Adversarial Mixup-Based Contrast Learning for Data-Driven Predictive Maintenance in Long-Tailed Recognition

Abstract

Deep neural networks have achieved remarkable success in various computer vision tasks. However, in real-world applications, such as the Internet of Things (IoT), these models often struggle due to the long-tailed data distributions. For instance, in scenarios such as Holographic Counterpart Integration in IoT-based predictive maintenance for home systems or smart repair services, common operational states are prevalent in the dataset. In contrast, rare failures, such as hardware malfunctions or system breakdowns, are represented by only a few samples. This imbalance severely impacts models, making it difficult to accurately predict rare failures, leading to costly downtime or unanticipated equipment failure. Current contrastive learning-based methods are effective at optimizing feature distributions but often overlook inter-class relationships and are highly sensitive to class imbalance, which limits their generalization ability. To address these challenges, we propose the Adversarial Mixup-based supervised contrast learning (AMCL) framework, which integrates Mixup-based data augmentation with contrastive learning and incorporates an adversarial-inspired sample policy generator. AMCL generates boundary samples via a dynamically optimized Mixup strategy to enhance inter-class relationship modeling and improve predictions on ambiguous boundaries. Furthermore, we introduce a new MixCo loss function to account for the non-one-hot distribution of Mixup-generated targets, ensuring better alignment with augmented data and improving optimization efficiency. AMCL is easy to implement and achieves a performance superior to recent approaches for long-tailed recognition across various datasets such as ImageNet-LT, iNaturalist18, CIFAR-10-LT, and CIFAR-100-LT.