Combining Process Monitoring with Text Mining for Anomaly Detection in Discrete Manufacturing

Abstract

One of the major challenges of today’s manufacturing industry is the reliable detection of process anomalies and failures in order to reduce unplanned downtimes and avoid quality issues. Process Monitoring (PM) requires the existence of a Normal Operating Condition (NOC) dataset that is used to train the respective algorithm. Obtaining such a NOC dataset involves extensive test runs aside from the actual production. Machine operators often collect a variety of unstructured process specific data in form of protocols, that contain valuable information about the process condition. We propose an approach that utilizes such text data to efficiently create the NOC dataset for a machining process in one of our learning factories. Using the NOC high-frequency machine sensor readings, we train a principal component analysis (PCA)-based model, which can identify anomalous process behavior. The model is consequently evaluated on a holdout test data set and shows promising results. Estimations of the process condition are visualized with two control charts allowing intuitive insights for the machine operator.