Deep learning (DL) models are now widely used across many engineering domains, including manufacturing. However, these models' performance is highly dependent on the data they are trained on, and they often fail silently when faced with out-of-distribution (OOD) data. This lack of uncertainty awareness presents a significant risk, potentially leading to operational failures, endangering the safety of human operators or the health of expensive machinery. To address this, uncertainty-aware DL models have emerged and been used recently in computer science, yet their application in manufacturing and industrial settings remains significantly under-explored. The central contribution of this thesis is a systematic uncertainty-aware reliability framework for industrial deep learning systems, which combines the construction of domain-realistic epistemic and aleatoric OOD benchmarks, an evaluation of widely adopted uncertainty-aware model families, and deployable uncertainty-based trustworthy filtering pipelines that enable practitioners to distinguish reliable from unreliable predictions. Unlike prior work that applies uncertainty estimation methods in isolation, this thesis provides unified empirical methodology and cross-domain evidence clarifying when uncertainty-aware approaches offer operational value in manufacturing settings. Presented as a collection of five papers, the thesis spans three application pillars: predictive maintenance through rotating machinery fault diagnosis, safe close human–robot interaction via vision-based hand segmentation, and human/machine localization through a novel deep learning framework for multi-source sound source localization. Across vibration, vision, and acoustic modalities, the work demonstrates how uncertainty-aware models more effectively flag untrustworthy predictions under specific OOD conditions, while highlighting practical constraints such as entropy-based uncertainty conflating epistemic and aleatoric sources. Overall, this thesis advances uncertainty-aware industrial AI by contributing reproducible evaluation protocols, methodological filtering pipelines, and a new multi-source sound localization framework, providing actionable guidance for building safer and more reliable deep learning systems in real-world manufacturing environments.
I modelli di deep learning (DL) sono oggi ampiamente utilizzati in numerosi settori ingegneristici, inclusa la manifattura. Tuttavia, le loro prestazioni dipendono fortemente dai dati su cui vengono addestrati e spesso falliscono in modo silenzioso quando si trovano ad affrontare dati fuori distribuzione (OOD). Questa mancanza di consapevolezza dell’incertezza rappresenta un rischio significativo, potenzialmente conducendo a guasti operativi, mettendo in pericolo la sicurezza degli operatori umani o l’integrità di macchinari costosi. Per affrontare tale problema, negli ultimi anni sono emersi modelli di DL consapevoli dell’incertezza, già applicati in informatica, ma il loro impiego nei contesti manifatturieri e industriali rimane ancora largamente inesplorato. Il contributo centrale di questa tesi consiste in un framework sistematico di affidabilità basato sull’incertezza per sistemi di deep learning industriali, che combina la costruzione di benchmark OOD epistemici e aleatorici realistici rispetto al dominio, la valutazione delle principali famiglie di modelli uncertainty-aware ampiamente adottate e pipeline di filtraggio affidabili basate sull’incertezza, implementabili nella pratica, che consentono ai professionisti di distinguere tra predizioni affidabili e non affidabili. A differenza dei lavori precedenti, che applicano metodi di stima dell’incertezza in modo isolato, questa tesi propone una metodologia empirica unificata e fornisce evidenze trasversali tra diversi domini, chiarendo quando gli approcci uncertainty-aware offrono un reale valore operativo in contesti manifatturieri. Presentata come una raccolta di cinque articoli, la tesi si articola in tre pilastri applicativi principali: la manutenzione predittiva tramite diagnosi dei guasti in macchinari rotanti, l’interazione sicura uomo–robot in prossimità mediante segmentazione delle mani basata su visione artificiale e la localizzazione uomo/macchina attraverso un nuovo framework di deep learning per la localizzazione di sorgenti sonore multi-sorgente. Attraverso modalità di dati vibrazionali, visivi e acustici, il lavoro dimostra come i modelli uncertainty-aware siano più efficaci nell’individuare predizioni non affidabili in specifiche condizioni OOD, evidenziando al contempo vincoli pratici, come la tendenza delle misure di incertezza basate sull’entropia a confondere le componenti epistemiche e aleatoriche.
Implementing uncertainty-aware deep learning models in manufacturing and industrial domain
Jalayer, Reza
2025/2026
Abstract
Deep learning (DL) models are now widely used across many engineering domains, including manufacturing. However, these models' performance is highly dependent on the data they are trained on, and they often fail silently when faced with out-of-distribution (OOD) data. This lack of uncertainty awareness presents a significant risk, potentially leading to operational failures, endangering the safety of human operators or the health of expensive machinery. To address this, uncertainty-aware DL models have emerged and been used recently in computer science, yet their application in manufacturing and industrial settings remains significantly under-explored. The central contribution of this thesis is a systematic uncertainty-aware reliability framework for industrial deep learning systems, which combines the construction of domain-realistic epistemic and aleatoric OOD benchmarks, an evaluation of widely adopted uncertainty-aware model families, and deployable uncertainty-based trustworthy filtering pipelines that enable practitioners to distinguish reliable from unreliable predictions. Unlike prior work that applies uncertainty estimation methods in isolation, this thesis provides unified empirical methodology and cross-domain evidence clarifying when uncertainty-aware approaches offer operational value in manufacturing settings. Presented as a collection of five papers, the thesis spans three application pillars: predictive maintenance through rotating machinery fault diagnosis, safe close human–robot interaction via vision-based hand segmentation, and human/machine localization through a novel deep learning framework for multi-source sound source localization. Across vibration, vision, and acoustic modalities, the work demonstrates how uncertainty-aware models more effectively flag untrustworthy predictions under specific OOD conditions, while highlighting practical constraints such as entropy-based uncertainty conflating epistemic and aleatoric sources. Overall, this thesis advances uncertainty-aware industrial AI by contributing reproducible evaluation protocols, methodological filtering pipelines, and a new multi-source sound localization framework, providing actionable guidance for building safer and more reliable deep learning systems in real-world manufacturing environments.| File | Dimensione | Formato | |
|---|---|---|---|
|
Reza_Jalayer_PhD_Thesis.pdf
accessibile in internet per tutti
Dimensione
31.72 MB
Formato
Adobe PDF
|
31.72 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/10589/255177