Quantitative model of near infrared spectroscopy based on pretreatment combined with parallel convolution neural network

Abstract

With the advancement of near infrared spectroscopy and deep learning technology, nondestructive quantitative testing plays a crucial role in various fields such as agriculture, petrochemicals, medicine, food, and forage. Currently, a high precision prediction model combined with spectral pretreatment is gaining popularity. In this paper, a quantitative analysis model of convolutional neural network including parallel network module (PaBATunNet) was proposed. PaBATunNet was composed of a one-dimensional convolutional layer, a parallel convolution module, a flattening layer, four fully connected layers and a parameter regulator (PR). The parallel convolution module was made up of five submodules and a Concatenate function. The linear and nonlinear multidimensional features of the spectra were extracted by five submodules and spliced by Concatenate function. The prediction accuracy of PaBATunNet was improved by optimizing the model parameters through the PR. Moreover, eight spectral pretreatment methods combined with PaBATunNet were tested on public datasets of beer, milk, and grain. The results indicated that the prediction accuracy of PaBATunNet with different spectral pretreatment increased by 4.83% to 28.40% on beer, 7.09% to 27.99% on milk and 4.96% to 25.31% on grain, compared to the PaBATunNet with original spectral. Among all models, the first derivative (D1) PaBATunNet (D1-PaBATunNet) performed the best. Compared with D1 partial least squares (D1-PLS), the prediction accuracy of D1-PaBATunNet increased by 34.89%, 65.04%, and 48.26% on beer, milk, and grain, respectively. Compared with D1 principal component regression (D1-PCR), the prediction accuracy increased by 34.17%, 63.98%, and 48.08% on beer, milk, and grain, respectively. Compared with D1 support vector machine (D1-SVM), the prediction accuracy increased by 39.29%, 61.78%, and 50.50% on beer, milk, and grain, respectively. Compared with D1 back propagation neural network (D1-BP), the prediction accuracy increased by 90.29%, 63.72%, and 44.75% on beer, milk, and grain, respectively. The problems of low prediction accuracy, poor stability, poor generalization ability and high risk of overfitting have been solved by D1-PaBATunNet. This study establishes an essential theoretical foundation for building a fast, nondestructive and high-precision quantitative analysis model of near infrared spectroscopy.