UNSUPERVISED AND SEMI SUPERVISED MACHINE LEANING FRAMEWORKS FOR MULTI CLASS TOOL WEAR RECOGNITION

Abstract

Tool condition monitoring is critical in manufacturing to ensure product quality and avoid downtime. Traditional supervised machine learning methods for multi-class tool wear recognition are limited by the availability of labeled data, which is costly and often scarce in real-world industrial settings. This work proposes novel unsupervised and semisupervised machine learning frameworks designed to accurately recognize multiple classes of tool wear under varying data constraints. The frameworks utilize advanced architectures including stacked autoencoders and self-training teacher-student models to leverage both unlabeled and sparsely labeled data, overcoming the challenges of data scarcity.The proposed frameworks are evaluated on benchmark tool wear datasets, demonstrating robust performance in distinguishing different wear states without requiring exhaustive annotations. By incorporating feature extraction, dimensionality reduction, and pseudo-labeling strategies, the methods provide scalable and adaptable solutions for real-world tool condition monitoring. The results indicate that these approaches can significantly reduce the dependency on labeled data while maintaining high recognition accuracy, thereby enhancing predictive maintenance capabilities in manufacturing processes. This research contributes to extending machine learning applicability in industrial fault diagnosis with practical, label-efficient models.