To address the scarcity of fault samples and dispersed multimodal features in transformer winding fault diagnosis, this paper proposes a diagnostic method combining a Wasserstein conditional generative adversarial network with gradient penalty (WCGAN‑GP) and deep multimodal feature fusion. First, WCGAN‑GP is employed to generate vibration and ultrasonic fault samples, alleviating the problems of small sample size and class imbalance. A dynamic time warping (DTW)‑based 1‑nearest neighbor (1NN) is introduced to quantitatively evaluate the quality of the generated samples. Then, differentiated feature extraction is applied for different modalities. Leakage flux signals are fed into a one‑dimensional improved residual network to extract local dynamic features. Vibration and ultrasonic signals are transformed into two‑dimensional feature maps and input in parallel into a two‑dimensional improved dense network to extract global sequence relationships and time‑frequency features. Finally, a layer‑wise cross‑modal interaction (LCMI) mechanism is designed to achieve deep fusion of the three modalities. Experiments show that the method effectively improves diagnostic performance under small‑sample conditions, providing a reliable solution for transformer winding fault diagnosis.