Datadriven fault diagnosis (DFD) refers to applications of machine learning and deep learning theories as machine fault diagnosis. This paper systematically summarizes the state of art theories and methods of DFD as well as their applications in spacecraft following the progress of machine learning, then offers a future perspective. DFD could be divided into three categories, traditional machine learningbased methods, deep learningbased methods, and transfer learningbased methods, according to the development of technology. Traditional machine learning based methods adopt advanced signal processing method and feature extraction, which requires a lot of contribution from human labor and extensive expert experience. Although it has excellent performance on small sample data, it is not suitable for processing big data. Over the recent years, the advent of deep learning, which encourages to construct an endtoend diagnosis model, further releases the human labor. Four deep learning models: Stack autoencoder, deep belief network, convolutional neural network, and recurrent neural network are introduced, their applications in diagnosing spacecraft faults are also summarized. Aiming to release the challenge that deep learning relies heavily on labeled data, transfer learning which attempts to use the diagnosis knowledge from one or multiple diagnosis tasks to other related ones, is introduced, and scenarios adapted to spacecraft applications are proposed, picturing a roadmap for the engineering application of DFD.