Wine sample selection and collection for near-infrared spectroscopy
The samples used in the experiment were from six different regions: Yinchuan and Zhenbeipu in Ningxia, Changli in Hebei, Yantai in Shandong, Turpan in Xinjiang, and Limari Valley in Chile. Thirty samples were collected from each producing region, resulting in a total of 180 samples. To ensure the reliability of the experimental results, all the wine samples used in the experiment were finished wines made from Cabernet Sauvignon grapes through temperature-controlled fermentation and aging in oak barrels.
The NIR test platform consisted of a NIR2500 (Ideaoptics Instruments Co., Ltd., China), a HL2000-12 halogen light source, a RIB-600-NIR direct fiber optic, an R4 color dish spectrum measurement stand, and Morpho software (version 3.2 12.2, available at http://www.ideaoptics.com ).
The experiment is carried out in a constant temperature and humidity environment. The wine samples are left to stand for 10 minutes before being opened and placed in a quartz cell, and the wine samples are collected using a near-infrared spectrometer. The wavelength range of the near-infrared spectra collected by the wine samples is 900-2500 nm, with a wavelength resolution of 3.2 nm, an integration time of 1 ms-120 s, and a signal-to-noise ratio of 7500:1. The collection time of each spectral scan is 10 seconds. The total collection time of near-infrared spectroscopy for all samples is 50 minutes.
UPLC-Q-TOF-MS experimental method
The Q-TOF instrument used in the experiment is an Agilent High Resolution Liquid Mass Spectrometry (HRLC-MS) system (Agilent Technologies, Santa Clara, CA, USA). The main components of wine are extracted and analyzed by MassHunter B.06.00 (Agilent Technologies, Inc. 2006–2019, Santa Clara, CA, USA) and Mass Profiler Professional 12.5 software (Agilent Technologies, Santa Clara, CA, USA) (version 12.5, available at www.agilent.com.cn) to extract and analyze 130 main components of wine. The experimental reagents include distilled water, ammonium formate (chromatographically pure), and methanol (chromatographically pure).
The experimental method is as follows: First, 1 mL of sample is precisely measured in a 1.5 mL centrifuge tube, centrifuged at 10,000 rpm for 10 min at 4 °C, and passed through a 0.22 μM microporous filter for on-machine detection. The chromatography column is an Agilent Eclipse Plus C18 (3 × 150 mm, 1.8 μM). Column temperature: 40 °C, autosampler temperature: 4 °C, input: 2 μL, flow rate: 0.3 mL/min, column balance time: 0.5 mL/min, analysis time: 20 min. The mobile phase is 5 mmol/L ammonium formate aqueous phase and methanol phase.
Each data collection cycle screens for parent ions with an intensity greater than 5000. The TOF-MS scan time is 150 ms, with a quality detection range of 50–1000 Da, and collected in high sensitivity mode.
Spectral data preprocessing
In addition to the characteristic data of the detected samples, the original near-infrared spectral data also contain many redundant variables and noise signals caused by external interference.20To eliminate interferences and establish reliable two-dimensional correlation infrared spectra, MSC+SG+FD was used to preprocess the original spectra from the experiments, subtracting the effects of instrument background and drift on the spectral signals, and eliminating the spectral differences caused by the scattering effect caused by the heterogeneity of particle sizes in the wine liquid during the spectral data collection process.twenty oneReduces spectral signal noise and improves the spectral signal-to-noise ratio.
Figure 2a shows the original average near-infrared spectra of six wine samples from different regions in the range of 900–2500 nm. Except for five peaks at 1123 nm, 1281 nm, 1592 nm, 1650 nm, and 1805 nm with slightly different absorbance values, the other bands are basically similar. Figure 2b shows the average near-infrared spectra of the six production regions after MSC + SG + FD pretreatment.
(One) Original near-infrared spectra from six production areas, (b) Mean near-infrared spectra after MSC+SG+FD pretreatment.
2D-COS image acquisition
The acquisition of generalized two-dimensional correlated infrared spectra refers to the process of recording the corresponding infrared spectra of a sample under disturbance (i.e., dynamic spectra), and then performing correlation analysis of a series of dynamic spectra when an external disturbance (electrical, magnetic, thermal, mechanical, chemical, concentration and composition changes, etc.) is applied to the sample being tested, and then the results are presented in a two-dimensional contour map or a three-dimensional diagram to obtain a two-dimensional correlated infrared spectral image.17.
Strictly speaking, the dynamic spectrum refers to the result of subtracting the reference spectrum from the spectrum of a sample in different states due to interferences. The dynamic spectrum including disturbances (electrical, magnetic, thermal, mechanical, chemical, concentration, composition changes, etc.) is recorded as shown in Equation (1).
$$y(v,t) = \left\{ \begin{gathered} x(v,t) – \overline{x}(v),T_{\min } \le t \le T_{\max } \hfill \\ 0,otherwise \hfill \\ \end{gathered} \right.$$
(1)
\(x(v,t)\) Represent the spectral intensity of the sample as a variable \(v\) Being obstructed \(t\)meanwhile \(\overline{x}(v)\) Represents the reference value of the spectral intensity in a variable \(v\)In general, the average spectral intensity value for a variable is \(v\) The sample spectrum obtained from the entire perturbation process ( \(t\) = \(T_{\min }\)To \(t\)= \(T_{\max}\) is) can be interpreted as follows: \(\overline{x}(v)\) Formula (2):
$$\overline{x}(v) = \frac{1}{{(T_{\max } – T_{\min } )}}\int_{{T_{\min } }}^{{T_{\max } }} {x(v,t)} dt$$
(2)
Or you can select a specific value \(\overline{x}(v)\) e.g., the spectral intensity values in the variables \(v\)In the sample spectrum \(t\)= \(T_{\min }\) or \(t\) = \(T_{\max}\) isAt this point, the reference point is the initial or final state of the experiment. If we simply set the reference point to 0, the dynamic spectrum will be the spectral intensity observed at this time.
Before calculating the generalized two-dimensional correlation spectrum, a Fourier transform needs to be performed on the dynamic spectrum. \(y(v,t)\) The calculation of the generalized two-dimensional correlation spectrum obtained from the cross-correlation analysis is shown in equation (3).
$$Y_{1} (\omega ) = \int_{-\infty }^{+\infty } {y(v_{1},t)}e^{-i\omega t}dt$$
(3)
$$\Phi (v_{1} ,v_{2} ) + i\Psi (v_{1} ,v_{2} ) = \frac{1}{{\pi (T_{\max } – T_{\min } )}}\int_{0}^{\infty } {Y_{1} (\omega )} Y_{2}^{*} (\omega )d\omega$$
(Four)
The real part \(\phi (v_{1} ,v_{2} )\) The imaginary part of equation (4) is called the synchronous correlation spectrum, and the imaginary part is \(\Psi (v_{1} ,v_{2} )\) This is called the asynchronous correlation spectrum.
The synchronous correlation spectrum represents the change in the spectral magnitude similarity between two variables as a function of a disturbance. The asynchronous correlation spectrum represents the change in spectral magnitude due to a disturbance, the difference in spectral magnitude between two variables, or the phase difference in spectral magnitude between two variables.
In a real experiment, it is necessary to transform the integral formula under finite and discrete experimental values. Assuming that m data points are measured at equal intervals under disturbance, \(t\) The synchronous correlation spectrum at this time is expressed as shown in equation (5).
$$\Phi (v_{1} ,v_{2} ) = \frac{1}{m – 1}\sum\limits_{j = 1}^{m} {y(v_{1} ,t_{j} )} \cdot y(v_{2} ,t_{j} )$$
(5)
The calculation of the asynchronous correlation spectrum is expressed as equation (6).
$$\Psi (v_{1} ,v_{2} ) = \frac{1}{m – 1}\sum\limits_{j = 1}^{m} {y(v_{1} ,t_{j} )} \cdot \sum\limits_{j = 1}^{m} {M \cdot y(v_{2} ,t_{j} )}$$
(6)
of \(M\) Equation (6) represents the nth order Hilbert-Noda matrix, and its expression is as follows:
$$M_{jk} = \left\{ {\left. \begin{gathered} 0,j = k \hfill \\ \frac{1}{\pi (k – j)},j \ne k \hfill \\ \end{gathered} \right\}} \right.$$
(7)
In the process of obtaining synchronous and asynchronous correlation spectral images of samples in the experiment, we first obtain the average spectrum of the samples, then select the characteristic band regions of the spectrum based on the Q-TOF results, compare the spectral data of the characteristic regions of each sample with the average spectrum, and use time as the disturbance variable to generate synchronous and asynchronous correlation spectral images of all samples.
Establishing a deep learning model
The deep learning model used in the experiment is a convolutional neural network (CNN). As a feedforward neural network with a convolutional structure, CNN extracts features of input data through convolution operations. As shown in Figure 3, the basic structure consists of an input layer, a convolution layer, a pooling layer, a fully connected layer, and an output layer, and has the structural characteristics of local area connection, weight sharing, and downsampling. Weight sharing and local area connection reduce the complexity of the network model and reduce the number of weights. The calculation formula for convolution is as follows:
$$X_{n}^{m} = f\sum\nolimits_{{i\hat{I}}} {(X_{n}^{m – 1} K_{in}^{m} + b_{n}^{m} )}$$
(8)
\(m\) is the number of convolutional layers, \(f()\) This is the activation function. By calculation, \(n\) Layer feature maps \(m\)Can be obtained,\(K\)represents the convolution kernel,\(\hat{I}\)represents a set of input images and represents the corresponding offset matrix.\(n\)Layer feature maps\(m\).
Diagram of the structure of a convolutional neural network model.
The structure of the CNN model is shown in Figure 4. The pooling layer in the structure can effectively reduce the size of the parameter matrix and improve the calculation speed and robustness of feature data. The activation function operates the output of the convolution layer nonlinearly to extract feature information. In this experiment, we selected the linear rectifier function (RELU) as the activation function. Compared with other activation functions, the sparsity of the linear rectifier function can accelerate learning and simplify the model. After applying the linear rectifier layer, the pooling layer performs parameter reduction to combine certain features of the convolution layer, avoid overfitting, and ensure a stable convolution process.
In this experiment, we use MATLAB 2020a to establish a CNN prediction model, construct a generalized 2D-COS image with filtered near-infrared characteristic wavelengths as input, and place neurons in the output layer to regress the classification results.
