Abstract: Aiming at the crucial problem that the current prevailing fault diagnosis model was difficult to realize the automation, high-precision and generalization, a novel fault diagnosis method based on feature compression with Discriminant Orthogonal Linear Local Tangent Space Alignment (DOLLTSA) was proposed. With this method, the time-frequency domain feature set was constructed to completely characterize the property of each fault. DOLLTSA was introduced to automatically compress the high-dimensional time-frequency domain feature sets into the low-dimensional eigenvectors with better discrimination, which were input into K-Nearest Neighbors Classifier (KNNC) to carry out fault identification. The time-frequency domain feature set could comprehensively and accurately reflect the fault features of rotating machinery. DOLLTSA not only combined the local geometry with class information for manifold decoupling, but also used spectral regression and subspace orthonormalization approach for solving the optimal low dimensional embedding subspace, which could improve fault identification accuracy. Fault diagnosis example of deep groove ball bearings and life state identification example of space bearing demonstrated the effectiveness of proposed method.

Key words: time-frequency domain feature set|discriminant orthogonal linear local tangent space alignment|feature compression|manifold learning|fault identification

摘要： 针对现有旋转机械故障诊断模式难以实现自动化、高精度和泛化性的关键问题，提出基于判别式正交线性局部切空间排列特征约简的故障辨识方法。该方法首先构造全面表征不同故障特性的时、频域特征集，再利用DOLLTSA将高维时、频域特征集自动约简为区分度更好的低维特征矢量，并输入到K近邻分类器中进行故障模式辨识。时、频域特征融集可较全面准确地反映旋转机械的故障特征；DOLLTSA综合利用局部几何结构和类判别信息进行流形解耦，并采用谱回归法和子空间正交化处理来优化低维嵌入子空间，提高了故障辨识精度。深沟球轴承故障诊断实例和空间轴承寿命状态辨识实例验证了所提方法的有效性。

关键词: 时、频域特征集|判别式正交线性局部切空间排列|特征约简|流形学习|故障辨识