基于正则化与遗忘因子的极限学习机及其在故障预测中的应用  

Extreme learning machine based on regularization and forgetting factor and its application in fault prediction

作　　者：杜占龙[1] 李小民[1] 郑宗贵 张国荣 毛琼[1] 

机构地区：[1]军械工程学院无人机工程系,石家庄050003 [2]第二炮兵研究院,北京100085 [3]厦门警备区,厦门361003

出　　处：《仪器仪表学报》2015年第7期1546-1553,共8页Chinese Journal of Scientific Instrument

基　　金：总装院校科技创新工程(ZYX12080008)项目资助

摘　　要：为了解决在线贯序极限学习机(OS-ELM)算法容易产生奇异矩阵、算法贯序更新过程中没有考虑训练样本时效性的问题,提出基于l2-正则化和自适应遗忘因子的OS-ELM(RFOS-ELM)算法。RFOS-ELM在初始阶段加入正则化机制,克服因矩阵奇异而降低OS-ELM泛化能力的缺点。在贯序更新阶段,RFOS-ELM通过引入自适应遗忘因子实时调整新旧训练样本所占比重,推导正则化条件下带遗忘因子RFOS-ELM的递推更新算法,提高其对动态变化系统的跟踪能力。某型无人机机载发射机故障预测实例表明,相比于传统OS-ELM和正则化OS-ELM算法,本文提出方法具有更高的预测精度。On-line sequential extreme learning machine （OS-ELM） algorithm is prone to generate singularity matrix, and the OS-ELM has no consideration about the training sample timeliness during the sequential updating process. To solve the problems mentioned above, an improved OS-ELM algorithm （RFOS-ELM） is presented based on 12-regularization and adaptive forgetting factor. Regulariza- tion mechanism is utilized in the RFOS-ELM initialization phase, which avoids the shortcoming of worsening the OS-ELM generalization ability due to the matrix singularity. During sequential updating phase, the RFOS-ELM introduces an adaptive forgetting factor to adjust the proportion between new training sample and the old one in real time. The RFOS-ELM recursive updating algorithm with forgetting factor is deduced under the regularization condition, which improves its tracking ability for dynamic varying system. The fault prediction case study on a certain unmanned aerial vehicle transmitter indicates that compared with conventional OS-ELM and regularized OS-ELM, the proposed method achieves higher prognostic accuracy.

关 键 词：故障预测 时间序列 在线贯序极限学习机 l2-正则化 遗忘机制 

分 类 号：TP206.3[自动化与计算机技术—检测技术与自动化装置] TH165.3[自动化与计算机技术—控制科学与工程]