Abstract: To address the issues of insufficient feature representation capability and degraded recognition performance faced by automatic modulation recognition in complex electromagnetic environments under low signal-to-noise ratio (SNR) conditions, this paper proposes an automatic modulation recognition model named RLANet, which integrates ResNet-LSTM with a temporal attention mechanism. The model takes raw I/Q time-domain sampled signals from the RML2016.10a dataset as input and achieves automatic recognition of 11 analog and digital modulation types without manual feature extraction. The model enhances the representation capability of key discriminative information in signals by combining local feature extraction, temporal dependency modeling, and temporal attention weighting, thereby improving recognition robustness under complex channel and low-SNR conditions. Experimental results show that RLANet achieves an average recognition accuracy of 93.39% under moderate to high SNR conditions and 65.54% across the full SNR range, outperforming several typical benchmark models overall, providing an effective solution for reliable signal recognition under low SNR conditions.