基于小样本数据增广的产品服役性能预测方法
Product Operating Performance Prediction Based on Small-sample Data Augmentation Method
刘振宇 1张楠 1裘辿 1谭建荣1
作者信息
- 1. 浙江大学计算机辅助设计与图形学国家重点实验室 杭州 310027
- 折叠
摘要
高精密机械产品往往生产规模有限,且各装配特征参数与性能之间存在错综复杂的耦合关系,难以用显性的映射关系表达.构建能够准确预测服役性能的映射模型,对于产品装调具有切实的指导意义.为此,提出一种小样本数据下产品服役性能多参数关联分析与预测方法.考虑重要组件装配结合面接触刚度特征,将接触刚度与原始装配数据相融合;通过随机森林对装配特征参数进行重要度排序,筛选出关键特征;针对小样本不完备数据,利用变分自动编码器学习先验分布知识,实现装配数据的模拟增广,并对增广数据进行离群值修正;将装配增广样本与原始装配数据相融合,通过集成核超限学习机实现服役性能的预测.最后以某型位标器陀螺仪稳速电流预测为例,验证提出方法的有效性.
Abstract
Generally,the production scale of high-precision mechanical products is limited,and there is a complex coupling relationship between assembly characteristic parameters and performances,which is difficult to express by explicit mapping relationship.Building a mapping model that can accurately predict operating performance has practical guiding significance for product assembly and adjustment.To this end,a product operating performance parameter correlation analysis and forecasting method for small-sample data is proposed:The contact stiffness of joint surface is considered to merged into assembly data,and the random forest is adopted to select key characteristics by calculating the importance of assembly feature parameters.The variational autoencoder learning is used to learn prior distribution knowledge of small sample data,then augmented assembly data is generated,and outlier correction is performed on augmented data.The assembly augmented sample is fused with the original assembly data,and the operating performance is predicted by integrated kernel extreme learning machine.Finally,an example of operating performance prediction of a position marker is taken to verify the effectiveness of the proposed method.
关键词
小样本/数据增广/接触特性/服役性能/超限学习机Key words
small sample/data augmentation/contact characteristic/operating performance/extreme learning machine引用本文复制引用
基金项目
科技创新2030-"新一代人工智能"重大项目(2021ZD0113100)
国家自然科学基金(52075480)
国家自然科学基金(51875517)
浙江省自然科学基金(LY20E050015)
出版年
2024
NETL
NSTL
万方数据