Abstract: Aiming at the challenging issues including insufficient coverage and credibility of sensor single-channel condition monitoring data,different data distribution under different operating conditions and high expert annotation costs,a multi-channel data fusion-based adaptive prototype comparative learning intelligent diagnosis method was proposed to realize bearing fault diagnosis based on multi-channel data under variable operating conditions and extremely few labels.By utilizing multi-channel vibration data from sensors,a sample enhancement strategy for multi-channel information fusion was established to improve sample quality;Kullback-Leibler divergence-guided adaptive optimization of concentration coefficient for prototype comparative learning was designed,thereby more representative cross-domain transferable features being extracted through self-supervised domain adaptation;semi-supervised bearing fault classification was achieved with extremely few labeled samples in source domain.The experimental results showed that with 1% labeled data in the source domain,the proposed method achieved an average diagnostic accuracy of 97.1% or higher in 12 cross-domain diagnostic tasks for 7 bearing fault categories,which was at least 15.5%,16.6%,13.5%,14.9%,19.1% and 12.7% higher than the other 6 comparative methods respectively.

Key words: domain adaptation, intelligent bearing diagnosis, information fusion, prototype contrastive learning

摘要： 针对传感器单通道状态监测数据信息涵盖量与可信度不足、不同工况下数据分布不同、专家标注成本高昂等挑战性问题,提出多通道数据融合的自适应原型对比学习智能诊断方法,实现极少量标签下基于多通道数据的变工况轴承故障诊断。首先,利用传感器多通道振动数据,建立多通道信息融合的样本增强策略,提升样本质量;然后,设计相对熵引导集中度系数自适应优化的原型对比学习,通过自监督域适应提取更具表征力的跨域迁移特征;最后,在源域样本极少量标签的半监督下,实现轴承故障分类。实验结果表明,源域有标签数据占比1%时,在面向轴承7种故障类别的12种跨域诊断任务中,所提方法的平均诊断准确率达97.1%及以上,相比其他6种对比方法至少分别提升了15.5%、16.6%、13.5%、14.9%、19.1%和12.7%。

关键词: 域适应, 智能轴承诊断, 数据融合, 原型对比学习