Selecting accurate Radar Cross Section (RCS) sequence historical data has an important impact on RCS-based spatial target attitude anomaly detection. RCS is related to factors such as satellite shape size, attitude and radar observation angle, and is a sensitive function of observation angle variation. In this paper, the influence of spatial observation geometric relationship on spatial target RCS sequence is analysed. Aiming at the orbital regression characteristics unique to spatial targets, a spatial target RCS historical data screening method based on dynamic time programming (DTW) distance is proposed. According to the relationship between the azimuth and the elevation angles of the radar line of sight, the orbit data with similar spatial geometry is found from the historical data. The orbital data of the system. This method can reduce the impact of Earth’s rotation on data observation and establish a more efficient database. The variational mode decomposition (VMD) of RCS is carried out, and the fractal box dimension of each component is extracted as the data feature, and the BP neural network is used as the classifier for the anomaly detection. The simulation results prove that compared with the traditional historical data collection method, this method can improve the attitude detection rate of space targets and has better robustness.