Fault　diagnosis　of　the　power　devices　in　inverters　is　crucial　for　improving　equipment　reliability.　However,　the　signal　fluctuations　caused　by　load　variations　during　actual　operation　pose　new　challenges　for　inverter　fault　diagnosis.　Existing　data-driven　fault　diagnosis　methods　are　designed　based　on　specific　system　fault　databases,　making　it　difficult　to　overcome　the　influence　of　system　parameter　changes.　In　addition,　existing　transfer　learning　methods　for　variable　working　conditions　often　require　a　large　amount　of　unlabeled　target　domain　data　for　model　training.　In　addition,　the　application　is　limited　by　the　sample　size　of　the　new　working　conditions.　To　tackle　these　challenges,　this　paper　presents　a　novel　approach　for　diagnosing　open-circuit　faults　in　three-phase　inverters　by　leveraging　transfer　learning.　In　this　approach,　the　output　voltage　of　different　three-phase　inverter　loads　is　used　as　the　fault　signal.　Then　a　one-dimensional　convolutional　neural　network　integrating　attention　mechanisms　and　global　average　pooling　layers　is　introduced　to　effectively　capture　the　channel　and　spatial　features　of　fault　characteristics.　Next,　a　domain　adversarial　neural　network　is　employed　to　enable　the　diagnostic　model　to　learn　domain-invariant　features,　so　that　the　target　domain　and　source　domain　cannot　be　distinguished.　Thus,　the　model　built　on　the　source　domain　can　adapt　to　changing　working　conditions.　Finally,　by　utilizing　an　iterative　pseudo-labeling　method　to　train　the　model,　high-precision　diagnostic　outcomes　are　achieved　and　a　limited　number　of　labeled　samples　from　the　target　domain　are　needed.　Experimental　results　show　that　the　proposed　method　achieves　an　average　diagnostic　accuracy　of　96.63%　in　transfer　diagnosis　tasks　across　different　systems,　and　exhibits　robustness　in　environments　with　various　types　of　noise.