Learning from Failure: Improving the Machine Learning Performance in Laser Cutting Defect Detection

In this paper a machine leaning model, more precisely a Convolutional Neural Network (CNN), is trained to implement a cut problem detection on a laser cutting machine. The model is trained on machine 1 but should be applied to machine 2. This task can be assigned to transfer learning. This leads to problems during the transfer, where the model systematically classifies only one class. The failure of the model is analysed with the help of SHapley Additive exPlanations (SHAP) and a new local post-hoc explainability method, called sector analysis. Using these explainable AI methods, it is discovered that the model is using features in the image for classification that are not ideal. It also reveals that a shift of the images leads to additional misclassification. There are several ways to deal with this problem. For example, pre-processing the images or using transfer learning. However, the ideal solution for the problem presented is to use a different model, namely a decision tree.