Starting from the biggest technology companies, there is a significant and continuous increase of investment in Artificial Intelligence. The main advantage of merging Machine Learning and industrial applications is mainly related to their performance and flexibility, also in complex contexts. This convergence set the basis for the fourth industrial revolution leading to the Industry 4.0 era and smart manufacturing. In this dissertation, the focus is on the development of approaches capable to investigate industrial processes in order to address tasks such as predictive maintenance and monitoring. Firstly, spatial analysis is shown to be effective in detecting and representing structured pattern of defects on integrated circuit fabrication. This issue is crucial to a production line since it can cause damages and yield loss. To address this hard-to-solve problem, the proposed approach is a concatenation of different methods, namely a p-value control chart, a clustering algorithm based on Mininum Spanning Tree and a graphical tool. The suggested procedure proves to be extremely fast and effective, allowing its implementation in-line during the fabrication process. Secondly, graphical models provide an effective tool to represent conditional independences among variables. Especially in manufacturing, heterogeneous data often occurs (e.g. continuous, categorical) and it is of interest to discover interactions between different settings or measurements. Therefore, the goal is to define a suitable model able to describe a joint distribution of mixed type variables. The novelty of the proposed methodology is the definition of Bayesian graphical model starting from a Conditional Gaussian distribution suitable both for parameter inference and structure learning. Additionally, an MCMC scheme is implemented for approximate posterior inference in two alternative parametrizations, and a structure learning algorithm for related undirected graph models. This method shows outperforming results when compared to alternative state-of-the-art approaches in a simulation environment.
A partire dalle più grandi aziende tecnologiche, si registra un notevole e continuo aumento degli investimenti in Intelligenza Artificiale. Il principale vantaggio dell'unione di Machine Learning e applicazioni industriali è principalmente legato alle loro prestazioni e flessibilità, anche in contesti complessi. Questa convergenza ha posto le basi per la quarta rivoluzione industriale che ha portato all'era Industria 4.0 e alla definizione di smart manufacturing. In questa tesi, il focus è sullo sviluppo di approcci in grado di indagare i processi industriali al fine di affrontare task come la manutenzione predittiva e il monitoraggio. In primo luogo, l'analisi spaziale si è dimostrata efficace nel rilevare e rappresentare pattern strutturati di difetti nella fabbricazione di circuiti integrati. Questo problema è cruciale per una linea di produzione poiché può causare danni e perdite in termini di resa. Per affrontare questo problema non banale, l'approccio proposto è una concatenazione di diversi metodi, quali una carta di controllo basata sul p-value, un algoritmo di clustering basato su Mininum Spanning Tree e uno strumento grafico. La procedura suggerita si rivela estremamente rapida ed efficace, consentendone l'implementazione in linea durante il processo di fabbricazione. In secondo luogo, i modelli grafici forniscono uno strumento efficace per rappresentare le indipendenze condizionali tra variabili. Soprattutto nella produzione, spesso si raccolgono dati eterogenei (ad es. continui, categoriali) ed è interessante scoprire le interazioni tra diverse setting e misurazioni. Pertanto, l'obiettivo è definire un modello adatto in grado di descrivere una distribuzione congiunta di variabili di tipo misto. La novità della metodologia proposta è la definizione di un modello grafico bayesiano a partire da una distribuzione gaussiana condizionale adatta sia per l'inferenza di parametri che per l'apprendimento della struttura grafica. Inoltre, viene implementato uno schema MCMC per l'inferenza a posteriori approssimativa in due parametrizzazioni alternative e un algoritmo di structure learning per i relativi modelli grafici non orientati. Questo metodo mostra risultati migliori rispetto ad approcci alternativi all'avanguardia in un ambiente di simulazione.
ADVANCED ANALYTICS AND MACHINE LEARNING FOR INDUSTRIAL MANUFACTURING APPLICATIONS
GALIMBERTI, CHIARA
2023
Abstract
Starting from the biggest technology companies, there is a significant and continuous increase of investment in Artificial Intelligence. The main advantage of merging Machine Learning and industrial applications is mainly related to their performance and flexibility, also in complex contexts. This convergence set the basis for the fourth industrial revolution leading to the Industry 4.0 era and smart manufacturing. In this dissertation, the focus is on the development of approaches capable to investigate industrial processes in order to address tasks such as predictive maintenance and monitoring. Firstly, spatial analysis is shown to be effective in detecting and representing structured pattern of defects on integrated circuit fabrication. This issue is crucial to a production line since it can cause damages and yield loss. To address this hard-to-solve problem, the proposed approach is a concatenation of different methods, namely a p-value control chart, a clustering algorithm based on Mininum Spanning Tree and a graphical tool. The suggested procedure proves to be extremely fast and effective, allowing its implementation in-line during the fabrication process. Secondly, graphical models provide an effective tool to represent conditional independences among variables. Especially in manufacturing, heterogeneous data often occurs (e.g. continuous, categorical) and it is of interest to discover interactions between different settings or measurements. Therefore, the goal is to define a suitable model able to describe a joint distribution of mixed type variables. The novelty of the proposed methodology is the definition of Bayesian graphical model starting from a Conditional Gaussian distribution suitable both for parameter inference and structure learning. Additionally, an MCMC scheme is implemented for approximate posterior inference in two alternative parametrizations, and a structure learning algorithm for related undirected graph models. This method shows outperforming results when compared to alternative state-of-the-art approaches in a simulation environment.File | Dimensione | Formato | |
---|---|---|---|
phd_unimib_771744.pdf
accesso aperto
Dimensione
2.31 MB
Formato
Adobe PDF
|
2.31 MB | Adobe PDF | Visualizza/Apri |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14242/75721
URN:NBN:IT:UNIMIB-75721