Abstract: Aiming at the problems that vibration signals of rolling bearings were greatly disturbed by random noises and representation errors of multi-scale entropy(MSE) for bearing degradation trend were relatively large. A degradation trend prediction method was proposed to solve the problems based on two-element multi-scale entropy(TMSE). Firstly, local characteristics-scale decomposition(LCD) was conducted to decompose the bearing vibration signals into several orders, and multi-element multi-scale theory was applied to compute the first two orders. In this way, TMSE was extracted to characterize bearing degradation trend. And then, mutual information method and false nearest neighbor method were used to optimize embedding dimensions and delay vectors. Lastly, ELM was applied as predictive model to predict the degradation trend which was represented by TMSE, and the prediction performance of different activation functions was compared. The results show that representation errors of TMSE are smaller than that of MSE, and the prediction performance of ELM is better when its activation function is sigmoid.

Key words: rolling bearing, multi-scale entropy, parameter optimization, degradation trend prediction, extreme learning machine(ELM)

摘要： 针对轴承振动信号随机噪声干扰大、多尺度熵表征轴承退化趋势偏差大的问题，提出了一种基于二元多尺度熵的滚动轴承退化趋势预测方法。首先对滚动轴承振动信号进行局部特征尺度分解，采用多元多尺度熵理论对二阶信号进行计算，提取了二元多尺度熵特征。然后采用互信息法和假近邻法对算法中的嵌入维数和延迟向量等参数进行了优化。最后采用极限学习机预测模型对二元多尺度熵退化趋势曲线进行预测，并对比了不同激活函数的预测性能。结果表明，相对于传统多尺度熵，二元多尺度熵偏差较小；激活函数为sigmoid时极限学习机模型预测精确度较高。

关键词: 滚动轴承, 多尺度熵, 参数优化, 退化趋势预测, 极限学习机