Transfer twin support matrix machine using rescaled pinball loss for roller bearing fault diagnosis

Abstract The support matrix machine (SMM) takes matrix-based data as input elements, which can protect the structural information of matrix-based data extracted from vibration signals and has good classification performance. However, in practical applications, due to the limitation of mechanical equipment and monitoring technology, there is often an insufficient number of annotation samples for SMM to find the optimal hyperplane. Therefore, a novel transfer twin SMM (TTSMM) is proposed in this article. In TTSMM, the structural information from the source data is transferred to the target model by simultaneously training on both the source and target data, which helps to establish a high-performance model using a limited number of labeled samples. Furthermore, a rescaled pinball loss function is used in TTSMM, which is bounded and utilizes quantile distance to further improve the generalization ability of the model. Finally, the proposed method is implemented on two kinds of roller bearing fault data sets including the variable speed vibration signal from Anhui University of Technology (AHUT) and the constant speed vibration signal from Case Western Reserve University (CWRU). The experimental results show that TTSMM effectively utilizes samples from both the source and target domains for modeling and has superior classification ability with a limited number of labeled samples.