The　aim　of　this　study　is　to　provide　a　fast　and　reliable　approach　to　detect　weak　arc　faults　that　do　not　noticeably　distort　the　bus　current,　with　the　minimum　possible　arc　duration　required　to　respond　in　low-voltage　AC　systems.　Progressive　singular-value　decomposition　is　utilized　to　filter　interference　components,　primarily　AC/DC　components.　Then,　the　signals　are　thoroughly　decomposed　by　empirical　analytic　tools　in　the　time-frequency　domain,　combined　with　the　fast　Fourier　transform　to　enhance　feature　extraction　in　the　frequency　domain.　The　features　are　passed　to　the　neural　networks,　where　the　networks　are　trained　and　validated　repetitively　by　datasets　that　are　randomly　selected　from　the　data　sampling.　The　comparison　experiments　demonstrate　the　excellent　performance　of　the　proposed　method　under　all　crucial　evaluation　criteria　of　arc-fault　detection.