An Integrated Active Learning Framework for the Deployment of Machine Learning Models for Defect Detection in Manufacturing Environments

The digitalization of factories has triggered a humongous number of applications of Machine Learning techniques to optimize manufacturing processes. In the last years, different methodologies have been suggested for tasks like predictive maintenance, defect detection, or advanced perception systems that might contribute to a better performance of entire production lines optimizing their different steps. However, one of the main issues for the deployment of these technologies in real manufacturing environments is the management of complex tools that need, in many cases, specific AI-related knowledge from workers that are not used to work with AI-based systems. In this work we present a set of tools for defect detection tasks within a real factory based on deep learning methods and we demonstrate their performance as parts of different quality control systems. In addition, to integrate all these tools, we introduce MINT a modular intelligent framework for deep learning model management. MINT uses active learning to learn efficiently from limited labeled data, allowing manufacturers to optimize production processes with higher accuracy and reduced data annotation efforts. This framework makes more accessible to non-experts steps like model training,performance evaluation and model traceability while providing an explainability module for classification models that encourage trust and adoption of AI between factory workers. Meaning that our tool contributes to the deployment of AI techniques in real manufacturing scenarios making more accessible the management of the entire life cycle of AI models.