Milk spoilage detection plays a pivotal role in safeguarding food safety and minimizing waste within the dairy sector, although conventional chemical assays remain labor-intensive, invasive, and expensive. The present investigation introduces a non-invasive microwave microstrip sensor coupled with a convolutional neural network (CNN) for real-time assessment of milk spoilage progression. The sensor, modeled and optimized using Advanced Design System (ADS) software to exhibit dual passbands (1807–2466 MHz and 3604–4426 MHz), was fabricated on an RT/Duroid 4003 substrate and evaluated using 10 commercial milk samples (3 % fat) procured sequentially over 10 days and maintained at 21 °C. Measurements of the S21 transmission parameter (101 frequency points per spectrum, with five replicates per sample yielding 50 spectra in total) demonstrated a substantial amplitude disparity, notably at 2166 MHz, where the difference between the freshest (day 10) and most spoiled (day 1) samples attained 7.02 dB—equivalent to approximately 105 times the mean standard deviation (0.067 dB)—facilitating robust differentiation of dielectric alterations attributable to microbial degradation. A one-dimensional CNN was trained on preprocessed spectral data augmented fivefold with white Gaussian noise using five variable standard deviations (σ = 0.20–0.60 dB) to simulate real-world measurement fluctuations, expanding the dataset from 50 to 250 spectra and attaining a training accuracy of 95.5 % and a validation accuracy of 90 %. This hybrid methodology surpasses traditional approaches in terms of rapidity and non-destructiveness, providing a viable framework for milk quality surveillance with applicability to other perishable commodities.
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号