LSTM and GNN Based Anti-Collision Algorithm for DFSA Labels

This work utilizes an advanced network model together with Graph Transformer and Long Short-Term Memory (LSTM) to address the tag collision problem in RFID systems. This approach improves upon traditional methods that rely on simple BP neural networks. First, a unified data set for training is employed, which eliminates the need to train separate models for each frame length and simplifies the data processing workflow. Second, a tag grouping algorithm was incorporated, enabling the system to maintain high throughput even under large-scale data conditions, with an average improvement of 10% when the number of tags exceeds 1000. Third, by leveraging Graph Transformer to extract complex spatial correlations among tags and LSTM to predict temporal features of tag collisions, optimized model achieves highly accurate estimations of pending tags, reducing prediction errors to the magnitude of le-4. These improvements not only enhance system performance and recognition efficiency but also enlarge the scalability and practicality of the RFID collision optimization process. This work demonstrates the potential of combining spatial temporal feature extraction with advanced neural networks to achieve superior results in large-scale RFID environments.