Non-acute methamphetamine use disorder (MUD) has a high risk of relapse but is often difficult to detect because the patient's cravings enter a latent state. Traditional methods such as clinical scales or hair/urine tests have detection limitations of being too subjective or having short retrospective periods. Deep learning based on electroencephalogram (EEG) allows for capturing abnormal brain discharges objectively, so this study proposes a filter-bank self-attention convolutional neural network (FBSA-CNN) for detecting patients with non-acute MUD. First, a natural frequency filtering module is designed in FBSA-CNN to realize the splitting of different frequency domains. Then, a spatial-temporal domain convolution module is designed to extract the spatio-temporal information of each frequency band. Finally, a self-attention (SA) mechanism is used to learn the global frequency domain dependence in EEG features, supplement the limited sensory field in the convolution module, and adaptively assign the weights of different frequency bands. The proposed algorithm makes up for the traditional framework that ignores the important information in different frequency domains. FBSA-CNN is tested on a 68subject resting-state EEG dataset with an accuracy of 85.92% (10-fold cross validation), which outperforms the other six classical algorithms. Furthermore, both feature visualization and ablation experiments demonstrate the good performance of FBSA-CNN.This method lays the foundation for the effective detection of non-acute MUD thus reducing the risk of recurrence.