局域极端学习机及其在状态在线监测中的应用

Local Extreme Learning Machine and Its Application to Condition On-line Monitoring

针对训练样本贯序输入时的极端学习机(ELM)训练问题,提出一种可实现在线训练的局域极端学习机(LELM).LELM以逐次增样训练与减样训练的方式实现在线训练,从而有效保持了简约的模型结构,同时利用分块矩阵求逆引理有效减小了多次模型训练的计算代价.混沌时间序列在线预测仿真表明,LELM的在线训练时间远小于ELM,且预测精度更高.基于时间序列预测的雷达发射机状态在线监测实例表明,相比于利用粒子群优化的自适应灰色模型方法,LELM具有更高的计算效率与预测精度,适用于电子系统状态在线监测.

To reduce the computational cost of extreme learning machine（ELM） on-line training,a new algorithm called local extreme learning machine（LELM） was proposed. LELM adopts the latest training sample and abandons the oldest training sample iteratively to insure that only the most relevant samples are applied to LELM on-line training.The output weights of LELM are determined recursively during each training procedure to reduce the computational cost of on-line training.The numerical experiments on chaotic time series prediction indicate that the on-line training computational cost of LELM is much less than that of ELM.The numerical experiments on radar transmitter condition on-line monitoring based on time series prediction indicate that LELM has better performance in on-line training computational cost and prediction accuracy in comparison with conventional electronic system condition on-line monitoring method（using） adaptive grey model.