Abstract
This paper proposed a novel rapid non-contact detection method for measuring frozen and thawed foods based on an innovative "data + neural network model" strategy. The purpose of the study is to develop a fast and effective detection method for frozen-thawed food, thereby overcoming critical industrial limitations such as slow sweeping speed and wide frequency range. The proposed model replaces the conventional full-spectroscopy scanning approach by requiring only the measurement signal ratio ( Im(∇V/V) ) at three fixed frequency points: 100 kHz, 500 kHz, and 1 MHz. A dataset comprising 4000 samples (labeled as "fresh" or "frozen-thawed") was constructed using finite element modeling and real measurement data. A fully connected neural network (FCNN) was employed as a binary classification model and trained on this dataset. The outcome of the study demonstrates that the proposed model achieves high classification accuracy on both the test set and physical validation samples. Compared to the conventional full-spectrum scanning method (7-8 min/sample), the new approach completes a single measurement in only 1-1.5 s. While some spectral information is sacrificed by not performing a full frequency sweep, the method significantly improves detection efficiency in the specific application scenario of frozen-thawed food identification, while substantially reducing instrument complexity and cost. These results provide both a theoretical foundation and a practical solution for implementing rapid, non-destructive, and non-contact identification of the frozen-thawed state of fresh ingredients in automated production lines, demonstrating considerable potential for industrial applications in the food sector.