5G NWDAF Feature Mapping for Training and Inference Data

The patent segments the dataset into two distinct parts: a training dataset and an inference dataset. This segmentation is achieved through the NWDAF's ability to collect data samples and divide them into separate sets, each with specific characteristics. The training dataset is used for model training while the inference dataset is used for generating analytics outputs, ensuring they remain distinct and preventing data leakage between phases.

The patent implements a multi-functional NWDAF architecture that can perform both model training and model inference operations. The NWDAF is designed with versatile capabilities to collect data samples, prepare training datasets, execute model training, and perform model inference using the same infrastructure. This universal design eliminates the need for separate dedicated hardware for training, reducing operational costs while maintaining reliable model training capability.

The patent extracts and separates the data collection and processing functions from individual NWDAF instances. Instead of every NWDAF collecting and processing all data, the system designates specific NWDAF instances to collect data samples and prepare training datasets. Other NWDAF instances can then utilize these prepared datasets for model training and inference, reducing redundant processing while maintaining broad data availability across the network.

The patent applies preliminary action by collecting and preparing data samples offline before model training. The NWDAF collects data samples from the network in advance, processes them into training datasets, and stores them for future use. This preliminary data preparation enables the model to be trained on extensive datasets without compromising real-time analytics capability, as the trained model can then rapidly process new data during inference phase.