Accurate　and　reliable　fault　analysis　is　essential　to　improve　the　performance　of　gas-insulated　switchgear　(GIS)　and　ensure　the　safe　and　stable　operation　of　the　power　system.　Advances　in　the　perception　and　measurement　technology　of　power　primary　equipment　have　increased　the　scale　of　GIS　operation　and　fault　data.　Driven　by　massive　samples,　deep　learning　has　brought　new　opportunities　for　GIS　fault　identification　and　intelligence　diagnosis.　As　an　unsupervised　learning　method,　stacked　autoencoder　(SAE)　can　automatically　extract　representative　expressions　and　in-depth　features　from　massive　unlabeled　data.　Consequently,　it　can　effectively　alleviate　the　problems　of　incomplete　data　category　labels　and　limited　data　samples.　Meanwhile,　transfer　learning　can　avoid　the　need　to　train　the　model　from　scratch　under　the　new　dataset.　For　this　reason,　this　paper　combines　SAE　with　transfer　learning　to　realize　partial　discharge　(PD)　pattern　recognition　of　GIS　in　order　to　solve　the　problem　of　complex　and　variable　and　large　randomness　of　the　on　partial　discharge　samples.　Unsupervised　greedy　layer-wise　pre-training　and　supervised　fine-tuning　are　leveraged　to　train　the　SAE　model.　The　experimental　data　was　generated　using　four　types　of　PD　in　the　laboratory.　The　optimal　SAE　structure　is　trained　under　the　source　domain,　and　the　GIS　PD　pattern　recognition　under　the　target　domain　is　realized　through　transfer　learning.　Experimental　results　show　that　the　proposed　method　has　stronger　robustness　and　generalization　while　improving　the　accuracy　of　partial　discharge　pattern　recognition　of　GIS.　©　2022　IEEE.