Biegel, Tobias (2023)
A self-supervised learning approach for multivariate statistical in-process control in discrete manufacturing processes.
Technische Universität Darmstadt
doi: 10.26083/tuprints-00026455
Dissertation, Erstveröffentlichung, Verlagsversion

Kurzbeschreibung (Abstract)

Monitoring discrete manufacturing processes to reliably detect anomalies, is of fundamental relevance to reduce unplanned downtimes and ultimately gain a competitive advantage. Multivariate Statistical In-Process Control (MSPC) represents one of the main areas of research that center around the detection of anomalies in process data and, as such, is heavily influenced by the developments of general Anomaly Detection (AD) research. Recent findings from various domains demonstrate that the incorporation of Self-Supervised Learning (SSL) into AD methods leads to unprecedented performance levels and is regarded as a major milestone in the AD literature. However, while these findings suggest that the incorporation of SSL might be beneficial for MSPC in discrete manufacturing, there is currently no research that incorporates SSL into MSPC for discrete manufacturing processes. The objective of this thesis is to close this gap and encompasses the development of an SSL approach for MSPC in discrete manufacturing processes. This approach is referred to as Self-Supervised Multivariate Statistical In-Process Control (SSMSPC). SSMSPC consists of three components: (1) Location + Transformation prediction pretext task, (2) AD downstream task, and (3) control chart extension. The objective of the Location + Transformation prediction pretext task is to classify the augmentation Tᵢ and the window Wⱼ based on a time series sample that is artificially augmented by one of k augmentation functions in one of p windows. In the AD downstream task, the learned representations from the pretext task are used to compute the anomaly score, which corresponds to the Hotelling’s T² statistic. The control limits of the control chart are determined with the help of a Kernel Density Estimation-based threshold selection scheme. The control chart extension provides an additional view to the conventional control chart, in which the anomalous regions in the process data are highlighted. The developed approach is validated both from a performance and a usability perspective. In terms of performance, SSMSPC is benchmarked against state-of-the-art shallow, deep, and self-supervised AD methods using the Bosch CNC milling dataset and the Center for Industrial Productivity Discrete Manufacturing Dataset (CiP-DMD). Both datasets represent real-world CNC milling processes in which high-frequency process data are collected. SSMSPC is shown to outperform leading state-of-the-art baselines on both datasets, achieving the highest overall score in terms of area under the receiver operating characteristic curve. The usability of SSMSPC is evaluated in the context of a usability study with human machine operators. In this study, the CNC milling process of the CiP-DMD is monitored with a deployed instance of SSMSPC and univariate post-process Statistical Process Control (SPC), which represents the accepted industry standard in discrete manufacturing. The usability is evaluated with the help of monitoring protocols as an objective evaluation measure and the System Usability Scale as a subjective evaluation measure. It is shown that the outstanding detection and localization capabilities provided by SSMSPC are not yet sufficient to effectively support the root-cause analysis. Instead, the findings of the usability study suggest to combine the control chart extension of SSMSPC with the insights obtained from a quality control step of the manufactured part to exceed the usability of univarate post-process SPC, both from an objective and subjective point of view.

Frage zum Eintrag

Redaktionelle Details anzeigen