Advancements in high-power laser cutting have positioned this technology as an indispensable component within the contemporary manufacturing landscape. Despite the widespread adoption of this technology, maximizing output quality and cutting efficiency remains challenging due to constraints arising from static parameters. This research thesis addresses these limitations by introducing robust virtual sensors deployable in the industrial context, which continuously monitor the melt pool and kerf region through a coaxial high-speed camera to estimate various quality indicators, thereby enabling diagnostic monitoring and prescriptive actions to optimize both cut quality and factory productivity in real time. The initial stage of the research involved developing a unified hardware and software platform, which enabled the generation of one of the largest multi-machine laser cutting datasets in the literature to date. The research involved next the improvement and optimization of traditional machine learning baseline pipelines to enable sub-millisecond inference latency. Finally, an innovative state-of-the-art neural model architecture and a novel framework enabling zero-shot generalization and weakly supervised fine-tuning have been developed. The first, CUT-Net, is designed to solve the real-time performance gap of neural architectures in the literature, achieving 153 μs inference latency on a single CPU core with a three-state laser cutting classification accuracy of 94.11%. CUT-Net achieves up to 450x higher pixel throughput and up to 590x lower memory usage compared to prior high-accuracy approaches in the literature, and introduces for the first time a module for temporal integration, resulting in a 67% reduction of prediction jitter. The second, GEN-Cut, presents a novel solution to the domain shift problem by attaining 85.5% zero-shot accuracy on unseen cutting machine data (a substantial improvement over near-random baselines of previous models), complemented by a framework for efficient data usage and weakly-supervised adaptation, which increases the operational accuracy to 92.7% using only two unlabeled calibration cuts gathered in-situ.
I progressi nel taglio laser hanno reso questa tecnologia un componente indispensabile nella manifattura contemporanea. Nonostante la diffusione, massimizzare qualità ed efficienza di taglio rimane complesso, a causa dei vincoli derivanti da parametri statici. Questa tesi di ricerca affronta tali limitazioni introducendo sensori virtuali che monitorano la zona di taglio tramite una telecamera coassiale ad alta velocità, per stimare vari indicatori di qualità, consentendo un monitoraggio diagnostico e prescrittivo che ottimizzi in tempo reale sia la qualità del taglio che la produttività dell'impianto. Inizialmente il progetto ha previsto lo sviluppo di una piattaforma hardware e software unificata, che ha permesso la generazione di uno dei più grandi dataset di taglio laser multi-macchina ad oggi presenti in letteratura. La ricerca ha poi riguardato il miglioramento e l'ottimizzazione delle tradizionali pipeline di apprendimento automatico per consentire una latenza di inferenza inferiore al millisecondo. Infine, sono stati sviluppati un'innovativo modello neurale e un nuovo framework che consente la generalizzazione zero-shot e il fine-tuning semi-supervisionato. Il primo, CUT-Net, è progettato per colmare il divario prestazionale delle architetture neurali presenti in letteratura, raggiungendo una latenza di inferenza di 153 μs su un singolo core CPU con un'accuratezza di classificazione del taglio laser a tre stati del 94,11%. CUT-Net raggiunge un throughput di pixel fino a 450 volte superiore e un utilizzo della memoria fino a 590 volte inferiore rispetto ai precedenti approcci presenti in letteratura, e introduce per la prima volta un modulo per l'integrazione temporale, con una conseguente riduzione del 67% del jitter di predizione. Il secondo, GEN-Cut, presenta una soluzione innovativa al problema del cambiamento di dominio, con un'accuratezza zero-shot dell'85,5% su dati di macchine da taglio non visti (un miglioramento sostanziale rispetto ai modelli precedenti), completata da un framework per un adattamento efficiente semi-supervisionato, che aumenta l'accuratezza operativa al 92,7% utilizzando solo due tagli di calibrazione non classificati.
Real-time and robust virtual sensing for laser manufacturing via low-latency deep learning and cross-machine adaptation
Restifo, Alessandro
2025/2026
Abstract
Advancements in high-power laser cutting have positioned this technology as an indispensable component within the contemporary manufacturing landscape. Despite the widespread adoption of this technology, maximizing output quality and cutting efficiency remains challenging due to constraints arising from static parameters. This research thesis addresses these limitations by introducing robust virtual sensors deployable in the industrial context, which continuously monitor the melt pool and kerf region through a coaxial high-speed camera to estimate various quality indicators, thereby enabling diagnostic monitoring and prescriptive actions to optimize both cut quality and factory productivity in real time. The initial stage of the research involved developing a unified hardware and software platform, which enabled the generation of one of the largest multi-machine laser cutting datasets in the literature to date. The research involved next the improvement and optimization of traditional machine learning baseline pipelines to enable sub-millisecond inference latency. Finally, an innovative state-of-the-art neural model architecture and a novel framework enabling zero-shot generalization and weakly supervised fine-tuning have been developed. The first, CUT-Net, is designed to solve the real-time performance gap of neural architectures in the literature, achieving 153 μs inference latency on a single CPU core with a three-state laser cutting classification accuracy of 94.11%. CUT-Net achieves up to 450x higher pixel throughput and up to 590x lower memory usage compared to prior high-accuracy approaches in the literature, and introduces for the first time a module for temporal integration, resulting in a 67% reduction of prediction jitter. The second, GEN-Cut, presents a novel solution to the domain shift problem by attaining 85.5% zero-shot accuracy on unseen cutting machine data (a substantial improvement over near-random baselines of previous models), complemented by a framework for efficient data usage and weakly-supervised adaptation, which increases the operational accuracy to 92.7% using only two unlabeled calibration cuts gathered in-situ.| File | Dimensione | Formato | |
|---|---|---|---|
|
2026_03_Restifo.pdf
non accessibile
Descrizione: Tesi Dottorato
Dimensione
54.54 MB
Formato
Adobe PDF
|
54.54 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/255117