Method for Detecting Non-Fat Milk Solids Content in Camel Milk Based on Dielectric Spectrum Technology

The difficulty in accurately detecting the nutritional components of camel milk severely restricts the development of the industry. Therefore, rapidly achieving non-destructive detection of the solids-not-fat content (SNFC) in camel milk is key to resolving this issue. This study was conducted on 400 camel milk samples collected from farms in five different regions of Xinjiang, China. Based on dielectric spectroscopy, this study proposes a quantitative detection method for camel milk SNFC that combines preprocessing with feature variable extraction. A network analyzer and an open-ended coaxial probe were used to collect the relative permittivity (ε′) and dielectric loss factor (ε″) of camel milk at 201 frequency points over the frequency range of 0.1–26.5 GHz. The raw dielectric spectra were preprocessed using standard normal variate (SNV), Savitzky–Golay smoothing (SGS), and Savitzky–Golay derivative (SGD) methods. Subsequently, feature variables were extracted from the preprocessed full dielectric spectra (FDS) using uninformative variable elimination (UVE), successive projections algorithm (SPA), and their combination (UVE-SPA). Then, partial least squares regression (PLSR), support vector machine (SVM), long short-term memory (LSTM), and Gaussian process regression (GPR) models were constructed to predict camel milk SNFC. Finally, the optimal model was externally validated. The results showed that the combination of preprocessing and feature variable extraction methods could improve the accuracy of the camel milk SNFC prediction models. Among all methods, SGD preprocessing combined with UVE-SPA feature variable extraction achieved the best overall performance, and the SGD-UVE-SPA-GPR model was identified as the optimal model for predicting camel milk SNFC. Compared with the model constructed from raw dielectric spectra, the model after preprocessing combined with feature extraction exhibited significant improvements in overall prediction accuracy. Specifically, the coefficient of determination of the prediction set RP2 increased from 0.9055 to 0.9685, representing a relative improvement of 6.96%; the root mean square error of prediction (RMSEP) decreased from 0.0899% to 0.0541%, corresponding to a reduction of 39.82%; and the residual predictive deviation of the prediction set (RPDP) increased from 3.2813 to 5.6780, representing an increase of 73.04%. These findings can provide a theoretical basis and technical support for the development of automated detection equipment for camel milk quality.