The　nonlinear　output　characteristics　of　PV　arrays　and　maximum　power　point　tracking　(MPPT)　techniques　bring　more　difficulties　to　fault　diagnosis.　The　fault　diagnosis　model　based　on　electrical　transient　time-domain　analysis　is　an　effective　method　for　solving　the　above　problems.　However,　existing　studies　using　transient　processes　usually　train　their　models　by　extensive　labeled　datasets,　and　some　approaches　apply　normalization　methods　with　environmental　condition　sensors　or　reference　PV　panels.　Therefore,　Fisher　discrimination　dictionary　learning　(FDDL)　for　sparse　representation　is　explored　for　diagnosing　PV　array　faults,　including　line-to-line　faults　(LLF),　open-circuit　faults　(OCF),　and　partial　shading　faults　(PSF),　with　a　small　labeled　dataset,　and　a　dynamic　normalization　method　without　additional　sensors　is　proposed　to　process　transient　data.　Moreover,　LLF　and　PSF　that　have　similar　characteristics　under　low　mismatch　should　be　further　distinguished.　The　proposed　model　is　designed　with　two　stages.　In　the　first　stage,　a　multiple　classifier　trained　using　small　labeled　datasets　with　all　fault　types　is　applied　to　diagnose　all　kinds　of　studied　PV　array　faults.　Then,　a　dictionary　only　for　PSF　and　LLF　is　learned　in　the　second　stage　to　further　identify　LLF　and　PSF.　Finally,　a　1.8　kW　rooftop　grid-connected　PV　system　with　$6\times3$　PV　arrays　is　applied　to　validate　the　performance　of　the　proposed　model.　The　comparison　result　shows　the　superiority　of　the　proposed　model.　©　2013　IEEE.