Vegetable oil detection method based on combination of DESI-MS and neural network

Abstract

The invention discloses a vegetable oil detection method based on the combination of DESI-MS and a neural network. The vegetable oil detection method comprises the following steps: sample collection, DESI-MS analysis, lipid fingerprint collection, analysis and evaluation based on the neural network, sample detection completion and sample type confirmation. The invention establishes a lipid fingerprint analysis method based on DESI-MS combined with an SCNN machine learning model, and the method is specially used for analysis of glyceride and realizes high-precision classification and identification of various vegetable oils. According to the method, the SCNN model and the DESI-MS technology are combined and optimized, so that the method is specially used for detecting the edible vegetable oil, and the method has the advantages of convenience and rapidness in detection and high detection precision.

Description

Technical Field

[0001] The invention relates to the field of detection technology, in particular to a vegetable oil detection method based on the combination of DESI-MS and neural network. Background Art

[0002] Different types of edible vegetable oils have significant differences in nutritional content and market value due to their different oilseed sources. Some unscrupulous vendors engage in fraudulent practices such as partial adulteration or complete substitution for economic gain. These practices not only pose a major threat to consumer health, but also undermine their economic interests. Therefore, it is crucial to develop reliable methods for the classification and identification of vegetable oils.

[0003] As a pioneer of ambient ionization methods, DESI-MS has been demonstrated to be a powerful tool for food analysis and classification. This method has several significant advantages: (1) it operates at atmospheric pressure, requiring little or no sample pretreatment; (2) its soft ionization mode preserves the structural integrity of the analytes; and (3) it is capable of rapid analysis, with detection times as short as a few seconds, thus facilitating high-throughput analysis.

[0004] Lipids, especially triacylglycerols (TGs), are key components of vegetable oils and form the basis for their classification and identification. However, the low ionization efficiency of non-polar TGs is a major challenge for their sensitive and accurate analysis using DESI-MS. Current DESI-MS methods for edible oil analysis are limited and the results are unsatisfactory.

[0005] Combining mass spectrometry with machine learning techniques has become a promising strategy to improve the accuracy of food analysis classification and identification. Convolutional neural network (CNN), a widely used deep learning model, performs well in data processing. Currently, various analytical chemistry techniques combined with CNN models are used for classification and identification of a variety of foods in other fields.

[0006] Although CNN models show significant potential and advantages over traditional chemometric methods, there is still a lack of research on their application in DESI-MS-based food testing, especially in the classification of edible vegetable oils.

[0007] In summary, there is still no effective detection method that can classify and identify edible vegetable oils conveniently and quickly. Summary of the invention

[0008] The purpose of the present invention is to provide a vegetable oil detection method based on the combination of DESI-MS and neural network. The present invention combines and optimizes the SCNN model with the DESI-MS technology so as to be specifically used for the detection of edible vegetable oils, and has the advantages of convenient and fast detection and high detection accuracy.

[0009] The technical solution of the present invention is a method for detecting vegetable oil based on the combination of DESI-MS and neural network, comprising the following process:

[0010] A. Collecting samples: Take 5 μL of sample onto a designated 1×1 cm square area on a glass slide to form a thin liquid film to be analyzed;

[0011] B. DESI-MS analysis: Perform DESI-MS analysis on the collected samples to obtain DESI-MS data;

[0012] C. Lipid fingerprint collection: lipid fingerprint of samples collected based on DESI-MS data;

[0013] D. Comparative classification based on neural network analysis: The lipid fingerprint of the collected samples is input into a shallow convolutional neural network for processing and output of the results;

[0014] E. Complete sample testing and confirm sample type.

[0015] In the aforementioned vegetable oil detection method based on the combination of DESI-MS and neural network, the DESI-MS analysis described in step B is performed using a quadrupole time-of-flight mass spectrometer, and the specific contents are as follows:

[0016] B1, calibration was performed using rhodamine 6G in positive ionization mode at m / z 443.2335;

[0017] B2, spraying the spray solvent onto the sample through the spray capillary at a flow rate of 1-5 μL / min, and the voltage of the spray capillary is 2-5 kV;

[0018] B3, extracting the ionized target molecules in the sample and transferring them to the mass spectrometer;

[0019] B5. Place the DESI probe on the glass slide away from the sample and record the background signal for 1.0 min;

[0020] B5. Full scan was performed in the mass range of m / z 50-1200 in positive ionization mode to obtain mass spectrometry data at a scan rate of 1 s / scan. The instrument parameters were spray impact angle 75°, distance between nozzle and glass slide surface 2 mm, distance between nozzle and ion transfer capillary hole 6 mm, distance between ion transfer capillary hole and glass slide surface 0.5 mm, capillary temperature 150 °C, nitrogen pressure 600 kPa, cone voltage 40 V.

[0021] B6. Aim the probe at the center of the oil film and collect mass spectra for 1.5-2.0 minutes;

[0022] B7, repeat the collection three times;

[0023] B8. Obtain DESI-MS data.

[0024] The aforementioned vegetable oil detection method based on the combination of DESI-MS and neural network is characterized in that:

[0025] The spray solvent is MeOH, AcN, tol in a ratio of 10:7:3 (v / v / v).

[0026] In the aforementioned vegetable oil detection method based on the combination of DESI-MS and neural network, an adjustment bracket for adjusting the angle and position is also provided; the adjustment bracket is detachably arranged in the mass spectrometer; it includes a seat body, a capillary support unit arranged on the upper left side of the seat body, an inlet tube support unit arranged on the right side of the seat body, a sample platform arranged above the middle of the seat body, and a fence surrounding the sample platform; the capillary support unit can raise and lower to adjust the height of the spray capillary and adjust the angle; the inlet tube support unit can move laterally to adjust the position of the inlet tube; the fence is provided with corresponding operating ports for the spray capillary and the inlet tube, and an openable and closable block is provided at the operating port; the sample platform can be rotatably arranged on the seat body.

[0027] In the aforementioned vegetable oil detection method based on the combination of DESI-MS and neural network, the capillary support unit includes a lifting column, a lifting platform with a lifting sleeve mounted on the lifting column, an adjustment block adjustably connected in front of the lifting platform, a connecting rod arranged on the adjustment block and a capillary mounting seat adjustably arranged on the top of the connecting rod; a lifting motor is arranged in front of the lifting column, and a threaded screw linked to the lifting motor is arranged in the lifting column; the lifting platform is linked to the threaded screw.

[0028] In the aforementioned vegetable oil detection method based on the combination of DESI-MS and neural network, the inlet pipe support unit includes a transverse seat, a transverse column that can be transversely moved on the transverse seat, and an inlet pipe mounting seat that can be adjustably connected to the transverse column; a transverse motor is provided in front of the transverse seat, and a transverse screw rod linked to the transverse motor is provided in the transverse seat; the transverse column is linked to the transverse screw rod.

[0029] In the aforementioned plant oil detection method based on the combination of DESI-MS and neural network, the lipid fingerprint collection described in step C is specifically as follows:

[0030] DESI-MS data were processed by Masslynx v4.1 software for background subtraction and mass drift correction;

[0031] Data normalization was performed using a normalization factor (NL), which was defined as the maximum peak intensity in each mass spectrum;

[0032] The glycerolipid species of the target peaks were identified based on the ionization behavior of triolein standards and the LIPIDMAPS prediction tool database.

[0033] Fatty acyl chains are expressed in the format TC:DB, where TC is the total number of carbon atoms and DB is the number of double bonds;

[0034] The characteristic percentages of the samples were calculated by peak area normalization;

[0035] SPSS23.0 software was used for statistical analysis, including mean, standard deviation and one-way analysis of variance;

[0036] Chemical structures and accurate molecular masses were generated using ChemDraw Professional 16.0 software;

[0037] Generate UpSet graphs and heat maps through TBtools software visualization;

[0038] The mass spectrum similarity index (SF) is used to measure the spectral similarity between any two oils and indirectly evaluate the similarity of their lipid components. The SF formula is:

[0039]

[0040] The mass spectrum consists of i m / z values, x i and i are the relative abundances of the ions corresponding to the i-th m / z value in the spectrum.

[0041] In the aforementioned vegetable oil detection method based on the combination of DESI-MS and neural network, the neural network analysis and evaluation described in step D has the following specific steps:

[0042] The mass spectrometry data matrix was used as a data set for machine learning analysis and a SCNN model was constructed. The data set was divided into a training set and a validation set. The training set accounted for 70% of the samples and the validation set accounted for 30% of the samples.

[0043] The SCNN model consists of a 15-layer structure, including an input layer, two convolutional layers, a pooling layer, a fully connected layer, a batch normalization layer, a ReLU layer, a Dropout layer, and a classification output layer;

[0044] The input variables of the training set and the test set are standardized to 1×53×1; the two convolutional layers are used to automatically extract key spectral feature information in the mass spectrometry data matrix;

[0045] The input variables first enter the first convolutional layer, which is configured with 16 convolution kernels, and then enter the second convolutional layer, which is configured with 32 convolution kernels; both convolutional layers enhance the stability of data distribution through batch normalization layers; after each convolutional layer, the pooling layer eliminates redundancy and reduces the risk of overfitting when extracting spectral feature information; the pooling layer uses an average pooling algorithm; the fully connected layer includes two hidden layers, containing 32 and 16 neurons respectively, and a regularization step is applied after the first hidden layer, with a dropout rate of 20%; the classification output layer is set to 9 neurons, corresponding to 9 types of vegetable oils;

[0046] In the SCNN model, the activation function is used to approximate any nonlinear function and fit various nonlinear models; the output layer uses the Softmax activation function, and the remaining layers use the ReLU activation function.

[0047] Compared with the prior art, the lipid fingerprint analysis method based on DESI-MS in the present invention is combined with the SCNN machine learning model, thereby achieving high-precision classification and identification of a variety of vegetable oils;

[0048] This application significantly enhances the signal intensity of DESI-MS by optimizing the spray solvent, flow rate and capillary conditions, and establishes a new method based on the ion characteristics of glycerol ionization.

[0049] [TG+Na]+, [TG+NH 4 ]+、[TG+K]+、[DG+HH 2 O]+ and [DG+Na]+ as the main ion types, which provides favorable conditions for lipid fingerprint analysis of plant oils;

[0050] The DESI-MS method of the present application can generate unique lipid fingerprints, and these lipid fingerprints correspond to the glycerolipid phenotypes across 9 kinds of vegetable oils (including corn oil, soybean oil, peanut oil, sesame oil, rice bran oil, sunflower oil, camellia oil, olive oil and walnut oil). The lipid fingerprint detected by the present application can be used to analyze and determine the type of vegetable oil;

[0051] Based on the DESI-MS method, this application established an SCNN model targeting the unique characteristics of the data set, optimized the network layers and hyperparameters, and performed rigorous cross-validation. The model showed a very low misclassification rate, with an accuracy of 98.5±2.2% for the training set and 97.4±3.1% for the test set, further improving the accuracy of detection;

[0052] By adopting this application, edible vegetable oils can be accurately, efficiently, stably and with high throughput to determine their true categories, providing a favorable technical means for the market to quickly identify whether vegetable oils are counterfeit, and further enhancing food safety.

[0053] In summary, the present invention combines and optimizes the CNN model with the DESI-MS technology so that it is specifically used for the detection of edible vegetable oils, and has the advantages of convenient and fast detection and high detection accuracy.

[0054] The present application sets an adjustment bracket, which includes a seat body, a sample platform, a capillary support unit and an inlet tube support unit. The sample platform can be rotatably set, the capillary support unit can be raised and lowered to adjust the capillary angle, and the inlet tube support unit can be laterally moved to adjust the position of the inlet tube;

[0055] The present application can better adjust the position and angle of the inlet tube of the capillary machine according to actual conditions to meet the detection needs. At the same time, by setting a rotatable sample table, the glass slide placed on it can be rotated, and the samples can be collected and analyzed from multiple angles, which is more comprehensive. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 It is a diagram of the SCNN model architecture based on DESI-MS data of the present invention;

[0057] Figure 2 It is a schematic diagram of the structure of the adjusting bracket in the present invention.

[0058] Figure 3 : is a comparison chart of the DESI-MS spray solvent flow rate in the embodiment; a, b, c indicate that the signal-to-noise ratio of m / z 887.73 is significantly different (p<0.05), and different letters in A, B, C, and D indicate that the signal-to-noise ratio of m / z 901.73 is significantly different (p<0.05);

[0059] Figure 4 : is a comparison diagram of DESI-MS capillary voltage in the embodiment; a, b, c indicate that the signal-to-noise ratio of m / z 887.73 is significantly different (p<0.05), and different letters in A, B, C, and D indicate that the signal-to-noise ratio of m / z 901.73 is significantly different (p<0.05);

[0060] Figure 5 It is the DESI-MS spectrum diagram in the embodiment.

[0061] Figure 6 is the DESI-MS spectrum of the triolein standard in the positive ionization state in the embodiment;

[0062] Figure 7 are nine kinds of plant oil lipid fingerprints and corresponding physical photos in the embodiments; wherein (B) is corn oil, (C) is soybean oil, (D) is peanut oil, (E) is sesame oil, (F) is rice bran oil, (G) is sunflower oil, (H) is camellia oil, (I) is olive oil and (J) is walnut oil;

[0063] Figure 8 is the UpSet graph of nine vegetable oils in the embodiment;

[0064] Fig. 9 is a hierarchical clustering heat map of nine vegetable oils in the embodiment;

[0065] Fig.10 is the mass spectrum similarity index of the nine vegetable oils in the embodiment;

[0066] Fig.11 : is the accuracy and loss curve of the SCNN model training set in the embodiment;

[0067] Fig.12 : is the classification result (B) and confusion matrix (C) of the SCNN model training set in the embodiment;

[0068] Fig.13 : is the classification result (D) and confusion matrix (E) of the SCNN model test set in the embodiment;

[0069] Fig.14 It is the ROC-AUC graph of the SCNN model in the embodiment.

[0070] Markings in the accompanying drawings: 1-seat body, 11-gasket, 2-lifting column, 21-lifting motor, 3-lifting platform, 31-threaded screw, 4-adjusting block, 41-connecting rod, 42-capillary mounting seat, 5-transverse seat, 51-transverse motor, 6-transverse column, 61-inlet tube mounting seat, 7-, 8-enclosure, 9-sample platform. DETAILED DESCRIPTION

[0071] The present invention is further described below in conjunction with the accompanying drawings and embodiments, but they are not intended to limit the present invention.

[0072] Embodiment. A method for detecting vegetable oil based on the combination of DESI-MS and neural network includes the following process:

[0073] A. Collecting samples: Take 5 μL of sample onto a designated 1×1 cm square area on a glass slide to form a thin liquid film to be analyzed;

[0074] B. DESI-MS analysis: Perform DESI-MS analysis on the collected samples to obtain DESI-MS data;

[0075] C. Lipid fingerprint collection: lipid fingerprint of samples collected based on DESI-MS data;

[0076] D. Comparative classification based on neural network analysis: The lipid fingerprint of the collected samples is input into a shallow convolutional neural network for processing and output of the results;

[0077] E. Complete sample testing and confirm sample type.

[0078] The DESI-MS analysis described in step B is performed using a quadrupole time-of-flight mass spectrometer, and the specific contents are as follows:

[0079] B1, calibration was performed using rhodamine 6G in positive ionization mode at m / z 443.2335;

[0080] B2, spraying the spray solvent onto the sample through the spray capillary at a flow rate of 1-5 μL / min, and the voltage of the spray capillary is 2-5 kV;

[0081] B3, extracting the ionized target molecules in the sample and transferring them to the mass spectrometer;

[0082] B5. Place the DESI probe on the glass slide away from the sample and record the background signal for 1.0 min.

[0083] B5. Full scan was performed in the mass range of m / z 50-1200 in positive ionization mode to obtain mass spectrometry data at a scan rate of 1 s / scan. The instrument parameters were spray impact angle 75°, distance between nozzle and glass slide surface 2 mm, distance between nozzle and ion transfer capillary hole 6 mm, distance between ion transfer capillary hole and glass slide surface 0.5 mm, capillary temperature 150 °C, nitrogen pressure 600 kPa, cone voltage 40 V.

[0084] B6. Aim the probe at the center of the oil film and collect mass spectra for 1.5-2.0 minutes;

[0085] B7, repeat the collection three times;

[0086] B8. Obtain DESI-MS data.

[0087] The spray solvent is MeOH, AcN, tol in a ratio of 10:7:3 (v / v / v).

[0088] like Figure 2 As shown, an adjustment bracket for adjusting the angle and position is provided; the adjustment bracket is detachably arranged in the mass spectrometer, comprising a base 1, a capillary support unit arranged on the upper left side of the base 1, an inlet tube support unit arranged on the right side of the base 1, a sample platform 9 arranged above the middle of the base 1, and an enclosure 8 surrounding the sample platform 9; the capillary support unit can be raised and lowered to adjust the height of the spray capillary and adjust the angle; the inlet tube support unit can be laterally moved to adjust the position of the inlet tube; the enclosure 8 is provided with corresponding operating ports for the spray capillary and the inlet tube, and an openable and closable block is provided at the operating port; the sample platform 9 can be rotatably arranged on the base 1.

[0089] The capillary support unit includes a lifting column 2, a lifting platform 3 that is lifted and sleeved on the lifting column 2, an adjustment block 4 that is adjustably connected to the front of the lifting platform 3, a connecting rod 41 that is arranged on the adjusting block 4, and a capillary mounting seat 42 that is adjustably arranged on the top of the connecting rod 41; a lifting motor 21 is arranged in front of the lifting column 2, and a threaded screw 31 that is linked to the lifting motor 21 is arranged in the lifting column 2; the lifting platform 3 is linked to the threaded screw 31.

[0090] The inlet pipe support unit includes a transverse seat 5, a transverse column 6 which can be transversely moved on the transverse seat 5, and an inlet pipe mounting seat 61 which can be adjustably connected to the transverse column 6; a transverse motor 51 is provided in front of the transverse seat 5, and a transverse screw rod which is linked to the transverse motor 51 is provided inside the transverse seat 5; the transverse column 6 is linked to the transverse screw rod.

[0091] The lipid fingerprint collection described in step C is specifically as follows:

[0092] DESI-MS data were processed by Masslynx v4.1 software for background subtraction and mass drift correction;

[0093] Data normalization was performed using a normalization factor (NL), which was defined as the maximum peak intensity in each mass spectrum;

[0094] The glycerolipid species of the target peaks were identified based on the ionization behavior of triolein standards and the LIPIDMAPS prediction tool database.

[0095] Fatty acyl chains are expressed in the format TC:DB, where TC is the total number of carbon atoms and DB is the number of double bonds;

[0096] The characteristic percentages of the samples were calculated by peak area normalization;

[0097] SPSS23.0 software was used for statistical analysis, including mean, standard deviation and one-way analysis of variance;

[0098] Chemical structures and accurate molecular masses were generated using ChemDraw Professional 16.0 software;

[0099] UpSet graphs and heat maps were generated by TBtools software visualization;

[0100] The mass spectrum similarity index (SF) is used to measure the spectral similarity between any two oils and indirectly evaluate the similarity of their lipid components. The SF formula is:

[0101]

[0102] The mass spectrum consists of i m / z values, x i andi are the relative abundances of the ions corresponding to the i-th m / z value in the spectrum.

[0103] The specific steps of the neural network analysis and evaluation described in step D are as follows:

[0104] The mass spectrometry data matrix was used as a data set for machine learning analysis and a SCNN model was constructed. The data set was divided into a training set and a validation set. The training set accounted for 70% of the samples and the validation set accounted for 30% of the samples.

[0105] The SCNN model consists of a 15-layer structure, including an input layer, two convolutional layers, a pooling layer, a fully connected layer, a batch normalization layer, a ReLU layer, a Dropout layer, and a classification output layer;

[0106] The input variables of the training set and the test set are standardized to 1×53×1; the two convolutional layers are used to automatically extract key spectral feature information in the mass spectrometry data matrix;

[0107] The input variables first enter the first convolutional layer, which is configured with 16 convolution kernels, and then enter the second convolutional layer, which is configured with 32 convolution kernels; both convolutional layers enhance the stability of data distribution through batch normalization layers; after each convolutional layer, the pooling layer eliminates redundancy and reduces the risk of overfitting when extracting spectral feature information; the pooling layer uses an average pooling algorithm; the fully connected layer includes two hidden layers, containing 32 and 16 neurons respectively, and a regularization step is applied after the first hidden layer, with a dropout rate of 20%; the classification output layer is set to 9 neurons, corresponding to 9 types of vegetable oils;

[0108] In the SCNN model, the activation function is used to approximate any nonlinear function and fit various nonlinear models; the output layer uses the Softmax activation function, and the remaining layers use the ReLU activation function.

[0109] The performance of the SCNN model is improved through the three indicators of recall rate, precision rate and accuracy rate. The specific contents are as follows:

[0110]

[0112] Experimental example

[0113] The effectiveness of the technology of the present invention is verified by conducting comparative experiments

[0114] 1. Prepare chemical reagents

[0115] Chromatographic grade methanol (MeOH), acetonitrile (AcN), toluene (tol), and dichloromethane (DCM) were purchased from Merck Life Science (Merck KGaA, Darmstadt, Germany). Triolein standards were purchased from Avanti Polar Lipids (Alabaster, AL, USA). The mass calibration sample rhodamine 6G was provided by Waters Corporation (Beijing, China). High-purity water with a resistivity of 18.2 MΩ·cm was provided by a Milli-Q water system (Millipore, Bedford, MA, USA).

[0116] 2. Sample collection

[0117] This assay was validated for analysis of nine different vegetable oils, including corn oil, soybean oil, peanut oil, sesame oil, rice bran oil, sunflower oil, camellia oil, olive oil, and walnut oil. All samples were purchased from reputable suppliers, and most brands are ISO 9001 and / or ISO 22000 certified to ensure product quality.

[0118] To verify the authenticity of the samples, lipid testing and principal component analysis (PCA) were performed. The PCA score plot showed significant clustering between oils of the same biological origin, indicating good authenticity. For brands that currently have no certification information available, such as Leto Krasno, Oh Chin Hing, Tsuno, Cookmod, Calena, Bizce, and Grandpa's Farm, they were also confirmed to be authentic samples.

[0119] A detailed list of samples, including brand name, geographical origin, extraction method, and number of samples, is shown in Table 1. The number of samples is expressed as “number of batches × number of replicates”. To minimize the influence of intraspecific differences on sample classification, 4–5 brands of each oil were purchased. In addition, sesame oil, rice bran oil, sunflower oil, olive oil, and walnut oil even came from different geographical regions. Each brand provided 2–3 batches of samples, and each batch was subjected to 3 technical replicates. All samples were quickly stored at 4 °C and returned to room temperature before testing.

[0120] Table 1. Information of vegetable oils used for DESI-MS analysis

[0121]

[0122]

[0123] 3. DESI-MS analysis

[0124] Triolein was used as a standard substance to study the ionization behavior of glycerolipids under experimental conditions. 5 μL of each sample was pipetted onto a designated 1 × 1 cm square area on a glass slide to form a thin liquid film for analysis without additional pretreatment.

[0125] The lipid fingerprints of all samples were obtained using a DESI source (Waters, Beijing) connected to a quadrupole time-of-flight mass spectrometer (Xevo G2-XS, Waters, Milford, USA).

[0126] The spray solvent is sprayed onto the sample through the spray capillary at a flow rate of 1-5μL / min. The capillary voltage is adjusted in the range of 2-5kV to ensure stable signal generation. The sample surface is bombarded by electrosprayed charged solvent droplets, and the target molecules are extracted and ionized, and then transferred to the mass spectrometer through secondary droplets. Mass spectral data are acquired in a full scan in the mass range of m / z 50-1200 in positive ionization mode with a scan rate of 1s / scan. The instrument parameters include: spray impact angle 75°, distance between nozzle and slide surface 2mm, distance between nozzle and ion transfer capillary hole 6mm, and distance between ion transfer capillary hole and slide surface 0.5mm. Capillary temperature 150℃, nitrogen pressure 600kPa, cone voltage 40V.

[0127] Prior to analysis, the instrument was calibrated using rhodamine 6G (m / z 443.2335 in positive ionization).

[0128] At the beginning of sample acquisition, the DESI probe was placed on the glass slide away from the sample to record the background signal for 1 minute. The probe was then aimed at the center of the oil film and mass spectra were collected for 1.5-2.0 minutes. Three technical replicates were performed for each sample.

[0129] To avoid cross contamination, the ion inlet tube was cleaned with water and MeOH in turn after each test. In addition, when switching between different oils, the ion inlet tube was ultrasonically cleaned with AcN / MeOH (1:1, v / v), water, and MeOH in turn.

[0130] 4. Fingerprint collection, processing and statistical analysis

[0131] DESI-MS data were processed by Masslynx v4.1 software for background subtraction and mass drift correction. Data normalization used a normalization factor (NL), which was defined as the maximum peak intensity in each mass spectrum. Ions with peak intensities less than 5% of NL were excluded from further analysis because of their negligible contribution to lipidomics analysis and classification model development. The glycerolipid species of the target peaks were identified based on the ionization behavior of triolein standards and the LIPIDMAPS prediction tool database (access date: November 7, 2024). Fatty acyl chains are represented in TC:DB format, where TC is the total number of carbon atoms and DB is the number of double bonds. The resulting data matrix contained 53 features (m / z intervals) and 270 observations (samples). The feature percentage for each sample was calculated by peak area normalization. Statistical analysis, including mean, standard deviation, and one-way ANOVA, was performed using SPSS23.0 software. Chemical structures and accurate molecular masses were generated using ChemDraw Professional16.0 software. UpSet graphs and heatmaps were visualized by TBtools software. The mass spectrum similarity index (SF) was used to measure the spectral similarity between any two oils, thereby indirectly assessing the similarity of their lipid compositions.

[0132] 5. Shallow Convolutional Neural Network Modeling and Evaluation

[0133] The above data matrix is ​​used for machine learning analysis, and the data set is divided into a training set (accounting for 70% of the samples, equivalent to 9 types of oil × 21 sets of fingerprint data) and a validation set (accounting for 30% of the samples, equivalent to 9 types of oil × 9 sets of fingerprint data). Due to the limited size of the data set, a shallow convolutional neural network (SCNN) was specially designed and implemented in MATLAB 2022a.

[0134] The SCNN model consists of 15 layers, including input layer, convolution layer, pooling layer, fully connected layer, batch normalization layer, ReLU layer, Dropout layer and classification output layer. The network structure is shown in the attached figure. Figure 1As shown. The input variables of the training set and the test set are standardized to 1×53×1. The model contains 2 convolutional layers, which are used to automatically extract key spectral feature information from the mass spectrometry data matrix. The input variables first enter the first convolutional layer, which is configured with 16 convolution kernels (3×1), and then enter the second convolutional layer, which is configured with 32 convolution kernels (3×1). Both convolutional layers enhance the stability of data distribution through the batch normalization layer (BatchNormalization). After each convolutional layer, a pooling layer is set to eliminate redundancy and reduce the risk of overfitting when extracting spectral feature information. This model uses the average pooling algorithm. A fully connected area is then connected, including two hidden layers, containing 32 and 16 neurons respectively. A regularization step (Dropout) is applied after the first hidden layer, with a Dropout rate of 20% to further reduce the risk of overfitting. Finally, the classification output layer is set to 9 neurons, corresponding to 9 vegetable oils.

[0135] In this SCNN framework, the activation function is used to approximate any nonlinear function and fit various nonlinear models. Only the output layer adopts the Softmax activation function, and the remaining layers use the "rectified linear unit" (ReLU) activation function.

[0136] Model performance is evaluated by recall, precision, and accuracy;

[0137] The model performance was also evaluated by the area under the receiver operating characteristic curve (ROC-AUC). To avoid sampling bias, a 10-fold cross validation was performed on the original dataset during the model validation process, and the average accuracy of the model was obtained. From a hardware perspective, all experiments in this validation were completed on a MacBook running the Monterey operating system (Mac OS12.3), equipped with a 1.3GHz Intel Core i5 processor, 16GB 1867MHz memory, and a 512GB hard drive.

[0138] 6. Optimization of detection parameters

[0139] 6.1 Optimization of DESI-MS conditions

[0140] When optimizing the conditions, soybean oil was used as the test object, and the signal-to-noise ratios of two representative ion peaks of m / z 877.73 and 901.73 were selected as the basis for screening.

[0141] The composition, ratio, and flow rate of the spray solvent used for DESI-MS ionization of glycerolipids can significantly affect the selective extraction of target analytes.

[0142] Since edible oils are mainly composed of non-polar triglycerides, three different solvents were compared. The first solvent was MeOH:AcN:tol=10:7:3 (v / v / v), which has been shown to be effective in detecting hydrophobic lipids such as cholesterol esters, diglycerides (DG), and triglycerides (TG) in biological samples in tissue imaging studies. The second solvent was MeOH:DCM=4:6 (v / v), which has been considered as the preferred solvent for lipidomics analysis of tissue samples in previous studies. The third solvent, AcN:H2O=6:4 (v / v), has been successfully used to detect a variety of TG molecular species through tissue extraction. These three spray solvents were not used to enhance the signal detection of edible oils with glycerolipids as the main component.

[0143] The flow rates of MeOH / AcN / tol (capillary voltage of 4 kV), MeOH / DCM (capillary voltage of 3 kV), and AcN / H2O (capillary voltage of 3 kV) were optimized as shown in the attached Figure 3 As shown. At a flow rate of 3 μL / min, the selected peak signal intensities of MeOH / AcN / tol and MeOH / DCM increased with increasing flow rate; further increasing the flow rate resulted in a significant decrease in the signal-to-noise ratio (p<0.05). This phenomenon can be attributed to the insufficient solution volume at low flow rates, which hindered the complete dissolution and extraction of the target analytes, while excessively high flow rates made the oil film surface unstable, resulting in signal attenuation. In contrast, for the AcN / H2O solvent system, the signal intensity reached a peak at a flow rate of 4 μL / min and then stabilized.

[0144] Therefore, the optimal flow rates of MeOH / AcN / tol, MeOH / DCM, and AcN / H2O were determined to be 3 μL / min, 3 μL / min, and 4 μL / min, respectively.

[0145] Capillary voltage also has a significant effect on ion detection. Figure 4 The optimization of capillary voltage for MeOH / AcN / tol (flow rate of 3μL / min), MeOH / DCM (flow rate of 3μL / min), and AcN / H2O (flow rate of 4μL / min) solvent systems is shown. As shown in the figure, the signal-to-noise ratio of all solvent systems increases with the capillary voltage, reaching a peak rapidly at high voltage and then dropping sharply (p<0.05). This indicates that low voltage leads to incomplete ionization, while excessively high voltage causes ion suppression due to enhanced background peaks in positive ionization mode. The fluctuations in the signal-to-noise ratio caused by changes in capillary voltage are more significant than those caused by changes in solvent flow rate, showing the dominant effect of voltage on the signal intensity of the target ion.

[0146] After comparison, the optimal capillary voltages of MeOH / AcN / tol, MeOH / DCM and AcN / H2O were determined to be 4 kV, 3 kV and 3 kV, respectively.

[0147] After optimizing the flow rate and capillary voltage, DESI-MS spectra of soybean oil were obtained and displayed, as shown in Figure 5 As shown. Spectra in the range of m / z 550-1200 are shown, as no significant peaks were observed below m / z 550. Both MeOH / AcN / tol and MeOH / DCM systems showed better performance, with NL values ​​of 1.88e5 and 1.82e5, respectively, which are two orders of magnitude higher than the AcN / H2O system (7.02e3). These results were confirmed by lipid phenotyping, where the overall ion composition was similar, but the AcN / H2O spectrum showed higher background noise. Among the three solvents, MeOH / AcN / tol had the highest signal-to-noise ratio at m / z 877.73 (54.98±1.63) and 901.73 (85.83±1.12), which may be due to the complementary effects of the solvent components. Polar solvents such as MeOH and AcN stabilize the electrospray signal, while non-polar components such as toluene and DCM enhance the extraction of non-polar molecules such as TG.

[0148] Based on these results, MeOH / AcN / tol and MeOH / DCM showed remarkable effectiveness in DESI-MS analysis of vegetable oils.

[0149] Comprehensive evaluation determined that the solvent system of MeOH:AcN:tol=10:7:3 (v / v / v) was the best choice for subsequent experiments, with a flow rate of 3 μL / min and a capillary voltage of 4 kV.

[0150] 6.2 DESI-MS lipid fingerprint analysis

[0151] 6.2.1 Optimization of ionization conditions

[0152] Figure 6 The full spectrum of the triolein standard in positive ionization mode is shown. TG shows a clear tendency to form adducts with ammonium sulfate, lithium, and sodium ions. By analyzing the ion species of the TG standard after DESI-MS ionization, the ions can be basically divided into two groups.

[0153] The first group includes three ion peaks ranging from m / z 900 to 950. The peak with the highest intensity, m / z 907.78, was identified as the sodium adduct ion of triglyceride [TG(54:3)+Na] + , calculated by accurate molecular mass. The other two peaks [TG(54:3)+NH 4 ] + (m / z 902.82) and [TG(54:3)+K] + (m / z 923.75) was also detected, but at a lower intensity.

[0154] The second group contains two fragment ions in the m / z range of 600-650. The peak at m / z 603.5 is due to the ammonium adduct of TG ([TG(54:3)+NH 4 -(C 17 H 33 COOH+NH 3 )] + ), losing a free fatty acid and an amino molecule to form [TG(54:3)+HC 17 H 33 COOH] + Since the structure of this ion is very similar to diglycerol, it is usually expressed in analysis as [DG+HH 2 O] + This is consistent with the observed ionization characteristics of TG under atmospheric pressure chemical ionization (APCI). In addition, the CO bond connecting the acyl chain to the glycerol backbone in the sodium adduct of TG is broken to form [DG(36:2)+Na] + (m / z 643.54), corresponding to C 18 H 32 The loss of O.

[0155] 6.2.2 Differences in lipidomic phenotypes

[0156] The lipid phenotypes of nine edible vegetable oil samples were analyzed using DESI-MS in positive ionization mode. Figure 7 BJ in the paper presents lipid fingerprints and corresponding photographs of plant oils from different biological sources. Except for sesame oil, which is brown, the other oils are transparent yellow liquids. Therefore, it is difficult to distinguish these oils by appearance alone unless a reference sample is available. In contrast, the DESI-MS lipid fingerprints of the nine oils showed obvious visual differences. Two types of glycerolipids, DG (mainly in the m / z 550-700 range) and TG (mainly in the m / z 800-1000 range), appeared as clusters in the mass spectra. The intensity of the TG peak is generally higher than that of the DG peak, which is consistent with the lipid distribution pattern observed in the analysis of triolein standard substances.

[0157] Based on lipid fingerprints and ionization behaviors, the lipid structures and relative abundances of all target peaks at m / z50-1200 were summarized. After excluding isotopic peaks, a total of 53 peaks were identified, including 11 DGs and 42 TGs. In terms of composition, corn oil has the advantage of OLL (13.98±1.36%), followed by LLL (10.99±0.59%), OOL (8.92±0.54%), PLL (8.71±0.86%), and POL (7.48±0.80%). Soybean oil contains a significant proportion of TG, including LLL (14.22±0.67%), OLL (9.14±1.68%), LLLn (8.66±1.30%), PLL (5.85±0.61%), and OOL (5.53±0.67%). The presence of linoleic acid in soybean oil contributes to a greater diversity of TGs containing C18:3 fatty acyl chains. Comparison of the TG compositions of peanut and sesame oils showed that the most prominent TG components of the two oils were very similar in composition and content, with both oils showing high levels of OOL (peanut oil 19%, sesame oil 20%), OLL (peanut oil 19%, sesame oil 20%), OOO (both oils 14%), and POL (peanut oil 11%, sesame oil 12%). This experiment determined that OOO (peanut oil 17.19 ± 1.42%, sesame oil 21.64 ± 8.21%) was the most abundant lipid in peanut oil and sesame oil, followed by OOL (peanut oil 10.17 ± 3.40%, sesame oil 15.24 ± 3.45%), POO (peanut oil 6.34 ± 0.99%, sesame oil 6.55 ± 2.16%), SOO (peanut oil 5.94 ± 0.71%, sesame oil 7.26 ± 2.49%) and OLL (peanut oil 4.29 ± 1.23%, sesame oil 7.42 ± 3.72%). Rice bran oil showed relatively high contents of POL (12.05 ± 0.77%), OOL (8.93 ± 2.03%), OOO (8.88 ± 1.40%), POO (8.37 ± 0.72%) and PLL (6.41 ± 0.45%). Sunflower oil showed a significantly high content of OLL (26.89 ± 6.37%). C18:1 and C16:1 showed obvious dominance in the fatty acid composition of camellia oil and olive oil, which are classified as woody oils, with a combined relative content of more than 80%. Among the TG molecular species, there were higher abundances of OOO (44.39 ± 4.11% for camellia oil and 36.44 ± 0.92% for olive oil) and POO (11.23 ± 1.16% for camellia oil and 16.78 ± 0.57% for olive oil), especially OOO. Walnut oil showed a unique lipid profile feature with abundant polyunsaturated fatty acids such as C18:2 and C18:3, especially C18:2. Among all the target lipids detected in this experiment, walnut oil contained almost all TGs containing C18:3 fatty acid chains, and their relative concentrations were higher than those of other oils.LLL (19.64 ± 1.06%) showed the highest relative content in walnut oil. Due to [TG + Na]. + The intensity is higher and less interfered by isotope peaks. The above content analysis is based on its ionization form.

[0158] These combined results indicate that DESI-MS can unambiguously provide lipid profiles of various plant oils, enabling reliable identification among them.

[0159] Figure 8 The UpSet graph shows the comparative analysis of lipid phenotype differences of 9 plant oils based on the summarized lipidomics data. Among them, corn oil, soybean oil, peanut oil, sesame oil, rice bran oil, sunflower oil and walnut oil contain 32, 29, 35, 34, 33, 22 and 26 lipid molecules, respectively. In contrast, camellia oil and olive oil detected fewer target lipid signals, with 12 and 13 molecules detected, respectively.

[0160] like Fig. 9 As shown, the nine vegetable oils were divided into two main clusters: soybean oil, sunflower oil and walnut oil formed one group, and the remaining six oils formed another group. Soybean oil and walnut oil as well as camellia oil and olive oil clustered most closely, with mass spectrum similarity indices as high as 0.97 in both groups. The high similarity between soybean oil and walnut oil may be attributed to similar TGs containing similar C18:3 fatty acyl chains, such as PLL, PLLn, LLL, LLLn and LLnLn. In addition, the high abundance of these TGs in their respective lipid compositions also contributed to their similarity. The high similarity of the lipid phenotypes of camellia oil and olive oil is consistent with the previous results of lipid composition and content analysis.

[0161] like Fig.10 As shown, a mass spectrum similarity index close to 1 indicates a high similarity between mass spectra. In addition to the above combinations, the mass spectrum similarity index between peanut oil and sesame oil is also as high as 0.94. In addition, a comprehensive analysis of 36 combinations shows that 25 have a mass spectrum similarity index exceeding 0.5, and 15 combinations have an index even exceeding 0.7.

[0162] 6.3 Classification Modeling Based on SCNN

[0163] Common chemometric methods, such as PCA and partial least squares discriminant analysis (PLS-DA), mainly extract linear components from spectral data and often ignore key nonlinear factors. This limitation can lead to a significant reduction in classification accuracy, especially when samples show high similarity. In contrast, CNN, as a nonlinear machine learning model, can capture complex relationships in the data in a single pass by learning key features. This ability makes the analysis of complex systems more accurate and efficient. Given the limited number of parameters in this experiment, SCNN was used to improve computational efficiency. By using fewer convolutional layers, SCNN reduces the risk of overfitting while maintaining high classification accuracy.

[0164] The network hyperparameters were optimized using the Adam optimizer algorithm to minimize the loss function and ensure fast and stable convergence of the target loss function curve. Key parameters include 100 iterations, a mini-batch size of 128, and an initial learning rate of 1×10 -3 , the L2 regularization parameter is 1×10 -4 , the learning rate reduction factor is 0.1, and the learning rate reduction period is 90.

[0165] The accuracy and loss curves of the training set are as follows Fig.11 As shown in the accuracy curve, the accuracy exceeds 90% at the 27th iteration and stabilizes at about 97% after the 62nd iteration. At the same time, the loss function curve gradually decreases with the increase in the number of iterations, decreasing rapidly in the first 30 iterations, and then decreasing more slowly, stabilizing at 0.11 after the 78th iteration. These results show that the SCNN model is effective in classifying vegetable oils, without overfitting, robust, and meaningful.

[0166] The classification results and corresponding confusion matrices of the training set are shown in Fig.12 Only one sesame oil sample was misclassified as camellia oil, and all other oils were accurately identified. The recall rate for sesame oil was 95.2%, and all other oils reached 100%. The precision rate was equally high, reaching 100% for all oils except camellia oil. The precision rate for camellia oil was 95.5%. The overall classification accuracy of the 9 vegetable oils was 99.5%.

[0167] The classification results and confusion matrix of the test set are recorded in Fig.13 In line with the training set, only one sesame oil sample was misclassified as camellia oil. The recall rate of sesame oil was 88.9%, and all other oils reached 100%. The precision rate of camellia oil was 90.0%, and the precision rate of all other oils was 100%. The total accuracy rate of the 9 varieties was 98.8%, and the misclassification rate was only 1.2%.

[0168] The SCNN model showed robust performance in both the training and test sets, highlighting its effectiveness in accurately classifying the nine types of vegetable oils. The SCNN model showed robust performance in both the training and test sets, highlighting its effectiveness in accurately classifying the nine types of vegetable oils.

[0169] The DESI-MS technique used in this experiment demonstrated high accuracy in obtaining lipid fingerprints. In addition, the SCNN model outperformed other machine learning models in terms of training time, with a training time of only about 20 seconds, which is a significant advantage for high-throughput detection applications.

[0170] In order to evaluate the overall performance of the model, the area under the ROC curve (ROC-AUC) was calculated. Fig.14 As shown. The ROC plot shows an unusually high AUC value of 0.9719, indicating the superior classification performance of the SCNN model. To avoid sampling bias, the dataset was randomly split and 10-fold cross validation was performed to optimize accuracy estimates and prevent overfitting. The accuracy of the training and test sets was generally high, and some test sets achieved 100% accuracy. The average accuracy of the training set was 98.5±2.2%, while the average accuracy of the test set was 97.4±3.1%. The volatility of the accuracy of the training and test sets was minimal, highlighting the remarkable stability and robustness of the SCNN model.

[0171] In summary, this experiment demonstrated the potential of combining DESI-MS with SCNN for rapid, high-throughput, and accurate identification of vegetable oils. The SCNN model significantly reduced the time and labor input associated with traditional data analysis methods and achieved high-throughput detection of label fraud in vegetable oils.

[0172] 8. Conclusion

[0173] The present invention establishes a lipid fingerprint analysis method based on DESI-MS combined with SCNN machine learning model, which is specifically used for the analysis of glycerolipids and realizes high-precision classification and identification of various vegetable oils. Through a series of optimizations, the spray solvent is optimized to a mixture of MeOH:AcN:tol=10:7:3 (v / v / v), the solvent flow rate is 3μL / min, and the capillary voltage is 4kV. [TG+Na] + , [TG+NH4] + , [TG+K] + 、[DG+H-H2O] + and [DG+Na] + The main ion types such as ions provide a solid theoretical basis for the analysis of plant oil lipid fingerprints.

[0174] This precise, simple, and rapid DESI-MS method generated unique lipid fingerprints representing the glycerolipid phenotypes in 270 samples across nine plant oils (including corn oil, soybean oil, peanut oil, sesame oil, rice bran oil, sunflower oil, camellia oil, olive oil, and walnut oil). A total of 53 target peaks were accurately identified, confirming the accuracy of the DESI-MS method in detecting TG-rich samples.

[0175] A SCNN model was built to target the unique characteristics of the dataset, optimizing the network layers, hyperparameters, and performing rigorous cross-validation. The model showed minimal misclassification rates, with an accuracy of 98.5±2.2% for the training set and 97.4±3.1% for the test set. These results highlight the superior classification capabilities of the SCNN model.

[0176] The optimized DESI-MS method overcomes the challenge of low TG ionization efficiency, thus expanding the applicability of the technique. Combining DESI-MS with SCNN modeling introduces a new, pretreatment-free approach that is accurate, efficient, and robust, with high-throughput classification capabilities.

[0177] The invention has great potential in detecting label fraud of edible vegetable oils, ensuring food safety and supporting the legitimate development of the food industry. In addition, because the CNN model can be continuously expanded and optimized with new data through continuous learning, it provides a simple, fast and sustainable classification and detection solution for an increasing number of vegetable oil types. This provides a comprehensive and dynamic approach to analyzing the authenticity of vegetable oils in the future.

Claims

1. A method for detecting vegetable oil based on the combination of DESI-MS and neural network, characterized in that , including the following processes: A. Collecting samples: Take 5 μL of sample onto a designated 1×1 cm square area on a glass slide to form a thin liquid film to be analyzed; B. DESI-MS analysis: Perform DESI-MS analysis on the collected samples to obtain DESI-MS data; C. Lipid fingerprint collection: lipid fingerprint of samples collected based on DESI-MS data; D. Comparative classification based on neural network analysis: The lipid fingerprint of the collected samples is input into a shallow convolutional neural network for processing and output of the results; E. Complete sample testing and confirm sample type.

2. A vegetable oil detection method based on DESI-MS combined with neural network according to claim 1, characterized in that: The DESI-MS analysis described in step B is performed using a quadrupole time-of-flight mass spectrometer, and the specific contents are as follows: B1, calibration was performed using rhodamine 6G in positive ionization mode at m / z 443.2335; B2, spraying the spray solvent onto the sample through the spray capillary at a flow rate of 1-5 μL / min, and the voltage of the spray capillary is 2-5 kV; B3, extracting the ionized target molecules in the sample and transferring them to the mass spectrometer; B5. Place the DESI probe on the glass slide away from the sample and record the background signal for 1.0 min; B5. Obtain mass spectrometric data by full scanning in the mass range of m / z 50-1200 in positive ionization mode with a scan rate of 1 s / scan; The instrument parameters were as follows: spray impact angle 75°, distance between nozzle and glass slide surface 2 mm, distance between nozzle and ion transfer capillary hole 6 mm, distance between ion transfer capillary hole and glass slide surface 0.5 mm, capillary temperature 150 °C, nitrogen pressure 600 kPa, cone voltage 40 V; B6. Aim the probe at the center of the oil film and collect mass spectra for 1.5-2.0 minutes; B7, repeat the collection three times; B8. Obtain DESI-MS data.

3. A vegetable oil detection method based on DESI-MS combined with neural network according to claim 2, characterized in that: The spray solvent is MeOH, AcN, tol in a ratio of 10:7:3 (v / v / v).

4. According to claim 2, a method for detecting vegetable oil based on the combination of DESI-MS and neural network is further provided with an adjustment bracket for adjusting the angle and position; the adjustment bracket is detachably arranged in the mass spectrometer, characterized in that: The invention comprises a base (1), a capillary support unit arranged on the left side above the base (1), an inlet pipe support unit arranged on the right side of the base (1), a sample platform (9) arranged above the middle of the base (1), and a surrounding block (8) arranged around the sample platform (9); the capillary support unit can be raised and lowered to adjust the height of the spray capillary and adjust the angle; the inlet pipe support unit can be laterally moved to adjust the position of the inlet pipe; the surrounding block (8) is provided with corresponding operation ports for the spray capillary and the inlet pipe, and an openable and closable stopper is provided at the operation port; the sample platform (9) can be rotatably arranged on the base (1).

5. A method for detecting vegetable oil based on the combination of DESI-MS and neural network according to claim 4, characterized in that: The capillary support unit comprises a lifting column (2), a lifting platform (3) liftingly sleeved on the lifting column (2), an adjustment block (4) adjustably connected to the front of the lifting platform (3), a connecting rod (41) arranged on the adjusting block (4), and a capillary mounting seat (42) adjustably arranged on the top of the connecting rod (41); a lifting motor (21) is arranged in front of the lifting column (2), and a threaded screw (31) linked to the lifting motor (21) is arranged inside the lifting column (2); the lifting platform (3) is linked to the threaded screw (31).

6. A method for detecting vegetable oil based on the combination of DESI-MS and neural network according to claim 4, characterized in that: The inlet pipe support unit comprises a transverse shift seat (5), a transverse shift column (6) which can be transversely shifted on the transverse shift seat (5), and an inlet pipe mounting seat (61) which can be adjustably connected to the transverse shift column (6); a transverse shift motor (51) is arranged in front of the transverse shift seat (5), and a transverse shift screw rod which is linked to the transverse shift motor (51) is arranged in the transverse shift seat (5); the transverse shift column (6) is linked to the transverse shift screw rod.

7. The vegetable oil detection method based on DESI-MS combined with neural network according to claim 1, characterized in that: The lipid fingerprint collection described in step C is specifically as follows: DESI-MS data were processed by Masslynx v4.1 software for background subtraction and mass drift correction; Data normalization was performed using a normalization factor (NL), defined as the maximum peak intensity in each mass spectrum; The glycerolipid species of the target peak were identified based on the ionization behavior of triolein standards and the LIPIDMAPS prediction tool database; Fatty acyl chains are expressed in the format TC:DB, where TC is the total number of carbon atoms and DB is the number of double bonds; The characteristic percentages of the samples were calculated by peak area normalization; SPSS23.0 software was used for statistical analysis, including mean, standard deviation and one-way analysis of variance; Chemical structures and accurate molecular masses were generated using ChemDraw Professional 16.0 software; UpSet graphs and heat maps were generated by TBtools software visualization; The mass spectrum similarity index (SF) is used to measure the spectral similarity between any two oils and indirectly evaluate the similarity of their lipid components. The SF formula is: The mass spectrum consists of i m / z values, x i and i are the relative abundances of the ions corresponding to the i-th m / z value in the spectrum.

8. The vegetable oil detection method based on DESI-MS combined with neural network according to claim 1, characterized in that: The specific steps of the neural network analysis and comparison classification described in step D are as follows: The mass spectrometry data matrix was used as a data set for machine learning analysis and a SCNN model was constructed. The data set was divided into a training set and a validation set. The training set accounted for 70% of the samples and the validation set accounted for 30% of the samples. The SCNN model consists of a 15-layer structure, including an input layer, two convolutional layers, a pooling layer, a fully connected layer, a batch normalization layer, a ReLU layer, a Dropout layer, and a classification output layer; The input variables of the training set and the test set are standardized to 1×53×1; the two convolutional layers are used to automatically extract key spectral feature information in the mass spectrometry data matrix; The input variables first enter the first convolutional layer, which is configured with 16 convolution kernels, and then enter the second convolutional layer, which is configured with 32 convolution kernels; both convolutional layers enhance the stability of data distribution through batch normalization layers; after each convolutional layer, the pooling layer eliminates redundancy and reduces the risk of overfitting when extracting spectral feature information; the pooling layer uses an average pooling algorithm; the fully connected layer includes two hidden layers, containing 32 and 16 neurons respectively, and a regularization step is applied after the first hidden layer, with a dropout rate of 20%; the classification output layer is set to 9 neurons, corresponding to 9 types of vegetable oils; In the SCNN model, the activation function is used to approximate any nonlinear function and fit various nonlinear models; the output layer uses the Softmax activation function, and the remaining layers use the ReLU activation function.

Images