Prior Knowledge–Guided Adversarial Augmentation for Small-Sample Fault Diagnosis of Wind Turbine Drivetrain

Abstract

Wind turbine drivetrains operate under variable aerodynamic loading, fluctuating speed, and harsh field environments, making condition monitoring difficult when fault samples are scarce and imbalanced. To address this challenge, this paper proposes a small-sample data augmentation method that fuses variational mode decomposition–particle swarm optimization–symmetric dot pattern (VMD–PSO–SDP) feature optimization with a prior knowledge–guided Wasserstein conditional generative adversarial network (PK-WCGAN). First, VMD is used to decompose vibration signals into informative modal components. In combination with an improved PSO algorithm for optimizing SDP parameters, high-order fault feature maps are generated to enhance feature transferability. Then, real feature maps are used to pretrain the discriminator as prior knowledge, guiding the Wasserstein conditional generative adversarial network (WCGAN) to learn the target domain distribution. A gradient penalty mechanism is introduced to improve the quality of generated samples and training stability. Finally, generated samples are fused with real training samples to construct a hybrid dataset. Experimental results demonstrate that the generated data effectively reduce small-sample distribution mismatch. When the target data samples are expanded from 50 to 1000, the average recognition accuracy of the five models increases from 54.6% to 94.0%, and the risk of overfitting is reduced by up to 26.5 percentage points under the 10-sample setting. This method provides an effective solution for small-sample fault diagnosis in the wind turbine drivetrain.