Traditional machine learning or deep learning relies on a sufficient amount of labeled data for the training of fault diagnosis models. However, for new wind farms, insufficient data and limited labels hinder the establishment of such models. In order to cope with these two challenges, we proposed a new domain adaptive method for wind turbine fault diagnosis: balanced joint adaptive network (BJAN), which can transfer wind turbine data from other wind farms to the target new wind farm. In this method, we proposed a new pseudo-label prediction method that combines the sub-domain majority voting and overall iterations (SMV-I) to label the unlabeled data. In addition, BJAN uses long short-term memory model (LSTM) to replace common convolutional neural network (CNN) as the feature extraction module to improve diagnosis efficiency. Moreover, we also proposed a new distributed adaptive distance for BJAN: balanced joint maximum mean discrepancy (BJMMD), which can balance the data of different states during the distributed adaptive process to improve diagnostic accuracy. Numerical experiments with real wind turbine data in three wind farms not only show that the proposed SMV-I has excellent pseudo-label prediction performance, but also verify that the proposed fault diagnosis model BJAN has higher diagnostic accuracy and efficiency.