Machine learning techniques have been widely investigated in gearbox diagnosis with the assumption that training data and test data follow the same distributions. However, various operating conditions inevitably cause dynamic changes of gearbox fault characteristics which pose significant challenges on gearbox diagnosis. To address this issue, this paper introduces a transfer learning method to mitigate the domain difference caused by various operating conditions in gearbox diagnosis. More specifically, a factor analysis based transfer learning method, named transfer factor analysis, is formulated and presented. It seeks the pivot features across different domains corresponding to various operating conditions, achieved by transferring the features into a low-dimensional latent space via feature selection to minimize domain difference and preserving data properties. The selected features by transfer factor analysis are then fed into a machine learning model (e.g. support vector machine) for gearbox diagnosis. Experimental studies on a gearbox have been performed to validate the effectiveness of transfer factor analysis method.