【摘要】 针对电力机车牵引变流器中故障率最高的单相脉宽调制(pulse width modulation, PWM)整流器，提出一种流形学习算法融合多域特征的故障诊断方法。根据整流器在不同工作状态下的时域、频域和时频域特征构建多域特征向量；采用Hessian局部线性嵌入(Hessian local linear embedding, HLLE)算法融合多域特征，根据故障样本数和聚类结果，解决高维数据中固有维数和最近邻数选取困难的问题，得到用于描述故障特征的最优低维特征向量，减少特征之间的冲突和冗余；采用支持向量机进行模式识别，实现对整流器的故障诊断。结果表明：对不同的输出电压，不同的训练和测试比，15种故障模式均具有较高的诊断率。与其他方法相比，本文方法具有较好的融合效果和较强的鲁棒性。更多还原

【Abstract】 In response to the single-phase pulse width modulation(PWM) rectifier with the highest failure rate in electric locomotive traction converters, a fault diagnosis method where a manifold learning algorithm was used to fuse multi-domain features. Firstly, multi-domain feature vectors were constructed based on the time domain, frequency domain and time-frequency domain features of the rectifier under different working conditions. Secondly, Hessian local linear embedding(HLLE) algorithm was adopted to fuse multi-domain features. Based on the number of fault samples and clustering results, the difficulty of selecting the intrinsic dimensionality and nearest neighbor number in high-dimensional data was solved, and the optimal low-dimensional feature vector was obtained for describing fault features, reducing conflicts and redundancy between features. Finally, support vector machine was used for pattern recognition to realize fault diagnosis of rectifier. The results show that for different output voltages and different training and testing ratios, all 15 fault modes have higher diagnostic rates. Compared with other methods, this method has better fusion effect and stronger robustness.更多还原