In view of the problems that samples are limited in size and difficult in establishing a deep model for predicting the number of faults to evaluate the support performance of new mechanical equipment at the stage of testing and identification, a Score evaluation index is proposed by adopting the “transfer learning” method. Large scale mature equipment data was used to assist in training the new equipment fault prediction model. Starting from the perspectives of samples, features, and models in transfer learning, with a focus on deep model-based transfer, this study conducts research on predicting the number of faults. The example shows that the precision of the fine-tuning based model deep transfer has increased by 46.55% and 164.87% respectively in the root mean square error and Score, while the standard deviation has decreased by 86.71% and 91.41% respectively. Far superior to the prediction methods based on sample and feature in design applications of transfer learning and seven typical comparative models, the data driven advantages of deep learning are fully utilized. With better performance in prediction accuracy, effectiveness, and stability, it is conducive to evaluating the support performance of new equipment and promoting the construction of equipment test and evaluation.