Scrap Float Detection in a Blanking Die Set with Multiple Retrofit Accelerometers Using the Mahalanobis–Taguchi System

Detection of scrap floating for a stamping die with 0.8 mm-thick A1050 aluminum sheets was conducted with multiple retrofit accelerometers attached to the outside of the stamping die-set. The accelerometers were attached to three locations on the side of the stripper plate and one location on the side of the punch plate of a 3-ϕ30 hole blanking die using a magnet-based jig. Anomaly detection technique using the Mahalanobis–Taguchi system was conducted with the gravity analysis of the waveform of the accelerometers’ signal. A total of 106 experiments without foreign objects (i.e., a scrap) were conducted to collect instances of the signal profile for the normal samples. In addition, 24 error samples with a foreign object were fabricated for anomaly detection tests. Only one of the four locations achieved 100% accuracy in detection using only one sensor. In detection using only one sensor, only one of the four locations achieved 100% accuracy. We attempted to improve the accuracy by increasing the amount of learning. However, the accuracy did not improve by increasing the amount of training except for the one sensor mentioned above. This result implies that machine learning, in which features are predefined by the user, cannot compensate for the disadvantage of sensor location by the amount of training. Then, combinations of the sensors were examined. Learning with all features of all 4 sensors (i.e., with 12 features) resulted in a still imperfect separation between normal and error samples. However, even if a single sensor causes false positives, it was possible to combine the influential features of multiple sensors, that were chosen by SN ratio analysis, to detect all anomalies without false positives. In future work, we would like to consider the detection of anomalies with multi-discipline features and combine anomaly detection systems with design and quality control systems.