基于稀疏核增量超限学习机的机载设备在线状态预测

Online condition prediction of avionic devices based on sparse kernel incremental extreme learning machine

为实现对机载设备工作状态的在线状态预测,提出了一种稀疏核增量超限学习机(ELM)算法。针对核在线学习中核矩阵膨胀问题,基于瞬时信息测量提出了一个融合构造与修剪策略的两步稀疏化方法。通过在构造阶段最小化字典冗余,在修剪阶段最大化字典元素的瞬时条件自信息量,选择一个具有固定记忆规模的稀疏字典。针对基于核的增量超限学习机核权重更新问题,提出改进的减样学习算法,其可以实现字典中任一个核函数删除后剩余核函数Gram矩阵的逆矩阵的前向递推更新。通过对某型飞机发动机的状态预测,在预测数据长度等于20的条件下,本文提出的算法将预测的整体平均误差率下降到2.18%,相比于3种流形的核超限学习机在线算法,预测精度分别提升了0.72%、0.14%和0.13%。

In order to achieve the online condition prediction for avionic devices, a sparse kernel incre- mental extreme learning machine （ELM） algorithm is presented. For the problem of Gram matrix expansion in kernel online learning algorithms, a novel sparsification rule is presented by measuring the instantaneous learn- able information contained on a data sample for dictionary selection. The proposed sparsification method com- bines the constructive strategy and the pruning strategy in two stages. By minimizing the redundancy of diction- ary in the constructive phase and maximizing the instantaneous conditional self-information of dictionary atoms in the pruning phase, a compact dictionary with predefined size can be selected adaptively. For the kernel weight updating of kernel based incremental ELM, an improved deeremental learning algorithm is proposed by using matrix elementary transformation and block matrix inversion formula, which effectively moderate the computational complexity at each iteration. In proposed algorithm, the inverse matrix of Gram matrix of the oth- er samples can be directly updated after one sample is deleted from previous dictionary. The experimental results of the aero-engine condition prediction show that the proposed method can make the whole average error rate reduce to 2.18% when the prediction step is equal to 20. Compared with three well-known kernel ELM online learning algorithms, the prediction accuracy is improved by 0.72%, 0.14% and 0.13% respectively.