As　the　terminal　component　of　the　power　system,　the　power　distribution　network　directly　supplies　electrical　energy　to　users　and　undertakes　the　crucial　task　of　ensuring　safe,　reliable,　and　high-quality　power　delivery.　Therefore,　effective　early　warning　for　risks　within　the　operation　of　the　distribution　network　is　particularly　crucial.　This　paper　proposes　an　early　warning　model　using　the　ECAPA-TDNN　audio　classification　model,　aiming　to　reduce　the　manpower　and　resources　required　for　network　operation　and　maintenance.　The　method　involves　on-site　automatic　recording　of　preliminary　datasets　using　a　Raspberry　Pi　and　audio　sensors.　After　multiple　preprocessing　steps,　model　training　is　conducted.　Additionally,　considering　the　periodic　nature　of　tools　during　construction　work　in　the　time　domain,　Wavegram　convolution　is　introduced　to　complement　spectrogram　features　by　combining　time-domain　and　frequency-domain　characteristics　for　optimization.　By　utilizing　the　mentioned　approach　to　identify　external　disruption　events　and　classify　these　events,　the　distribution　network＇s　risk　warning　capability　can　be　significantly　improved　compared　to　the　previous　manual　fault　location　identification,　resulting　in　substantial　savings　in　manpower　and　resources.　©　2024　IEEE.