Dynamic　effective　connectivity　(DEC)　contains　abundant　temporal　and　spatial　dynamic　information,　which　can　characterize　the　formation　and　dissolution　of　distributed　directional　functional　patterns　over　time.　Recently,　learning　DEC　from　functional　magnetic　resonance　imaging　(fMRI)　time-series　data　has　become　a　new　hot　spot　in　the　field　of　neuroinformatics.　However,　current　DEC　learning　methods　are　hard　to　effectively　estimate　the　transition　of　brain　states　and　accurately　learn　the　network　structure　of　DEC.　In　this　article,　we　propose　a　novel　DEC　learning　method　based　on　nonparametric　state　estimation　and　generative　adversarial　network,　named　nPSE-GAN.　The　nPSE-GAN　first　employs　nonparametric　state　estimation　(nPSE)　to　automatically　infer　the　number　of　brain　states　and　transition　time.　In　detail,　the　nPSE　uses　dual　extended　Kalman　filtering　(dEKF)　to　obtain　state　features　and　employs　hierarchical　clustering　to　estimate　the　transition　of　brain　states.　Then,　the　proposed　method　uses　generative　adversarial　network　(GAN)　to　learn　the　network　structure　of　DEC.　Specifically,　GAN　takes　the　transition　information　and　original　fMRI　time-series　data　as　input,　which　trains　the　generator　and　discriminator　simultaneously.　The　experimental　results　on　simulated　datasets　show　that　nPSE-GAN　can　effectively　estimate　the　transition　of　brain　states　and　is　superior　to　other　state-of-art　methods　in　learning　the　network　structure　of　DEC.　The　experimental　results　on　real　datasets　show　that　nPSE-GAN　can　better　reveal　abnormal　patterns　of　brain　activity　and　has　a　good　application　potential　in　brain　network　analysis.