Mixed-Up Experience Replay for Adaptive Online Condition Monitoring
Mixed-Up Experience Replay for Adaptive Online Condition Monitoring
复制标题
DOI：
10.1109/tie.2023.3260351
复制
发表时间：
2024-02
影响因子：
7.7
通讯作者：
Matthew Russell;P. Wang;Shaopeng Liu;I. S. Jawahir
中科院分区：
文献类型：
--
作者：
Matthew Russell;P. Wang;Shaopeng Liu;I. S. Jawahir
数据驱动的预测性维护减少了制造停机时间，复杂的过程传感关系鼓励使用深度学习来自动提取特征。然而，标记的训练数据往往是缺乏的，新的故障条件可能会发生。实际部署必须从未标记的数据中学习，适应新出现的条件，并且在事先不知道条件何时变化的情况下这样做。将最先进的自监督学习（SSL）与持续学习（CL）相结合，有助于在观察到新条件时进行适应。本文提出了一种基于Barlow Twins SSL和新的混合经验重放（MixER）的自适应在线状态监测框架，用于无监督CL。Barlow Twins专为一维传感数据量身定制，可有效地对未标记的数据进行聚类。当与MixER相结合时，该系统在运动健康状况数据集上的表现优于最先进的无监督CL，达到92.4%的分类准确率。未来的工作将证明人在回路集成真实的制造环境。
Data-driven predictive maintenance reduces manufacturing downtime, and complex process-sensing relationships encourage the use of deep learning to automatically extract features. However, labeled training data are often lacking, and novel fault conditions may occur. Practical deployments must learn from unlabeled data, adapt to emerging conditions, and do so without prior knowledge of when the condition changes. Combining state-of-the-art self-supervised learning (SSL) with continual learning (CL) facilitates adaptation as new conditions are observed. This article proposes a framework for adaptive online condition monitoring based on Barlow Twins SSL and novel Mixed-Up Experience Replay (MixER) for unsupervised CL. Tailored for 1-D sensing data, Barlow Twins effectively clusters unlabeled data. When combined with MixER, the system outperforms state-of-the-art unsupervised CL on a motor health condition dataset, reaching 92.4% classification accuracy. Future work will demonstrate human-in-the-loop integration for real manufacturing environments.