An imbalance-aware masked autoencoder framework for percussion-based bolt-ball joint looseness monitoring under few-shot and long-tailed conditions

Abstract

This study addresses sample scarcity and long-tailed class distributions in percussion-based bolt-ball joint looseness monitoring by proposing an imbalance-aware continuous wavelet transform-based masked autoencoder (CWT-MAE) framework. Each single percussion signal is uniformly converted into a CWT time–frequency image and fed to a Vision Transformer (ViT) backbone. MAE self-supervised pretraining is performed on CWT time–frequency images to initialize the ViT. A class-balanced mixed cross-entropy loss (CBCE mix) is introduced for imbalance-aware optimization, and the pretrained CWT-MAE is used to reconstruct tail-class time–frequency images, which are linearly mixed with the original images to generate augmented tail-class samples for supervised training. Experimental results demonstrate the effectiveness and superiority of the proposed framework for bolt-ball percussion monitoring. On a constructed dataset with scarce samples, the framework achieves high accuracy close to that under a near-complete data setting and it maintains stable advantages on medium-scale and large-scale datasets.