改进的多变量极限学习机在滚动轴承故障预测中的应用

Application of Improved Multivariable Extreme Learning Machine in Fault Prediction of Rolling Bearing

传统滚动轴承故障预测仅对单个故障特征频率做时间序列预测,而滚动轴承故障由多个故障频率共同表征。为了全面的表征整个频谱的结构,并且不破坏各个频率间的内部联系,提出奇异值分解和极限学习机相结合的多变量时间序列预测方法。首先通过全矢谱方法得到振动信号频谱,然后以整个频谱的各个频率作为输入变量,构建多变量时间序列。最后通过多变量极限学习机和奇异值分解相结合的方法构建训练和测试样本,对频谱进行预测。采用该方法对全寿命滚动轴承数据进行验证,实验结果表明了该方法的有效性。

Traditional rolling bearing fault prediction only performs time series prediction for single fault characteristic frequency,while rolling bearing faults are characterized by multiple fault frequencies.In order to comprehensively characterize the structure of the whole spectrum without destroying the internal relations between the frequencies,a multivariate time series prediction method combining singular value decomposition and extreme learning machine is proposed.Firstly,the spectrum of the vibration signal is obtained by the full vector method,and then the multivariate time series is constructed by using the respective frequencies of the entire spectrum as input variables.Finally,the training and test samples are constructed by combining the multivariate extreme learning machine and the singular value decomposition to predict the spectrum.The method is used to verify the data of the full-life rolling bearing,The experimental results show the effectiveness of the method.