Low-field NMR-based deep learning for non-destructive quality assessment of frozen model foods

Gel model foods with 90 % and 80 % water content were prepared for non-destructive quality testing of samples frozen under gradient temperature conditions. First, the drip loss and textural indices of the frozen samples were measured, and the nucleation process of the samples under each freezing condition was analyzed to generate freezing curves. Subsequently, the transverse relaxation times (T₂) of all samples were collected in the frozen state using low-field nuclear magnetic resonance (LF-NMR). Finally, two machine learning methods, partial least squares regression (PLSR) and back-propagation artificial neural network (BP-ANN), were employed to construct distinct models: the LF-NMR-PLSR model and the LF-NMR-BP-ANN model. The results of model parameters (R², RMSE, and RPD) indicated that the LF-NMR-BP-ANN model performed significantly better than the LF-NMR-PLSR model, demonstrating its effectiveness in non-destructive quality assessment of frozen products.