Emotion recognition is crucial for human-computer interaction, enhancing user experience and safety. Although Electroencephalogram (EEG)-based methods have made progress, they often capture redundant information and fail to extract key emotional features from frequency and spatial domains. They also overlook the use of richer multimodal data that could enhance model performance. Physiological signals such as EEG, Electrocardiogram (ECG), and Galvanic Skin Response (GSR) exhibit inherent differences in signal properties and feature distributions. These heterogeneities lead to cross-modal feature inconsistency, making it difficult to achieve precise alignment and effective fusion. To this end, we propose a joint multi-level attention and consistency aligned method (MLACA) for multimodal emotion recognition. A multi-level channel-spatial attention to mine important emotional information across multiple layers in both the spatial and frequency domains of the EEG modality, enabling the extraction of rich emotional features. To further achieve fine-grained cross-modal alignment of physiological signals, we design a multimodal cross-attention alignment method that effectively captures local dynamic associations between modalities, facilitating consistent alignment and interaction of inter-modal sequence features. Extensive experimental results on three datasets demonstrate that our method outperforms the state-of-the-art methods.