Machine tool process monitoring by segmented timeseries anomaly detection using subprocess-specific thresholds

Abstract

Time series data generated by manufacturing machines during processing is widely used in mass part production to assess if processes run without errors. Systems that make use of this data use machine learning approaches for flagging a time series as a deviation from normal behaviour. In single part production, the amount of data generated is not sufficient for learning-based classification. Here, methods often focus on global signal variance but have trouble finding anomalies that present as local signal deviations. The referencing of the process states of the machine is usually performed by state indexing which, however, is not sufficient in highly flexible production plants. In this paper, a system that learns granular patterns in time series based on mean shift clustering is used for detecting processing segments in varying machine conditions. An anomaly detection then finds deviating patterns based on the previously identified processing segments. The anomalies can then be labeled by a human-in-the-loop approach for enabling future anomaly classification using a combination of machine learning algorithms. The method of anomaly detection is validated using an industrial machine tool and multiple test series.