Aluminum production is an energy-intensive process and small improvements in the process can lead to great impacts in terms of sustainability and operational efficiency. This research introduces innovative computer vision techniques for real-time monitoring of aluminum melting furnaces, aiming to enhance process control and efficiency. An industrial camera is designed for continuous operation in high-temperature environments. It is enclosed in a refractory case with thermal shields and features an air-cooling mechanism to withstand the conditions of the furnace. The camera system was tested in a laboratory muffle furnace at 800°C. With a heating rate measured at 1.4°C/min, the camera operated safely for 2 hours without cooling. The camera sensor end at the rear remained at ambient temperature, showing effective heat transfer minimization. The camera has been installed in 3 industrial furnaces and images from the furnace interior were used for process monitoring: algorithms were developed to measure the melt level, to detect thermite flame activity, and to oversee the scrap melting process. A vision-based algorithm has been developed for non-contact measurement of melt levels, with a measurement uncertainty of less than 3 mm. An image processing algorithm that uses level thresholding in different image planes (HSV+RGB) has been implemented to detect thermite reaction flames in the melt bath, enabling prompt intervention to prevent dross formation. Multiple solutions were developed to monitor the melting process. A standard image processing approach was developed for scenarios where the availability of furnace image data is limited: the algorithm detects the solid scrap boundaries within the melt bath and quantifies the percentage of unmelted scrap content present in the melt bath. Based on this quantification, the algorithm provides an automatic indication of whether the furnace is ready to process the next batch, ensuring efficient operation even in data-scarce environments. Sequential Convolutional Neural network models were used to identify the melting status of the furnace: the models achieved 93 – 99% accuracy in indicating the furnace readiness using relatively low-resolution images. An AI-based scrap monitoring model was proposed to enable localized segmentation of scrap pixels, offering real-time information regarding the distribution between unmelted solid scrap and melted liquid aluminum present in the melt bath. Furthermore, the model estimates the remaining process time with an uncertainty of ±5 minutes. The developed solutions operate in real-time, with processing speeds ranging between 0.15 and 0.7 s per frame. Their implementation in real production environments has significantly reduced the need to open the furnace door during operation, thereby minimizing thermal losses associated with frequent door openings and enhancing operator safety through automated monitoring. By improving process efficiency and reducing energy consumption, this research contributes towards sustainable practices in secondary aluminum production. Additionally, it advances the literature in the field of furnace image processing and computer vision applications for furnace monitoring.
La produzione di alluminio è un processo ad alta intensità energetica e piccoli miglioramenti possono avere un grande impatto in termini di sostenibilità ed efficienza operativa. Questa ricerca introduce tecniche innovative di visione artificiale per il monitoraggio in tempo reale dei forni di fusione dell’alluminio, con l’obiettivo di migliorare il controllo del processo e l’efficienza operativa. Una telecamera industriale è stata progettata per operare continuamente in ambienti ad alta temperatura. È racchiusa in un involucro refrattario con scudi termici e dispone di un sistema di raffreddamento ad aria per resistere alle condizioni del forno. Il sistema di telecamera è stato testato in un forno a muffola da laboratorio a 800°C. Con una velocità di riscaldamento misurata a 1,4°C/min, la telecamera ha funzionato in sicurezza per 2 ore senza raffreddamento. L'estremità del sensore della telecamera, situata nella parte posteriore, è rimasta a temperatura ambiente, dimostrando un’efficace minimizzazione del trasferimento di calore. La telecamera è stata installata in tre forni industriali e le immagini dell’interno del forno sono state utilizzate per il monitoraggio del processo: sono stati sviluppati algoritmi per misurare il livello del metallo fuso, rilevare l’attività delle fiamme di termite e controllare il processo di fusione dei rottami. È stato sviluppato un algoritmo basato sulla visione artificiale per la misurazione senza contatto del livello del metallo fuso, con un’incertezza inferiore a 3 mm. Un algoritmo di elaborazione delle immagini che utilizza la soglia di livello in diversi piani immagine (HSV+RGB) è stato implementato per rilevare le fiamme di reazione della termite nel bagno di fusione, consentendo interventi tempestivi per prevenire la formazione di scorie. Sono state sviluppate diverse soluzioni per monitorare il processo di fusione. Un approccio standard di elaborazione delle immagini è stato implementato per scenari in cui la disponibilità di dati visivi del forno è limitata: l’algoritmo rileva i contorni dei rottami solidi all’interno del bagno di fusione e quantifica la percentuale di rottame non fuso presente. Sulla base di questa quantificazione, l’algoritmo fornisce un’indicazione automatica della prontezza del forno a elaborare il lotto successivo, garantendo un funzionamento efficiente anche in ambienti con dati limitati. Modelli di reti neurali convoluzionali sequenziali sono stati utilizzati per identificare lo stato di fusione del forno: i modelli hanno raggiunto un’accuratezza compresa tra il 93% e il 99% nell’indicare la prontezza del forno utilizzando immagini a bassa risoluzione. È stato inoltre proposto un modello di monitoraggio dei rottami basato sull’intelligenza artificiale, che consente la segmentazione localizzata dei pixel dei rottami, offrendo informazioni in tempo reale sulla distribuzione tra rottami solidi non fusi e alluminio liquido nel bagno di fusione. Inoltre, il modello stima il tempo di processo rimanente con un’incertezza di ±5 minuti. Le soluzioni sviluppate operano in tempo reale, con velocità di elaborazione comprese tra 0,15 e 0,7 secondi per fotogramma. La loro implementazione in ambienti di produzione reali ha ridotto significativamente la necessità di aprire la porta del forno durante il funzionamento, minimizzando così le perdite termiche associate alle aperture frequenti e migliorando la sicurezza degli operatori grazie al monitoraggio automatizzato. Migliorando l’efficienza del processo e riducendo il consumo energetico, questa ricerca contribuisce alle pratiche sostenibili nella produzione secondaria di alluminio. Inoltre, avanza lo stato dell’arte nel campo dell’elaborazione di immagini per forni e delle applicazioni di visione artificiale per il monitoraggio dei forni.
Computer vision solutions for real-time furnace monitoring and sustainability in aluminum melting
RAVI, YUVAN SATHYA
2024/2025
Abstract
Aluminum production is an energy-intensive process and small improvements in the process can lead to great impacts in terms of sustainability and operational efficiency. This research introduces innovative computer vision techniques for real-time monitoring of aluminum melting furnaces, aiming to enhance process control and efficiency. An industrial camera is designed for continuous operation in high-temperature environments. It is enclosed in a refractory case with thermal shields and features an air-cooling mechanism to withstand the conditions of the furnace. The camera system was tested in a laboratory muffle furnace at 800°C. With a heating rate measured at 1.4°C/min, the camera operated safely for 2 hours without cooling. The camera sensor end at the rear remained at ambient temperature, showing effective heat transfer minimization. The camera has been installed in 3 industrial furnaces and images from the furnace interior were used for process monitoring: algorithms were developed to measure the melt level, to detect thermite flame activity, and to oversee the scrap melting process. A vision-based algorithm has been developed for non-contact measurement of melt levels, with a measurement uncertainty of less than 3 mm. An image processing algorithm that uses level thresholding in different image planes (HSV+RGB) has been implemented to detect thermite reaction flames in the melt bath, enabling prompt intervention to prevent dross formation. Multiple solutions were developed to monitor the melting process. A standard image processing approach was developed for scenarios where the availability of furnace image data is limited: the algorithm detects the solid scrap boundaries within the melt bath and quantifies the percentage of unmelted scrap content present in the melt bath. Based on this quantification, the algorithm provides an automatic indication of whether the furnace is ready to process the next batch, ensuring efficient operation even in data-scarce environments. Sequential Convolutional Neural network models were used to identify the melting status of the furnace: the models achieved 93 – 99% accuracy in indicating the furnace readiness using relatively low-resolution images. An AI-based scrap monitoring model was proposed to enable localized segmentation of scrap pixels, offering real-time information regarding the distribution between unmelted solid scrap and melted liquid aluminum present in the melt bath. Furthermore, the model estimates the remaining process time with an uncertainty of ±5 minutes. The developed solutions operate in real-time, with processing speeds ranging between 0.15 and 0.7 s per frame. Their implementation in real production environments has significantly reduced the need to open the furnace door during operation, thereby minimizing thermal losses associated with frequent door openings and enhancing operator safety through automated monitoring. By improving process efficiency and reducing energy consumption, this research contributes towards sustainable practices in secondary aluminum production. Additionally, it advances the literature in the field of furnace image processing and computer vision applications for furnace monitoring.File | Dimensione | Formato | |
---|---|---|---|
2025_03_Ravi.pdf
non accessibile
Descrizione: PhD Thesis - Computer Vision Solutions for Real-Time Furnace Monitoring and Sustainability in Aluminum Melting
Dimensione
6.62 MB
Formato
Adobe PDF
|
6.62 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/10589/237237