Twin robust matrix machine for intelligent fault identification of outlier samples in roller bearing

In the industrial processes, the intelligent fault diagnosis related to signal analysis and pattern recognition is an important step to ensure the health of mechanical equipment. A popular intelligent monitor method as it has been, support matrix machine (SMM) enables to use the two-dimensional features extracted from vibration signals to build model. The core of the SMM is to extract structure information within matrix by minimizing nuclear norm to approximate the rank of the matrix. However, the nuclear norm has limited performance to eliminate the noise contained in structure information. Furthermore, features extracted from vibration signals often become outliers, and SMM is sensitive to outliers in training data. Therefore, a novel nonparallel classifier called twin robust matrix machine (TRMM) is proposed and applied to roller bearing fault diagnosis. TRMM can not only fully leverage the low-rank structure information, but also has the following novelties. First, TRMM uses the truncated nuclear norm as the low-rank constraint, to pay more attention to the large singular values related to the main structure information. Further, the ramp loss is used in TRMM as the loss function, which reduces the loss penalty for outlier sample and make TRMM insensitive to outlier samples. Finally, the accelerated proximal gradient (APG) is devised to solve the resulting optimization problem. Experimental results show that the proposed method has excellent fault diagnosis performance, especially in the case of existing outlier samples. •A twin robust matrix machine (TRMM) is proposed for classification with outliers.•The truncated nuclear norm makes TRMM can capture strong correlated information.•The ramp loss is introduced to make TRMM insensitive to outlier samples.