APPLICATION OF ARTIFICIAL INTELLIGENCE METHODS IN THE ANALYSIS OF GEAR SURFACE DAMAGE AFTER SCUFFING TESTS

This article presents an approach for applying machine learning algorithms to the recognition of damage types on the working surfaces of cylindrical gear teeth after scuffing tests. The identification process was based on a dataset containing images of damage generated during scuffing tests carried out using the FZG scuffing test method on the T-12U test rig, designed for examining cylindrical gears.The aim of this study is to evaluate the effectiveness of selected deep learning models in classifying damage on the working surfaces of gear teeth resulting from scuffing tests. Three architectures representing different approaches to image processing were compared: the Residual Network (ResNet) convolutional architecture, EfficientNet - a model based on compound scaling of network depth and width - and BLIP (Bootstrapping Language-Image Pretraining), a multimodal model combining a Vision Transformer (ViT) visual encoder with a language component that enables the integration of textual descriptions into the classification process. Although the applied models differ in their capacity for generalization and their feature extraction mechanisms, all demonstrate stable performance under data variability and effectively identify subtle differences between damage types characteristic of scuffing processes.