Fault　monitoring　can　find　out-of-control　conditions　of　equipment　operation　in　a　timely　manner,　which　is　essential　for　eliminating　faults　and　for　stable　operation　of　industrial　systems　in　batch　processes.　Many　conventional　data-driven　fault　detection　methods　focus　less　on　the　non-Gaussian　and　Multi-stage　characteristics　of　batch　process　data,　which　may　result　in　degradation　of　monitoring　performance.　In　this　paper,　a　Multi-stage　Fourth　Order　Moment　Staked　Autoencoder　(M-FOM-SAE)　is　designed　to　solve　the　above　problems.　The　proposed　method　firstly　automatically　determines　the　number　of　clusters　and　divides　the　batch　process　into　multiple　stages.　After　that,　the　FOM-SAE　model　is　established　in　each　sub-stage,　which　can　not　only　effectively　learn　the　nonlinear　features　of　process　data,　but　also　extract　the　non-Gaussian　information.　The　proposed　strategy　is　applied　to　real-world　industrial　processes.　Experimental　results　indicate　that　it　can　better　capture　the　non-Gaussian　and　Multi-stage　characteristics　of　process　data,　and　improve　the　ability　to　monitor　abnormalities.