A Machine-Learning-Based Product Defect Diagnosis Method Using Multi-Sensor Information in Press Production

The diverse components that make up automobiles are manufactured through press forming processes using metal sheets. These processes utilize press production equipment and dies to form the parts. As press forming is an upstream manufacturing process, defects occurring at this stage can propagate to subsequent processes and adversely affect overall production. Hence, accurate defect diagnosis is essential for maintaining product quality. This study aims to develop a diagnostic method that utilizes production data obtained from multiple sensors installed on press production equipment. We verified the effectiveness of the machine-learning-based defect diagnosis method using actual production data. Abnormalities in the die cushion pressure and slide velocity signals were diagnosed and visualized using the reconstruction error of the proposed autoencoder-based method. To verify the effectiveness of the proposed method, we also visualized and compared the features used to determine abnormalities with a one-dimensional convolutional neural network combined with Grad-CAM. As a result, both methods consistently identified the characteristics of defects, objectively demonstrating the effectiveness of the proposed method.