融合图标签传播和判别特征增强的工业机器人关键部件半监督故障诊断方法

Semi-supervised Fault Diagnosis Method via Graph Label Propagation and Discriminative Feature Enhancement for Critical Components of Industrial Robot

RV减速器作为工业机器人关键部件之一,其机械故障将造成整机定位精度下降。围绕RV减速器开展健康状态监测与智能故障诊断具有重要意义。基于故障标记数据充足的假设,数据驱动的智能诊断方法可以有效建立监测信号与健康状态的非线性映射关系。然而,在工程实际中收集大量故障标记样本需要昂贵的标记代价和人力成本。针对上述问题,提出了一种融合图标签传播和判别特征增强的RV减速器半监督故障诊断方法。首先,利用标签传播算法赋予无标记样本伪标签。然后,通过信息熵定量评估伪标签置信度,降低误标记对模型半监督学习的干扰。同时,在深度特征嵌入空间下优化少量标记样本度量损失,构造更具判别力的特征图,提升伪标签质量。最后,采用实际工业机器人RV减速器故障数据进行方法验证。结果表明,所提半监督故障诊断方法可以对无标记样本精准地传播标签,仅利用少量标记样本获取更优的故障识别精度。

RV reducer is the critical component of industrial robot.Its mechanical faults will reduce the machine performance.The monitoring and intelligent fault diagnosis is of great significance.Traditional fault diagnosis methods assume that sufficient labeled data are available,while labeling the fault data is labor-consuming in practice.To solve this problem,a novel semi-supervised fault diagnosis method via graph label propagation and discriminative feature enhancement is proposed for RV reducer.First,the pseudo labels are produced by label propagation algorithm for unlabeled data.By using entropy,the pseudo labels are associated with a weight reflecting its certainty,so as to reduce the effect of pseudo label noise.Then,by optimizing the metric learning loss in deep embedding space for few labeled samples,the discriminative ability of feature graph is enhanced.The effectiveness of proposed method is demonstrated in the fault dataset of actual industrial robot RV reducer.The results show that the proposed semi-supervised method can produce accuracy pseudo labels,and achieve superior fault identification rate with few labeled samples.