Pinball transfer support matrix machine for roller bearing fault diagnosis under limited annotation data

Abstract

As a classical matrix classification technology, support matrix machine (SMM) takes the matrix as the input element, so that the structure information between matrix samples is maximized to establish an accurate classification model. However, in practical industrial practice, it is difficult to collect enough annotation samples for SMM to find the optimal hyperplane. To solve this issue, a new pinball transfer support matrix machine (PTSMM) method is proposed in this paper. In the case of limited annotation sample, PTSMM trains a model using a large amount of labeled source domain data, and then fine-tunes the prior model using a small amount of target domain data, so as to obtain an optimal model for target domain. Besides, the pinball attribute is applied to the objective function to make the direct distance of the hyperplane larger, which weakens the influence of noise on the hyperplane. The two different roller bearing experiment results show that PTSMM effectively uses the samples of source domain and target domain for modeling, and the diagnostic accuracy is more than 4% higher than SMM and its improved algorithms.