Hybrid DCNN-Enabled Depolarizing Chipless RFID: Improving Tag Detection Across Varying Lossy Surfaces and Shapes

This paper presents a comprehensive design and implementation approach for robust detection of depolarizing chipless RFID (CRFID) tags. Depolarizing tags are advantageous compared to co-polar CRFID tags due to their improved performance on RF-lossy materials. This work introduces the application of deep learning (DL) regression modelling to a specialised dataset of depolarised Radar Cross Section (RCS) measurements of a custom 3-bit CRFID tag, acquired through an extensive robot-based data acquisition method. A dataset of 12,600 depolarised Electromagnetic (EM) RCS signatures were collected using an automated data acquisition system to train and validate a 1-dimensional Convolutional Neural Network (1D CNN) architecture. A novel hybrid 1D CNN with Bi-LSTM and attention mechanism architecture was also implemented to visualize the model attention and improve detection performance. We present, for the first time reported in literature, a comprehensive design and AI implementation approach for reliably detecting identification (ID) information from depolarized signals. Also, we report the first instance of describing the impact of surface permittivity variations, tag deformations, tilt angles, and read ranges, all integrated into model training for enhanced robustness in detecting ID information. The developed models facilitate real-time identification and recording of objects, enhancing IoT applications in varied environments. It was observed that both models were able to generalize well to given data, with Model-1 achieving a low RMSE of 0.040 (0.66%) on an unseen test dataset. However, the hybrid model reduced the error further by 27.5% with a test RMSE of 0.029 (0.48%).