Deep Learning-based Recognizing Liquids Using a Multi-resonance Microwave Sensor

In this paper, a microwave sensor is investigated by etching multiple different split ring resonators (SRRs) on the ground layer, thus obtaining multiple complementary split ring resonators (CSRRs). In order to improve the sensitivity, we control the size of the CSRRs between λ/4 to λ/2, so that the the stored energy is large near the resonances. The model uses a two-port microstrip line coupled with the CSRRs to extract the insertion loss (S
21
) in the range of 5–8 GHz, rather than the usual analysis just near a single resonant points. In addition, we find that the size of the CSRRs and the distance between the CSRRs and microstrip line have great influences on its resonant frequency and Q-factor of the sensor. We detect unknown liquids in two cases. First, different types of liquids including water, ethanol, ether, and methanol are investigated, and we find that the characteristics of S
21
are distinctive. Secondly, a challenge task that detecting different concentrations of ethanol solutions is carried out. In this case, the changes in their permittivity are small, indicating small changes in the characteristics of S
21
. Considering the fact that most of liquid containers on the market have grooves, hollow bottom, and different thickness of walls, the measured S
21
of liquids is inconsistent, let alone the placement position error between bottles and the sensing area in rapid test. To solve the problem, we use a feedforward neural network (FNN) by the backpropagation method to process the measured S
21
in a wide band. The results show that the accuracy of detecting different liquids by the deep learning method is 100% and of detecting the same liquid with different concentrations is 99%. Compared with the traditional method, the proposed one is insusceptible to random errors. The nondestructive liquid detection proposed in this work has good future in detecting suspicious liquids in the market.