This thesis aims to develop and compare two innovative tools designed to support operators during the start-up phase of an injection molding process. The first approach, based on predictive artificial intelligence, involves the development of a machine learning model built upon a Latin Hypercube Sampling (LHS) experimental design. The purpose of this model is to predict component quality performance as a function of selected process parameters and to identify, through optimization strategies, an optimal set of parameters to ensure defect-free production. The second approach relies on a Retrieval-Augmented Generation (RAG) system, which integrates a general-purpose Large Language Model (LLM) with test reports prepared by experienced operators. This tool selects, extracts, and organizes the information contained in the documents into operational recommendations, thus guiding less experienced operators in identifying process parameters that ensure high-quality molded components. To provide an objective assessment of component defectiveness, a computer vision model was also developed to classify product quality, reducing the subjectivity associated with human evaluation. The comparison between the two approaches was carried out by considering both the intrinsic cost associated with the development of these tools and the number of operations required by the operator to obtain a component of required quality.
Il presente elaborato si propone di sviluppare e confrontare due strumenti innovativi a supporto dell’operatore nella fase di avvio di un processo di stampaggio a iniezione. Il primo approccio, basato su intelligenza artificiale predittiva, prevede la realizzazione di un modello di machine learning fondato su un piano sperimentale di tipo Latin Hypercube Sampling (LHS). Tale modello è finalizzato a prevedere le prestazioni in termini di qualità dei componenti in funzione dei parametri di processo e a individuare, mediante strategie di ottimizzazione, un set di parametri ottimale per ottenere componenti esenti da difetti. Il secondo approccio si fonda sull’utilizzo di un sistema di tipo Retrieval-Augmented Generation (RAG), il quale integra un Large Language Model (LLM) generalista con i documenti di collaudo redatti da operatori esperti. Questo strumento seleziona, estrae e organizza le informazioni contenute nei documenti sotto forma di suggerimenti operativi, volti a guidare l’operatore meno esperto nell’ottenimento di parametri di processo ottimali per la qualità dei componenti stampati. Per garantire un giudizio oggettivo sul grado di difettosità dei componenti, è stato inoltre sviluppato un modello di computer vision in grado di classificare la qualità del prodotto, riducendo la soggettività legata alla valutazione umana. Il confronto tra i due approcci è stato condotto considerando sia il costo intrinseco associato allo sviluppo degli strumenti, sia il numero di operazioni richieste all’operatore per ottenere un componente di qualità soddisfacente.
Confronto tra AI predittiva e RAG nel supporto all’avvio dello stampaggio a iniezione
BUSATO, ROBERTO
2024/2025
Abstract
This thesis aims to develop and compare two innovative tools designed to support operators during the start-up phase of an injection molding process. The first approach, based on predictive artificial intelligence, involves the development of a machine learning model built upon a Latin Hypercube Sampling (LHS) experimental design. The purpose of this model is to predict component quality performance as a function of selected process parameters and to identify, through optimization strategies, an optimal set of parameters to ensure defect-free production. The second approach relies on a Retrieval-Augmented Generation (RAG) system, which integrates a general-purpose Large Language Model (LLM) with test reports prepared by experienced operators. This tool selects, extracts, and organizes the information contained in the documents into operational recommendations, thus guiding less experienced operators in identifying process parameters that ensure high-quality molded components. To provide an objective assessment of component defectiveness, a computer vision model was also developed to classify product quality, reducing the subjectivity associated with human evaluation. The comparison between the two approaches was carried out by considering both the intrinsic cost associated with the development of these tools and the number of operations required by the operator to obtain a component of required quality.| File | Dimensione | Formato | |
|---|---|---|---|
|
Busato_Roberto.pdf
Accesso riservato
Dimensione
6 MB
Formato
Adobe PDF
|
6 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/20.500.12608/99949