There　is　a　limitation　in　the　process　of　acoustic　signal　detection,　which　lies　in　serious　background　noise　interference　and　the　complex　correlation　between　acoustic　signal　characteristics　and　operation　states.　Integrating　the　denoising　model　and　feature　classification　model,　a　method　of　transformer　acoustic　signal　anti-interference　detection　and　operation　state　detection　based　on　deep　learning　is　proposed　in　this　paper.　Through　tests　in　anechoic　rooms,　acoustic　signals　of　transformers　in　the　normal　state　or　under　harmonic　load　are　acquired.　Combining　these　signals　with　the　background　noise,　a　dataset　containing　12000　samples　of　acoustic　signals　is　constructed.　To　implement　anti-interference　detection,　Conv-TasNet　is　utilized　to　get　the　transformer　acoustic　signal　and　environmental　noise　separated;　then,　ResNet　is　utilized　to　classify　the　operation　states　of　the　transformer　accurately.　Results　show　that　compared　with　the　blind　source　separation　method　through　RNN　and　FastICA,　the　denoising　model　established　in　this　paper　improves　Si-SDRi　parameters　by　37.4dB　and　17.53dB　respectively,　and　the　transformer　operation　state　classification　model　established　in　this　paper　classifies　the　test　dataset　with　an　accuracy　of　97.7%,　thus　providing　an　effective　method　for　the　extraction　of　transformer　acoustic　signal　and　diagnosis　of　transformer　operation　states　in　complex　environments.　©　Published　under　licence　by　IOP　Publishing　Ltd.