l_(p)-范数约束下MKL-OC-ELM的装备故障检测

MKL and OC-ELM fault detection based on l_(p)-norm constraint

针对列装时间短的现役装备故障样本匮乏、现有算法故障检测准确率较低的问题,将多核学习(multiple kernel learning, MKL)与一类超限学习机(OC-ELM)相结合,提出l_(p)-范数约束下多核学习一类超限学习机(l_(p)-MKOCELM)的检测模型.在l_(p)-范数约束下,定义了将MKL与OC-ELM相结合的数学优化形式,推导出基核组合权重与Lagrange乘子的更新方式;为方便故障检测的实施,基于l_(p)-MKOCELM定义了统计检验量与检测阈值;通过实验验证了不同范数的约束形式的近似等价性.将所提出方法应用于常用的UCI数据集和某型装备的测试数据,实验结果表明,相比于传统的SVDD、PCA、OC-SVM、OC-KELM等方法,所提出方法在平衡漏警、虚警的同时,能够显著提升检测精度.

Aiming at the problems of the shortage of fault samples for active new equipment and the low accuracy of existing algorithms for fault detection, the multiple kernel learning(MKL) and the one-class extreme learning machine(OC-ELM) are combined, and the l_(p)-norm constrainted multiple kernel learning one-class ELM(l_(p)-MKOCELM) is proposed. Under the l_(p)-norm constraint, a mathematical optimization form combining the MKL and the OC-ELM is defined, and the update method of combination weights of the base kernel and Lagrange multipliers are derived. To facilitate the implementation of fault detection, the test statistic and detection threshold based on the l_(p)-MKOCELM are defined. The approximate equivalence of different norm constraints is confirmed through experiments. The proposed method is applied to the commonly used UCI data set and test data of an equipment. The experimental results show that,compared with the traditional SVDD, PCA, OC-SVM, and OC-KELM, the proposed method can significantly improve the detection accuracy while balancing missing alarm and false alarm.