基于流形子带特征映射的转子复合故障特征提取方法

Rotor compound fault feature extraction based on a manifold sub-band feature mapping method

针对复合故障特征易被噪声信号淹没,传统时频分析和流形学习方法不能完整有效的挖掘故障潜在信息和进一步实现故障特征提取。在流形学习的基础上提出了一种流形子带思想并将其应用到转子复合故障特征提取研究中,进而得出了一种基于流形子带特征映射的转子复合故障特征提取方法。对故障原始信号序列进行相空间重构,结合小波包对噪声的强烈抑制性和对信号分辨率高的特点,将重构信号分解成不同频带即子带。将同故障多种工况下的同一频带融合成频带矩阵并估计其本征维数,并通过拉普拉斯特征映射算法以本征维数为依据将子带降维获取低维特征向量并提取信息熵,进一步实现故障特征提取。实验表明,相对于经典的局部线性嵌入和拉普拉斯特征映射等算法,流形子带特征映射算法不仅对单故障而且对复合故障特征进行了更完整有效的挖掘和提取。

; Compound fault features can be easily submerged by noise signals. Traditional time-frequency analysis and manifold learning methods cannot effectively mining the potential failure information for further fault feature extraction. The concept of manifold learning was proposed based on the manifold sub-band method and was applied to study compound rotor faults. Then a manifold subband feature mapping method was obtained. First of all, phase space reconstruction was performed for fault original signal sequence. Combining with wavelet packet strong inhibitory to noise and the characteristics of high resolution, reconstructing signal was decomposed into different frequency bands. The same frequency band of the same fault and many conditions were intergrated into the band matrix and estimated the intrinsic dimension. At last, low dimensional feature vector and subband dimension reduction were obtained by Laplace feature mapping algorithm based on the intrinsic dimension and the information entropy was extracted. Then the fault feature extraction was further realized. Experiments show that compared with the classical local linear embedding and Laplace feature map algorithm, etc. , the proposed method is not only effective for single fault but also for compound fault.