As an essential component of many mission-critical equipment, mechanical bearings need to be monitored to identify any traces of abnormal conditions. Most of the latest data-driven methods applied to bearing anomaly detection are trained using a large amount of fault data collected a priori. However, in many practical applications, it may be unsafe and time-consuming to collect enough data samples for each fault category, which brings challenges to training a robust classifier. This paper proposes a few-shot learning framework for bearing anomaly detection based on model-agnostic meta-learning (MAML), which aims to train an effective fault classifier using very limited data. In addition, it can use training data and learn to more effectively identify new fault conditions. A case study on the generalization of new artificial faults shows that this method can achieve up to 25\% overall accuracy when compared to a benchmark study based on the Siamese network. Finally, the generalization ability of MAML is also competitive when compared with some state-of-the-art few-shot learning methods in terms of identifying realistic bearing damages using a sufficient amount of training data from artificial damages.
翻译:作为许多关键任务设备的基本组成部分,需要监测机械轴承,以发现异常状况的任何痕迹。大多数用于异常检测的最新数据驱动方法都是使用事先收集的大量故障数据进行培训的。然而,在许多实际应用中,收集每个故障类别足够的数据样本可能不安全而且耗费时间,这对培训一个强有力的分类员提出了挑战。本文件提议了一个以模型 -- -- 遗传元学为基础,用微小的学习方法来测出异常现象,目的是用非常有限的数据培训一个有效的故障分类员。此外,它可以利用培训数据并学习更有效地查明新的故障条件。关于新的人为故障的概括化案例研究表明,与以西亚姆斯网络为基础的基准研究相比,这种方法可以达到25 ⁇ 的总体准确度。最后,在使用足够数量的人工损害培训数据确定实际承受损害方面,MAML的通用能力与一些最先进的学习方法相比,也具有竞争力。