An Efficient Data Generation Method Based on Flow Matching for Bearing Fault Diagnosis Under Imbalanced Data Conditions

Abstract

The process of fault diagnosis is greatly constrained by imbalanced datasets, as a lack of fault data results in models with poor diagnostic performance. To overcome this issue, generative adversarial networks (GANs) have been utilized. However, they often suffer from training instability and are prone to mode collapse, where this critical issue causes the generator to produce a very limited variety of outputs. Consequently, this article proposes an efficient data augmentation framework based on flow matching for the domain of bearing fault diagnosis under imbalanced data. First, the continuous wavelet transform of vibration signals is employed to train a conditional flow matching model. After training, the model can generate new synthetic data by transforming a random Gaussian noise vector into a high-fidelity time–frequency spectrum. A key advantage of this approach is its computational efficiency, where the model produces diverse and authentic samples with as few as 5-15 generation steps, significantly reducing the generation time compared to traditional diffusion models. Furthermore, to comprehensively evaluate the diversity and fidelity of synthesized samples, various evaluation techniques, such as kernel inception distance (KID), the GAN-test/GAN-train metric and t-distributed stochastic neighbor embedding (t-SNE) visualizations of the learned feature space, are employed. Ultimately, a series of imbalanced fault diagnosis experiments was conducted to validate the practical effectiveness of the proposed method, resulting in final diagnosis accuracies of 95.90% and 97.31% on two widely used datasets and 98.81% on an actual test bed.