Machine Learning-Based Feature Evaluation for Scrap Float Detection with Accelerometers in Stamping

This study entailed an evaluation of the features observed in the retrofitted-accelerometer signal obtained for scrap float detection in stamping. The author employed a machine-learning technique with retrofitted accelerometers. Three accelerometers (hereafter we call them sensors A, B, and C) were mounted on three different locations on the side surface of the stripper plate of a three-holes-stamping die set. A1050 aluminum sheets of thickness 0.8 mm were used as blank materials. Six features in the accelerometer signal were determined for the detection. Half of them were associated with the downward journey of the press slide, and another three were associated with its return journey. Next, the author applied a machine learning technique using the Mahalanobis-Taguchi system. For sensor C, unknown and unexpected events unrelated to the scrap floating occurred two times to cause false detections of normal samples without scraps. By contrast, the unit spaces which contain normal samples for sensors A and B were completely separated from the error samples with scraps. Moreover, the author conducted the S/N ratio analysis to investigate the influential features on the detection of the error samples for each sensor. Depending on the above, the author conducted machine learning with selected three features for each sensor. As a result, for sensors A and B, normal and error samples were separated as completely as the detection with six features. For sensor C, false positives for two normal samples when six features were used were now also correctly detected to be normal.