Answers are grounded in this paper and cite their sources.
In real-time food safety monitoring, detection of pigment content in cream using near-infrared spectroscopy is severely compromised by temperature-induced non-linear spectral shifts, which render linear calibration methods ineffective. This study introduces a Fusion Attention Dual-Target Network under Dynamic Loss Modulation (FADNet-DLM) to address this challenge. By simultaneously predicting temperature and pigment content, the model effectively balances regression accuracy, preventing the dominant thermal signal from suppressing subtle pigment features. FADNet-DLM integrates multi-scale feature fusion and dynamic weight adjustment, outperforming traditional chemometric methods and standard neural networks. Feature enhancement via multi-output symbolic regression revealed that specific vibrational overtones exert varying influences on the targets; encapsulating these weights before input reduced the coefficient of determination difference between targets to 0.0143. Stratified analysis confirmed exceptional robustness across low, medium, and high temperature ranges, demonstrating the capability to decouple interference across phase transitions. Overall, the model achieved a pigment determination coefficient of 0.9617 and an RPD of 3.9677 across the 5–35 ° Celsius range, providing a methodological framework for decoupling physical thermal interference from trace chemical signals in complex emulsion matrices.
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: