TOOL WEAR PREDICTION BY DEEP LEARNING FROM AUGMENTABLE VISIBILITY GRAPH REPRESENTATION OF TIME SERIES DATA

Abstract

Tool wear prediction has a crucial role for improving manufacturing quality and reliability due to optimizing tool replacement schedules, reducing downtime, and improving overall production efficiency. Deep learning models, having the ability to analyze large and complex datasets, can extract relevant information, and make accurate predictions about the condition of cutting tools. We propose a smart detection methodology based on converting the available sensory data collected from a CNC milling machine into a visibility graph representation. Due to the high dimensionality of the data with 44 attributes related to machining, a multilayer visibility graph representation is achieved after this conversion procedure, resulting in a 44-layered 128x128 adjacency matrix formation. A novel data augmentation technique specifically applicable to graph representation is also employed to increase the data size originally composed of 18 experiments into 360, each one represented as a multilayer graph. Augmented graph representations are further input to a custom CNN deep learning architecture with a split of 70% train, 10% validation and 20% test instances. Results indicate that Augmented Graph-induced classification of CNC mill tool with custom CNN model (GA-CNN) yields full accuracy for detecting whether the tool is worn or not.