The J wave in the electrocardiogram indicates that the patient may suffer from malignant arrhythmia, ventricular tachycardia and other diseases, and the probability of sudden death is greatly increased. Due to the low incidence of J wave syndrome, J wave data in public databases are scarce, which hinders the research progress of J wave detection. Therefore, this study aims to automatically identify J wave signals, construct relevant datasets, and combine transfer learning with the pre-trained VGG16 model to detect J wave signals. Given the limited availability of J wave data, we utilized a local transformation method to simulate the generation of J wave signals. Subsequently, the simulated J wave signals underwent a continuous wavelet transform, converting them into grayscale images suitable for input into the pre-trained network. Finally, we fine-tuned the VGG16 network structure, which was initially trained on ImageNet, and modified the network for binary classification output for J wave data training. According to experimental results, this method achieves a 99.34 % accuracy rate, 99.35 % sensitivity, and 99.32 % specificity for the identification of J wave data. The aforementioned results outperform existing J wave research methodologies, effectively facilitating the identification of J wave signals.