The　accuracy　of　machine　learning-based　arc　fault　diagnosis　methods　relies　on　the　quality　of　the　selected　features.　First,　an　experimental　system　is　constructed　for　collecting　realistic　arc　fault　current　signals,　and　different　multi-domain　feature　extraction　methods　are　described.　Then,　the　feature　relationships　are　investigated　using　various　correlation　methods　to　reveal　the　negative　impact　of　weakly　relevant　or　redundant　features　on　the　machine　learning　model.　Based　on　the　combination　of　feature　clustering　(FC)　and　maximal　information　coefficient　(MIC),　this　paper　proposes　a　feature　selection　(FS)　strategy.　The　strategy　searches　for　cluster　centers　through　FC　and　merging　processes　to　obtain　non-redundant　features　with　the　highest　representational　abilities.　The　elimination　of　weakly　relevant　clusters　and　the　selection　of　the　most　relevant　features　within　clusters　are　performed　based　on　the　MIC　scores.　Subsequently,　a　valid　feature　subset　with　the　highest　representativeness　and　relevance　is　selected　to　improve　the　accuracy　of　arc　fault　diagnosis.　Finally,　the　proposed　method　is　evaluated　against　several　popular　FS　strategies　using　three　typical　classifiers　on　different　arc　fault　feature　datasets,　and　the　results　show　that　the　strategy　effectively　improves　the　accuracy　of　fault　diagnosis.　In　addition,　the　cross-validation　and　hardware　tests　verify　the　stability,　interference　immunity,　and　feasibility　of　the　proposed　method　for　practical　applications.