关键词: 非负矩阵分解, 聚类, 迭代算法, 特征选择

Abstract: With the increasing complexity of electromechanical system state informations, traditional feature extraction and selection methods were unable to meet the needs. According to the characteristics of conventional non-negative matrix factorization(NMF) algorithm, a NMF method for monitoring and fault diagnosis was proposed based on the clustering property of NMF. By comparing classification accuracy and iteration efficiency, an improved alternating least square iterative algorithm with sparsity and correlation constraints was selected, and the low-dimensional embedded dimension and iterative initialization method were also determined. Experimental results to UCI test datasets and fault diagnosis of Tennessee-Eastman process(TEP) systems show that this approach is more effective to extract the fault features, and enhance the failure pattern capabilities.

Key words: non-negative matrix factorization（NMF）, clustering, iterative algorithm, feature selection