Product Quality Monitoring in Hydraulic Presses Using a Minimal Sample of Sensor and Actuator Data

Abstract

Machine learning and artificial intelligence provide methods and algorithms to take advantage of sensor and actuator data in automated production systems. Product quality monitoring is one of the promising applications available for data-driven modeling, particularly in cases where the quality parameters cannot be measured with reasonable effort. This is the case for defects such as cracks in workpieces of hydraulic metal powder presses. However, the variety of shapes produced at a powder press requires training of individual models based on a minimal sample size of unlabeled data to adapt to changing settings. Therefore, this article proposes an unsupervised product quality monitoring approach based on dynamic time warping and non-linear regression to detect anomalies in unlabeled sensor and actuator data. A preprocessing step that isolates only the relevant intervals of the process is further introduced, facilitating efficient product quality monitoring. The evaluation on an industrial dataset with 37 samples, generated in test runs, shows a true-positive rate for detected product quality defects of 100% while preserving an acceptable accuracy. Moreover, the approach achieves the output within less than 10 seconds, assuring that the result is available before the next workpiece is processed. In this way, efficient product quality management is possible, reducing time- and cost-intensive quality inspections.