Sparse filtering based domain adaptation for mechanical fault diagnosis

Abstract

Recently, machine learning has achieved considerable success in the field of mechanical fault diagnosis. Nevertheless, in many real-world applications, the original vibration data usually collected under different work conditions which lead to large distribution divergences. As a result, the performances of many machine learning methods may drop dramatically. To overcome this deficiency, domain adaptation is introduced by adapting the regression model or classifier trained in the source domain for use in the distinct but related target domain. Particularly, a novel sparse filtering based domain adaptation approach (SFDA) is proposed for the mechanical fault diagnosis. Comparing with the previous researches, two main contributions of SFDA are concluded as follows: (1) the domain adaptation is applied to the sparse filtering algorithm. (2) The ℓ1-norm and ℓ2-norm are employed to the maximum mean discrepancy (MMD). (3) SFDA is easy to be implemented, and high classification accuracy can be obtained. The bearing and gear dataset are utilized to testify the validity and reliability of SFDA.