The　automatic　detection　of　epileptic　EEG　signals　based　on　deep　learning　has　garnered　significant　attention　due　to　its　ability　to　mitigate　the　influence　of　human　factors.　The　effectiveness　of　epilepsy　detection　primarily　hinges　on　the　architecture　of　deep　neural　network　models.　Thus,　this　study　proposes　a　Multi-Scale　Multi-Level　Temporal-Spatial　Attention　CNN　model　(MSMLTSACNN),　which　includes　the　Multi-Scale　Multi-Level　Attention　Module　(MSMLAM),　the　Spatial-Temporal　Feature　Extraction　Module　(STFEM),　the　Scale　Feature　Reconstruction　Module　(SFRM)　and　the　Classification　Module　(CM).　Firstly,　MSMLAM　is　used　to　extract　the　multi-scale　features　of　epileptic　EEG　signals.　Then,　for　each　scale　feature,　the　information　flow　between　electrodes　is　acquired　at　three　hierarchical　levels,　and　the　three-level　features　of　the　current　scale　are　integrated　by　using　the　attention-weighted　mechanism.　Furthermore,　STFEM,　SFRM　and　CM　are　employed　successively　to　extract,　rerepresent　and　classify　the　multi-scale　temporal-spatial　features.　Through　five-fold　cross-validation　experiments　on　the　CHB-MIT　dataset,　the　MSMLTSACNN　model　achieved　favorable　results　in　terms　of　sensitivity　(98.97%),　F1　score　(98.68%),　accuracy　(98.78%),　and　precision　(98.57%).　The　results　indicate　that　MSMLTSACNN　can　effectively　improves　epileptic　seizure　detection　performance　by　extracting　more　representative　multi-scale　features　based　on　the　multi-level　information　flow　between　electrodes　of　EEG　signals.　©　2024　Technical　Committee　on　Control　Theory,　Chinese　Association　of　Automation.