Digital identification method for natural incense raw materials in functional incense base

Abstract

The invention discloses a digital identification method for natural incense raw materials in a functional incense base. The method comprises the following steps: collecting nuclear magnetic fingerprints of a natural incense raw material and a functional incense base by using a nuclear magnetic resonance spectrum, and establishing a nuclear magnetic resonance spectrum database of the natural incense raw material; the method for recognizing component information in the mixture based on the combination of 1H NMR spectrogram data and deep learning is used for recognizing the inclusion relation between a material base formed by mixing natural incense raw materials and a natural incense raw material mixture. The two convolution modules respectively learn the characteristics of the natural perfume raw material and the blended perfume base, the characteristics are spliced, the spliced characteristics pass through a two-dimensional convolution layer and a spatial pyramid pooling layer and then are input into a full connection layer, and the possibility that the blended perfume base contains the natural perfume raw material is predicted. And evaluating the accuracy, sensitivity and specificity of the model by using the test set. According to the method, the tobacco material base can be distinguished, auxiliary manual work is replaced to carry out formula design on tobacco material liquid essence, and the method has the advantages of being rapid, efficient and high in accuracy.

Description

Technical Field

[0001] The invention belongs to the technical field of digital identification of liquid essences, and in particular relates to a method for digitally identifying natural fragrance raw materials in a functional fragrance base. Background Art

[0002] Functional aroma bases are mainly used to increase the taste of cigarettes, reduce irritation, and cover up impurities. The base is composed of a variety of natural aroma raw materials according to a certain formula. Natural aroma raw materials are mixtures. The base is a mixture of mixtures. Identifying the natural aroma raw materials in the base is to identify the mixture in the mixture. Therefore, this process is called mixture identification. Digital identification of base aroma is the basis of digital flavoring and the key to independent control of key technologies of flavors and fragrances in the tobacco industry. The natural aroma raw materials in the functional aroma base formula are mostly water-soluble and non-volatile components, with complex components, very few volatile components, and the characteristic information in the volatile components is not obvious. It is difficult to complete the digital characterization of functional aroma bases and natural aroma raw materials using conventional chromatography-mass spectrometry. The digital characterization of fingerprints of natural aroma raw materials and liquid flavors has always been a difficult problem, that is, there is no method that can comprehensively characterize all component information in natural aroma raw materials and functional aroma bases, and it is even more impossible to realize the digital identification of functional aroma base component information. Relying solely on the experience and judgment of formulators is not only time-consuming, but also inefficient. Summary of the invention

[0003] In order to solve the above technical problems existing in the prior art, the present invention provides a method for digitally identifying natural aroma raw materials in functional aroma bases. The method of the present invention can effectively realize the digital identification and traceability of functional aroma bases for tobacco.

[0004] The technical solution of the present invention to solve the above technical problems is: a method for digitally identifying natural fragrance raw materials in functional fragrance bases, characterized in that it comprises the following steps:

[0005] 1) Establishing a nuclear magnetic resonance spectrum database of natural fragrance raw materials using nuclear magnetic resonance spectrum;

[0006] 2) Perform baseline correction, zeroing, normalization and data amplification on each natural fragrance raw material;

[0007] 3) Establish a deep learning DeepMID model to detect whether the characteristic peaks or characteristic peak clusters of each natural fragrance raw material exist in the nuclear magnetic resonance spectrum of the functional material base, so as to realize the digital identification of natural fragrance raw materials in the material fragrance base.

[0008] Furthermore, the nuclear magnetic resonance spectrum includes hydrogen spectrum, carbon spectrum, and one-dimensional and two-dimensional nuclear magnetic resonance spectra of nitrogen spectrum, phosphorus spectrum, and sulfur spectrum containing heterocyclic rings;

[0009] Furthermore, the characteristic peaks or characteristic peak clusters of the nuclear magnetic resonance spectrum data in step (1) include chemical shift and abundance information of the nuclear magnetic resonance spectrum.

[0010] Furthermore, in the step (1), a nuclear magnetic resonance spectrum database of natural fragrance raw materials is established by using nuclear magnetic resonance spectroscopy, and the nuclear magnetic resonance characteristics of natural fragrance raw materials are collected as follows: 200 mg of natural fragrance raw materials are taken in a centrifuge tube, and the fragrance raw materials are fully dissolved by vortexing, sonicating, centrifuging and other operations with deuterated water, deuterated chloroform, deuterated methanol, and deuterated DMSO, and 550 μL is taken in a nuclear magnetic injection tube; the injection time is 3 min-8 h.

[0011] Furthermore, the characteristic peak or characteristic peak cluster of each fragrance raw material is selected manually or by machine learning, and the baseline and chemical shift corrections are performed.

[0012] Furthermore, the baseline correction method includes a baseline correction method based on an adaptive iterative reweighted penalty least squares method and a baseline correction method in the nmrglue software package.

[0013] Furthermore, in the step (3), the deep learning model uses cross-correlation to match characteristic peaks, and the threshold setting step is to use a deep learning method to establish an identification model for each component in the database, detect whether the characteristic peak or characteristic peak cluster of each natural fragrance raw material exists in the nuclear magnetic resonance spectrum of the functional material base, and determine whether the natural fragrance raw material exists in the functional fragrance base.

[0014] Furthermore, the deep learning DeepMID model consists of four parts: data augmentation, pseudo-companion convolutional neural network, model-based component recognition, and spatial pyramid pooling layer.

[0015] Furthermore, the data amplification refers to superimposing several NMR spectra sampled at random ratios from a spectral database with random noise to generate negative and positive spectral pairs, which are divided into training, validation and test sets.

[0016] Furthermore, the threshold range for predicting the presence of natural fragrance raw materials in fragrance bases is 0.4-1.

[0017] Furthermore, to predict whether natural fragrance raw materials exist in the fragrance base, the NMR spectrum of the fragrance base is represented as a vector s, and the NMR spectra of all natural fragrance raw materials in the spectral library can be represented as a matrix D(N×P). Each spectrum pair is predicted by the DeepMID model to obtain the probability of the existence of natural fragrance raw materials in the fragrance base.

[0018] Furthermore, the evaluation indicators of the prediction results of natural fragrance raw materials in fragrance bases include accuracy, sensitivity and false negative rate.

[0019] The technical effects of the present invention are as follows: the present invention uses nuclear magnetic resonance spectra to establish a nuclear magnetic resonance spectrum database of natural fragrance raw materials, and performs baseline and chemical shift correction processing; uses a deep learning DeepMID model to establish an identification method for component information in functional fragrance bases, detects natural fragrance raw materials in functional fragrance bases, and effectively realizes the traceability of fragrance raw materials. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a flow chart of the present invention.

[0021] Figure 2 Flow chart of NMR data preprocessing in the present invention.

[0022] Figure 3 The figure shows the comparison of NMR spectra of the carob extract in the present invention before and after baseline correction.

[0023] Figure 4 The schematic diagram is a schematic diagram of the digital identification method of natural fragrance raw materials in the functional fragrance base of the present invention.

[0024] Figure 5 600MHz of tobacco Maillard reaction, Roman chamomile extract and valerian root extract from top to bottom 1 H NMR spectrum. DETAILED DESCRIPTION

[0025] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0026] like Figure 1 As shown, a digital identification method of natural fragrance raw materials in a functional fragrance base of the present invention comprises the following steps:

[0027] 1) Using nuclear magnetic resonance spectroscopy to collect nuclear magnetic resonance spectroscopy data of natural incense raw materials, and establishing a nuclear magnetic resonance spectroscopy database of natural incense raw materials. The nuclear magnetic resonance spectrum includes hydrogen spectrum, carbon spectrum, and one-dimensional and two-dimensional nuclear magnetic resonance spectra of nitrogen spectrum, phosphorus spectrum, and sulfur spectrum containing heterocyclic rings.

[0028] 2) Baseline correction and zeroing of solvent peaks, normalization and data amplification are performed on the nuclear magnetic resonance spectrum data of each natural fragrance raw material. The characteristic peaks or characteristic peak clusters of the nuclear magnetic resonance spectrum include chemical shift and abundance information of the nuclear magnetic resonance spectrum. The baseline correction method includes a baseline correction method based on the adaptive iterative reweighted penalty least squares method and the nmrglue software package. Adaptive iterative reweighted penalty least squares smoothing is used for baseline correction, and the smoothness of the fitted baseline is better controlled by iteratively estimating the noise level and adjusting the weight accordingly. The baseline correction of the nuclear magnetic spectrum is preliminarily completed. Most natural fragrance raw materials contain large solvent peaks. In order to further highlight the characteristic peaks or characteristic peak clusters of the fragrance, some solvent peaks need to be zeroed, and then the spectrum is shifted as a whole by setting the 3-(trimethylsilyl) propionic acid-D_4 sodium salt peak to 0.00 to achieve calibration of possible data drift of the resonance spectrum. After completing the calibration, data normalization is then performed, and the peak position containing useful information is sliced ​​to reduce the interference of useless information data, and data amplification is performed on this basis. For the specific preprocessing process, see Figure 2 .

[0029] The data augmentation process is to superimpose several NMR spectra sampled at random ratios from an established NMR spectrum database with random noise to generate negative and positive spectral pairs, which are divided into training, validation, and test sets.

[0030] 3) A deep learning DeepMID model was established, and the NMR spectrum pairs formed by data amplification were used as input. Features were extracted through two 9-layer convolution and pooling sub-networks, and then the features were concatenated for two-dimensional convolution. After entering the SPP layer, the model entered the fully connected layer to output the probability of predicting whether the characteristic peak or characteristic peak cluster of each natural fragrance raw material exists in the NMR spectrum of the functional material base. The model was continuously optimized by adjusting hyperparameters such as the number of layers, learning rate, and number of generations to achieve digital identification of natural fragrance raw materials in the fragrance base of the material.

[0031] The schematic diagram of the DeepMID model is shown in Figure 4 It mainly consists of four parts: data augmentation, pseudo-companion convolutional neural network, model-based component recognition, and spatial pyramid pooling layer. The NMR spectra in the natural fragrance raw material database are Figure 4 The data shown in (a) are augmented to form NMR spectrum pairs which are input into the network. Figure 4 (b) is the network structure of the DeepMID model, which consists of two sub-networks. Each sub-network consists of 9 convolutional layers. The extracted features are concatenated in series for two-dimensional convolution and then input into the SPP layer and then predicted through the fully connected layer. Figure 4(c) is the component identification based on the DeepMID model. The inclusion relationship between each natural fragrance raw material (mixture) and fragrance base (mixture composed of mixtures) in the database is predicted by the DeepMID model. The data augmentation procedure is to superimpose several NMR spectra sampled at random ratios from the spectral database with random noise to generate negative and positive spectral pairs, which are divided into training, validation and test sets. DeepMID takes NMR spectral pairs as its input, extracts high-level features using convolutional layers, and predicts the probability of pure compounds in the mixture using fully connected layers. For each compound in the database, model-based component identification predicts its probability in the mixture. By filtering the predicted probabilities using thresholds, the possible components of the mixture can be obtained.

[0032] The method for identifying natural aroma raw materials in liquid aroma bases requires preprocessing of the original data before data amplification, so that the model can better learn the characteristics from natural aroma raw materials with more complex components such as extraction and animal and plant metabolites. The original data needs to be preprocessed, including baseline correction, zeroing solvent peaks, normalization, internal standard peak calibration, slicing of the peak part of the spectrum (0.300-10.700ppm), etc., and then data amplification is performed on this basis. The sample set is amplified from the following aspects to improve the diversity of the data. The data enhancement operations are: (1) random scaling of the height of the spectrum within a certain range (0.2-1.0); (2) adding random Gaussian noise to the spectrum; (3) slicing the chemical shift of the nuclear magnetic resonance spectrum within a certain range (-3×10 -4 ~3×10 -4 ppm) for left and right translation.

[0033] In the method for identifying natural fragrance raw materials in liquid fragrance base, when establishing the DeepMID model, each NMR data can be regarded as a one-dimensional vector, and the dual input of the convolutional neural network is a pair of NMR data. A 9-layer one-dimensional convolutional neural network is used for each sub-network to learn the features of NMR data, each convolutional layer contains 32 convolution kernels of size 5×1, the activation function is ReLU, and the He normal distribution initializer is introduced to perform weight initialization to make the network more stable. After each convolutional layer, there is a maximum pooling layer, which can help solve the offset of the nuclear magnetic resonance spectrum. The two convolutional sub-networks extract the corresponding features of the dual-input NMR spectra respectively, and then the features extracted from the two inputs are vertically spliced ​​to form a two-dimensional feature map, and then through a two-dimensional convolutional layer containing 128 5×5 convolution kernels, and then through a (4×4, 3×3, 2×2, 1×1) multi-layer window pooling SPP layer is input into the fully connected layer to predict the possibility of natural fragrance raw materials in the liquid fragrance base. Spatial pyramid pooling can make the features of CNN no longer single-scale. It can produce fixed output for input of any size, improve the scale invariance of the image, and reduce overfitting. The activation function of the fully connected layer is ReLU, and the activation function of the output layer is Sigmoid; the optimizer used is the Adam optimizer with high computational efficiency and low memory requirements. The cross entropy of the binary classification is selected as the loss function.

[0034] The method for identifying natural fragrance raw materials in liquid fragrance base can be used to predict the formula of fragrance base after the DeepMID model is established. The specific identification process is that the NMR spectrum of the fragrance base is represented by a vector s. The NMR spectra of all natural fragrance raw materials in the spectral library can be represented by a matrix D(N×P), where N is the number of natural fragrance raw materials in the library and P is the number of data points in each NMR spectrum. The NMR spectrum s is compared with each NMR spectrum in the database (D i ) can be combined to obtain N spectral pairs (s, D 1 ),...,(s,D N ). Each spectral pair is predicted by the DeepMID model to obtain the probability of a natural fragrance raw material that may exist in the fragrance base. In order to control the number of false positives, a threshold (e.g. 0.5) is set for this probability. If the probability of a natural fragrance raw material in the fragrance base is greater than the threshold, it can be determined as a candidate. After completing the prediction and filtering of N spectral pairs, a candidate list of natural fragrance raw materials present in the fragrance base is obtained.

[0035] 4) Evaluation of the recognition results: Three indicators were used to evaluate the accuracy of the recognition results of natural fragrance raw materials in the fragrance base, namely accuracy, sensitivity, and specificity.

[0036] The evaluation indexes of the recognition results of natural fragrance raw materials in the functional fragrance base are the accuracy rate (ACC), sensitivity (TPR) and false negative rate (FPR) used in the mixture analysis of the proposed model. The calculation formulas of ACC, TPR and FPR are as follows:

[0037]

[0038]

[0039]

[0040] TP, TN, FP and FN are true positive, true negative, false positive and false negative respectively. In the binary classification problem, the sample is either classified as positive or negative. True positive means that both the predicted value and the true value are positive; true negative means that both the predicted value and the true value are negative; false positive means that the predicted value is positive and the true value is negative (taking the false as true); false negative means that the predicted value is negative and the true value is positive (taking the true as false).

[0041] Specific implementation process

[0042] This embodiment will illustrate the specific process and results of identifying the functional fragrance bases mixed with natural fragrance raw materials.

[0043] 1. Instruments used to collect NMR spectra

[0044] Bruker AVANCE NEO 600M NMR spectrometer (Bruker Scientific Instruments, Switzerland); IKA MS3 vortex mixer; KQ5200DE CNC ultrasonic cleaner; VL-165R high-speed desktop refrigerated centrifuge; AL104 electronic balance; and PHS-3C pH meter.

[0045] 2. Establish a nuclear magnetic resonance spectrum database. All natural fragrance raw material samples adopt a unified pre-treatment method.

[0046] Dissolve 1 mL of NaOD (40%, 10 M) in 9 mL of D 2 Prepare NaOD (1.0 M) in 1 L D O; take 0.4000 g of 3-(trimethylsilyl) propionic acid-D_4 sodium salt and dissolve it in 1 L D 2 Prepare 3-(trimethylsilyl)propionic acid-D_4 sodium salt (0.04%, w / v) in water. Weigh 1.232 g KH 2 PO 4and 10mg 3-(trimethylsilyl) propionic acid-D_4 sodium salt were dissolved in 100mL D2O and fully dissolved. The pH of the above solution was adjusted to 6.0 using a pH meter and NaOD (1.0M), and a phosphate buffer (90mM, pH6.0) was prepared. 200mg of natural fragrance raw materials were weighed in a 2mL centrifuge tube using an electronic balance, and 0.6mL CD3OD and 0.6mL PB (phosphate buffer) were added. The sample was vortexed at 1500 rpm for 60s, then ultrasonically dissolved for 15min, and centrifuged at 13000rpm for 10min at 15°C. 550μL of the supernatant was taken and placed in a nuclear magnetic resonance tube for nuclear magnetic resonance testing.

[0047] 3. NMR data acquisition and measurement conditions

[0048] The natural fragrance raw materials were measured at 25°C on a 600M NMR instrument using a low temperature probe (-196°C). The measurement frequency is 1 HNMR 600.12MHz, signal accumulation 128 times, noesypprld sequence for water peak suppression; 3-(trimethylsilyl) propionic acid-D_4 sodium salt as internal standard; MeOD for lock. Carbon spectrum determination frequency 13 C NMR 400.12MHz, signal accumulation 1024 times, the measurement time for each sample is 58 minutes, and other test parameters are set the same 1 H NMR.

[0049] 4. Experimental equipment information

[0050] The deep neural network and related modules were implemented in Python (version 3.8.13), Tensorflow GPU package (version 2.5.0), and scikit-learn (version 1.0.0). The NMR spectra were read using the nmrglue package (version 0.8.dev0). The computational tasks were performed on an Inspur TS10000 high-performance computing (HPC) cluster using the Slurm workload manager (version 20.02.3). This HPC cluster has 1022 central processing unit (CPU) nodes, 10 fat nodes, and 26 graphics processing unit (GPU) nodes. The DeepMID model was trained on a GPU node with 2×Intel(R) Xeon(R) Gold 6248R processors, 2×Nvidia Tesla V100s, 384G DDR4 memory, and CentOS 7.5 operating system.

[0051] 5. Preparation of functional fragrance base

[0052] The accuracy of this method was tested by using a self-blended functional fragrance base with known composition of natural fragrance raw materials.

[0053] Prepare respectively two-component natural aroma raw material aroma base A1, A2, A3, A4; three-component natural aroma raw material aroma base mixture B1, B2, B3, B4; four-component natural aroma raw material aroma base mixture C1, C2, C3, C4; five-component natural aroma raw material aroma base mixture D1, D2, D3, F1. During the preparation, an electronic balance is used to weigh 200 mg of natural aroma raw materials of equal mass in a 2mL centrifuge tube. The pretreatment method of the functional aroma base sample and the NMR data acquisition and measurement conditions are the same as above.

[0054] In order to study the prediction results of the model when using samples of different masses for NMR testing, two groups of known aromatic bases with the same formula and different dosages were prepared, namely B2 and B2_2; C1 and C1_2, where the total mass of B2 and C1 is 200 mg, the total mass of B2_2 is 300 mg, and the total mass of C1_2 is 400 mg. The obtained spectra are input into the model for prediction, and the prediction results can be obtained as shown in Table 1. The specific prediction results can be viewed in Appendix 1. It can be seen that even if the sample masses used in the preparation are different, the model can still give correct prediction results.

[0055] Table 1 Prediction results of the model for different sample dosages of known fragrance bases

[0056]

[0057] Due to the limitations of sensitivity and detection limit of nuclear magnetic resonance spectrum, for diluted samples, as the concentration decreases, the amount of peak information decreases. Some fragrance bases are diluted with a large amount of propylene glycol, and too high solvent peaks may make it difficult to identify characteristic peaks. Therefore, a gradient experiment of propylene glycol dilution of fragrance bases was carried out to determine the recognition ability of this model for diluted fragrance bases, and the results were input into the model for prediction. The prediction results are shown in Table 22. It can be seen that although propylene glycol dilution will cause certain difficulties in model recognition, the final accuracy rate is still above 80%.

[0058] Table 2 Prediction results of the propylene glycol dilution gradient experiment

[0059]

[0060] Table 3 Identification results of natural fragrance raw materials in fragrance base

[0061]

[0062]

[0063]

[0064] Note: HF, QY, BL, JL, JM, and ZF are manufacturers of fragrance raw materials. The same type of fragrance raw materials may have different manufacturers.

Claims

1. A method for digitally identifying natural fragrance raw materials in functional fragrance bases, characterized in that: The following steps are involved: (1) Establishing a nuclear magnetic resonance spectrum database of natural fragrance raw materials using nuclear magnetic resonance spectroscopy; (2) performing baseline correction, zeroing, normalization and data amplification on the nuclear magnetic resonance spectrum data of each natural fragrance raw material; (3) A deep learning DeepMID model was established to detect whether the characteristic peaks or characteristic peak clusters of each natural fragrance raw material exist in the nuclear magnetic resonance spectrum of the functional material base, so as to realize the digital identification of natural fragrance raw materials in the functional material base.

2. The method for digital characterization of functional fragrance bases and digital identification of natural components as claimed in claim 1: characterized in that, The nuclear magnetic resonance spectrum in step (1) includes hydrogen spectrum, carbon spectrum, and one-dimensional and two-dimensional nuclear magnetic resonance spectra of nitrogen spectrum, phosphorus spectrum and sulfur spectrum containing heterocyclic rings.

3. The method for digital characterization of functional fragrance bases and digital identification of natural components as claimed in claim 1, characterized in that: The characteristic peaks or characteristic peak clusters of the nuclear magnetic resonance spectrum data in step (1) include chemical shift and abundance information of the nuclear magnetic resonance spectrum.

4. The method for digital characterization of functional fragrance bases and digital identification of natural components according to claim 1, characterized in that: The step (1) uses nuclear magnetic resonance spectroscopy to establish a nuclear magnetic resonance spectrum database of natural fragrance raw materials, and collects the nuclear magnetic resonance characteristics of natural fragrance raw materials as follows: take 200 mg of natural fragrance raw materials in a centrifuge tube, use deuterated water, deuterated chloroform, deuterated methanol, deuterated DMSO to fully dissolve the fragrance raw materials through vortexing, ultrasound, centrifugation and other operations, and take 550 μL into a nuclear magnetic injection tube; the injection time is 3 min-8 h.

5. The method for digital characterization of functional fragrances and digital identification of natural components as described in claim 3, characterized in that: The characteristic peak or characteristic peak cluster of each fragrance raw material is selected manually or by machine learning, and the baseline and chemical shift are corrected.

6. The method for digital characterization of functional fragrances and digital identification of natural components as described in claim 5, characterized in that: The baseline correction method includes a baseline correction method based on an adaptive iterative reweighted penalty least square method and a baseline correction method in a nmrglue software package.

7. The method for digital characterization of functional fragrances and digital identification of natural components as described in claim 1, characterized in that: In the step (3), the deep learning DeepMID model uses TMS and 3-(trimethylsilyl) propionic acid-D_4 sodium salt as internal standards to calibrate the nuclear magnetic resonance spectrum and then match the characteristic peaks, set a threshold, and use a deep learning method to establish an identification model for each component in the database, detect whether the characteristic peak or characteristic peak cluster of each natural fragrance raw material exists in the nuclear magnetic resonance spectrum of the functional material base, and determine whether the natural fragrance raw material exists in the functional fragrance base.

8. The method for digital representation of functional fragrance bases and digital identification of natural components according to claim 7, characterized in that: The deep learning DeepMID model consists of four parts: data augmentation, pseudo-companion convolutional neural network, model-based component recognition, and spatial pyramid pooling layer.

9. The method for digital representation of functional fragrance bases and digital identification of natural components according to claim 8, characterized in that: The data amplification refers to superimposing several NMR spectra sampled at random ratios from a spectrum database with random noise to generate negative and positive spectrum pairs, which are divided into training, validation and test sets.

10. The method for digital representation of functional fragrance bases and digital identification of natural components according to claim 7, characterized in that: The threshold range for predicting the presence of natural fragrance raw materials in fragrance base is 0.4-1.

11. The method for digital representation of functional fragrance bases and digital identification of natural components according to claim 10, characterized in that: Predict whether natural fragrance raw materials exist in the fragrance base. The NMR spectrum of the fragrance base is represented as a vector s. The NMR spectra of all natural fragrance raw materials in the spectral library can be represented as a matrix D(N×P). Each spectrum pair is predicted by the DeepMID model to obtain the probability of the existence of natural fragrance raw materials in the fragrance base.

12. The method for digital representation of functional fragrance bases and digital identification of natural components according to claim 11, characterized in that: The evaluation indicators of the prediction results of natural fragrance raw materials in fragrance bases include accuracy, sensitivity and false negative rate.

Images