Sawtooth instability is a common and influential MHD phenomenon in tokamak plasmas, closely associated with core transport and the onset of larger instabilities. Therefore, accurate and real-time identification of sawtooth cycles is essential for plasma control and operational safety. In this study, we propose a real-time identification algorithm based on a hybrid long short-term memory and convolutional neural network architecture. Unlike traditional methods, our approach extracts both temporal and spatial features from diagnostic signals such as soft x-ray and ECE. Trained on over 10 000 annotated events, the model classifies sawtooth phases into three categories. Experiments on HL-3 data show 92.5% real-time accuracy and 95.6% post-processed accuracy. The model outperforms four mainstream algorithms in accuracy, noise robustness (SNR ⩾ 5 dB) and inference speed (<2 ms on GPU) enabling real-time deployment. This work provides a scalable and effective solution for intelligent sawtooth control in fusion devices.