AC　series　arc　faults　(SAFs)　are　one　of　the　leading　causes　of　electrical　fires　in　buildings,　and　the　development　of　arc　fault　detection　devices　(AFDDs)　can　effectively　reduce　the　fire　risk　caused　by　arc　faults.　To　address　the　issue　of　unsatisfactory　detection　performance　for　SAFs　and　frequent　false　positives　in　existing　AFDDs　when　dealing　with　unknown　load　combinations,　this　paper　proposes　an　adaptive　SAF　detection　system.　The　system　is　based　on　the　remote　interaction　between　AFDD　and　cloud　server,　which　enables　the　AFDD　to　update　its　SAF　detection　model　for　unknown　load　combinations,　thereby　improving　its　generalization　performance.　First,　a　lightweight　neural　network　model　for　SAF　detection　based　on　depth-wise　separable　convolution　and　inverted　residual　block　was　designed　and　ported　to　the　K210　chip,　combined　with　peripheral　circuits　to　create　the　AFDD.　The　AFDD　collects　high-frequency　coupling　signals　from　the　circuit　at　a　sampling　rate　of　100　kHz,　achieving　real-time　SAF　detection　with　a　detection　cycle　of　80　ms.　The　cloud　server　receives　and　filters　false　positive　and　SAF　data　uploaded　by　the　AFDD　during　operation,　and　updates　the　detection　model　on　the　AFDD　through　data　augmentation　and　transfer　learning　to　improve　its　generalization　capability.　Experimental　results　show　that　the　normal　state　recognition　rate　of　the　updated　AFDD　for　unknown　load　combinations　increased　from　98.87%　to　99.92%,　and　the　SAF　recognition　rate　improved　from　96.26%　to　98.16%.　The　results　demonstrate　that　the　adaptive　SAF　detection　system　significantly　improves　the　AFDD’s　performance　in　reducing　false　positives　and　missed　detections　for　unknown　load　combinations.　©　2024　IOP　Publishing　Ltd.　All　rights,　including　for　text　and　data　mining,　AI　training,　and　similar　technologies,　are　reserved.