Corrosion of turbine blades is a critical reliability and safety concern in turboshaft engines, yet automated detection remains difficult due to the very limited availability of annotated data. This work develops a data-efficient deep-learning approach capable of identifying corrosion under such constraints. A lightweight synthetic data pipeline is introduced to generate simplified but realistic corrosion patches that are blended into clean blade images, forming a large training and validation set. Real corrosion samples are reserved exclusively for testing to avoid data leakage. A dual-head deep learning model architecture is proposed, combining a classification branch for corrosion detection with a segmentation branch that improves feature learning from synthetic patterns. Experiments on blades from 13 engines, comprising 39 source images and 7020 subimages, show that the proposed model achieves a classification accuracy of 90% despite extreme data scarcity. These results demonstrate that even simple synthetic augmentation, when combined with a joint classification-segmentation design, substantially improves generalisation and provides interpretable outputs suitable for turboshaft blade inspection workflows.
Kobrlova et al. (Mon,) studied this question.