In the context of cognitive load recognition, precise identification across different domains is essential for improving the robustness and adaptability of models. At present, methods for evaluation based on EEG have emerged as a leading area of research. However, because of the intrinsic variability and non-stationary nature of EEG signals, there is a need to improve the generalizability of EEG measurements to effectively achieve cognitive load recognition through EEG signals. This study presents a deep learning model that integrates an advanced attention mechanism with Long Short-Term Memory (LSTM) networks and Domain-Adversarial Networks. The proposed model incorporates a wavelet entropy-based attention mechanism within the framework of a domain adversarial network to facilitate the extraction of time-frequency features from the signals. The attention mechanism dynamically focuses on these features, thereby improving the model’s generalization capability in recognition tasks. After constructing the model, we compared it with existing methods and achieved promising results, demonstrating the superiority of the model in EEG recognition.