涡轮泵试车数据单类支持向量机检测算法  

Support vector machines detection method for turbopump test data analysis

作　　者：胡雷[1] 胡茑庆[1] 秦国军[1] 

机构地区：[1]国防科技大学机电工程与自动化学院,湖南长沙410073

出　　处：《推进技术》2008年第2期244-248,共5页Journal of Propulsion Technology

基　　金：国家自然科学基金(50375153);高等学校全国优秀博士学位论文专项资金(200434)

摘　　要：为了在缺乏故障样本的情况下检测某型液体火箭发动机涡轮泵故障,实现基于不完整信息的状态决策,建立了基于v-支持向量分类器的单类支持向量机新异类检测模型。在分析了模型决策边界、支持向量和约束条件之间关系的基础上,为单类支持向量机引入并改进了序贯最小优化算法,提高了训练效率,解决了大样本训练问题。通过对某型液体火箭发动机涡轮泵历史试车数据的分析,结果表明,所建模型的训练速度得到了很大提高,对涡轮泵状态的检测效果良好。For lacking of fault samples, it is very difficult to detect the faults of a Liquid Rocket Engine （LRE） turbopump and make decision based on incomplete information. To solve this problem, a v-support vector machine novelty detection model was founded. Taking into account of the relationship between decision boundary, support vectors and constraints, a training algorithm based on Sequential Minimal Optimization （SMO） was introduced and improved for One-Class Support Vector Machines （OCSVM）. With the analysis of LRE historical test data, it showed that SMO algorithm improves the training efficiency evidently and enables the model to deal with large training data. And this model trained by SMO can detect the faults of the LRE turbopump well.

关 键 词：液体推进剂火箭发动机 涡轮泵 新异类检测模型 单类支持向量机 序贯最小优化 故障检测 

分 类 号：V434.21[航空宇航科学与技术—航空宇航推进理论与工程]