost vision system for the in-line measurement of the warpage of a flat ABS switch plate. A camera films edge-on the side face of the parts on the conveyor, a neural network segments its profile and the deflection is derived from a geometric descriptor, calibrated on nine plates measured with a laser scanner. Of the three networks compared (U-Net++, YOLO26n-seg and RF-DETR-Seg), RF-DETR was selected: it achieved the highest mean IoU (0.885), although its difference from U-Net++ (0.859) is not statistically significant. Finally, the system was also trained and calibrated in a virtual environment built in Blender from the 3D model of the part. Under production conditions, with the parts in motion, the measurement error is 0.11 mm (0.10 mm on parts never seen during calibration), with 5.5 distinct categories, slightly above the MSA threshold of 5; combining two acquisitions reduces it to 0.09 mm. The error is almost entirely systematic: the repeatability, about 0.017 mm, makes the system best suited to relative measurements. A network trained only on synthetic images, without manual annotation, shows on real footage an accuracy comparable to that of the network trained on real images (0.112 vs 0.111 mm), albeit with a lower repeatability; on the moving conveyor the calibration slope estimated in the virtual environment proved transferable, and on the plant it was sufficient to set its offset with three plates. Within the limits of a single component and nine reference plates, the results indicate that a low-cost camera and a segmentation network can measure warpage in line with an error of about 0.1 mm, and that image annotation and network training can largely be carried out before installation. The system is a first step towards extending to geometric defects the preventive approach adopted for aesthetic ones: fed by process simulations, the virtual environment could provide the data for a model that corrects the molding parameters.
Nello stampaggio a iniezione la messa a punto del processo, soprattutto in fase di avviamento, si basa ancora in larga parte sull’esperienza degli operatori. I sistemi di supporto decisionale basati sull’apprendimento automatico finora proposti riguardano per lo più i difetti estetici. Per il warpage il problema è più complesso: la deformazione dipende in modo non lineare da processo, materiale e geometria del pezzo, e la sua misura in linea richiede strumenti precisi e veloci, spesso costosi o poco flessibili. Questa tesi sviluppa un sistema di visione a basso costo per misurare in linea l’inarcamento di una placchetta piana in ABS per interruttori. Una telecamera riprende di taglio la faccia laterale dei pezzi sul nastro, una rete neurale ne segmenta il profilo e la freccia è ricavata da un descrittore geometrico, tarato su nove placchette misurate con scanner laser. Fra U-Net++, YOLO26n-seg e RF-DETR-Seg è stata scelta RF-DETR, che ha ottenuto la IoU media più alta (0,885), anche se la differenza rispetto a U-Net++ (0,859) non è statisticamente significativa. Il sistema è stato infine addestrato e tarato anche in un ambiente virtuale, costruito con Blender dal modello 3D del pezzo. In condizioni di produzione, con i pezzi in movimento, l’errore di misura è di 0,11 mm (0,10 mm su pezzi mai visti in taratura), con 5,5 categorie distinguibili, poco sopra la soglia di 5 del manuale MSA; combinando due acquisizioni scende a 0,09 mm. L’errore è quasi interamente sistematico: la ripetibilità, circa 0,017 mm, rende il sistema adatto soprattutto a misure relative. Una rete addestrata solo su immagini sintetiche, senza annotazioni manuali, mostra sulle riprese reali un’accuratezza comparabile a quella della rete addestrata su immagini reali (0,112 contro 0,111 mm), pur con una ripetibilità inferiore; a nastro in movimento la pendenza della taratura stimata in ambiente virtuale è risultata trasferibile, e sull’impianto è bastato fissarne lo zero con tre placchette. Pur nei limiti di un solo componente e di nove placchette di riferimento, i risultati indicano che una telecamera economica e una rete di segmentazione possono misurare il warpage in linea con un errore dell’ordine di 0,1 mm, e che annotazione delle immagini e addestramento della rete possono svolgersi in larga parte prima dell’installazione. Il sistema è un primo passo per estendere ai difetti geometrici l’approccio preventivo adottato per quelli estetici: alimentato da simulazioni di processo, l’ambiente virtuale potrebbe fornire i dati per un modello che corregga i parametri di stampaggio.
Rilevamento di deformazioni di componenti stampati ad iniezione tramite Deep Learning
MALACHIN, MATTEO
2025/2026
Abstract
In injection molding, process set-up, especially during start-up, still relies largely on operator experience. The machine-learning-based decision support systems proposed so far mostly address aesthetic defects. For warpage the problem is more complex: the deformation depends non-linearly on process, material and part geometry, and measuring it in line requires precise and fast instruments, often expensive or inflexible. This thesis develops a low-cost vision system for the in-line measurement of the warpage of a flat ABS switch plate. A camera films edge-on the side face of the parts on the conveyor, a neural network segments its profile and the deflection is derived from a geometric descriptor, calibrated on nine plates measured with a laser scanner. Of the three networks compared (U-Net++, YOLO26n-seg and RF-DETR-Seg), RF-DETR was selected: it achieved the highest mean IoU (0.885), although its difference from U-Net++ (0.859) is not statistically significant. Finally, the system was also trained and calibrated in a virtual environment built in Blender from the 3D model of the part. Under production conditions, with the parts in motion, the measurement error is 0.11 mm (0.10 mm on parts never seen during calibration), with 5.5 distinct categories, slightly above the MSA threshold of 5; combining two acquisitions reduces it to 0.09 mm. The error is almost entirely systematic: the repeatability, about 0.017 mm, makes the system best suited to relative measurements. A network trained only on synthetic images, without manual annotation, shows on real footage an accuracy comparable to that of the network trained on real images (0.112 vs 0.111 mm), albeit with a lower repeatability; on the moving conveyor the calibration slope estimated in the virtual environment proved transferable, and on the plant it was sufficient to set its offset with three plates. Within the limits of a single component and nine reference plates, the results indicate that a low-cost camera and a segmentation network can measure warpage in line with an error of about 0.1 mm, and that image annotation and network training can largely be carried out before installation. The system is a first step towards extending to geometric defects the preventive approach adopted for aesthetic ones: fed by process simulations, the virtual environment could provide the data for a model that corrects the molding parameters.| File | Dimensione | Formato | |
|---|---|---|---|
|
Malachin_Matteo.pdf
Accesso riservato
Dimensione
3.69 MB
Formato
Adobe PDF
|
3.69 MB | Adobe PDF |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/118685