Прогнозування дефектів із застосуванням машинного навчання

Abstract

This paper explores advanced approaches to automating quality control processes in manufacturing by implementing machine learning algorithms. The study focuses on the transition from traditional manual inspection to proactive defect forecasting based on real-time data analysis. Key stages of data processing, including acquisition, noise filtration, and dataset augmentation, are thoroughly analyzed to address the challenge of class imbalance. The research provides a comparative evaluation of classical statistical models against deep learning architectures, specifically Convolutional Neural Networks, for identifying structural and surface defects. Special attention is given to the mathematical foundations of model training and the minimization of error functions. The results demonstrate that integrated machine learning systems significantly enhance production reliability, reduce scrap rates, and optimize resource allocation within the framework of «Industry 4.0» standards. The practical implications suggest that utilizing models like YOLO can achieve classification accuracy exceeding 95% in industrial environments. Future research directions involve model optimization for edge computing devices