Defects Detection for Wire Drawing Machine Using Deep Learning

As the digitalization in the manufacturing industry is progressing, the use of machine learning is expanding. Among these, deep learning has been widely studied because of its potential high performance. However, there are challenges in applying deep learning techniques to production systems, such as cost, collection of large amounts of data, and rapid updating of models in response to changes of, e.g., the production system. The purpose of this study is to propose a guideline for applying deep learning techniques to production systems. To achieve the purpose, we take defect detection of products in a wire drawing machine as an example and extract valid guidelines from it. While implementing a defect classification model that classifies various defects in wire from images in the wire drawing machine, we aim to solve typical problems in applying deep learning techniques, such as high computational cost of the model for real-time processing and insufficient amount of defect data. For reducing the computational cost, we selected models with small FLOPs (Floating Point Operations Per Second) and widely used in the field of image recognition. To address the insufficient amount of data, we employed fine tuning and data expansion. Furthermore, Score-CAM was applied to each convolutional layer of the model to visualize focal areas of classification. This identified the convolutional layers that were particularly useful for classification. Based on this result, we improved the model to emphasize the feature maps output from these useful layers. As a result, the target accuracy was achieved, indicating that the approach proposed in this study is effective for image classification tasks, which poses inherent issues in applying deep learning techniques to production systems.