Deep learning assisted three-dimensional fluorescence spectroscopy for detection of contaminated vegetable oils

Abstract

The present application belongs to the technical field of food safety detection, and particularly relates to a detection method for contaminated vegetable oil by deep learning assisted three-dimensional fluorescence spectrum. The present application firstly predicts the optimal adsorption condition of magnesium silicate on vegetable oil based on a linear regression model and an L-BFGS-B algorithm, realizes the targeted separation of mineral oil in contaminated vegetable oil through the adsorption of magnesium silicate on vegetable oil under the optimal condition. Then, based on the separated mineral oil pollutant sample, a plurality of pre-training network models are applied to extract characteristic fingerprint spectra of different mineral oil categories from the collected three-dimensional fluorescence spectrum, so as to distinguish several common mineral oil pollutant categories. In combination with a parallel factor data dimension reduction algorithm, the fluorescence components of the corresponding pollutants in the mineral oil are decomposed and determined. Finally, based on the corresponding relationship between the fluorescence signals and the concentrations of each pollutant component, a support vector regression model of the corresponding component is built, and then the quantitative detection of various pollution components in the mineral oil is realized.

Description

Technical Field

[0001] This invention belongs to the field of food safety testing technology, specifically relating to a deep learning-assisted three-dimensional fluorescence spectroscopy method for detecting contaminated vegetable oils. Background Technology

[0002] Edible vegetable oils contaminated with mineral oil pose a serious threat to food safety and public health. Gas chromatography-mass spectrometry (GC-MS) is typically the preferred method for analyzing mineral oil contaminants in vegetable oils. This method offers high sensitivity and accuracy in the laboratory, effectively identifying and quantifying mineral oil contaminants and their components. However, vegetable oils themselves are complex matrices, containing abundant fatty acids and other organic compounds. These substances are similar in composition to mineral oils, easily causing matrix effects that can affect the accuracy of GC-MS results. Furthermore, the method's complex and time-consuming operation hinders the rapid response required for food safety testing. Therefore, developing an efficient detection method is extremely important.

[0003] Highly alkylated aromatic compounds in mineral oils exhibit fluorescence response due to the large π bonds in their molecular structure. Excitation-emission matrix fluorescence (EEMF) technology has been increasingly used in food analysis due to its high sensitivity and ability to obtain rich information on fluorescence excitation wavelengths, emission wavelengths, and fluorescence intensities of fluorescent compounds in a short time. However, the collected fluorescence data are numerous and complex, and simple methods such as linear regression and multivariate statistical analysis cannot fully mine and effectively utilize this fluorescence information. Therefore, a more sophisticated high-dimensional data processing and complex pattern recognition method is needed.

[0004] The rapid development of artificial intelligence (AI) is reshaping almost every field of work. As a subfield of AI, deep learning possesses powerful information mining capabilities, and its application in food safety testing has been increasingly evident in recent years. Xia et al., based on three-dimensional fluorescence spectroscopy combined with convolutional neural networks (CNNs), achieved a classification accuracy of 97.5% for different pesticide residue types on the surfaces of vegetables and fruits. Zhang's research team successfully classified heavy metals such as lead, cadmium, and arsenic in different food samples using the same method, achieving a classification accuracy of 98.2%. On the other hand, by predicting the optimal conditions for magnesium silicate adsorption of vegetable oil through machine learning, targeted separation of mineral oil can be achieved, largely eliminating matrix effects and collecting more effective spectral data, thereby improving the accuracy of classification and quantification. Currently, deep learning has achieved good results in classification research. However, consumers are more concerned about which components in mineral oil contaminants are harmful to the human body and whether the content of these components is below the safe threshold. Therefore, the analysis of harmful components and their content in various mineral oil contaminants is extremely important.

[0005] Parallel factor analysis (PARAFAC) is a multivariate data analysis technique that decomposes fluorescence signals into independent individual fluorescence phenomena. This technique can track the fluorescent components in different parts of mineral oil and separate fluorescence signals with specific excitation and emission spectra. However, the relationship between fluorescence intensity and concentration obtained from the analysis is not always linearly correlated due to factors such as self-absorption and fluorescence quenching. Support Vector Regression (SVR) is a regression model extended from Support Vector Machine (SVM). Based on kernel functions, it can handle large-scale nonlinear relationships well and is currently frequently used by researchers in nonlinear relational environments.

[0006] Although the issue of mineral oil contamination in vegetable oils has gradually attracted attention in the field of food safety, existing research mainly focuses on the source and total amount of mineral oil. A search revealed no relevant studies or reports on the classification of mineral oil in contaminated vegetable oils or the qualitative and quantitative analysis of its contaminating components.

[0007] In view of this, the inventors aim to provide a deep learning-assisted three-dimensional fluorescence spectroscopy method for detecting contaminated vegetable oils, which can be used to quickly identify the contamination categories of mineral oils in contaminated vegetable oils (gasoline, diesel, industrial boiler oil, engine oil, double neutral industrial boiler oil, and aviation kerosene), and simultaneously determine the harmful components and their contents in each mineral oil contaminant. Summary of the Invention

[0008] The purpose of this invention is to overcome the aforementioned problems in traditional technologies and provide a method for detecting contaminated vegetable oils using deep learning-assisted three-dimensional fluorescence spectroscopy.

[0009] To achieve the above-mentioned technical objectives and effects, the present invention is implemented through the following technical solution:

[0010] This invention provides a method for detecting contaminated vegetable oil using deep learning-assisted three-dimensional fluorescence spectroscopy, comprising the following steps:

[0011] Step 1: Prediction of the optimal adsorption conditions for magnesium silicate on different vegetable oils;

[0012] Step 2: Sample preparation and data preprocessing;

[0013] Step 3: Qualitative analysis of the types of mineral oil contaminants;

[0014] Step 4: Qualitative and quantitative analysis of mineral oil contaminants.

[0015] Furthermore, in step one, using the adsorption data of magnesium silicate on different vegetable oils under different adsorption conditions, the adsorption amount is used as the dependent variable and the different adsorption conditions are used as independent variables to establish a linear regression model, fit the regression equation and calculate the determination coefficient of the model; then, using an optimization algorithm, the experimental conditions are limited, and through the optimization process, the optimal conditions that meet the adsorption target are finally obtained.

[0016] Furthermore, the regression equation takes the form of:

[0017] y=β0+β1x1+β2x2+β3x3+β4x4

[0018] Where y is the adsorption amount, x1, x2, x3, and x4 represent the amount of magnesium silicate used, time, water content, and temperature, respectively, and β0, β1, β2, β3, and β4 are regression coefficients;

[0019] To find the optimal experimental conditions, the L-BFGS-B optimization algorithm was used to solve the regression equation, set the adsorption target, and minimize the deviation between the predicted adsorption amount and the target value.

[0020] The range of optimized variables is limited to the actual experimental conditions. Through the optimization process, the optimal conditions for meeting the adsorption target are gradually approached.

[0021] After optimization, the optimal experimental conditions for each vegetable oil included four variables: magnesium silicate dosage, time, water content, and temperature.

[0022] Further, in step two, mineral oils of different concentration gradients are added to different vegetable oils, and then a certain amount of n-hexane is added to each mixture. Under the predicted optimal adsorption conditions of magnesium silicate for different vegetable oils, effective adsorption of the vegetable oils is achieved. After adsorption, the magnesium silicate is filtered to obtain the corresponding mineral oil test solution. The test solution is transferred to a quartz cuvette, and spectral data is acquired using a fluorescence spectrophotometer. First, the background signal of the solvent is removed from the acquired fluorescence spectral data to eliminate the interference of the solvent on the spectral data. Based on the characteristic that the excitation wavelength and the emission wavelength are close and easily scatter, a specific wavelength region clipping method is used to remove scattering noise to obtain a higher signal-to-noise ratio. Then, interpolation is used to supplement missing data to ensure data integrity. The data of mineral oil samples of the same type and concentration are averaged to improve the stability and representativeness of the data. Data augmentation techniques are used to enhance the generalization ability of the model.

[0023] Furthermore, in step three, based on the separated mineral oil contaminant samples, various pre-trained network models are used to extract characteristic fingerprints of different mineral oil categories from the collected three-dimensional fluorescence spectra. Through hyperparameter optimization and comparison of the performance of each model, the optimal model is finally selected to identify several common mineral oil contaminant categories.

[0024] Furthermore, various pre-trained network models, including Artificial Neural Networks (ANN), Convolutional Neural Networks (CNN), and Recurrent Neural Networks (RNN), were applied to extract characteristic fingerprints of different mineral oil categories from the collected three-dimensional fluorescence spectra. Hyperparameters, including learning rate, batch size, number of training epochs, optimizer type, and activation function, through grid search, random search, and Bayesian optimization techniques. The performance of each model was compared, and the optimal model was finally selected to identify several common mineral oil contaminant categories.

[0025] Furthermore, in step four, the fluorescent components of corresponding pollutants in mineral oil are decomposed and determined by combining parallel factor data dimensionality reduction algorithm; a support vector regression model of the corresponding components is built based on the correspondence between the fluorescence signals and concentrations of each pollutant component, thereby realizing the quantitative detection of various pollutant components in mineral oil.

[0026] Furthermore, outlier detection was performed on the preprocessed three-dimensional fluorescence spectra using parallel factor analysis, and the optimal fraction of fluorescent components was screened and verified through core consistency, variance explained rate, factor matching score, and split-half analysis. Based on the correspondence between the excitation and emission wavelengths of each fluorescent component obtained from the parallel factor analysis and the standard substances of mineral oil, the main mixed contaminant categories in mineral oil were qualitatively determined. Furthermore, a support vector regression model was built based on the correspondence between the fluorescence signals and concentrations of each contaminant component, thereby achieving quantitative detection of various contaminants in mineral oil.

[0027] The beneficial effects of this invention are:

[0028] This invention provides a deep learning-assisted three-dimensional fluorescence spectroscopy method for detecting contaminated vegetable oils, enabling rapid identification and qualitative and quantitative analysis of mineral oil types and contaminant components in contaminated vegetable oils. Machine learning algorithms are used to predict the optimal adsorption conditions of magnesium silicate for vegetable oils. By utilizing the adsorption of magnesium silicate under optimal conditions, targeted separation of mineral oils from contaminated vegetable oils is achieved, effectively avoiding complex matrix effects in the vegetable and mineral oil environments. A deep learning pre-trained network model is used to extract characteristic fingerprints of different mineral oil categories from the three-dimensional fluorescence spectrum, enabling accurate identification of mineral oil contaminant categories in vegetable oils. Combined with a parallel factorial data dimensionality reduction algorithm, the fluorescent components of corresponding contaminants in the mineral oil are decomposed and determined. Based on the correspondence between the fluorescence signals and concentrations of each contaminant component, a support vector regression model is built for the corresponding component, enabling quantitative detection of various contaminant components in mineral oils. This method helps to shift the field of food safety testing from experience-based to data-driven intelligent and automated testing. It is expected to achieve early warning and precise control of mineral oil contamination in food production, transportation, and testing, providing strong protection for food safety.

[0029] Of course, any product implementing this invention does not necessarily need to achieve all of the above advantages at the same time. Attached Figure Description

[0030] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 This is a schematic diagram illustrating the qualitative and quantitative methods of the present invention for rapidly identifying the types of mineral oils and their contaminating components in contaminated vegetable oils.

[0032] Figure 2 A schematic diagram showing the effect of different experimental conditions on the adsorption capacity of magnesium silicate;

[0033] Where a-magnesium silicate content, b-time, c-moisture content of magnesium silicate, d-temperature;

[0034] Figure 3 Three-dimensional fluorescence spectra of six mineral oils;

[0035] Among them, a-95 gasoline, b-diesel, c-industrial boiler oil, d-engine oil, e-double neutral industrial boiler oil, and f-aviation kerosene;

[0036] Figure 4 A comparison chart of the accuracy of six deep learning models;

[0037] Among them, ResNet18 is the best model, with a classification accuracy of 96.21% for the six mineral oils;

[0038] Figure 5 The SVR regression curve is shown for the support vector regression model based on standard materials.

[0039] Among them, a, b, c, and d correspond to components 1 to 4, respectively, and their corresponding standard substances are naphthalene, anthracene, phenanthrene, and fluorene. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] This invention provides a deep learning-assisted three-dimensional fluorescence spectroscopy method for detecting contaminated vegetable oils, used to quickly identify the contamination category of mineral oil in contaminated vegetable oils and simultaneously determine the harmful components and their content in each mineral oil contaminant. First, machine learning algorithms are used to optimize the conditions for magnesium silicate adsorption of vegetable oil to achieve targeted separation of mineral oil. The adsorbed and desorbed mineral oil contaminants are then subjected to fluorescence testing to obtain a three-dimensional fluorescence spectrum (EEM). Background and scattering removal techniques are employed, combined with interpolation to process the data, and data enhancement methods are used through rotation, cropping, and noise injection. Next, six pre-trained network models are applied to extract characteristic fingerprints of different mineral oil categories from the three-dimensional fluorescence spectra. Hyperparameters are optimized through grid search, and the performance of each model is compared. Finally, the optimal model is selected for qualitative analysis of different types of mineral oil contaminants. Simultaneously, parallel factor analysis (PARAFAC) is used to detect outliers in the pre-processed three-dimensional fluorescence spectra. The optimal number of fluorescent components in the three-dimensional fluorescence spectral data is selected based on core consistency, variance explained rate, and minimum factor similarity. The compositional information of the corresponding contaminants in the mineral oil is analyzed based on the excitation and emission wavelength information of each component. Finally, a support vector regression model for each pollutant component was built based on the correspondence between fluorescence signals and concentrations, thereby enabling the quantitative detection of various pollutants in mineral oil.

[0042] Specific embodiments of the present invention:

[0043] Example 1

[0044] First, the adsorption amounts of magnesium silicate on vegetable oils under different conditions were obtained in 300 experiments. A linear regression model was established for each vegetable oil, with adsorption amount as the dependent variable and magnesium silicate dosage, time, water content, and temperature as independent variables. The regression equation was fitted, and the coefficient of determination of the model was calculated. The regression equation is typically in the form y = β0 + β1x1 + β2x2 + β3x3 + β4x4, where y is the adsorption amount, x1, x2, x3, and x4 represent magnesium silicate dosage, time, water content, and temperature, respectively, and β0, β1, β2, β3, and β4 are regression coefficients. To find the optimal experimental conditions, the L-BFGS-B optimization algorithm was used to solve the regression equation, setting the target adsorption amount to 100 mg and minimizing the deviation between the predicted adsorption amount and the target value. The optimized variable range was limited to the actual experimental conditions: magnesium silicate dosage 1–6 g, time 10–50 minutes, water content 0–10%, and temperature 15–40℃. Through optimization, the optimal conditions for achieving the adsorption target are gradually approached. After optimization, the optimal experimental conditions for each vegetable oil will include four variables: magnesium silicate dosage, time, water content, and temperature.

[0045] Example 2

[0046] Six different vegetable oils (peanut oil, soybean oil, rapeseed oil, corn oil, sesame oil, and olive oil) were mixed with six different mineral oils (diesel, gasoline, engine oil, aviation kerosene, industrial boiler oil, and double-neutral industrial boiler oil) at 28 concentration gradients, and 10 mL of n-hexane was added to each mixture. Next, magnesium silicate under optimal adsorption conditions was added to adsorb the mineral oil components in the vegetable oils, and the mixture was allowed to stand and separate. The supernatant was filtered through a glass funnel lined with glass wool, which had been pre-washed with n-hexane and dried. The filtrate was collected in colorimetric tubes for subsequent measurements. This process yielded 1008 standard samples containing six mineral oil contaminants at 28 concentration gradients. These samples simulated actual pollution conditions, constructing standard samples for known contaminants and providing a foundation for subsequent data acquisition and analysis.

[0047] Example 3

[0048] The acquired samples were first processed to remove interference from the blank solvent, and then noise in the spectral data was removed by cropping first- and second-order Rayleigh scattering to ensure greater accuracy. Next, cubic spline interpolation was used to complete the blank values, ensuring data integrity. During this process, samples belonging to the same mineral oil and concentration gradient were summed and averaged to reduce experimental error and improve data stability and representativeness. This step reduced the number of samples from 1008 to 168 new samples (after averaging), laying the foundation for subsequent analysis. Subsequently, data augmentation techniques were applied to further increase the sample size to 672. The augmented samples provided a richer sample size for model training, improving the model's generalization ability and prediction accuracy.

[0049] Example 4

[0050] In the classification of mineral oils, transfer learning technology was employed, combined with multiple deep learning models, to extract rich fingerprint information from three-dimensional fluorescence spectra, aiding in accurate identification of mineral oils. Specifically, six deep learning models were applied, including SimpleCNN, LeNet-5, AlexNet, GoogLeNet, VGG16, and ResNet18. Grid search technology was used to optimize the batch size and learning rate of all models, and the performance of different models was compared. Ultimately, the optimal model was selected for qualitative analysis of mineral oils.

[0051] In this process, transfer learning enables small-sample learning, reducing reliance on large amounts of labeled data and thus lowering the cost and workload of data collection. Simultaneously, it improves the model's generalization ability, enabling it to achieve high accuracy even with limited samples. Through this classification model, researchers can quickly and accurately identify the types of gasoline, diesel, engine oil, aviation kerosene, industrial boiler oil, and doubly neutral industrial boiler oil, significantly improving experimental efficiency, reducing human error, and providing strong technical support for subsequent analysis.

[0052] Example 5

[0053] In the quantitative analysis of mineral oil components, parallel factor analysis (PARAFAC) was first used to decompose the preprocessed three-dimensional fluorescence spectral data. Outlier detection, core consistency assessment, variance explanation rate analysis, and factor matching scoring were performed on the data. The optimal component score was then determined using a split-half analysis method, thereby extracting the main spectral characteristics of various pollutant components in the mineral oil. Next, based on the correspondence between the excitation and emission wavelengths of each fluorescent component obtained from the parallel factor analysis and the mineral oil standard substances, the main mixed pollutant component categories in the mineral oil were qualitatively determined. Furthermore, support vector regression models were constructed based on the correspondence between the fluorescence signals and concentrations of each pollutant component, thus achieving the quantitative detection of various pollutant components in the mineral oil. This not only improves the efficiency of pollutant detection but also provides a feasible technical path for pollutant monitoring and environmental protection in related fields.

[0054] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to specific implementations. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for detecting contaminated vegetable oil using deep learning-assisted three-dimensional fluorescence spectroscopy, characterized in that, Includes the following steps: Step 1: Prediction of the optimal adsorption conditions for magnesium silicate on different vegetable oils; Step 2: Sample preparation and data preprocessing; Step 3: Qualitative analysis of the types of mineral oil contaminants; Step 4: Qualitative and quantitative analysis of mineral oil contaminants; In step one, using the adsorption data of magnesium silicate on different vegetable oils under different adsorption conditions, the adsorption amount is used as the dependent variable and the different adsorption conditions are used as independent variables to establish a linear regression model, fit the regression equation and calculate the determination coefficient of the model; then, using an optimization algorithm, the experimental conditions are limited, and through the optimization process, the optimal conditions that meet the adsorption target are finally obtained. The regression equation is in the form of: ； Where y is the adsorption amount. , , , These represent the amount of magnesium silicate used, time, moisture content, and temperature, respectively. These are the regression coefficients; To find the optimal experimental conditions, the L-BFGS-B optimization algorithm was used to solve the regression equation, set the adsorption target, and minimize the deviation between the predicted adsorption amount and the target value. The range of optimized variables is limited to the actual experimental conditions. Through the optimization process, the optimal conditions for meeting the adsorption target are gradually approached. After optimization, the optimal experimental conditions for each vegetable oil included four variables: magnesium silicate dosage, time, water content, and temperature. In step two, mineral oils of different concentration gradients were added to different vegetable oils, and then a certain amount of n-hexane was added to each mixture. Under the predicted optimal adsorption conditions for different vegetable oils by magnesium silicate, effective adsorption of the vegetable oils was achieved. After adsorption, the magnesium silicate was filtered to obtain the corresponding mineral oil test solution. The test solution was transferred to a quartz cuvette, and spectral data were collected using a fluorescence spectrophotometer. First, the background signal of the solvent was removed from the collected fluorescence spectral data to eliminate the interference of the solvent on the spectral data. Based on the characteristic that the excitation wavelength and the emission wavelength are close and easily scatter, a specific wavelength region clipping method was used to remove scattering noise to obtain a higher signal-to-noise ratio. Then, interpolation was used to supplement missing data to ensure data integrity. The data of mineral oil samples of the same type and concentration were averaged to improve the stability and representativeness of the data. Data augmentation techniques were used to enhance the generalization ability of the model.

2. The method for detecting contaminated vegetable oil using deep learning-assisted three-dimensional fluorescence spectroscopy according to claim 1, characterized in that, In step three, based on the separated mineral oil contaminant samples, various pre-trained network models are used to extract characteristic fingerprints of different mineral oil categories from the collected three-dimensional fluorescence spectra. Through hyperparameter optimization and comparison of the performance of each model, the optimal model is finally selected to identify several common mineral oil contaminant categories.

3. The method for detecting contaminated vegetable oil using deep learning-assisted three-dimensional fluorescence spectroscopy according to claim 2, characterized in that, The study employs various pre-trained network models, including Artificial Neural Networks (ANN), Convolutional Neural Networks (CNN), and Recurrent Neural Networks (RNN), to extract characteristic fingerprints of different mineral oil categories from the collected three-dimensional fluorescence spectra. Hyperparameters, including learning rate, batch size, number of training epochs, optimizer type, and activation function, through grid search, random search, and Bayesian optimization techniques. The performance of each model is compared, and the optimal model is finally selected to identify several common mineral oil contaminant categories.

4. The method for detecting contaminated vegetable oil using deep learning-assisted three-dimensional fluorescence spectroscopy according to claim 3, characterized in that, In step four, a parallel factor data dimensionality reduction algorithm is used to decompose and determine the fluorescent components of corresponding pollutants in mineral oil; a support vector regression model for the corresponding components is built based on the correspondence between the fluorescence signals and concentrations of each pollutant component, thereby realizing the quantitative detection of various pollutant components in mineral oil.

5. The method for detecting contaminated vegetable oil using deep learning-assisted three-dimensional fluorescence spectroscopy according to claim 4, characterized in that, Parallel factor analysis was used to detect outliers in the preprocessed three-dimensional fluorescence spectra. The optimal fractions of fluorescent components were screened and validated using core consistency, variance explained rate, factor matching score, and split-half analysis. Based on the correspondence between the excitation and emission wavelengths of each fluorescent component obtained from parallel factor analysis and various mineral oil standard substances, the main mixed contaminant categories in the mineral oil were qualitatively determined. Furthermore, support vector regression models were constructed based on the correspondence between the fluorescence signals and concentrations of each contaminant component, thereby enabling the quantitative detection of various contaminants in the mineral oil.

Images