Data Augmentation for Discharging Time Prediction of Particle Flow: A Deep Learning and Style Transfer Approach

Abstract Recent years have witnessed an increasing interest in utilizing deep learning models trained on large datasets for particle flow research. The discharging time of particle flow characteristics can be well predicted by the deep learning methods. However, the collection of a substantial amount of experimental data (real data) can be burdensome due to the significant manpower and equipment expenses involved. To address the challenge of small and insufficient samples, we have developed a data augmentation strategy with transfer learning that effectively combines real and simulated data, thereby synthesizing data to expand the dataset at a minimal cost. In this specific study focusing on discharging time prediction, we generate synthetic data that combines the flow characteristics from simulated data with the image style from experimental images since the time required for simulation and synthetic data generation is negligible compared to the effort involved in obtaining real data. The results indicate that the style transfer model can successfully generate synthetic data with realistic styles, significantly reducing data collection burden. Besides, models trained with synthetic data have the ability to predict discharging time for real scenes images of particle flow. Our proposed strategy provides a feasible data augmentation approach for particle flow research, and further efforts are needed on data quality and network capabilities in the future.