Abstract: Functional Magnetic Resonance Imaging (fMRI) has shown considerable potential in brain disease identification due to its high-dimensional data characteristics. However, existing deep learning methods lack sensitivity to multi-scale temporal features. To address this, we propose a multi-level Blood Oxygenation Level Dependent signal feature interaction fusion model. The model divides BOLD signals into regions of interest based on a brain atlas to construct time series. After encoding, sliding window and multi-head self-attention mechanisms are used to capture global features and a distribution-specific attention mechanism is employed to extract local features. The multi-scale information extracted from both global and local features is then fused and input into a linear layer for disease recognition. Additionally, an explainability analysis based on clustering reveals the most relevant brain regions in the process of brain disease identification. Experiment results show that the proposed method improves accuracy by approximately 3.9% and 2.5% compared to the optimal baseline methodon the on the ADNI and ABIDE I datasets, respectively. Furthermore, the abnormal brain regions identified by the proposed model are highly consistent with the clinical manifestations of Alzheimer's disease and Autism Spectrum Disorder.