CSI Data Augmentation for Neural Network Training

The patent applies data augmentation techniques to create synthetic copies of limited real CSI data through various transformations (phase rotation, amplitude scaling, frequency shifts). These copied data samples expand the training dataset without requiring additional physical measurements, thereby reducing data collection cost while providing sufficient training data for neural network model development

The patent transforms existing CSI data by modifying its parameters such as phase, amplitude, frequency offset, and time delays. These parameter changes generate diverse training samples from a single real measurement, enabling the neural network to learn from varied channel conditions without requiring actual measurements under all possible conditions

The patent generates multiple synthetic copies of limited real CSI data through data augmentation techniques including phase rotation, amplitude scaling, and frequency shifts. This copying approach creates a large training dataset from minimal real measurements, eliminating the need for difficult and time-consuming collection of massive real-world channel state information under diverse conditions

The patent performs data augmentation and synthetic data generation in advance before neural network training. By preparing expanded training datasets through parameter transformations and synthetic copying beforehand, the system eliminates the need for complex real-time data collection during model development, thereby improving compression efficiency while reducing measurement difficulty