二次聚合个性化联邦的不同工况下滚动轴承寿命预测方法  

Life prediction method for rolling bearings under different operating conditions based on secondary aggregation personalized federated learning

作　　者：康守强[1] 杨得济 王玉静[1] 王庆岩[1] 谢金宝 KANG Shouqiang;YANG Deji;WANG Yujing;WANG Qingyan;XIE Jinbao(Heilongjiang Province Key Laboratory of Pattern Recognition and Information Perception,Harbin University of Science and Technology,Harbin 150080,China;College of Physics and Electronic Engineering,Hainan Normal University,Haikou 571158,China)

机构地区：[1]哈尔滨理工大学模式识别与信息感知黑龙江省重点实验室,哈尔滨150080 [2]海南师范大学物理与电子工程学院,海口571158

出　　处：《振动与冲击》2025年第2期254-266,共13页Journal of Vibration and Shock

基　　金：国家自然科学基金(52375533);黑龙江省自然科学基金(PL2024E022);山东省自然科学基金(ZR2023ME057);哈尔滨市制造业科技创新人才项目(2023CXRCCG017)。

摘　　要：针对不同工况下滚动轴承振动数据分布差异大,单一用户数据量少且多个用户间数据不共享的问题,提出一种二次聚合个性化联邦的滚动轴承寿命预测方法。该方法用不同深度的自编码器提取多尺度特征信息并压缩为散点图,实现特征增强;利用无监督二元回归模型确定第一预测时间,构建分段退化标签;提出二次聚合个性化联邦学习算法,各用户构建改进的卷积神经网络-长短时记忆网络模型,并将其参数上传至服务端,服务端采用多任务学习框架,一次聚合多用户同种工况模型参数;在此基础上,利用批量归一化层参数统计信息计算一次聚合模型间相似度,引入权重更新机制指导模型参数二次聚合,减少不同工况模型间的负迁移现象并学习有益的全局知识,最终形成针对各工况的个性化预测模型。经试验验证,所提方法在保障数据隐私的前提下,可实现不同工况下滚动轴承寿命预测,并且预测的平均得分与不考虑数据隐私的集中式学习方法相当、相较于联邦平均算法平均得分提高0.2197。For rolling bearings,addressing the issues of significant differences in vibration data distribution under various operating conditions,limited data for individual user and the data not able to share among multiple users,a life prediction method for rolling bearings under different operating conditions was proposed based on secondary aggregation personalized federated learning.Auto-encoders with different depth were utilized to extract the multi-scale feature information and compress it into a nested-scatter plot to achieve feature enhancement.An unsupervised binary regression model was employed to determine the first prediction time and construct segmented degradation labels.A secondary aggregation personalized federated learning algorithm was proposed.For each user,an enhanced convolutional neural network-long short term memory model was constructed and its parameters were uploaded to the server.The server adopted a multi-task learning framework,primarily aggregated model parameters under the same operating conditions from multiple users,furthermore,the server utilized batch normalization layer parameter statistics to calculate the similarity among one-time-aggregation-models.A weight updating mechanism was introduced to guide the secondary aggregation of model parameters.Negative transfer effects among models were reduced under different operating conditions,and the learning of beneficial global knowledge was facilitated.Ultimately,personalized prediction models tailored to specific operating conditions were established.Experimental validation demonstrates that the proposed method can achieve rolling bearing life prediction under different operating conditions while ensuring data privacy.The prediction average score is comparable to the centralized learning method without considering data privacy,and increases by 0.2197 compared to the federated average algorithm.

关 键 词：滚动轴承 多尺度特征提取 联邦学习 个性化 剩余寿命预测 

分 类 号：TH165[机械工程—机械制造及自动化] TP18[自动化与计算机技术—控制理论与控制工程]