Sparse　auto　encoder(SAE)　can　reduces　information　loss　and　extract　the　meaningful　feature　by　learning　the　deep　structure　of　complex　data.　This　paper　presents　a　novel　SAE　based　semi-supervised　feature　learning　method　for　fault　diagnosis　of　batch　process　which　includes　two　stages,　namely,　unsupervised　pre-training　stage　and　supervised　tuning　stage.　At　the　unsupervised　pre-training　stage,　denoising　SAE(DSAE)　is　utilized　by　introducing　denoising　auto　encoder　into　SAE　to　improve　the　robustness　of　network.　At　the　supervised　tuning　stage,　the　pretrained　DSAE　netwrok　is　optimized　using　back　propagation　algorithm　to　improve　the　accuracy　of　classification.　The　proposed　method　is　validated　on　penicillin　fermentation　simulation　experiment　and　Escherichia　coli　fermentation　experiment.　Experimental　results　show　that　the　proposed　approach　achieves　good　fault　diagnostic　performance　and　is　superirior　to　the　traditional　fault　diagnosis　method.　©　2018　IEEE.