A transferable bearing fault diagnosis method by FM-WGAN-GP under limited samples

Complex industrial working conditions and diverse operational parameters give rise to significant challenges in data collection scalability and feature adaptability across different scenarios. Transfer learning offers a solution, yet lacks a standard distance-evaluation method. This paper proposes a CNN-integrated, source-generated bearing task-driven diagnosis method. Here, a generative adversarial network model (FM-WGAN-GP) is adopted to solve limited real fault sample and low-quality generated samples. Further, an improved training mechanism is used to enhance the quality and diversity of samples. The generator of this model maps noise into vibration signals, while the discriminator extracts features to distinguish between real and fake signals. Based on the generated samples, a source diagnostic task is constructed. Subsequently, a CNN classifier is used to perform fault classification for the target task through cosine annealing fine-tuning transfer. Experimental results show that both the creativity index and diversity index exceed 0.8, indicating that this method significantly expands the scale of the training dataset and enhances the diversity and complexity of the data without increasing the actual data collection cost. In addition, this method outperforms other generative methods in target bearing diagnosis tasks with limited samples.