A novel imbalance fault diagnosis method based on data augmentation and hybrid deep learning models

In recent years, although a large number of intelligent fault diagnosis methods have been proposed, their effectiveness has mostly been validated through balanced datasets. However, the data accumulated in real engineering scenarios tend to show a long-tail distribution, with a much larger number of healthy samples compared to faulty samples. This phenomenon makes the classifiers of deep learning models prone to learn features from samples of the majority class, which affects the model’s accuracy in identifying faulty samples. To solve the above imbalanced data problem, a novel hybrid model is proposed. First, to reduce the dimensionality of the dataset, time and frequency domain feature indicators are used instead of the raw vibration signals. Subsequently, the k-nearest neighbor maximum trend diffusion algorithm optimized through grid search is employed to generate high quality fault samples for balancing the dataset. Finally, the temporal correlation features and spatial features of the data are extracted and fused to predict the fault classes by the hybrid deep learning model. The effectiveness of the proposed method is verified on two different gearbox datasets. The experimental results indicate that the proposed model can generate synthetic samples that align with the distribution of real samples. And it also achieves the highest diagnostic accuracy and F1 score under different imbalance ratios, which are superior than other models.