Portal de Periódicos da CAPES

Non-parallel least squares support matrix machine for rolling bearing fault diagnosis

2019; Elsevier BV; Volume: 145; Linguagem: Inglês

10.1016/j.mechmachtheory.2019.103676

1873-3999

Xin Li, Yu Yang, Haiyang Pan, Jian Cheng, Junsheng Cheng,

Machine Learning in Bioinformatics

For rolling bearing fault classification, the input samples can be naturally expressed as two-dimensional matrices in some cases, such as time-frequency grayscale diagram and multichannel vibration signals. To make full use of the structure information of matrix data, a novel matrix data classifier called non-parallel least squares support matrix machine (NPLSSMM) is proposed and applied to rolling bearing fault diagnosis with wavelet time-frequency grayscale diagram. To construct NPLSSMM, we design a pair of novel matrix-based objective functions to obtain two non-parallel hyperplanes. Every hyperplane is required to be as close as possible to the samples of one class while being as far as possible from other samples. In each objective function, the matrix-form squares loss terms are used to simplify NPLSSMM, and decrease the computation complexity. The nuclear norm term is added to control the structure information extracted from the matrix data input. Moreover, an effective solution is provided for NPLSSMM with the alternating direction method of multiplier (ADMM) method. The results of two rolling bearing datasets show that NPLSSMM has great classification performance for rolling bearing fault diagnosis, but also has great advantage in running time over other matrix data classifiers.