Multi-Task Learning-based Structure Damage Identification With Diffusion Data Augmentation

Deep learning has become popular in aircraft engine fault diagnosis. Since aircraft engine fault diagnosis typically requires both identifying the location of structural damage and quantifying its size, it can be viewed as a multi-task learning problem. However, practical applications are often constrained by small sample datasets, as obtaining sufficient damaged and undamaged signal samples is inherently difficult. In order to address this limitation, we integrate diffusion models into the deep learning pipeline, enhancing both the quantity and diversity of damaged sample features. Moreover, we propose a multi-task joint learning framework leveraging a Dilated-CNN architecture, tailored to optimize fault diagnosis performance. Experimental results across multiple evaluation metrics demonstrate the superiority of our approach compared to traditional models employing conventional data augmentation techniques.