In　this　paper,　a　cross-subject　emotion　recognition　network　based　on　semi-supervised　and　domain　adversarial　learning　with　electroencephalogram　(EEG)　signals　(SEDA-EEG)　is　proposed,　which　addresses　the　challenge　of　high　EEG　variability　from　different　subjects　in　brain-computer　interface　(BCI)　research.　By　employing　the　differential　entropy　features　of　EEG　signal　processed　by　the　linear　dynamic　systems　as　input,　influences　of　noises　can　be　greatly　reduced　and　the　essential　changes　in　EEG　can　be　well　reflected　while　capturing　the　implicit　signal　features　over　time.　Simultaneously,　a　deep　neural　network　is　designed　to　match　similarity　relationships　between　samples　and　labels　by　capturing　key　patterns　in　the　data,　which　improves　the　performance　of　recognizing　EEG　signals.　Moreover,　a　domain　transfer　module　based　on　domain　adversarial　loss　and　offline　feature　decomposition　semi-supervised　learning　is　proposed　to　enhance　inter-domain　knowledge　generalization,　which　achieves　cross-domain　feature　alignment　and　enables　the　model　more　adapt　to　the　target　data.　Experimental　results　show　that　the　proposed　SEDA-EEG　yields　the　state-of-the-art　performance,　which　outperforms　other　advanced　models　by　12.90%　on　the　SEED-IV　dataset　with　stronger　robustness,　indicating　the　potential　of　applying　into　the　EEG-oriented　cross-subject　emotion　recognition.