IEEE Transactions on Automation Science and Engineering

Streaming recognition of robotic-arm degradation is challenging due to severe class imbalance, sparse labels, and the progressive emergence of new states. Existing stream-based active learning methods often rely on fixed feature spaces and static class sets, leading to redundant queries during healthy-only phase and slow adaptation when new states appear. Although contrastive learning can improve representation quality under imbalance, it is typically trained offline and remains decoupled from query selection, limiting the responsiveness to evolving tasks. To address these limitations, we propose a contrastive ensemble active learning (ConEAL) framework that integrates staged contrastive representation learning with contextual and rewardadaptive acquisition. The encoder first learns invariant and compact feature embeddings by training with an unsupervised contrastive loss on healthy state data. As new classes emerge, an anchor-based triplet loss with dynamic anchor updates is applied to improve the separability between classes. Based on the encoder output and predicted posteriors, ConEAL collects normalized acquisition scores from multiple agents. It then updates the agent weights online using a contextual bandit strategy, guided by the reward from each query, while respecting both the global labeling budget and the temporal order of the data stream. Evaluated on an accelerated-wear robotic dataset, ConEAL achieves a superior utility-cost (i.e., classification accuracy and labeling cost) trade-off compared to strong stream-based baselines, attaining earlier regime adaptation and sustained three-state recognition with substantially fewer annotations. The results indicate that the joint adaptation of representations and query policies is an effective strategy for label-efficient degradation monitoring in nonstationary robotic settings.