White spirit flavor identification method based on machine learning

Abstract

The invention discloses a white spirit flavor recognition method based on machine learning, and belongs to the technical field of food detection and artificial intelligence. Aiming at the problems of high subjectivity of manual evaluation, low efficiency of mass spectrometry, noise sensitivity of a machine learning model and the like in the prior art, the method provides a solution integrating solvent background deduction and feature weight screening. The method specifically comprises the following steps: diluting a white spirit sample with methanol according to a volume ratio of 1: 10, collecting mass spectrum data through GC-MS, and dynamically deducting a methanol background peak; a BP neural network (GABP) optimized by a genetic algorithm is utilized to automatically analyze a weight coefficient of each molecular peak to flavor classification, and key features are screened; a classification model is constructed based on XGBoost, and parameters are optimized through cross validation, so that automatic judgment of the flavor type, authenticity and quality of the white spirit is realized. According to the method, through data dimension reduction and model collaborative optimization, the overfitting problem caused by high noise and high redundancy of mass spectrum data is solved, the classification accuracy is remarkably improved, and the method can be extensively applied to white spirit brand identification, process optimization and market quality supervision.

Description

Technical Field

[0001] The present invention relates to the fields of food quality testing and artificial intelligence technology. Specifically, it is an automated liquor flavor recognition method based on the combination of mass spectrometry data preprocessing and machine learning algorithms. The method can be applied to liquor flavor classification, brand identification, authenticity recognition and quality grading. Background Art

[0002] As a traditional Chinese brewed beverage, the flavor characteristics of baijiu are formed by the interaction of hundreds of trace volatile components. Currently, flavor evaluation in the industry mainly relies on the following two methods:

[0003] Manual sensory evaluation: Professional tasters subjectively rate the flavor of wines through their senses of smell and taste. This method suffers from significant individual variability, susceptibility to environmental interference, difficulty in standardizing evaluation criteria, and an inability to quantify the contribution of specific flavor molecules.

[0004] Chromatography-mass spectrometry analysis: The components of liquor are separated by gas chromatography (GC) or liquid chromatography (LC), and molecular information is detected by mass spectrometry (MS). Although chemical composition data can be obtained, there are the following limitations:

[0005] (1) The detection process is cumbersome and requires manual screening of target peaks and elimination of solvent background interference;

[0006] (2) It is difficult to establish a correlation model between multidimensional molecular characteristics and flavor attributes;

[0007] (3) The accuracy of distinguishing complex fragrances (such as mixed fragrances) is insufficient.

[0008] In recent years, although some studies have attempted to introduce machine learning into liquor analysis, the following technical bottlenecks still exist:

[0009] (a) Data noise interference: The methanol solvent used in the pretreatment of liquor samples introduces background peaks, and traditional threshold filtering methods cannot accurately remove interfering signals;

[0010] (b) Feature redundancy: Mass spectrometry data has high dimensionality and contains a large number of non-flavor-related peaks, so direct modeling can easily lead to overfitting.

[0011] (c) Poor model interpretability: Existing algorithms have difficulty in automatically identifying key flavor molecules and their weights, which restricts the industrial application of flavor regulation. Summary of the Invention

[0012] To address the above technical deficiencies, the present invention proposes a liquor flavor recognition method that integrates solvent background subtraction, feature weight screening, and machine learning classification. Its core innovations include:

[0013] Methanol background dynamic filtering technology: By establishing a solvent background peak database, it can accurately eliminate interference signals and improve the signal-to-noise ratio of mass spectrometry data;

[0014] Flavor molecule weight self-learning mechanism: Utilizes machine learning models to automatically analyze the contribution of each molecular feature to flavor classification and screen key flavor markers;

[0015] Hierarchical modeling strategy: Solve the generalization problem of high-dimensional data modeling through the process of data cleaning → feature dimensionality reduction → classifier optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 : Schematic diagram of the process of the present invention;

[0017] Figure 2 : Mass spectra of 7 kinds of liquor (including methanol background mass spectrum);

[0018] Figure 3 : The weight determination results of each molecule;

[0019] Figure 4 : Prediction accuracy of liquor flavor determination;

[0020] Figure 5 : PCA principal component analysis was used to determine the flavor results of 7 kinds of liquor. DETAILED DESCRIPTION

[0021] Example: Liquor Flavor Recognition Based on XGBoost and GABP Neural Network

[0022] 1. Sample Pretreatment and Data Acquisition

[0023] Seven batches of Luzhou-flavor, Maotai-flavor, and Qing-flavor liquors were selected and diluted with methanol at a volume ratio of 1:10. Mass spectral data were collected using gas chromatography-mass spectrometry (GC-MS), with a total of 30 mass spectra collected for each batch. Solvent interference signals in the samples were dynamically subtracted by comparing them against a database of blank methanol solvent background peaks, retaining the characteristic molecular peaks of the liquor.

[0024] 2. Data standardization and feature screening

[0025] After converting the mass spectrometry data into a standardized matrix, a genetic algorithm-optimized BP neural network (GABP) was used to analyze the weight coefficients of each molecular peak. Key flavor molecules were screened by setting contribution thresholds, significantly reducing the data dimension.

[0026] 3. Classification model construction and verification

[0027] Based on the identified key molecular features, an XGBoost classification model was constructed, with cross-validation used to optimize parameters. Once the model was trained, it could input mass spectrometry data in real time and output results on liquor flavor, authenticity, and quality grade.

[0028] 4. Application Scenarios

[0029] After being deployed on the cloud-based analysis platform, this method can support rapid sampling of winery production lines, authenticity screening by market supervision departments, and traceability by scanning codes at consumer terminals.

[0030] Technical Effects

[0031] The model training efficiency is improved through dynamic subtraction of solvent background and feature weight screening; the quantitative weight analysis of key flavor molecules provides data support for the optimization of liquor processing; the classification model supports multi-scenario expanded applications, and the recognition accuracy is significantly higher than traditional methods.

Claims

1. A liquor flavor recognition method based on machine learning, characterized in that: The following steps are involved: (a) diluting the liquor sample to be tested with methanol solvent according to a preset ratio; (b) using a mass spectrometer to collect data on the diluted liquor sample to obtain an original mass spectrum; (c) by comparing the mass spectrum background peak data of methanol solvent, subtracting the methanol interference peak in the original mass spectrum to generate pure mass spectrum information of the liquor sample; (d) The pure mass spectrum information of each liquor is used as an independent data unit and input into the machine learning model for training; (e) automatically analyzing the weight coefficients of each molecular characteristic peak for flavor recognition through the machine learning model to screen out key flavor-related molecular information; (f) Based on the screened molecular information, the training set and test set are divided, and a liquor flavor classification model is constructed to realize the automatic recognition of liquor flavor.

2. The method according to claim 1, characterized in that The ratio of methanol dilution in step (a) is a volume ratio of the liquor sample to methanol of 1:5 to 1:

20.

3. The method according to claim 1, characterized in that The mass spectrometer in step (b) is a gas chromatography-mass spectrometer (GC-MS) or a liquid chromatography-mass spectrometer (LC-MS), and the mass spectrometry acquisition mode is a full scan mode.

4. The method according to claim 1, wherein The method for subtracting the methanol background peak in step (c) includes: establishing a background peak database through the mass spectrum data of blank methanol solvent, or experimentally measuring the mass spectrum characteristic peaks of methanol solvent and performing dynamic matching to eliminate them.

5. The method according to claim 1, wherein The machine learning model in step (d) is selected from one of random forest, support vector machine (SVM), convolutional neural network (CNN), GA-BP or gradient boosted decision tree (GBDT).

6. The method according to claim 1, characterized in that The analysis of the weight coefficients in step (e) includes determining the contribution of each molecular peak to the flavor classification by feature importance ranking, principal component analysis (PCA) or Lasso regression algorithm.

7. The method according to claim 1, characterized in that The ratio of the training set to the test set in step (f) is 7:3 to 8:2, and cross-validation is used to optimize the model parameters.

8. The method according to any one of claims 1 to 7, characterized in that The method is further used for liquor brand identification, authenticity recognition or flavor quality grading.

Images