A study on quantitative analysis methods for mixed solution concentrations based on deep learning.

To explore efficient and non-destructive detection methods for ion concentrations in mixed salt solutions, this study proposes a deep learning model integrating a feature-extraction-based convolutional neural network (CNN) with an attention mechanism for predicting concentrations in multi-component salt solutions using near-infrared spectroscopy (NIRS). Three-component mixed salt solution samples containing NaCl, KCl, and MgCl2 were prepared, and their NIRS data at various concentrations were collected. In the subsequent analysis, after outlier removal, smoothing differentiation preprocessing, and feature selection were applied to the spectra, a CNN-Attention prediction model was constructed. This model was comparatively analyzed with other models, such as Partial Least Squares Regression (PLSR) and Random Forest (RF). Concurrently, the modeling performance using water-subtracted spectra versus raw spectra (without water subtraction) was compared. The results indicated that in mixed solutions with severe spectral overlap, the predictive performance of models built on raw spectra comprehensively surpassed that of models using water-subtracted data. The proposed CNN-Attention model demonstrated significant superiority over traditional benchmark models, achieving prediction coefficients of determination (R2) for K+, Na+, and Mg2+ of 0.939, 0.897, and 0.853, respectively. Finally, the SHAP (SHapley Additive exPlanations) method was employed to validate the rationality of the model's decisions and the effectiveness of the selected features. This study concludes that the proposed CNN-Attention model can effectively address the challenge of spectral overlap, demonstrating the potential of deep learning for the quantitative analysis of complex mixtures via NIRS.