Publication
Anomaly detection as a quality control tool in an industrial context
| Summary: | This thesis reports on the development of an automated anomaly detection (AD) systemfor industrial quality control (QC), addressing inefficiencies of manual inspections and current practices. Leveraging machine learning and computer vision techniques, this work optimizes QC using AD algorithms. The research includes a review of AD techniques, algorithm development, hyperparameter optimization, performance evaluation, real-time implementation, and API backend creation. Central to this work is a framework built on an autoencoder (AE) architecture, optimized with a custom objective function that leverages data augmentation techniques to be trained using only defect-free images. Designed as a plug-and-play solution, it can be deployed across various production lines via API calls. This study assessed fourmethodologies, each building upon the one before, except for the variational approach, which showed no improvements. The best strategy involved a refined hyperparameter optimization coupled with advanced data augmentation techniques. The overall results were primarily attributed to the developed post-processing techniques and the use of the BCE metric for evaluation. In academic evaluations on theMetal Nut subset of the MVTec AD dataset, thiswork surpassedSOTA in the MSE AE regime, achieving an AUC of 0.87 and an IoU of 0.37, outperforming handcrafted methods in the literature by ≈ 1.2% and ≈ 42.3%, respectively. Although only slightly outperformed in AUC by the top SOTA method, which achieved 0.89, the methodology still delivered highly competitive results. In collaboration with industrial partners, the system successfully detected anomalous parts using a real-world dataset. The results validate the novel methodology and showcases the potential of data-augmented NOKs as an alternative to real-world defects. Aligned with Industry 4.0 objectives, this work contributes to the advancement of smarter, data-driven manufacturing processes, facilitating the development of more automated quality control systems, reducing industrywaste, and improving market competitiveness. |
|---|---|
| Main Authors: | Peixoto, Miguel Caçador |
| Subject: | Anomaly Detection Quality Control Computer Vision Autoencoders Waste Reduction Machine Learning Real-Time Implementation HyperparameterOptimization Semi-Supervised Learning Defective Products Machine Learning Backend Deteção de Anomalias Controlo de Qualidade Visão Computacional Redução de Desperdício AprendizagemAutomática ImplementaçãoemTempoReal Otimização de Hiperparâmetros AprendizagemSemi-Supervisionada Produtos Defeituosos Backendde Aprendizagem Automática Engenharia e Tecnologia::Engenharia Eletrotécnica, Eletrónica e Informática |
| Year: | 2024 |
| Country: | Portugal |
| Document type: | master thesis |
| Access type: | open access |
| Associated institution: | Universidade do Minho |
| Language: | English |
| Origin: | RepositóriUM - Universidade do Minho |
| Summary: | This thesis reports on the development of an automated anomaly detection (AD) systemfor industrial quality control (QC), addressing inefficiencies of manual inspections and current practices. Leveraging machine learning and computer vision techniques, this work optimizes QC using AD algorithms. The research includes a review of AD techniques, algorithm development, hyperparameter optimization, performance evaluation, real-time implementation, and API backend creation. Central to this work is a framework built on an autoencoder (AE) architecture, optimized with a custom objective function that leverages data augmentation techniques to be trained using only defect-free images. Designed as a plug-and-play solution, it can be deployed across various production lines via API calls. This study assessed fourmethodologies, each building upon the one before, except for the variational approach, which showed no improvements. The best strategy involved a refined hyperparameter optimization coupled with advanced data augmentation techniques. The overall results were primarily attributed to the developed post-processing techniques and the use of the BCE metric for evaluation. In academic evaluations on theMetal Nut subset of the MVTec AD dataset, thiswork surpassedSOTA in the MSE AE regime, achieving an AUC of 0.87 and an IoU of 0.37, outperforming handcrafted methods in the literature by ≈ 1.2% and ≈ 42.3%, respectively. Although only slightly outperformed in AUC by the top SOTA method, which achieved 0.89, the methodology still delivered highly competitive results. In collaboration with industrial partners, the system successfully detected anomalous parts using a real-world dataset. The results validate the novel methodology and showcases the potential of data-augmented NOKs as an alternative to real-world defects. Aligned with Industry 4.0 objectives, this work contributes to the advancement of smarter, data-driven manufacturing processes, facilitating the development of more automated quality control systems, reducing industrywaste, and improving market competitiveness. |
|---|