Aiming at the problem that it is difficult to be detected in time by satellite tiny faults or tiny anomalous maneuvers, a method for satellite anomalous power detection based on directly acquired observation data is proposed. Combining the Bayesian optimization method to adaptively adjust the size of the convolution kernel, using the convolutional neural network to extract the local spatio-temporal dependence features of the observation data, and after learning and training with a large amount of simulation data, an adaptive convolutional neural network model is obtained that can quickly detect the abnormal satellite dynamics. The results show that after learning a large amount of orbital observation data, the adaptive convolutional neural network model can quickly determine whether the target satellite has abnormal dynamics and produces abnormal orbit change.