Beschreibung
Monitoring process errors and tool condition in single item production is challenging, as a teach-in is not possible due to a missing reference process. An approach to this problem is anomaly detection, e.g. based on motor currents or axis position signals from metal cutting processes. However, with no references anomaly detection struggles to detect failures from signals, because failure patterns are often too similar to regular process dynamics. While single items inherently constitute an anomaly by themselves, they do contain repetitive elements, like boreholes or milled pockets. These elements are similar, what provides an anomaly detection with additional information on regular processes. Hierarchical K-Means clustering combined with Dynamic Time Warping (DTW) and Barycenter Averaging (DBA) enables the identification of similar process elements. The algorithm allows ordering similar process segments by similarity in a tree structure. The introduced method supports querying subsequences from any given cutting process, for which it returns the closest cluster in the tree. This allows to (a) improve the data basis for anomaly detection and (b) to transfer labels with errors between processes. The article demonstrates the transfer of labels (for errors) from a turning process, to a single item milling process.
