This paper proposes a multi-scale Convolutional Neural Network(CNN) online pre-earthquake electromagnetic anomaly detection model which is applied in noisy environment. Based on the powerful feature extraction ability of CNN, cooperating with the characteristics of long-term and short-term ground-space electromagnetic spectrum, the pre-earthquake electromagnetic anomaly detection is performed in multi-dimensional and multi-perspective. At the same time, the adaptive Variational Mode Decomposition(VMD) noise reduction method is introduced to extract the effective information in the observation signal. Combined with online learning strategy, the continuous learning of possible changes of pre-earthquake electromagnetic anomaly mode is realized. The simulation results show that the multi-scale model can maintain high accuracy under low Signal-to-Noise Ratio(SNR), and the online learning strategy can effectively reduce the model update time, which proves the effectiveness of the model.