An efficient detection method for camel milk quality based on dielectric spectroscopy technology

To address the long-standing dual challenges in the camel milk industry, namely, the difficulty in identifying adulteration and the severe lag in nutritional component detection, this study analyzed the variation patterns of the dielectric constant (ε') and dielectric loss factor (ε") of camel milk fat, protein, and lactose. Based on the dielectric spectrum, detection models were constructed using partial least squares (PLS), convolutional neural network (CNN), long short-term memory network (LSTM), CNN-LSTM, and an optimized CNN-LSTM fused with the Crested Porcupine Optimizer (CPO) and squeeze-and-excitation network (CPO-CNN-LSTM-SEN) for camel milk quality detection. The results indicate that across the 0.1 to 26.5 GHz frequency range, ε' exhibits a gradual decreasing trend with increasing frequency, and ε" demonstrates an initial decrease followed by an increase. Both ε' and ε" decrease with elevated fat, protein, and lactose content, showing strong negative correlations. Among CNN, LSTM, and CNN-LSTM, CNN-LSTM achieved optimal performance in discriminating camel milk adulteration with water and similar species milk (e.g., cow, goat, and mixed milk), attaining a 100% discrimination rate. Among PLS, CNN, LSTM, CNN-LSTM, and CPO-CNN-LSTM-SEN, CPO-CNN-LSTM-SEN performed the best in simultaneous quantitative determination of the 3 nutritional components of camel milk (fat, protein, and lactose). Its prediction set determination coefficients R_P² reached 0.9334, 0.9602, and 0.9703, respectively, with prediction set root mean square error below 0.1392%. Validation of the optimal CPO-CNN-LSTM-SEN model yielded R² values of 0.9186, 0.9507, and 0.9456, respectively, with root mean square error under 0.2231%.