Generalized Statistical Process Control via 1D-ResNet Pretraining

Statistical Process Control (SPC) suffers from high false positive rates for non-normally distributed quality characteristics as stability criteria are too sensitive. This makes SPC uneconomic when applied consequently due to production downtimes. To overcome limitations of SPC, we develop an approach limiting the false postitives without changing the quality inspection workflow. Based on synthetic data subject to an approach-specific definition of stability, a 1D-Residual Neural Network (1D-ResNet) is pretrained. The pretrained model can subsequently be applied to various use cases without the need of large amounts of data. A benchmark against SPC shows a significant decrease in false positives.