Detection and Handling of Laser Cutting Parameter Changes during the Deployment of Machine Learning Models

In this paper, a convolutional neural network is trained to classify images to detect cut interruptions during laser cutting. It is noticed that the accuracy of the model decreases when the trained model is applied to test data containing laser parameter changes. This is problematic because some laser parameter changes can occur unnoticed during operation. It follows that a machine learning model would fail unnoticed during operation. The investigated parameter changes have been selected by laser cutting experts as the most likely to be relevant, these are: focus position and image brightness. Various data drift detection methods are applied to the test data. These have been validated to detect data drift caused by the parameter changes. In addition, data augmentation is used to retrain a model that is less sensitive to parameter changes. This is partially successful. A data drift detection method is still required. The paper concludes with a data drift detection process for the laser cutting machine. This process acts as a safety check and monitors whether the trained model can be used or not.