Large, high-dimensional, heterogeneous multi-sensor data analysis approach for process yield optimization in polymer film industry

Abstract

Today's advanced complex manufacturing processes from microelectronics to pharmaceutical industries provide huge datasets (big data) from heterogeneous multi-sensory monitoring. Analytical tools for such high-dimensional and large datasets are available to support manufacturers in quality and yield optimization, but latent problems efficiently linking process data to these tools exist, the available methods are not fully satisfactory, and the real-time support is very limited. In this work, one particular industrial process from polymer industry was investigated as a research vehicle for the development of methods for efficient process interfacing, process status classification and prediction, as well as online result visualization. Novelty filtering, anomaly detection, and one-class classification (OCC) methods have been in the focus of the investigation. These methods are of particular importance as expert knowledge from the production lines discloses that the process shows unexpected states and behavior during production runs, which lead to low yield and high cost. Causes for this abnormal behavior are assumed sensor faults, environmental influences, and unexpected raw material properties' deviations. Thus, novelty filters, anomaly detection, and OCC methods are particularly suitable for such cases. A process line from standard production with approximately 160 sensory channels has been monitored for 3 months in 1-min intervals. This provided 21,900 process data mostly normal condition samples of about 160 dimensions each for the conducted experiments. Accuracies of 99 % in OCC were achieved, and a first prototype of an advanced novelty detection GUI for process monitoring was conceived, which supports the process monitoring by the responsible staff. The achieved results allow predicting process yield problems within 2 h in advance of occurrence despite unknown objects. Future work will focus on exploiting this for advanced process control and yield optimization.