Abstract: The basic assumption of the traditional fault diagnosis model is that the training and test data have the same probability distribution, however, in the diagnosis of industrial reciprocating compressors, problems such as sensor failure and change of installation location often occur, resulting in the inability to obtain sufficient effective sample data to satisfy the requirements of the existing fault diagnosis model for input samples. This study proposes a feature migration-based shale gas reciprocating compressor fault diagnosis method under small samples, and fault vibration data are collected through a self-developed experimental platform to conduct fault diagnosis experiments. The experimental results show that the common migration component obtained by transfer component analysis (TCA) can reduce the distribution difference between the source and target domain data in the high-dimensional regenerated kernel Hilbert space (RKHS), thus obtaining better fault diagnosis accuracy. Compared with the traditional small-sample machine learning algorithm and the deep belief network (DBN) model, the proposed method shows stronger expressive and generalization abilities when dealing with data collected from sensors with different orientations, and the fault identification accuracy reaches 92%. The technique balances the advantages of migration learning and deep learning. It provides a theoretical reference for the migration fault diagnosis of complex dynamic equipment in the case of small samples.