A network ultrasonic Wasserstein generative adversarial network (US-WGAN), which can generate ultrasonic guided wave signals, is proposed herein to solve the problem of insufficient datasets for pipe ultrasonic nondestructive testing based on deep neural networks. This network was trained with pre-enhanced and US-WGAN-enhanced datasets with 3000 epochs; the ultrasound signals generated by the US-WGAN were proved to be of high quality (peak signal-to-noise ratio scores in the range of 30–50 dB) and belong to the same population distribution as the original dataset. To verify the effectiveness of the US-WGAN, a fully connected neural network with seven layers was established, and the performances of the network after data enhancement using the US-WGAN and popular virtual defects were verified for the same network parameters and structures. The results show that adoption of the US-WGAN effectively suppresses the overfitting phenomenon while training the network and increases the dataset size, thereby improving the training and testing accuracies (>97%). Additionally, we noted that a simple, fully connected shallow neural network was sufficient for achieving high-accuracy defect classification using the US-WGAN data enhancement method.