Abstract: o address the challenge of fault detection caused by the strong non-stationary nature and scarce fault samples of power transformer acoustic-vibration signals, a fault detection method integrating compressed sampling, multi-scale time-frequency analysis, and the wasserstein Generative adversarial network (WGAN) was proposed. Firstly, an improved Hadamard compressed sampling matrix was employed to eliminate redundancy and reduce dimensionality of high-dimensional vibration signals. Subsequently, wavelet packet decomposition was utilized to extract multi-band energy features, constructing a time-frequency feature set that characterizes faults. Furthermore, a WGAN with a gradient penalty mechanism was introduced to achieve stable modeling and data enhancement of acoustic-vibration features under limited fault samples. Experimental results demonstrate that the proposed method achieves an area under the curve (AUC) of 0.957 and an overall accuracy of 98.76% in transformer fault detection. An average recognition accuracy of 97.63% for four condition types is attained on actual substation operational data, confirming its engineering applicability and robustness under complex working conditions. Furthermore, an accuracy of 91.64% is maintained under 40 dB noise interference, and strong cross-scenario generalization capability is exhibited. The discriminability and robustness of fault features are significantly enhanced by this method, providing an effective technical solution for online condition monitoring and intelligent diagnosis of power transformers.