Automatic Modulation Recognition (AMC) emerges as a pivotal innovation in non-cooperative communication systems, evolving rapidly alongside breakthroughs in deep learning. The amalgamation of Convolutional Neural Networks (CNN) and Long Short-Term Memory networks (LSTM) now stands at the forefront of network architecture in the field. Addressing the limitations of existing CNN+LSTM network architectures, this paper proposes a dual-stream CNN+LSTM network structure based on an attention mechanism. The network utilizes IQ signals and their fourth-order cumulant (FOC) as inputs, marking a significant advancement in signal processing. The network employs a Multi-Dimensional Compensatory CNN (MD-CNN) module to extract independent and interactive features of the signal, which are then effectively merged and optimized. Subsequently, the LSTM network captures the temporal characteristics of the signal and, combined with the Bahdanau attention mechanism, dynamically selects the most critical feature information, reducing the impact of less significant features. Comparative experiments show that the proposed network structure outperforms various existing neural network schemes in modulation recognition accuracy and adapts to different signal-to-noise ratio environments. This network structure offers an efficient solution for AMC technology in non-cooperative communication systems, with significant theoretical value and practical application prospects.