基于核主元分析的滚动轴承故障混合域特征提取方法

Mixed-domain feature extraction approach to rolling bearings faults based on kernel principle component analysis

为有效利用时域、频域、时-频域中各类具有显著类别差异信息的非平稳统计特征,提高滚动轴承状态监测和故障诊断的性能和效率,提出一种基于核主元分析的混合域特征提取方法。通过对原始信号分别生成时域、频域状态特征,并利用多分辨率小波分解生成时-频域状态特征,构建出144个表征原始振动信号特征的混合域特征集。采用核主元分析方法对其中能敏感地反映故障特性的特征进行二次非线性特征提取,按累计贡献率大于90%的标准,选取前11个核主元作为主要特征量,将其输入支持向量机分类器进行状态识别。仿真结果表明:混合域特征集比单个特征、单域特征能更全面准确地反映故障特性,核主元分析方法能有效降低输入特征维数,并确保输出特征具有较高的反映轴承运行状态的敏感性和适于模式识别的可分性;与通常使用的基于小波分解的特征提取方法相比,本文方法能更加准确有效地提取不同运行条件下滚动轴承不同类型不同程度的故障特征。

In order to effectively use the various nonstationary statistical features with significant differences from time domain,frequency domain and time-frequency domain,a novel mixed-domain feature extraction approach was proposed,which was based on kernel principle component analysis to improve the performance and efficiency for condition monitoring and fault diagnosis of rolling bearings.At first,the time-domain and frequency-domain features which were generated by the original signal,and time-frequency-domain features which were generated by the multi-resolution wavelet decomposition were extracted.The mixed-domain features set including 144 features were composed to characterize the original vibration signals.Then the kernel principle component analysis method was used to secondary extract the features which reflected sensitively the failure characteristics in the mixed-domain features set.According to the accumulated contribution rate of more than 90%,the first 11 nonlinear principal components were extracted as primary feature vector for support vector machine classifier to recognize.The results show that the mixed-domain features set can reflect the failure characteristics more comprehensively and accurately than a single feature or single-domain features.Kernel principle component analysis method can effectively reduce the input feature dimensions,and ensure the output features to be of high sensitivity to reflect the operational status of bearings and high separability for pattern recognition.Compared to the common feature extraction method based on wavelet decomposition,this proposed method becomes more apparent to extract fault feature of rolling bearing in different types and degrees under different operating conditions.