Meta-task interpolation-based data augmentation for imbalanced health status recognition of complex equipment

In the research of health status detection technology for complex equipment such as liquid rocket engines, the extreme working environment hinders the widespread conduct of fault experimental simulations, leading to data scarcity and imbalance. Consequently, the performance of intelligent models deteriorates rapidly with direct training. To address this issue, this paper proposes a meta-task feature space interpolation network model. Firstly, the model uses an encoder to map randomly selected task pairs to a more discriminative feature space, and then interpolates corresponding features and labels within this latent feature space to generate additional tasks, increasing the distribution density of tasks and alleviating the problem of insufficient training tasks. Furthermore, the model leverages self-distillation to improve the learning of label information. By integrating soft labels with supervised labels, it captures the hidden category information of newly interpolated tasks, thereby reducing the impact of class imbalance on model performance. The effectiveness of the proposed method is validated through a series of experiments conducted across three different scenarios. The results demonstrate that the proposed method achieves an average accuracy of 97.91% on the turbopump bearing dataset, which is a significant improvement over the comparative methods.