In　recent　years,　many　supervised　learning　algorithms　have　been　successfully　applied　for　photovoltaic　(PV)　fault　diagnosis.　In　practice,　it　is　not　possible　to　effectively　obtain　labels　of　large　samples,　limiting　the　engineering　application　of　these　algorithms.　As　for　the　unsupervised　learning　algorithm,　it　is　completely　adaptive　learning,　requiring　a　large　number　of　samples　to　better　learn　the　potential　features　in　the　data.　To　address　the　above　problems,　an　improved　online　fault　diagnosis　method　is　proposed,　which　uses　a　small　number　of　labeled　samples　to　train　the　semisupervised　ladder　network　(SSLN)　fault　diagnosis　model　to　realize　the　diagnosis　of　line-to-line　faults,　open-circuit　faults,　partial　shadow　faults,　and　hybrid　faults.　In　the　proposed　method,　only　the　real-time　operating　voltage　and　current　of　PV　array　are　needed　for　fault　diagnosis.　The　sequential　voltage　and　current　of　the　PV　array　are　first　normalized,　and　the　sequential　power　waveforms　are　obtained　through　numerical　calculation.　Then,　the　SSLN　is　used　to　extract　the　fault　features　from　the　sequence　power　waveforms.　Finally,　the　classification　is　realized　using　the　SSLN＇s　noiseless　encoder.　To　eliminate　overfitting　and　improve　convergence,　the　activation　function,　optimizer,　and　loss　function　of　the　SSLN　is　studied　and　improved.　Meanwhile,　numerical　simulations　and　measured　data　verify　that　the　proposed　method　provides　strong　anti-interference,　and　the　diagnostic　accuracies　of　both　exceed　98%.　Comparative　experiments　show　that　the　proposed　method　outperforms　algorithms　such　as　squared-loss　mutual　information　regularization,　semisupervised　support　vector　machine,　graph-based　semisupervised　learning,　and　semisupervised　extreme　learning　machine.　©　2011-2012　IEEE.