Series　arc　fault　(SAF)　is　one　of　the　most　harmful　faults　during　the　operation　of　photovoltaic　(PV)　systems.　It　is　a　challenging　task　to　find　SAFs　promptly　for　avoiding　the　PV　fire.　Aiming　at　the　SAFs　under　different　operating　conditions,　a　novel　detection　algorithm　by　combining　the　Hankel　singular　value　decomposition　(Hankel-SVD)　denoising　method　and　the　improved　empirical　wavelet　transform-twin　support　vector　machine　(IEWT-TWSVM)　is　proposed.　The　Hankel-SVD　algorithm　is　used　to　denoise　the　dc-bus　current,　which　alleviates　the　influence　of　switching　frequency　and　irrelevant　background　noise　effectively.　Then,　the　denoised　current　is　decomposed　by　the　IEWT,　and　the　composite　multiscale　permutation　entropy　of　each　frequency　band　is　input　into　the　TWSVM　classifier　of　the　salp　swarm　optimization　for　completing　the　fault　detection.　The　proposed　algorithm　not　only　can　detect　SAFs　at　various　fault　locations,　but　also　resist　dynamic　shading,　inverter　startup,　strong　wind,　and　other　interference　phenomena.　Moreover,　the　detection　performance　due　to　the　arc　transient　process,　a　long-line　fault,　a　single　string　array　fault,　a　multiply　string　array　fault,　and　different　sampling　rates　of　this　model　is　verified　in　this　study,　and　the　corresponding　results　are　relatively　ideal.　Experimental　results　show　that　the　detection　accuracy　of　the　proposed　method　is　as　high　as　98.10%　for　the　measured　data,　which　is　more　superior　than　other　methods,　such　as　wavelet　decomposition,　empirical　mode　decomposition,　statistics　methods　including　mean,　standard　deviation,　and　entropy.