Traditional manufacturing processes are heavily dependent on experienced workers and human judgment. When these workers leave the manufacturer, a huge wealth of knowledge, know-how and understanding are lost. Moreover, as human judgment is bounded and biased, the decisions and operations can be far from optimal. With endless and unceasing progress in science and technology, the manufacturing systems are constantly upgrading and transforming from the traditional form to intelligent manufacturing which consists of evidence-based decision-making systems and can result in sustainable and optimized operations. Thanks to new developments in high-precision sensors, the availability of stronger computational powers and the emergence of novel Artificial Intelligence (AI) techniques, today’s manufacturing operations integrate resources in a more complicated manner and a wider scale to enhance its performance and reduces its costs. Intelligent Condition Monitoring and Automated Visual Inspection (AVI) systems, known as Fault Detection and Diagnosis (FDD), are among the most important research areas, which can significantly cut the manufacturing costs and contribute to the quality assurance and reliability of the production line. An intelligent FDD system is thus highly in demand for smart manufacturing systems. This thesis aims at proposing novel decision-support systems, utilizing AI and signal and image processing algorithms to improve intelligent FDD systems. It presents a novel condition monitoring method that employs deep learning and signal processing for rotating machinery such that the manufacturer can practice an error-free diagnosis system. Then it develops a hybrid generative deep learning algorithm that can create realistic synthetic samples, efficiently learn from rare fault patterns and achieves state-of-the-art performance in different imbalanced and noisy conditions. The evaluation results on an industrial dataset with different noise and imbalance severities demonstrate the superiority of the proposed model over the novel models selected for the comparison panel. In the third paper, we propose a novel generative framework for data oversampling and automatic annotation of the defects alongside a hybrid deep learning-based image object detection algorithm to improve AVI systems. After some experimental results on two industrial datasets, show that the proposed data augmentation model and defect detection algorithm outperforms the other state-of-the-art approaches. The thesis results are promising, and the developed frameworks can be applied to different industries with similar facilities. In the last part, the thesis discusses the findings and envisions some research paths, research opportunities and future possibilities to achieve an efficient intelligent fault detection and diagnosis for manufacturing systems.
I processi di produzione tradizionali dipendono fortemente da lavoratori esperti e dal giudizio umano. Quando questi lavoratori lasciano il produttore, si perde un enorme patrimonio di conoscenze, know-how e comprensione. Inoltre, poiché il giudizio umano è limitato e distorto, le decisioni e le operazioni possono essere tutt'altro che ottimali. Con i progressi incessanti nella scienza e nella tecnologia, i sistemi di produzione si aggiornano e si trasformano costantemente dalla forma tradizionale allo smart manufacturing, che consiste in sistemi decisionali basati sull'evidenza e che può portare a operazioni sostenibili e ottimizzate. Grazie ai nuovi sviluppi nel campo dei sensori ad alta precisione, alla maggiore capacità computazionale e all'emergere di nuove tecniche di intelligenza artificiale (AI), le operazioni di produzione odierne integrano le risorse in modo più complesso e su scala più ampia per migliorarne le prestazioni e ridurre i costi. I sistemi Intelligent Condition Monitoring e Automated Visual Inspection (AVI), noti come Fault Detection and Diagnosis (FDD), sono tra le aree di ricerca più importanti, e possono ridurre significativamente i costi di produzione e contribuire alla garanzia della qualità e all'affidabilità della linea di produzione. Un sistema FDD intelligente è quindi molto richiesto per i sistemi di produzione smart. Questa tesi mira a proporre nuovi sistemi di supporto alle decisioni utilizzando l'intelligenza artificiale e algoritmi di elaborazione di segnali e immagini per migliorare i sistemi FDD intelligenti. In particolare, presenta un nuovo metodo di monitoraggio delle condizioni che utilizza tecniche di deep learning e di elaborazione dei segnali per le macchine rotanti in modo che il produttore possa praticare un sistema di diagnosi privo di errori. La tesi sviluppa, quindi, un algoritmo di deep learning generativo ibrido in grado di creare campioni sintetici realistici, apprendere in modo efficiente da pattern di guasto rari e raggiungere prestazioni all'avanguardia in diverse condizioni di imbalance e rumore. I risultati della validazione su un set di dati industriali con diverse gravità di rumore e sbilanciamento dimostrano la superiorità del modello proposto rispetto ai nuovi modelli selezionati per il confronto. Nel terzo articolo, la tesi propone un nuovo framework generativo per il sovra-campionamento dei dati e l'annotazione automatica dei difetti insieme ad un algoritmo ibrido per l’object detection basato sul deep learning per migliorare i sistemi AVI. Risultati sperimentali su due set di dati industriali mostrano che il modello di data augmentation proposto e l'algoritmo di identificazione dei difetti superano gli altri approcci all'avanguardia. I risultati della tesi sono promettenti e i framework sviluppati possono essere applicati a diversi settori con strutture simili. Nell'ultima parte, la tesi discute i risultati e prevede alcuni percorsi di ricerca, opportunità di ricerca e possibilità future per ottenere un efficiente rilevamento e diagnosi intelligente dei guasti per i sistemi di produzione.
Applying artificial intelligence to fault detection and diagnosis in manufacturing
Masoud, Jalayer
2021
Abstract
Traditional manufacturing processes are heavily dependent on experienced workers and human judgment. When these workers leave the manufacturer, a huge wealth of knowledge, know-how and understanding are lost. Moreover, as human judgment is bounded and biased, the decisions and operations can be far from optimal. With endless and unceasing progress in science and technology, the manufacturing systems are constantly upgrading and transforming from the traditional form to intelligent manufacturing which consists of evidence-based decision-making systems and can result in sustainable and optimized operations. Thanks to new developments in high-precision sensors, the availability of stronger computational powers and the emergence of novel Artificial Intelligence (AI) techniques, today’s manufacturing operations integrate resources in a more complicated manner and a wider scale to enhance its performance and reduces its costs. Intelligent Condition Monitoring and Automated Visual Inspection (AVI) systems, known as Fault Detection and Diagnosis (FDD), are among the most important research areas, which can significantly cut the manufacturing costs and contribute to the quality assurance and reliability of the production line. An intelligent FDD system is thus highly in demand for smart manufacturing systems. This thesis aims at proposing novel decision-support systems, utilizing AI and signal and image processing algorithms to improve intelligent FDD systems. It presents a novel condition monitoring method that employs deep learning and signal processing for rotating machinery such that the manufacturer can practice an error-free diagnosis system. Then it develops a hybrid generative deep learning algorithm that can create realistic synthetic samples, efficiently learn from rare fault patterns and achieves state-of-the-art performance in different imbalanced and noisy conditions. The evaluation results on an industrial dataset with different noise and imbalance severities demonstrate the superiority of the proposed model over the novel models selected for the comparison panel. In the third paper, we propose a novel generative framework for data oversampling and automatic annotation of the defects alongside a hybrid deep learning-based image object detection algorithm to improve AVI systems. After some experimental results on two industrial datasets, show that the proposed data augmentation model and defect detection algorithm outperforms the other state-of-the-art approaches. The thesis results are promising, and the developed frameworks can be applied to different industries with similar facilities. In the last part, the thesis discusses the findings and envisions some research paths, research opportunities and future possibilities to achieve an efficient intelligent fault detection and diagnosis for manufacturing systems.File | Dimensione | Formato | |
---|---|---|---|
Masoud Jalayer - PhD_thesis.pdf
accesso solo da BNCF e BNCR
Dimensione
17.35 MB
Formato
Adobe PDF
|
17.35 MB | Adobe PDF |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14242/204581
URN:NBN:IT:POLIMI-204581