Abstract: Auroral substorm is a geomagnetic disturbance resulting from the interaction between Earth’s magnetic field and the solar wind. The accurate identification of the onset times is crucial for a deep understanding of the underlying physical mechanisms. The existing machine methods for auroral substorm identification differ from manual identification standards and typically require complex image preprocessing and parameter tuning by manual. To achieve a machine model consistent with manual identification standards, this paper designs two identification strategies aimed at addressing the issue of variable image sequence lengths encountered in replicating manual standards. Based on deep learning methods, this paper proposes an EfficientNet model featuring CBAM attention as a key component for model construction. The model is trained using ultraviolet auroral images from the Polar satellite between 1996 and 1998 and tested on image data from 1999 to 2000. The model achieves an identification accuracy of up to 0.98 and an efficiency of 36.93 frames per second. This model not only eliminates the reliance on image preprocessing present in existing models but also adapts to real-world observations with unequal image sequence lengths and extreme imbalances in the number of samples between substorm and non-substorm sequences, demonstrating its high practicality.