基于深度特征域适应的滚动轴承多状态同尺度定量评估方法研究

State assessment of rolling bearings is the important way of achieving condition-based maintenance management and ensuring the equipment safe service. In order to extract the sensitive deep features of multi-state vibration signals of rolling bearings under variable “working condition”, here, multi-state includes the normal state, different fault locations and different performance degradation degrees, using convolutional neural network and Gauss distribution characteristics, a deep belief network feature extraction model is proposed based on convolutional Gauss. Aiming to the practical problems that vibration data of different “working condition” is not the same distribution, the target domain data is lacking and the label is unknown, the domain adaptation method of transfer learning is researched, and two domain adaptation classification models are proposed, which are respectively double constraint multi-kernel support vector machine (SVM), and weighted combined-kernel joint adaptation combined with SVM. The spatial distance from each state of rolling bearings to classification hyperplane is deduced, the relative distance measure which is relevance to normal state is built. Then, combining vibration transmission distance measure and feature vector direction measure, the multi-state same scale quantitative assessment index and model of rolling bearings under variable “working condition” can be constructed, the same scale assessment curve can be drawn, and the state of rolling bearings under variable “working condition” will be same scale quantitatively assessed more accurately. The representation relationship between each state of rolling bearings under variable “working condition” and vibration transmission path, the performance degradation law of rolling bearings also can be revealed. The reliability of the equipment can be improved, and it has important scientific theory significance and engineering value.

滚动轴承状态评估是实现视情维修管理和保障设备安全服役的重要手段。为提取变“工况”下滚动轴承正常、不同故障位置、不同性能退化程度的多状态振动信号的敏感深度特征，借鉴卷积神经网络和高斯分布特性，提出卷积高斯深度信念网络特征提取模型；针对实际中不同“工况”振动数据不同分布、目标域数据稀缺且标签未知的问题，研究迁移学习中域适应方法，提出双约束多核支持向量机（SVM）和加权混合核联合适配结合SVM的两种域适应分类模型；推导滚动轴承各状态到分类超平面的空间距离，建立与正常状态相关的相对距离测度，再结合振动传递距离测度及特征向量方向测度，构建变“工况”下滚动轴承多状态同尺度定量评估指标与模型，绘制同尺度评估曲线，期望更准确地同尺度定量评估变“工况”下滚动轴承的状态。研究可揭示变“工况”下滚动轴承各状态与振动传递路径的表征关系以及滚动轴承的性能退化规律，提高设备的可靠性，具有重要的科学理论意义和工程价值。

项目为实现不同“工况”下滚动轴承正常、不同故障位置、不同性能退化程度的多状态同尺度定量评估，揭示变“工况”下滚动轴承的性能退化规律，提出了一套滚动轴承多状态定量评估方法。该套方法中包含滚动轴承各状态振动信号深度特征提取、智能分类和定量评估模型建立的具体方法。借鉴卷积神经网络和高斯分布特性，构建了卷积高斯深度信念网络（CGDBN）特征提取模型；提出将AlexNet卷积层的卷积核进行修改，构建领域共享的深度卷积网络，有效提取了变“工况”下滚动轴承振动数据的深度特征。针对实际中变“工况”振动数据分布差异大、目标域数据稀缺的问题，提出了利用加权混合核代替单核的联合分布适配（WKJDA）方法，基于此结合CGDBN构建了变“工况”下滚动轴承多状态分类模型。构建了基于支持向量数据描述（SVDD）的评估模型，提出多状态隶属度的思想，建立变“工况”下多状态定量评估指标，绘制了同尺度评估曲线，实现了滚动轴承多状态定量评估。作为扩展研究，提出了一种基于深层迭代特征级联CatBoost的滚动轴承剩余寿命预测新方法，获得了较好效果。通过理论和实验研究相结合的手段，验证了所提评估方法中的振动信号深度特征提取、智能分类和评估模型建立的有效性，为实现滚动轴承及其它设备的主动维护奠定了理论基础。