Tool Wear Classification Based on Convolutional Neural Network in Micro Drilling

Abstract

Tool wear significantly impacts both machining performance and product quality. While sensor-based approaches are prohibitively expensive and manual verification is prone to human error, this study proposes a cost-effective computer vision-based classification system using Convolutional Neural Networks (CNN). High-resolution images of tool wear were acquired using a CMO Stereo Microscope at high magnification, with the tool positioned vertically to allow for a 90° top-down image recording of the rake face. The dataset consists of two primary wear categories: Built-up Edge (BUE) and Chipping. To evaluate the system's robustness, the study acknowledges a real-world class imbalance within the dataset, specifically noting that light chipping samples are underrepresented compared to BUE. These images were processed via four distinct frameworks: Gray_Edge, Gray_MorphEdge, RGB_Enhanced, and ContourOnly. The CNN model, adapted from the AlexNet architecture and optimized using the Adam optimizer, was evaluated based on precision, recall, and F1-score. Performance results, including a validation accuracy of up to 99.26% for the ContourOnly method, demonstrate that proper image preprocessing is essential for accurate, low-cost tool wear classification in smart manufacturing environments.