Attention deficit hyperactivity disorder (ADHD) is a common neurodevelopmental behavioral disorder. Its traditional diagnostic methods are highly subjective and prone to misdiagnosis. At present, ADHD diagnosis is mainly based on subjective evaluation, and the accuracy of objective diagnosis is low. To address this problem, this paper proposes an ADHD automatic classification network BASA-FMGNet based on EEG signals. The network uses FMGNet as the backbone network, and its core module FMGConv can effectively extract the time domain and frequency domain features of EEG signals. In addition, the network introduces a bilinear adaptive sparse attention (BASA) structure, extracts inter-channel features through sparse calculation, and uses bilinear adaptive calculation to fuse local and non-local attention, thereby enhancing feature expression. Experimental results on the ADHD children’s dataset show that the F1 score of the proposed BASA-FMGNet network reaches 97.90%, which is significantly better than that of the current mainstream network. Ablation experiments further confirm the effectiveness of the BASA module in improving the performance of different networks, and verify that it can effectively enhance EEG features, providing strong support for the objective diagnosis of ADHD.