Deep learning assisted robust detection techniques for a chipless RFID sensor tag

TL;DR

This paper introduces a novel deep learning approach for robust detection of identification and sensor data from chipless RFID tags, utilizing a large dataset of electromagnetic signatures collected via robotic measurements.

Contribution

It is the first to apply ML and DL regression models to chipless RFID data, incorporating variations in tag surface, tilt, and range for improved robustness.

Findings

1D CNN models outperform traditional ML models in detection accuracy.

Achieved low RMSE of 0.061 for ID detection and 0.0241 for sensing.

Robust detection is feasible across varying tag conditions.

Abstract

In this paper, we present a new approach for robust reading of identification and sensor data from chipless RFID sensor tags. For the first time, Machine Learning (ML) and Deep Learning (DL) regression modelling techniques are applied to a dataset of measured Radar Cross Section (RCS) data that has been derived from large-scale robotic measurements of custom-designed, 3-bit chipless RFID sensor tags. The robotic system is implemented using the first-of-its-kind automated data acquisition method using an ur16e industry-standard robot. A large data set of 9,600 Electromagnetic (EM) RCS signatures collected using the automated system is used to train and validate four ML models and four 1-dimensional Convolutional Neural Network (1D CNN) architectures. For the first time, we report an end-to-end design and implementation methodology for robust detection of identification (ID) and sensing data using ML/DL models. Also, we report, for the first time, the effect of varying tag surface shapes, tilt angles, and read ranges that were incorporated into the training of models for robust detection of ID and sensing values. The results show that all the models were able to generalise well on the given data. However, the 1D CNN models outperformed the conventional ML models in the detection of ID and sensing values. The best 1D CNN model architectures performed well with a low Root Mean Square Error (RSME) of 0.061 (0.87%) for tag ID and 0.0241 (3.44%) error for the capacitive sensing.