Prompt, Accurate, and Noncontact Material Identification Using a Single Microwave Sensor With Machine Learning Analysis

In this article, a novel approach to microwave sensors is proposed and demonstrated that allows for prompt, accurate, and noncontact material identification even with arbitrary lift-off distances between the sensor and the material under test (MUT). A multilayer perceptron (MLP) is trained to directly learn the relation of the measured reflection spectra of a homemade open-ended coaxial cable resonator probe with respect to different MUTs and then achieve one-to-one mapping between a measured spectrum and the MUT. As a proof-of-concept demonstration, the performance of the MLP model is tested using easily accessible materials, including wood, glass, water, and metal. Reflection spectra of the probe for different MUTs with different lift-off distances are acquired using a vector network analyzer (VNA) and are utilized to train and test the MLP model. High classification accuracy is realized. More importantly, the MUT can be identified using the well-trained model with an arbitrary lift-off distance and even a tilt angle with respect to the probe end face, as long as the lift-off distance is within the model calibration range. This preliminary work is expected to inspire researchers in the field of microwave sensors and instrumentation to develop a new generation of intelligent systems with expanded functionalities that could find potential applications in diverse fields.