In　resonant　grounding　systems,　most　single-phase-to-ground　faults　evolve　from　IAFs　(Intermittent　Arc　Faults).　Earlier　detection　of　IAFs　can　facilitate　fault　avoidance.　This　work　proposes　a　novel　method　based　on　machine　learning　for　detecting　IAFs　in　three　steps.　First,　the　feature　of　zero-sequence　current　is　automatically　extracted　and　selected　by　a　newly-designed　FINET　(＂For　IAFs,　Neuron　Elaboration　Net＂),　instead　of　traditional　feature　selection　based　on　time-frequency　decomposition.　Moreover,　data　of　the　zero-sequence　current　divided　by　different　time　windows　are　successively　input　into　the　trained　FINET.　A　proposed　PSF　(principal-subordinate　factor)　analyses　the　results　obtained　from　FINET　to　improve　anti-interference　in　the　mentioned　IAF　detection　algorithm.　Experiments　using　PSCAD/EMTDC　software　simulation　data　show　the　proposed　method　is　feasible　and　highly　adaptable.　In　addition,　the　detection　result　of　on-site　recorded　data　demonstrates　the　effectiveness　of　the　proposed　method　in　practical　resonant　grounding　systems.