Knowledge-based data augmentation of small samples for oil condition prediction

Due to insufficient monitoring data, the reliability and accuracy of oil condition predictions are not guaranteed. Data-driven models provide data augmentation with small samples to solve this problem. However, the absence of degradation mechanisms would introduce unpredictable uncertainties in a long-term prediction. To address this, a data augmentation method is proposed for improved prediction by integrating degradation mechanisms and monitoring data. Primarily, a degradation model is established considering the degradation mechanisms. The model parameters are estimated with the time-vary probability distribution of the monitoring data. Therefore, the evidential variables are used to describe parameters with small samples. Then, the detailed parameters are estimated by integrating small-sample and prior parameters. With this well-trained model, the augmented data can be obtained with a particle filtering method for prediction. For validation, both the sparse and truncated samples from real-world monitoring are used to demonstrate the superiority of the proposed method. The high predicted accuracy demonstrates that the reliability of oil condition prediction can be guaranteed even with small samples.