一种基于深度迁移学习的滚动轴承早期故障在线检测方法

A New Deep Transfer Learning-based Online Detection Method of Rolling Bearing Early Fault

近年来,深度学习技术已在滚动轴承故障检测和诊断领域取得了成功应用,但面对不停机情况下的早期故障在线检测问题,仍存在着早期故障特征表示不充分、误报警率高等不足.为解决上述问题,本文从时序异常检测的角度出发,提出了一种基于深度迁移学习的早期故障在线检测方法.首先,提出一种面向多域迁移的深度自编码网络,通过构建具有改进的最大均值差异正则项和Laplace正则项的损失函数,在自适应提取不同域数据的公共特征表示同时,提高正常状态和早期故障状态之间特征的差异性;基于该特征表示,提出一种基于时序异常模式的在线检测模型,利用离线轴承正常状态的排列熵值构建报警阈值,实现在线数据中异常序列的快速匹配,同时提高在线检测结果的可靠性.在XJTU-SY数据集上的实验结果表明,与现有代表性早期故障检测方法相比,本文方法具有更好的检测实时性和更低的误报警数.

In recent years,deep learning techniques have been successfully applied to fault detection and diagnosis for rolling bearings.However,for online detection of incipient fault without system halt,these techniques still have some shortcomings such as insufficient feature representation of incipient fault and high false alarm rate.To solve such problems,this paper presents a new deep transfer learning-based online detection approach on the perspective of temporal anomaly detection.First,a new deep auto-encoder network with multi-domain transferring is proposed by constructing a new loss function with the maximum mean discrepancy regularizer and Laplace regularizer.This model can adaptively extract the common feature representation among the data of different domains,and effectively improve the feature difference between normal state and early fault state as well.Second,with the obtained feature representation,a new online detection model based on temporal anomaly pattern is proposed.By utilizing the permutation entropy of normal state of offline bearings to build an alarm threshold,this model can match quickly anomaly sequence of the online monitoring data,and then improve the detection reliability.The experimental results on the XJTU-SY bearings dataset demonstrates that the proposed approach obtains better real-time detection performance and lower false alarm rate compared to some state-of-the-art methods of incipient fault detection.