Liquid aluminum leakage was the direct cause that resulted in explosion accidents in aluminum deep well-casting. In order to solve the problems that the judgment methods of liquid aluminum leakage had large lag, low accuracy and limited monitoring range in practical engineering, a sound recognition method for liquid aluminum leakage based on improved EfficientNetV2 was proposed. This method improved the monitoring range by using the sound characteristics of liquid aluminum leakage as the judgment basis, and improved the EfficientNetV2 structure by simplifying the stacking factor and introducing the ECA attention mechanism to improve the recognition rate and accuracy. First, the sound data was collected by the pickup, and seven kinds of sound scenes were constructed. Then, the logarithmic Mel-spectrogram was extracted from the sound signal as the feature set, which was input into the improved EfficientNetV2 for training and verification to obtain the sound recognition model of aluminum liquid leakage. The experimental results showed that the recognition accuracy of the improved EfficientNetV2 reached 95.48%. Compared with the original EfficientNetV2、ResNet、RegNet and DenseNet, the PLOPs were reduced to 12.34%, 8.64%, 11.14%, and 10.80% of the above models, the Params were reduced to 11.37%, 9.55%, 15.95%, and 17.24% of the above models, and FPS (CPU) was increased to 6.53 times, 6.14 times, 4.41 times, and 8.00 times of the above models. It showed that the improved EfficientNetV2 had accurate and fast recognition performance. In addition, based on the sound recognition method, an early warning mechanism for liquid aluminum leakage was proposed and applied to the real-time risk monitoring of the casting unit, which verified the effectiveness of the proposed method and provided a reference for the prevention of explosion accidents in aluminum deep well-casting.