Tea oil authenticity identification method based on in-situ mass spectrum-machine learning coupling technology

Abstract

The invention relates to the technical field of food quality detection, in particular to a tea oil authenticity identification method based on an in-situ mass spectrum-machine learning coupling technology. Comprising the following steps: S1, preparing a sample; s2, acquiring mass spectrum data; s3, feature extraction: performing dynamic binning processing on the mass spectrum data, extracting feature peak information, and performing time dimension feature aggregation to form an initial feature matrix; s4, data preprocessing; s5, training a machine learning model; and S6, intelligent identification and visual output: inputting the mass spectrum original data of the tea oil sample to be detected into the trained multi-model integrated system, outputting an identification result through weighted voting or an optimal model strategy, and giving a confidence interval to quantify the reliability of the result.

Description

Technical Field

[0001] This invention relates to the field of food quality testing technology, and in particular to a method for identifying the authenticity of tea oil based on in-situ mass spectrometry-machine learning coupling technology. Background Technology

[0002] Camellia oil, a high-quality woody edible oil, is rich in various nutrients such as monounsaturated fatty acids, vitamin E, squalene, and sterols. It offers numerous health benefits, including lowering cholesterol and preventing cardiovascular disease, making it highly sought after in the market. As market demand for camellia oil continues to grow, its economic value is becoming increasingly prominent. However, this has also led to a more serious problem of adulteration in the market. Unscrupulous merchants, seeking higher profits, often mix inexpensive other vegetable oils with camellia oil, passing it off as high-quality, seriously harming consumers' interests and disrupting the normal order of the camellia oil market.

[0003] Traditional methods for authenticating tea oil mainly include sensory evaluation, physicochemical analysis, and chromatography-mass spectrometry. Sensory evaluation relies on human senses such as sight, smell, and taste to judge the color, aroma, and flavor of tea oil. This method is highly subjective, greatly influenced by the experience and expertise of the evaluators, and is difficult to accurately identify trace adulteration. Physicochemical analysis, such as the determination of acid value, peroxide value, and iodine value, can reflect the quality of tea oil to some extent, but these indicators are not unique to tea oil and lack specificity for identifying adulteration between different vegetable oils.

[0004] While chromatography-mass spectrometry (GC-MS) can separate and identify chemical components in tea oil with high accuracy and sensitivity, its sample pretreatment process is complex and time-consuming, making it difficult to meet the needs of rapid, high-throughput detection. Furthermore, existing technologies suffer from a significant drawback: excessively long detection times.

[0005] On September 17, 2025, a search was conducted in the China Patent Publication Database using "in-situ mass spectrometry and machine learning and tea oil and doping" as the abstract keywords and with the option to allow synonym expansion. No relevant literature was found.

[0006] On September 17, 2025, an abstract search was conducted on CNKI (China National Knowledge Infrastructure) for the keywords "in-situ mass spectrometry and machine learning and tea oil and doping", but no relevant literature was found. Summary of the Invention

[0007] Purpose of the invention: To provide a more effective method for identifying the authenticity of tea oil based on in-situ mass spectrometry-machine learning coupling technology. Specific objectives are detailed in the specific implementation section, which outlines several substantive technical effects.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: A method for identifying the authenticity of tea oil based on in-situ mass spectrometry-machine learning coupling technology, characterized by the following steps: S1. Sample preparation: Collect pure tea oil samples and other vegetable oil samples from different origins, and mix the other vegetable oils with tea oil in different proportions to prepare binary, ternary, quaternary and quinary adulterated oil samples. S2. Mass Spectrometry Data Acquisition: An Orbitrap Exploris mass spectrometer equipped with a SICRIT ion source was used to perform full scan / data-dependent mass spectrometry analysis on the sample in positive ion mode to obtain raw mass spectrometry characteristic data. S3. Feature Extraction: Dynamically bin the mass spectrometry data, extract feature peak information, and aggregate time-dimensional features to form an initial feature matrix; S4. Data preprocessing: Perform data cleaning and transformation, data partitioning and amplification, and dimensionality reduction on the mass spectrometry data. Dimensionality reduction includes PCA (principal component analysis) and UMAP (uniform manifold approximation and projection). S5. Machine Learning Model Training: Five algorithms, namely Multilayer Perceptron (MLP), Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVM) and Extreme Gradient Boosting (XGB), are used to build a multi-model classification system. The model performance is verified through training and test sets and finally integrated into the software encapsulation module. S6. Intelligent identification and visualization output: Input the raw mass spectrometry data of the tea oil sample to be tested into the trained multi-model integration system, and output the identification results through weighted voting or optimal model strategy, and give the confidence interval to quantify the reliability of the results.

[0009] A further technical solution of the present invention is that the other vegetable oils mentioned in step S1 are any one or more of soybean oil, corn oil, sunflower seed oil, and olive oil. The binary component refers to tea oil mixed with any one of the above samples, the ternary component refers to tea oil mixed with any two of the above samples, the quaternary component refers to tea oil mixed with any three of the above samples, and the pentagonal component refers to tea oil mixed with any four of the above samples.

[0010] A further technical solution of the present invention is that the adulteration ratio of the adulterated oil sample in step S1 is 3%, 5%, and increases sequentially from 10% to 100% in a gradient of 10%. v / v Three parallel samples were prepared for each doping gradient; the mass spectrometry acquisition parameters in step S2 were set as follows: scan range 75–1000 m / z, first-order mass spectrometry resolution 60,000 FWHM, and second-order mass spectrometry resolution 15,000 FWHM.

[0011] A further technical solution of the present invention is that the dynamic binning process in step S3 includes: (1) A total of 14 m / z characteristic intervals were set for different statistical measures; (2) Dynamically divide each interval into bins, count the signal strength in each preset feature interval, and calculate the maximum value, mean, median, standard deviation and sum of statistical features based on the feature importance analysis results; (3) Combining the eight statistical features of the chromatogram, the feature vector dimension of each sample is "the sum of the bin feature numbers of all statistics + 8".

[0012] A further technical solution of the present invention is that the data preprocessing in step S4 includes: (1) Data cleaning: Set all negative values ​​less than or equal to -1 to 0 to avoid NaN values ​​generated by subsequent logarithmic transformation; (2) Data transformation and normalization: Logarithmic transformation is performed on the cleaned data, followed by normalization; (3) Label conversion: The mass spectrometry data labels are unified into two categories (pure tea oil is 1, adulterated tea oil is 0). (4) Data set partitioning and augmentation: The training set (80%) and test set (20%) were partitioned by random shuffling and stratified sampling. The positive class samples in both the training set and the test set were replicated and augmented by 13 times to ensure that the two classes of samples were evenly distributed. (5) Dimensionality reduction: Dimensionality reduction is performed by PCA and UMAP respectively.

[0013] A further technical solution of the present invention is that the training process of the machine learning model in step S5 includes: (1) MLP model: using Adam optimizer, loss function is binary cross-entropy, early stopping strategy and learning rate decay mechanism are set; (2) XGB, RF, SVM, LR models: Use default parameters or optimized parameters, save the model through joblib and output confusion matrix and classification report; (3) Model evaluation: All models were trained on the original features, PCA dimensionality reduction features, and UMAP dimensionality reduction features respectively. The performance was evaluated by confusion matrix, accuracy, precision, recall, F1 score and AUC value.

[0014] A further technical solution of the present invention is that the identification result output mechanism in step S6 includes: (1) Perform three independent mass spectrometry acquisitions on the sample to be tested, generate three sets of feature data and input them into the model group respectively. If the prediction is "adulterated" more than twice, it is determined to be adulterated. (2) When using model groups for identification, the prediction results of multiple models are integrated by majority voting or weighted average method to improve the reliability of identification.

[0015] A further technical solution of the present invention is that the multiple vegetable oils include at least two of soybean oil, corn oil, sunflower seed oil, and olive oil, and a complex adulteration scenario sample library is constructed by mixing multiple components.

[0016] A further technical solution of the present invention is to select MLP, RF, XGB and NON or PCA dimensionality reduction schemes to give full play to the adaptation advantages of the model and features.

[0017] A further technical solution of the present invention is that the accurate identification model for low-proportion adulteration (3%-5%) adopts a combination of RF model (nestimators=100) and PCA dimensionality reduction. For the identification of complex multi-component adulterant systems, a multi-model prediction approach (MLP+RF+XGB) was adopted. The present invention, employing the above technical solution, offers the following advantages over existing technologies: In-situ mass spectrometry enables direct analysis of chemical components in samples without complex sample pretreatment, offering advantages such as speed, sensitivity, and in-situ accuracy. Machine learning, on the other hand, possesses powerful data processing and pattern recognition capabilities, enabling the extraction of potential information and patterns from large amounts of data. Coupled with in-situ mass spectrometry, machine learning has led to the development of a rapid, accurate, and high-throughput method for identifying the authenticity of tea oil, providing strong technical support for the regulation and quality control of the tea oil market.

[0018] The beneficial effects of this invention are as follows: 1. High efficiency: No complicated pretreatment steps are required, and a single mass spectrometry analysis can be completed in just 1 minute, which greatly improves detection efficiency; 2. Accuracy: By combining a multi-adulteration system with multiple algorithm models, the generalization ability of the model is effectively improved, and a high accuracy rate can be achieved on the test set. 3. Universality: Applicable to the identification of tea oil and various adulterated vegetable oils (soybean oil, corn oil, etc.), covering a wide range of adulteration ratios of over 3%. Attached Figure Description

[0019] To further illustrate the present invention, the following description is provided in conjunction with the accompanying drawings: Figure 1 A diagram illustrating the implementation steps of the invention; Figure 2 The accuracy index values ​​for adulteration identification of five machine learning models with different processing methods; Figure 3 The results are for identifying samples with a low proportion of adulteration. Figure 4 The results are for the identification of complex multi-component adulterant systems. Detailed Implementation

[0020] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0021] To address the shortcomings of existing technologies, this invention proposes a method for identifying the authenticity of tea oil based on in-situ mass spectrometry-machine learning coupling technology. Through innovative feature extraction methods and multi-model integration strategies, it achieves rapid and accurate identification of adulterated tea oil.

[0022] The technical solution of the present invention includes the following steps: S1. Sample preparation: Collect pure tea oil samples and other vegetable oil samples from different origins, and mix the other vegetable oils with tea oil in different proportions to prepare binary, ternary, quaternary and quinary adulterated oil samples. S2. Mass Spectrometry Data Acquisition: An Orbitrap Exploris mass spectrometer equipped with a SICRIT ion source was used to perform full scan / data-dependent mass spectrometry analysis on the sample in positive ion mode to obtain raw mass spectrometry characteristic data. S3. Feature Extraction: Dynamically bin the mass spectrometry data, extract feature peak information, and aggregate time-dimensional features to form an initial feature matrix; S4. Data preprocessing: Perform data cleaning and transformation, data partitioning and amplification, and dimensionality reduction on the mass spectrometry data. Dimensionality reduction includes PCA (principal component analysis) and UMAP (uniform manifold approximation and projection). S5. Machine Learning Model Training: Five algorithms, namely Multilayer Perceptron (MLP), Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVM) and Extreme Gradient Boosting (XGB), are used to build a multi-model classification system. The model performance is verified through training and test sets and finally integrated into the software encapsulation module. S6. Intelligent identification and visualization output: Input the raw mass spectrometry data of the tea oil sample to be tested into the trained multi-model integration system, and output the identification results through weighted voting or optimal model strategy, and give the confidence interval to quantify the reliability of the results.

[0023] Furthermore, the other vegetable oils in step S1 are soybean oil, corn oil, sunflower seed oil, and olive oil.

[0024] Furthermore, in step S1, the adulteration ratio of the adulterated oil sample is 3%, 5%, and increases sequentially from 10% to 100% in increments of 10%. v / v Three parallel samples were prepared for each doping gradient.

[0025] Furthermore, in step S2, the mass spectrometry acquisition parameters are set as follows: scan range 75–1000 m / z, first-stage mass spectrometry resolution 60,000 FWHM, and second-stage mass spectrometry resolution 15,000 FWHM.

[0026] Furthermore, the dynamic binning process in step S3 includes: (1) A total of 14 m / z characteristic intervals were set for different statistical measures; (2) Dynamically divide each interval into bins, count the signal strength in each preset feature interval, and calculate the maximum value, mean, median, standard deviation and sum of statistical features based on the feature importance analysis results; (3) Combining the eight statistical features of the chromatogram, the feature vector dimension of each sample is "the sum of the bin feature numbers of all statistics + 8".

[0027] Furthermore, the data preprocessing in step S4 includes: (1) Data cleaning: Set all negative values ​​less than or equal to -1 to 0 to avoid NaN values ​​generated by subsequent logarithmic transformation; (2) Data transformation and normalization: Logarithmic transformation is performed on the cleaned data, followed by normalization; (3) Label conversion: The mass spectrometry data labels are unified into two categories (pure tea oil is 1, adulterated tea oil is 0). (4) Data set partitioning and augmentation: The training set (80%) and test set (20%) were partitioned by random shuffling and stratified sampling. The positive class samples in both the training set and the test set were replicated and augmented by 13 times to ensure that the two classes of samples were evenly distributed. (5) Dimensionality reduction: Dimensionality reduction is performed by PCA and UMAP respectively.

[0028] Furthermore, the training process of the machine learning model in step S5 includes: (1) MLP model: Construct a network structure that includes convolutional branches, MLP branches and attention mechanism, adopt Adam optimizer, loss function is binary cross-entropy, set early stopping strategy and learning rate decay mechanism, and monitor indicators including accuracy, AUC, precision and recall. (2) XGB, RF, SVM, LR models: Use default parameters or optimized parameters, save the model through joblib and output confusion matrix and classification report; (3) Model evaluation: All models were trained on the original features, PCA dimensionality reduction features, and UMAP dimensionality reduction features respectively. The performance was evaluated by confusion matrix, accuracy, precision, recall, F1 score and AUC value.

[0029] Furthermore, the identification result output mechanism in step S6 includes: (1) Input the original mass spectrometry data of the measured sample into the model; (2) When using model groups for identification, the prediction results of multiple models are integrated by majority voting or weighted average method to improve the reliability of identification.

[0030] Furthermore, the plurality of vegetable oils includes at least two of soybean oil, corn oil, sunflower seed oil, and olive oil.

[0031] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention. Example

[0032] 1. Sample Preparation: Samples of pressed pure camellia oil from major camellia oil producing areas were collected, along with soybean oil, corn oil, sunflower seed oil, and olive oil. Each variety of oil sample was mixed to prepare blended camellia oil, blended soybean oil, blended corn oil, blended sunflower seed oil, and blended olive oil. Then, the blended soybean oil, blended corn oil, blended sunflower seed oil, and blended olive oil were added at 3%, 5%, and 10% increments, increasing sequentially from 10% to 100%. v / v The adulterated tea oil was mixed with other plant oils in equal proportions, and adulterated samples of the binary system were prepared. Three samples were prepared for each adulteration gradient. Two or more other plant oils were mixed in equal proportions, and the adulteration was increased from 10% to 100% in gradients of 3%, 5%, and 10%. v / v The proportions of each component were mixed into pure tea oil to prepare ternary, quaternary, and quinary tea oil adulteration samples.

[0033] 2. Mass Spectrometry Data Acquisition: An Orbitrap Exploris 120 mass spectrometer equipped with a SICRIT ion source was used to acquire mass spectrometry data for tea oil, other vegetable oils, and adulterated oil samples. Each sample was acquired 10 times. Ion source parameters were set as follows: acquisition voltage 1,500 V, frequency 15,000 Hz. Mass spectrometry acquisition employed full scan / data-dependent mass spectrometry (Full MS / dd-MS2) mode, with a scan range of 75–1000 m / z. The primary mass spectrometry resolution was 60,000 FWHM, the secondary mass spectrometry resolution was 15,000 FWHM, and the ion transmission tube temperature was 320℃.

[0034] 3. Feature extraction: The acquired mass spectrometry data is dynamically binned, 14 feature regions are set, the signal intensity in each bin is statistically analyzed, and statistical features such as maximum value, mean, median, standard deviation and sum are calculated. At the same time, 8 chromatographic statistical features are extracted, and finally the feature vector dimension of each sample is 22.

[0035] 4. Data Preprocessing: The mass spectrometry data were labeled as binary (pure tea oil as 1, adulterated tea oil as 0). Random shuffling and stratified sampling were used to divide the data into training (80%) and test sets (20%). Positive samples in both the training and test sets were amplified 13-fold to alleviate class imbalance. All data underwent various feature processing, including raw data processing, PCA dimensionality reduction (5D), and UMAP dimensionality reduction (5D). Data normalization was performed using log1p transformation followed by MinMaxScaler normalization.

[0036] 5. Machine Learning Model Training: A multi-model classification system was constructed using five algorithms: Multilayer Perceptron (MLP), Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVM), and Extreme Gradient Boosting (XGB). The MLP model was built using the Adam optimizer with a binary cross-entropy loss function, and early stopping and learning rate decay were implemented. The XGB, RF, SVM, and LR models used default or optimized parameters and were saved via joblib. All models were trained on the original features, PCA-reduced features, and UMAP-reduced features, respectively. Performance was evaluated using the confusion matrix, accuracy, precision, recall, F1 score, and AUC value, and finally integrated into the software encapsulation module.

[0037] 6. The accuracy index values ​​for adulteration detection of five machine learning models using different dimensionality reduction methods are as follows: Figure 2 As shown in the figure, in the anti-spoofing models, MLP and RF consistently demonstrate stable and efficient accuracy across various dimensionality reduction scenarios. With no dimensionality reduction (NON), the training set accuracy reaches 100%, and the test set accuracy is 99.28% and 98.43%, respectively. PCA maintains outstanding performance even after dimensionality reduction (test set accuracy is 98.07% and 97.87%), exhibiting excellent adaptability and discrimination efficiency, making it a reliable first choice. XGB maintains high performance under NON, PCA, and UMAP dimensionality reduction. Without dimensionality reduction, the test set accuracy is 99.40%, and after PCA and UMAP dimensionality reduction, it still reaches 98.23% and 97.79%, respectively, with stable core metrics, making it a strong alternative. SVM is sensitive to dimensionality reduction; after UMAP dimensionality reduction, the test set accuracy drops sharply to 72.08%, requiring optimization of dimensionality reduction adaptation or fine-tuning of parameters. LR's accuracy drops to 87.46% after UMAP dimensionality reduction, necessitating improvements in feature learning adaptability. In practical applications, to ensure high discrimination accuracy, MLP, RF, and XGB are preferred in combination with NON or PCA dimensionality reduction schemes to give full play to the advantages of model and feature adaptation.

[0038] 7. The classification performance of five machine learning models with different dimensionality reduction methods on adulterated samples is as follows: Figure 1As shown, in the pure tea oil scenario, SVM and LR exhibit more misclassifications, while MLP, RF, and XGB demonstrate superior robustness without dimensionality reduction or with PCA dimensionality reduction. In the medium adulteration range, MLP, XGB, and RF outperform SVM and LR in misclassification control, with PCA dimensionality reduction contributing to stable performance. At high adulteration levels (80%-100%), all models show reduced misclassifications, with MLP and XGB demonstrating outstanding performance by enhancing feature discrimination through PCA or UMAP. In summary, MLP, XGB, and RF exhibit more reliable adulteration detection performance across different dimensionality reduction and adulteration scenarios.

[0039] Binary system: 10 samples each of adulterated soybean oil (3%), olive oil (3%), corn oil (5%), and sunflower seed oil (5%). The ternary system consists of 10 parts each of 3% (soybean oil + corn oil, 1:1) and 5% (soybean oil + olive oil, 1:1).

[0040] 2. Model Selection The results of three parallel experiments were analyzed by combining the RF model (nestimators=100) with PCA dimensionality reduction. 3. Using the identification results as follows Figure 3 As shown: It can be seen that the RF model combined with PCA dimensionality reduction exhibits good performance in identifying adulterated edible oils with low proportions (3%-5%). In the binary system, the correct identification count for adulterated samples of 3% soybean oil, 3% olive oil, 5% corn oil, and 5% sunflower oil was 10, achieving an accuracy of 100%, with a stable confidence interval of 98.5%-100%. This indicates that the method is accurate and reliable when identifying adulteration of a single oil type at low proportions. In the ternary system, the accuracy of identifying adulterated samples of 5% (soybean oil + olive oil, 1:1) was 100%, comparable to the binary system. The accuracy of adulterated samples of 3% (soybean oil + corn oil, 1:1) was 90%, slightly lower but still at a high level, indicating that the method also has a certain ability to identify adulterated mixed oil types, and the identification effect is better as the adulteration proportion increases. Overall, this model combination can be effectively applied to the accurate identification of adulterated edible oils with low proportions, providing technical support for the quality supervision of edible oils.

[0041] Five-element system: 10% (soybean oil + corn oil + sunflower seed oil + olive oil, 1:1:1:1) 10 parts; Real blind samples: 10 known real oil samples.

[0042] 2. Model Selection Multiple models (MLP+RF+XGB) were used for prediction, and the results of three parallel experiments were analyzed comprehensively. The identification results are as follows Figure 4 As shown: It can be seen that the MLP+RF+XGB multi-model approach demonstrates excellent identification capabilities for complex multi-component adulteration systems. With 10 samples of a five-component blended oil (soybean oil, corn oil, sunflower oil, and olive oil in a 1:1:1:1 ratio, with an adulteration rate of 10%), the accuracy rate reached 100%, with an average confidence interval of 99.5±0.3%. This indicates that the multi-model combination can accurately identify adulteration in complex multi-component blended oils and effectively capture the differences in the characteristics of the multi-component components. With 10 real blind samples, the accuracy rate reached 90% (after hypothesis correction), with an average confidence interval of 99.8±0.2%, demonstrating that the model also possesses high reliability in identifying real, unknown samples. The multi-model fusion strategy provides an efficient and practical technical path for identifying adulteration in complex multi-component oils, and can strongly support adulteration identification in complex scenarios in oil quality supervision.

[0043] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claims.

Claims

1. A method for identifying the authenticity of tea oil based on in-situ mass spectrometry-machine learning coupling technology, characterized in that, Includes the following steps: S1. Sample preparation: Collect pure tea oil samples and other vegetable oil samples from different origins, and mix the other vegetable oils with tea oil in different proportions to prepare binary, ternary, quaternary and quinary adulterated oil samples. S2. Mass Spectrometry Data Acquisition: An Orbitrap Exploris mass spectrometer equipped with a SICRIT ion source was used to perform full scan / data-dependent mass spectrometry analysis on the sample in positive ion mode to obtain raw mass spectrometry characteristic data. S3. Feature Extraction: Dynamically bin the mass spectrometry data, extract feature peak information, and aggregate time-dimensional features to form an initial feature matrix; S4. Data preprocessing: Perform data cleaning and transformation, data partitioning and amplification, and dimensionality reduction on the mass spectrometry data. Dimensionality reduction includes PCA (principal component analysis) and UMAP (uniform manifold approximation and projection). S5. Machine Learning Model Training: Five algorithms, namely Multilayer Perceptron (MLP), Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVM) and Extreme Gradient Boosting (XGB), are used to build a multi-model classification system. The model performance is verified through training and test sets and finally integrated into the software encapsulation module. S6. Intelligent identification and visualization output: Input the raw mass spectrometry data of the tea oil sample to be tested into the trained multi-model integration system, and output the identification results through weighted voting or optimal model strategy, and give the confidence interval to quantify the reliability of the results.

2. The method for identifying the authenticity of tea oil based on in-situ mass spectrometry-machine learning coupling technology as described in claim 1, characterized in that, The other vegetable oils mentioned in step S1 are any one or more of soybean oil, corn oil, sunflower seed oil, and olive oil. The binary refers to tea oil mixed with any one of the above samples, the ternary refers to tea oil mixed with any two of the above samples, the quaternary refers to tea oil mixed with any three of the above samples, and the pentagonal refers to tea oil mixed with any four of the above samples.

3. The method for identifying the authenticity of tea oil based on in-situ mass spectrometry-machine learning coupling technology as described in claim 1, characterized in that, The adulteration ratio of the adulterated oil samples in step S1 is 3%, 5%, and increases sequentially from 10% to 100% in increments of 10%. v / v Three parallel samples were prepared for each doping gradient; the mass spectrometry acquisition parameters in step S2 were set as follows: scan range 75–1000 m / z, first-order mass spectrometry resolution 60,000 FWHM, and second-order mass spectrometry resolution 15,000 FWHM.

4. The method for identifying the authenticity of tea oil based on in-situ mass spectrometry-machine learning coupling technology as described in claim 1, characterized in that, The dynamic binning process described in step S3 includes: (1) A total of 14 m / z characteristic intervals were set for different statistical measures; (2) Dynamically divide each interval into bins, count the signal strength in each preset feature interval, and calculate the maximum value, mean, median, standard deviation and sum of statistical features based on the feature importance analysis results; (3) Combining the eight statistical features of the chromatogram, the feature vector dimension of each sample is "the sum of the bin feature numbers of all statistics + 8".

5. The method for identifying the authenticity of tea oil based on in-situ mass spectrometry-machine learning coupling technology as described in claim 1, characterized in that, The data preprocessing in step S4 includes: (1) Data cleaning: Set all negative values ​​less than or equal to -1 to 0 to avoid NaN values ​​generated by subsequent logarithmic transformation; (2) Data transformation and normalization: Logarithmic transformation is performed on the cleaned data, followed by normalization; (3) Label conversion: The mass spectrometry data labels are unified into two categories (pure tea oil is 1, adulterated tea oil is 0). (4) Data set partitioning and augmentation: The training set (80%) and test set (20%) were partitioned by random shuffling and stratified sampling. The positive class samples in both the training set and the test set were replicated and augmented by 13 times to ensure that the two classes of samples were evenly distributed. (5) Dimensionality reduction: Dimensionality reduction is performed by PCA and UMAP respectively.

6. The method for identifying the authenticity of tea oil based on in-situ mass spectrometry-machine learning coupling technology as described in claim 1, characterized in that, The training process of the machine learning model described in step S5 includes: (1) MLP model: using Adam optimizer, loss function is binary cross-entropy, early stopping strategy and learning rate decay mechanism are set; (2) XGB, RF, SVM, LR models: Use default parameters or optimized parameters, save the model through joblib and output confusion matrix and classification report; (3) Model evaluation: All models were trained on the original features, PCA dimensionality reduction features, and UMAP dimensionality reduction features respectively. The performance was evaluated by confusion matrix, accuracy, precision, recall, F1 score and AUC value.

7. The method for identifying the authenticity of tea oil based on in-situ mass spectrometry-machine learning coupling technology as described in claim 1, characterized in that, The identification result output mechanism in step S6 includes: (1) Perform three independent mass spectrometry acquisitions on the sample to be tested, generate three sets of feature data and input them into the model group respectively. If the prediction is "adulterated" more than twice, it is determined to be adulterated. (2) When using model groups for identification, the prediction results of multiple models are integrated by majority voting or weighted average method to improve the reliability of identification.

8. A method for identifying the authenticity of tea oil based on in-situ mass spectrometry-machine learning coupling technology according to any one of claims 1-8, characterized in that, The various vegetable oils include at least two of soybean oil, corn oil, sunflower seed oil, and olive oil, and a complex adulteration scenario sample library is constructed by mixing multiple components.

9. The method for identifying the authenticity of tea oil based on in-situ mass spectrometry-machine learning coupling technology as described in claim 1, characterized in that, Choose MLP, RF, or XGB combined with NON or PCA dimensionality reduction schemes to fully leverage the advantages of model and feature adaptation.

10. The method for identifying the authenticity of tea oil based on in-situ mass spectrometry-machine learning coupling technology as described in claim 1, characterized in that, For accurate identification of low-proportion adulteration (3%-5%), the RF model (nestimators=100) + PCA dimensionality reduction combination was selected. For the identification of complex multi-component adulteration systems, a multi-model prediction method (MLP+RF+XGB) is adopted.

Images