关键词: 粮食, 水分含量检测, 射频传感, 多模态, 域对抗神经网络

Abstract: Moisture detection is an indispensable part of grain storage and trade, and non‑destructive and fast grain moisture detection can be achieved by using various RF sensing technologies, such as WiFi, Radio Frequency Identification (RFID), and radar. However, existing solutions are developed based on a single type of RF signal, and corresponding detection models need to be trained for different RF signals, which can lead to additional manpower costs. Therefore, this paper proposes a grain moisture detection method fusing multimodal RF signals, named RF—Grain, which can be applied to multiple RF sensing technologies. Firstly, a WiFi channel state information (CSI) data preprocessing method is proposed to address the noise problem caused by multipath environment and hardware defects. Secondly, a domain‑adversarial neural network model is proposed to eliminate the differences in the distribution of grain moisture features extracted from different types of RF signals. Finally, three different RF sensing technologies were designed to detect grain moisture. The convolutional neural network was used as a comparison to evaluate the performance of the proposed method and compare it with existing methods. The experiments showed that the proposed method could effectively detect grain samples with five different moisture levels, with overall accuracy rates of 98.87%, 96.22% and 96.56%, better than traditional convolutional neural networks. The proposed method has the advantages of high accuracy and good generalization, and it provides powerful technical support for non‑destructive moisture detection of grain.

Key words: grain, moisture content detection, radio frequency sensing, multi?modal, domain adversarial neural