Composite acoustic hole segmentation by semi-supervised learning for robotic multi-spindle drilling of aero-engine nacelle acoustic liners

Abstract

The large-scale tiny acoustic holes densely distributed on acoustic liners are essential in aero-engine noise reduction. Accurate segmentation of those holes is fundamental for a robotic multi-spindle drilling system. This paper introduces a novel semi-supervised segmentation method for acoustic holes on composite acoustic liners. This method uses perturbation consistency loss to ensure output consistency while solving the problem of data volume imbalance. Afterward, a multi-reliability enhancement method and a pseudo-label reliability enhancement module are utilized to enhance the model’s robustness and accuracy. The segmentation experiments show that our method is superior to UniMatch and FixMatch. When using only 30% of labeled data, our method achieves an IoU of 96.39%, and the porosity differs only 0.038% from the ground truth, which is better than the fully-supervised segmentation method using all data. Our method can meet the accuracy and efficiency requirements of aero-engine nacelle production in China with only limited labeled data.