Long-tailed representation learning algorithm based on adaptive prototypes and semantic awareness

Label scarcity and long-tailed distribution imbalance are significant challenges in industrial equipment monitoring. Currently, self-supervised learning methods are affected by sample quantity bias and semantic confusion under complex operating conditions, which limits their ability to represent sparse critical states. To address these issues, we propose a co-evolutionary prototypical contrastive learning (EPCL) framework. Through progressive learning from coarse-grained semantic discovery to fine-grained discriminative enhancement, this framework enables an in-depth analysis of the intrinsic structure of long-tailed data. Specifically, an adaptive prototype-based clustering algorithm based on optimal transport theory is introduced, thereby achieving unbiased representation learning through data-driven dynamic priors. Furthermore, a semantic-aware and hierarchical negative sample weighting scheme is designed to optimize discriminative boundaries while mitigating class imbalance by enforcing prototype consistency constraints and employing an adaptive weighting strategy. Extensive experiments were conducted on several public long-tailed visual benchmarks, including CIFAR10-LT, CIFAR100-LT, and ImageNet-100-LT, as well as the industrial fault diagnosis dataset. The results demonstrated that the EPCL achieved better performance than fifteen mainstream self-supervised methods (e.g., SimCLR and SwAV) in both linear evaluation and few-shot classification tasks. On the CIFAR100-LT dataset, the EPCL improved the tail-class accuracy by 4.56% compared to SimCLR. Ablation studies and visualization results verified the effectiveness and generalization ability of the framework. This work offers a promising insight and practical solution for representation learning from unlabeled long-tailed measurement data.