Soft　sensing　techniques　have　been　extensively　employed　to　monitor　water　quality　in　the　Urban　Wastewater　Treatment　Process.　Wastewater　data　often　exhibit　complex　features,　including　nonlinearity,　time　correlation,　and　non-Gaussianity.　Therefore,　to　establish　an　accurate　soft　sensing　model,　extracting　complex　features　from　wastewater　data　during　the　modeling　process　is　essential.　Considering　the　characteristics　of　wastewater　data,　a　feature-augmented　extraction　broad　learning　system　(FAE-BLS)　is　proposed　for　soft　sensing　applications.　Inspired　by　the　structure　of　recurrent　networks,　a　recurrent　cascading　improvement　of　the　feature　window　in　the　broad　learning　system　(BLS)　is　implemented　by　FAE-BLS　to　extract　the　time　correlation　features　of　wastewater　data.　Furthermore,　a　feature　window　based　on　overcomplete　independent　component　analysis　(OICA)　is　proposed　to　extract　the　inherent　non-Gaussian　features　of　wastewater　data.　Finally,　a　method　for　fast　online　model　updating　is　developed　to　address　the　decline　in　model　accuracy　under　the　harsh　conditions　of　wastewater　treatment　processes.　Experimental　results　on　the　wastewater　simulation　platform　validate　the　effectiveness　and　superiority　of　the　proposed　approach.