Nowadays, non-destructive techniques play a fundamental role in the production industry since the early detection of defects can reduce possible costs and avoid catastrophic failures. Under these aspects, all methods for fast and reliable inspection deserve special attention. This thesis aims to correctly detect manufacturing defects or other damage mechanisms without compromising the original condition of the material (non-destructive testing, NDT) through the use of the active IR thermopgragy and automatic semantic segmentation. In computer vision, semantic segmentation refers to the task of identification and exact classification of all the objects in a scene based on learned concepts. It is one of the fundamental tasks since it is the basis of other following tasks such as scene understanding and scene parsing. [1] The segmentation of defects in composite materials is achieved by using a deep learning algorithm on a high-variance dataset from lock-in thermography performed under five different heat source configurations. The specimen considered is made of Carbon Fiber Reinforced Polymer (CFRP) enriched by several inclusions of foreign materials. A deep network has been adopted to analyse the thermal signals. During a training phase, several examples from defect areas have been fed up to the network in order to automatically acquire knowledge to be used in recognizing the defects in the test phase. The experiments performed in this thesis showed that the best architecture of the neural network is characterized by a Stem Cell and ten layers of alternated Temporal Base Blocks (TBBs) and max-pooling layers. The Intersection over Union (IOU) metric has been used to evaluate the results demonstrating that the proposed solution reaches an average value of 0.685. The qualitative results of the proposed approach show that the network can detect most of the defects, providing valuable support to specialized operators, hence optimizing quality control both in terms of execution time and detection accuracy.
Al giorno d'oggi, le tecniche non distruttive giocano un ruolo fondamentale nell'industria manifatturiera poiché la diagnosi precoce dei difetti consente di ridurre i possibili costi ed evitare guasti. Sotto questi aspetti, tutti i metodi per un'ispezione rapida e affidabile meritano un'attenzione particolare. L'obiettivo di questa tesi è infatti quello di apprendere come rilevare correttamente i difetti di fabbricazione o altri meccanismi di danneggiamento, senza compromettere le condizioni originarie del materiale. Per raggiungere questo scopo, la tesi si occupa dell'analisi di immagini termografiche per i controlli non distruttivi di materiali compositi attraverso l’applicazione di tecniche automatiche di segmentazione semantica. Nell’ambito della visione computazionale, la Semantic Segmentation mira ad identificare e classificare tutti gli oggetti in una scena. Questo processso è fondamentale poiché costituisce una sostanziale pre-elaborazione per altre attività, tra cui la comprensione e l'analisi delle scene. [1] Nell’ambito di questa tesi si è dimensionato un algoritmo di Machine Learning in grado di evidenziare automaticamente difetti interni a partire da un set di dati ottenuto dall’applicazione di termografia attiva lock-in con cinque diverse configurazioni di eccitazione. I test sono stati condotti su un campione noto di Carbon Fiber Reinforced Polymer (CFRP) arricchito con inclusioni di materiale tali da simulare la presenza di difetti. Per analizzare i segnali termografici è stata adottata una deep neural network addestrata attraverso esempi di aree difettose con l’obiettivo di generalizzare le conoscenze acquisite e riconoscere i difetti nelle successive fasi di test. Gli esperimenti condotti in questa tesi hanno dimostrato che la migliore architettura di rete è caratterizzata da una Stem cell e dieci strati di Temporal Base Blocks (TBBs) alternati ad altrettanti strati max-pooling. In questo caso, la metrica Intersection over Union (IOU) ha raggiunto un valore medio di 0.685 dimostrando come la rete sia in grado di rilevare la maggior parte dei difetti, fornendo supporto alle decisioni per gli operatori specializzati, ottimizzando così il controllo di qualità sia in termini di tempo di esecuzione che di accuratezza di rilevamento.
ANALISI DI IMMAGINI TERMOGRAFICHE PER I CONTROLLI NON DISTRUTTIVI DI MATERIALI COMPOSITI
GUARAGNELLA, GIOVANNA
2019/2020
Abstract
Nowadays, non-destructive techniques play a fundamental role in the production industry since the early detection of defects can reduce possible costs and avoid catastrophic failures. Under these aspects, all methods for fast and reliable inspection deserve special attention. This thesis aims to correctly detect manufacturing defects or other damage mechanisms without compromising the original condition of the material (non-destructive testing, NDT) through the use of the active IR thermopgragy and automatic semantic segmentation. In computer vision, semantic segmentation refers to the task of identification and exact classification of all the objects in a scene based on learned concepts. It is one of the fundamental tasks since it is the basis of other following tasks such as scene understanding and scene parsing. [1] The segmentation of defects in composite materials is achieved by using a deep learning algorithm on a high-variance dataset from lock-in thermography performed under five different heat source configurations. The specimen considered is made of Carbon Fiber Reinforced Polymer (CFRP) enriched by several inclusions of foreign materials. A deep network has been adopted to analyse the thermal signals. During a training phase, several examples from defect areas have been fed up to the network in order to automatically acquire knowledge to be used in recognizing the defects in the test phase. The experiments performed in this thesis showed that the best architecture of the neural network is characterized by a Stem Cell and ten layers of alternated Temporal Base Blocks (TBBs) and max-pooling layers. The Intersection over Union (IOU) metric has been used to evaluate the results demonstrating that the proposed solution reaches an average value of 0.685. The qualitative results of the proposed approach show that the network can detect most of the defects, providing valuable support to specialized operators, hence optimizing quality control both in terms of execution time and detection accuracy.I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14253/95314