A novel data augmentation strategy for aeroengine multitask prognosis based on degradation behavior extrapolation and diversity-usability trade-off

For aeroengine multitask prognosis, dataset's quantity and quality significantly affect the prediction performance. Due to the insufficiency and high redundancy of collected data, data augmentation techniques are widely utilized in industrial scenarios. However, traditional methods struggle to balance the degradation behavior diversity along with the usability of generated data. To tackle these challenges, this study proposes a novel data augmentation framework for aeroengine multitask prognosis. A novel First Predicting Time (FPT) identification method is proposed to identify the degradation starting point through Health Indictor (HI) volatility. Then, an optimal data augmentation strategy is designed based on Dual Discriminator Time-series Generative Adversarial Network (DDTGAN) and Negative Sample Elimination (NSE), which can enrich samples by extrapolating degradation behavior with multi-scale temporal features, and eliminating unqualified samples to obtain optimal generated samples through diversity-usability trade-off. Based on these, an adaptive Transformer-Multi-gate Mixture-Of-Experts (T-MMOE) multitask prognosis model with gradient normalization is constructed to predict Remaining Useful Life (RUL) and diagnose faults simultaneously with dynamic weights trade-off between two tasks. The proposed framework was compared with other models through the C-MAPSS dataset. Comparison results manifest that the proposed framework is not only able to generate realistic high-quality time-series data but also outperforms the other prognosis models.