Wine composition detection utilizing 1DCNN and the self-attention mechanism

Abstract

This study proposes a one-dimensional convolutional autoencoder model that incorporates self-attention mechanisms—1DCNN-ATTENTION-SAE. This model solves the problem of unstable prediction performance in quantitative modeling of multiple components in infrared spectroscopy, especially when dealing with complex nonlinear problems involving severe overlap of characteristic peak bands and difficulty in capturing high-dimensional nonlinear features. The model effectively captures long-term dependencies in infrared spectral data and is particularly suitable for the rapid detection of key components such as pH, total phenols, total sugars, and alcohol in wine. On the ATR-FTIR dataset of dry red wine, the proposed model demonstrates robust performance, achieving a root mean square error (RMSE) of 2.017 g/L and a coefficient of determination (R²) of 0.967 g/L. The RMSE represents the average prediction error across the chemical properties measured (pH, total phenols, total sugars, and alcohol). Similarly, the R² value reflects the overall predictive accuracy of the model for these properties. Additionally, the 1DCNN-ATTENTION-SAE model was further optimized by integrating the DeepHealth algorithm, which is based on the TRANSFORMER structure, forming the hybrid DeepHealth & 1DCNN-ATTENTION-SAE feature fusion model. When applied to the near-infrared spectral dataset of open-source pharmaceuticals to predict bioactivity values, the hybrid model achieved an RMSE of 3.262 g/L and an R² of 0.914 g/L, validating its transfer learning capability in handling "cross-instrument, cross-wavelength" infrared spectral data.