Method for recovering AFB1 inherent fluorescence in vegetable oil complex matrix based on one-dimensional convolutional neural network

Abstract

The invention discloses a method for recovering AFB1 intrinsic fluorescence in a vegetable oil complex matrix based on a one-dimensional convolutional neural network. The method comprises the following steps: firstly, acquiring AFB1 polluted vegetable oil samples with different matrixes and different concentrations; acquiring absorption, scattering and fluorescence spectrums of the oil sample, and measuring the AFB1 standard solution to obtain fluorescence intensity as inherent fluorescence; performing repeatability, linearity and precision verification on the optical platform through the phantom, and performing correction to obtain correction parameters; the method comprises the following steps: extracting, correcting and calculating absorption, reduced scattering and diffusion coefficients under excitation wavelength and absorption and diffusion coefficients under fluorescence emission wavelength, and establishing an inherent fluorescence regression model by using a one-dimensional convolutional neural network according to the five parameters; and finally, establishing a quantitative prediction model of the AFB1 in the vegetable oil by using the unrecovered fluorescence intensity and the recovered fluorescence intensity. The AFB1 intrinsic fluorescence recovery method provided by the invention can effectively recover severely distorted AFB1 characteristic fluorescence from a vegetable oil complex matrix, and belongs to the field of food quality safety rapid nondestructive testing.

Description

Technical Field

[0001] The present invention belongs to the technical field of food light transmission property detection, and relates to the AFB of vegetable oil 1 spectral non-destructive detection method, and particularly relates to a method for recovering the inherent fluorescence of AFB in a complex matrix of vegetable oil based on a one-dimensional convolutional neural network 1 recovery method. Background Technique

[0002] For vegetable oils, especially peanut oil, the over-standard incidents of aflatoxin B 1 (AFB 1 ) have occurred frequently in recent years. Due to its carcinogenic, teratogenic, and highly toxic properties, it has been listed as a Group 1 carcinogen by the International Agency for Research on Cancer (IARC). Therefore, effective monitoring of AFB in vegetable oils 1 is of great significance for ensuring consumer health. At present, the mature detection methods for AFB in vegetable oils 1 are difficult to be used in the general screening process because they more or less require sample pretreatment, complex and expensive instruments, and professional testers, etc.

[0003] AFB 1 will emit blue fluorescence (424 nm) under ultraviolet excitation. Therefore, the laser-induced fluorescence spectroscopy technology can obtain the characteristic fluorescence of AFB 1 and is expected to achieve effective rapid non-destructive detection. However, the matrix of vegetable oil samples is complex, with a large number of chromophores and endogenous fluorescent components, which will cause serious distortion to the excitation light and the characteristic fluorescence emitted by AFB 1 When predicting AFB1 in vegetable oils using only the fluorescence signal, it highly depends on the representativeness of the sample data in the model calibration process and the prior knowledge of the oil samples, such as the raw material variety, brand, origin, processing method, etc. of the oil samples. These prior knowledges determine the main physicochemical indexes of vegetable oil samples.

[0004] At present, although in the field of medical testing, some scholars have determined the distortion law of fluorescence in different human tissues from a theoretical or simulation perspective and proposed corresponding fluorescence recovery methods, there is no relevant research on the fluorescence non-destructive detection of foods or agricultural products. Moreover, the food medium is very different from human tissues. Therefore, developing a method for recovering the inherent fluorescence of AFB 1 in turbid food media such as vegetable oils is of great significance for the laser-induced fluorescence non-destructive detection of AFB 1 content. Summary of the Invention

[0005] The object of the present invention is: when using the laser-induced fluorescence spectroscopy technology to detect the AFB 1 concentration in vegetable oils with continuously changing complex matrices, AFB 1The characteristic fluorescence is seriously distorted, which makes it difficult to implement nondestructive testing using laser induced fluorescence technology. The model is poor in robustness and prone to failure. A one-dimensional convolutional neural network is proposed to detect AFB in complex matrix of vegetable oil. 1 The method of restoring intrinsic fluorescence belongs to the field of rapid nondestructive testing of food quality and safety, and can greatly improve the effect of fluorescence nondestructive prediction.

[0006] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0007] A one-dimensional convolutional neural network based on AFB in complex vegetable oil matrix 1 The method for restoring intrinsic fluorescence comprises the following steps:

[0008] Obtain different matrix gradients and AFB with different concentration gradients 1 Contaminated vegetable oil samples;

[0009] The absorption, scattering and fluorescence spectra of vegetable oil samples were obtained through the optical platform, and the AFB 1 The fluorescence intensity of the standard solution at different concentrations is measured as the intrinsic fluorescence;

[0010] Extract or calculate AFB 1 Absorption coefficient, reduced scattering coefficient, diffusion coefficient at the excitation wavelength and absorption coefficient and diffusion coefficient at the fluorescence emission wavelength;

[0011] According to the five parameters and the measured distorted fluorescence intensity, a one-dimensional convolutional neural network was used to establish an intrinsic fluorescence regression model to calculate the restored AFB. 1 Characteristic fluorescence;

[0012] Based on the restored AFB 1 Characteristic fluorescence, AFB was obtained by linear regression analysis 1 Quantitative calculation formula for content.

[0013] As a preferred embodiment of the present invention, the vegetable oil sample acquisition process includes:

[0014] Four vegetable oils with different intrinsic fluorescent components and pigments (vitamin E, polyphenols, carotenoids, chlorophyll) in the matrix were selected from the market as crude oils, namely, 'Golden Dragon Fish' soybean oil, 'Golden Dragon Fish' corn oil, 'Golden Dragon Fish' peanut oil and 'Fulinmen' peanut oil;

[0015] Mix 'Golden Arowana' soybean oil with 'Golden Arowana' corn oil, 'Golden Arowana' soybean oil with 'Fulinmen' peanut oil, 'Golden Arowana' corn oil with 'Golden Arowana' peanut oil, and 'Golden Arowana' soybean oil with 'Golden Arowana' peanut oil respectively. The mixing ratios are 0%, 25%, 50%, 75%, and 100% respectively. There are 8 oil samples for each ratio, and the volume of each oil sample is 40 mL, with a total of 128 oil samples;

[0016] Weigh 1 mg of AFB 1 standard product into a 100 mL volumetric flask, dissolve it with acetonitrile and make up the volume to prepare AFB 1 stock solution (10000 μg·kg -1 );

[0017] Add 0, 9, 18, 36, 72, 108, 144, 180 μL of AFB 1 stock solution to each of the 8 oil samples of each ratio with a pipette and stir evenly to prepare AFB 1 contaminated vegetable oil samples with contamination levels of: 0, 2.5, 5, 10, 20, 30, 40, 50 μg·kg -1 respectively.

[0018] As an optimization of the present invention, for the acquisition of absorption, scattering and fluorescence spectra, the process of obtaining the intrinsic fluorescence of AFB 1 includes:

[0019] By using the double integrating sphere technology combined with the inverse doubling algorithm, calculate the absorption coefficient (μ a ) and the reduced scattering coefficient (μ' s ) spectra from the transmittance and reflectance of the oil sample, and obtain the fluorescence intensity spectrum of the oil sample through the laser-induced fluorescence spectroscopy technology;

[0020] Absorb 1 mL of AFB 1 stock solution into a 100 mL volumetric flask, add acetonitrile to make up the volume to obtain 100 μg·kg -1 AFB 1 solution. Respectively absorb different volumes of 100 μg·kg -1 AFB 1 solution and add them to 25 mL volumetric flasks, make up the volume with acetonitrile to obtain 0, 0.1, 1, 10, 20, 30, 40, 50 μg·kg -1 AFB 1 working solution, and obtain the fluorescence intensity of the AFB 1 working solution as the intrinsic fluorescence at different concentrations through the laser-induced fluorescence spectroscopy technology. Through linear fitting, with the AFB 1 concentration (c' AFB1 ) as the independent variable and the fluorescence intensity (F int,AFB1 ) at 424 nm as the dependent variable, obtain AFB1 Intrinsic fluorescence calculation formula:

[0021] F int,AFB1 = 21.405·c' AFB1 + 1261.3

[0022] Substitute c’ AFB1 = 0, 2.5, 5, 10, 20, 30, 40, 50 μg·kg -1 into the formula to calculate the AFB in the oil sample 1 theoretical value of intrinsic fluorescence (F int,AFB1 ).

[0023] As a preference of the present invention, the method for restoring AFB 1 intrinsic fluorescence in the complex matrix of vegetable oil based on a one-dimensional convolutional neural network is characterized in that the verification and calibration process of the repeatability, linearity and precision of the optical platform includes:

[0024] Prepare 2 groups of liquid phantoms containing India Ink and TiO 2 . For the first group, the volume concentrations of India Ink are 0.005%, 0.010%, 0.020%, 0.040%, 0.080% respectively, and the volume concentration of TiO 2 is 0.020%. For the second group, the volume concentrations of TiO 2 are 0.004%, 0.008%, 0.012%, 0.016%, 0.020%, 0.024%, 0.028%, 0.032%, 0.036%, 0.040% respectively, and the volume concentration of India Ink is 0.020%. A total of 15 phantoms are prepared;

[0025] Measure the collimated transmittance of the pure India Ink solution at 5 concentrations, and calculate the μ a reference values of India Ink at different wavelengths and different concentrations according to the Lambert-Beer law (the following formula)

[0026] μ a,ref = lg(1 / T) / b

[0027] where T is the measured collimated transmittance and b is the optical path length of the cuvette.

[0028] Use the Mie algorithm to calculate the μ’ 2 reference values of TiO s at different wavelengths and different concentrations;

[0029] The transmission intensity and reflection intensity of the standard diffuser and one of the phantoms (India Ink: 0.010%, TiO2: 0.020%) were detected at different time points by the double integrating sphere technique, and the coefficient of variation of the transmission intensity and reflection intensity at each wavelength was calculated to verify the stability of the platform;

[0030] By combining the double integrating sphere technique with the inverse doubling algorithm, μ was calculated based on the transmittance and reflectance of the phantom a and μ' s measurement values. Taking the concentrations of India Ink and TiO 2 as independent variables and the μ a and μ' s measurement values as dependent variables, the correlation coefficients at each wavelength were calculated to verify the linearity of the platform;

[0031] The μ a , μ' s measurement values of the phantom were compared with the reference values, the relative errors at each wavelength were calculated, and μ a and μ' s were corrected respectively by fitting with a cubic and a linear polynomial curve to obtain the correction parameters of μ a and μ' s . Then the corrected values of μ a and μ' s were compared with the reference values, and the relative errors at each wavelength were calculated as the parameters for finally verifying the accuracy of the platform.

[0032] As a preference of the present invention, the method for restoring the intrinsic fluorescence of AFB in the complex matrix of vegetable oil based on a one-dimensional convolutional neural network is characterized in that the excitation wavelength and fluorescence emission wavelength of the AFB 1 are 375 and 424 nm respectively, and the diffusion coefficient calculation formulas at 375 nm (D 1 ) and 424 nm (D ex ) are as follows: em where μ'

[0033]

[0034] and μ' t,ex are the transport coefficients at 375 and 424 nm respectively, μ' t,em and μ' s,ex are the reduced scattering coefficients at 375 and 424 nm respectively, and μ s,em and μ a,ex are the absorption coefficients at 375 and 424 nm respectively. a,em

[0035] ​Preferably, the one-dimensional convolutional neural network of the present invention includes: 1 input layer, 2 convolutional layers, 2 batch normalization layers, 2 ReLU layers, 1 pooling layer, 1 dropout layer, 1 fully connected layer, and 1 regression output layer. The input parameters of the input layer are 6, namely: absorption coefficient at 375 nm, reduced scattering coefficient, diffusion coefficient, absorption coefficient at 424 nm, diffusion coefficient, and fluorescence intensity. The sizes of the convolutional kernels are all 3*1, and the numbers of convolutional kernels are 16 and 100 respectively. The pooling layer is a max pooling layer. The optimization algorithm is the ADAM algorithm.

[0036] Preferably, for the one-dimensional convolutional neural network of the present invention, before model training, all oil samples are divided into a calibration group and a verification group according to a ratio of 3:1. The calibration group is used for training the one-dimensional convolutional neural network model, and the verification group is used for verifying the model effect. The input parameters of the oil samples and the true values of the intrinsic fluorescence are respectively normalized before being used for model training, and the calculated values of the intrinsic fluorescence are denormalized when the model training is completed and the calculated values are output.

[0037] Preferably, the AFB 1 The process of obtaining the simple linear calculation formula for the concentration is as follows: Through linear fitting, with the theoretical value of the intrinsic fluorescence of AFB 1 as the independent variable and the actual concentration of AFB 1 as the dependent variable, the AFB 1 concentration calculation formula based on the intrinsic fluorescence of AFB 1 is obtained as follows:

[0038] c AFB1 = 0.0467·F' int,AFB1 - 58.925

[0039] When this formula is actually applied, F’ int,AFB1 is the restored value of the intrinsic fluorescence, and c’ AFB1 is the calculated value of the AFB 1 concentration.

[0040] The beneficial effect of the present invention is that the method for restoring the intrinsic fluorescence of AFB 1 in the complex matrix of vegetable oil based on a one-dimensional convolutional neural network of the present invention solves the problem that when using fluorescence spectroscopy for non-destructive detection of vegetable oil samples with complex matrices and large differences in AFB 1 , the characteristic fluorescence of AFB 1 is severely distorted, resulting in no linear relationship between the fluorescence intensity at the characteristic wavelength and the AFB 1 content, and the prediction model established by relying on complex chemometric methods has low accuracy and is prone to failure. This method determines that the parameters required for the intrinsic fluorescence of AFB 1 are: AFB 1The absorption coefficient, reduced scattering coefficient, and diffusion coefficient at the excitation wavelength, as well as the absorption coefficient and diffusion coefficient at the fluorescence emission wavelength. Further, based on these five parameters, a one-dimensional convolutional neural network is used to establish an intrinsic fluorescence regression model; finally, based on the restored intrinsic fluorescence intensity, a simple linear calculation formula is established for AFB in vegetable oil samples with different matrix gradient changes. It belongs to the field of rapid non-destructive detection of food quality and safety and can significantly improve the detection effect. 1 BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The drawings herein are incorporated into the specification and form a part of this specification, showing embodiments consistent with the present application, and are used together with the specification to explain the principles of the present application.

[0042] Figure 1 is a flowchart of a method for restoring the intrinsic fluorescence of AFB in complex matrices of vegetable oil based on a one-dimensional convolutional neural network proposed by the present invention. 1

[0043] Figure 2 is the fluorescence intensity spectra of 4 crude oils in the 200 - 1000 nm wavelength range at different AFB 1 concentrations.

[0044] Figure 3 is the fluorescence intensity spectra of oil samples with different mixing ratios in the 200 - 1000 nm wavelength range at the same AFB 1 concentration (20 μg·kg -1 ).

[0045] Figure 4 is the spectra of μ 1 and μ’ -1 in the 200 - 1000 nm wavelength range of oil samples with different mixing ratios at the same AFB a concentration (20 μg·kg s ).

[0046] Figure 5 is the μ 1 value and the intrinsic fluorescence spectra in the 200 - 1000 nm wavelength range of AFB working solutions with different concentrations at 375 nm. a

[0047] Figure 6 is the fitting graph of the detected value of the intrinsic fluorescence intensity of AFB working solutions with different concentrations at 424 nm and the AFB 1 concentration, as well as the fitting graph between the AFB 1 concentration and the reference value of the intrinsic fluorescence at 424 nm. 1

[0048] Figure 7 ​is the μ of the prepared phantom in the 200 - 1000 nm band range a and μ’ s reference value.

[0049] Figure 8 are the transmission and reflection intensity spectra measured by the double - integrating - sphere technique for air, a standard diffuser, and one of the phantoms (India Ink: 0.010%, TiO2: 0.020%) at different time points.

[0050] Figure 9 are the coefficient of variation of the air transmission intensity, the standard diffuser reflection intensity, and the phantom transmission and reflection intensities at different time points.

[0051] Figure 10 is the μ of the first group of phantoms in the 200 - 1000 nm band range a and the μ’ of the second group of phantoms s measurement values.

[0052] Figure 11 is the correlation coefficient between the μ of the first group of phantoms in the 200 - 1000 nm band range a and the India Ink concentration, and the correlation coefficient between the measurement value of the μ’ of the second group of phantoms s and the TiO 2 concentration.

[0053] Figure 12 is the relative error between the measurement value of the μ of the first group of phantoms in the 200 - 1000 nm band range a and the reference value, and the relative error between the measurement value of the μ’ of the second group of phantoms s and the reference value.

[0054] Figure 13 is the relative error between the corrected value of the μ of the first group of phantoms in the 200 - 1000 nm band range a and the reference value, and the relative error between the corrected value of the μ’ of the second group of phantoms s and the reference value.

[0055] Figure 14 is the schematic diagram of the one - dimensional convolutional neural network model structure for AFB 1 intrinsic fluorescence recovery.

[0056] Figure 15 are the un - recovered fluorescence v.s. the theoretical value of the intrinsic fluorescence, and the recovered intrinsic fluorescence v.s. the theoretical value of the intrinsic fluorescence measured for the phantoms in the calibration group and the validation group 1

[0057] Figure 16 is the AFB 1 content calculated from the recovered intrinsic fluorescence value for the phantoms in the validation group v.s. AFB 1 ​Content reference value. Detailed implementation manners

[0058] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0059] The present invention proposes a method for restoring the inherent fluorescence of AFB in complex matrices of vegetable oils based on a one-dimensional convolutional neural network, as 1 shown, and its implementation process is as follows: Figure 1 shown, and its implementation process is as follows:

[0060] (1) Oil sample preparation

[0061] In this embodiment, four vegetable oils with large differences in endogenous fluorescent components and pigments (vitamin E, polyphenols, carotenoids, chlorophyll) in the matrix are selected from the market as crude oils, namely: 'Jinlongyu' soybean oil ('JLY' soybean oil), 'Jinlongyu' corn oil ('JLY' corn oil), 'Jinlongyu' peanut oil ('JLY' peanut oil), and 'Fulinmen' peanut oil ('FLM' peanut oil). The contents of the main endogenous fluorescent components and pigments in the four vegetable oils are shown in Table 1.

[0062] Table 1 Contents of main endogenous fluorescent components and pigments in four vegetable oils

[0063]

[0064] Respectively mix 'Jinlongyu' soybean oil with 'Jinlongyu' corn oil, 'Jinlongyu' soybean oil with 'Fulinmen' peanut oil, 'Jinlongyu' corn oil with 'Jinlongyu' peanut oil, and 'Jinlongyu' soybean oil with 'Jinlongyu' peanut oil. The mixing ratios are 0%, 25%, 50%, 75%, and 100% respectively. There are 8 oil samples for each ratio, and the volume of each oil sample is 40 mL, for a total of 128 oil samples;

[0065] Weigh 1 mg of AFB 1 standard product into a 100 mL volumetric flask, dissolve it with acetonitrile and make up the volume to prepare an AFB 1 mother liquor (10000 μg·kg -1 );

[0066] Add 0, 9, 18, 36, 72, 108, 144, 180 μL of AFB 1 mother liquor to each of the 8 oil samples at each ratio through a pipette and stir evenly to prepare AFB1 The pollution levels were: 0, 2.5, 5, 10, 20, 30, 40, 50 μg·kg -1 of vegetable oil samples, and a total of 128 samples were obtained.

[0067] (2) Spectral acquisition

[0068] Through the double integrating sphere technique combined with the inverse doubling algorithm, μ in the wavelength range of 200 - 1000 nm was calculated based on the transmittance and reflectance of the oil sample a and μ’ s spectra. The fluorescence intensity spectra in the wavelength range of 200 - 1000 nm of the oil sample were obtained by laser-induced fluorescence spectroscopy. For the fluorescence intensity spectra of 4 types of crude oil at different AFB 1 concentrations (0, 2.5, 5, 10, 20, 30, 40, 50 μg·kg -1 ), as shown Figure 2 in the figure. It can be found from the figure that for the samples within the 4 types of crude oil groups, due to the consistent sample matrix, only the AFB 1 concentration is different, and the fluorescence intensity increases with the increase of the AFB 1 concentration, showing a significant positive correlation. For the samples between different crude oil groups, due to the great difference in the sample matrix, the shapes and intensities of the fluorescence intensity spectra are very different between groups, indicating that the influence of different types of crude oil matrices on fluorescence is more significant than that of AFB 1 . Figure 3 For oil sample fluorescence intensity spectra at the same AFB 1 concentration (20 μg·kg -1 ) and different mixing ratios (0%, 25%, 50%, 75 / 100%), as shown in the figure, it can be found that the sample matrix seriously distorts the characteristic fluorescence of AFB 1 . If it is not restored, using the fluorescence spectra of samples with different matrices to establish an AFB 1 concentration prediction model will result in extremely poor model accuracy prediction effect and poor stability and robustness.

[0069] Therefore, in order to obtain the intrinsic fluorescence of AFB 1 , 1 mL of the AFB 1 stock solution was pipetted into a 100 mL volumetric flask, and acetonitrile was added to make up the volume to obtain a 100 μg·kg -1 AFB 1 solution. Different volumes of the 100 μg·kg -1 AFB 1 solution were respectively pipetted into 25 mL volumetric flasks, and acetonitrile was added to make up the volume to obtain 0, 0.1, 1, 10, 20, 30, 40, 50 μg·kg -1 AFB 1 working solutions, and the AFB fluorescence spectra were obtained by laser-induced fluorescence spectroscopy1 The fluorescence intensity of the working fluid is regarded as the inherent fluorescence at different concentrations. Figure 5 The right figure shows AFB at different concentrations obtained. 1 The inherent fluorescence spectrum of the working fluid in the wavelength range of 200 - 1000 nm. As Figure 6 The left figure shows AFB at different concentrations. 1 The detected value of the inherent fluorescence intensity of the working fluid at 424 nm and the fitting graph of AFB 1 concentration. Further, through linear fitting, with AFB 1 concentration (c’ AFB1 ) as the independent variable and the fluorescence intensity (F int,AFB1 ) at 424 nm as the dependent variable, as Figure 6 the right figure shows the fitting graph between AFB 1 concentration and the reference value of the inherent fluorescence at 424 nm, and the calculation formula for the inherent fluorescence of AFB is obtained: 1 The formula for inherent fluorescence:

[0070] F int,AFB1 = 21.405·c' AFB1 + 1261.3

[0071] Substitute c’ AFB1 = 0, 2.5, 5, 10, 20, 30, 40, 50 μg·kg -1 into the formula to calculate the theoretical value of the inherent fluorescence of AFB in the oil sample (F 1 ). int,AFB1 )

[0072] (3) Platform verification and calibration

[0073] Prepare 2 groups of liquid phantoms containing India Ink (Talens, Netherlands) and TiO 2 (T104950 - 100g, Aladdin, China). For the first group, the volume concentrations of India Ink are 0.005%, 0.010%, 0.020%, 0.040%, 0.080% respectively, and the volume concentration of TiO 2 is 0.020%. For the second group, the volume concentrations of TiO 2 are 0.004%, 0.008%, 0.012%, 0.016%, 0.020%, 0.024%, 0.028%, 0.032%, 0.036%, 0.040% respectively, and the volume concentration of India Ink is 0.020%. A total of 15 phantoms are prepared;

[0074] Measure the collimated transmittance of the pure India Ink solution at 5 concentrations and calculate μ of India Ink at different wavelengths and different concentrations according to the Lambert - Beer law (the following formula). aReference value

[0075] μ a,ref = lg(1 / T) / b

[0076] where T is the measured collimated transmittance and b is the optical path length of the cuvette

[0077] Using the Mie algorithm, the μ' at different wavelengths and different concentrations of TiO 2 is calculated. The input parameters include: the volume concentration of the added TiO s , the refractive index of TiO 2 (2.52) and the wavelength values (200 - 1000 nm); 2 a

[0078] Finally, the μ a and μ' s reference values of 50 phantoms in the wavelength range of 200 - 1000 nm are measured and calculated. As Figure 7 shown, it can be found from the figure that the reference values have an obvious positive correlation with the concentrations of the added India Ink and TiO 2 at each wavelength

[0079] The transmission intensity and reflection intensity of air, a standard diffuse reflection plate (IOC - CW - 30, Wuling Optics) and one of the phantoms (India Ink: 0.010%, TiO2: 0.020%) are detected at different time points by the double - integrating - sphere technique. As Figure 8 shown, the measured air transmission and the reflection intensity spectrum of the standard diffuse reflection plate are presented. It can be found from the figure that the measured reflection and transmission intensities at different time points basically coincide, indicating good repeatability. Further, the coefficient of variation of the transmission intensity and reflection intensity at each wavelength is calculated to verify the stability of the platform Figure 9 The coefficient of variation of the air transmission intensity, the reflection intensity of the standard diffuse reflection plate, and the transmission and reflection intensities of the phantom at different time points are shown. It can be found from the figure that the coefficient of variation is less than 0.05 in the full wavelength range of 200 - 1000 nm, indicating good stability of the double - integrating - sphere system

[0080] By combining the double - integrating - sphere technique with the inverse doubling algorithm, the measured values of μ a and μ' s are calculated according to the transmittance and reflectance of the phantom. As Figure 10 shown, the measured and calculated μ a of the first group of phantoms and the μ' s measured values of the second group of phantoms in the wavelength range of 200 - 1000 nm are presented. It can be found from the figure that the calculated values have a certain relationship with the added India Ink and TiO 2The concentration has an obvious positive correlation at most wavelengths (400 - 1000). The noise is relatively large at 200 - 400 nm due to the low signal-to-noise ratio of the spectrometer. However, comparing Figure 7 with Figure 10 it can be found that μ a , μ’ s The measured value spectrum and μ a , μ’ s The waveform and numerical differences of the reference value spectra are relatively large. Taking the concentrations of IndiaInk and TiO 2 as independent variables, and the measured values of μ a and μ’ s as dependent variables, calculate the correlation coefficients at each wavelength to verify the linearity of the platform. The calculation results of the correlation coefficients are as Figure 11 shown. It can be found from the figure that the correlation coefficients at most wavelengths (400 - 1000 nm) are close to 1, indicating that the linearity of the double integrating sphere system is relatively high. It is expected to improve the detection accuracy of the system for μ a and μ’ s through further calibration.

[0081] Compare the measured values and reference values of the phantom μ a , μ’ s Calculate the relative errors at each wavelength. The calculation results are as Figure 12 shown. It can be found from the figure that the relative errors between the uncalibrated measured values of μ a , μ’ s and their reference values in the non-noise region (400 - 1000 nm) are extremely large, reaching 300% and 1400% respectively. However, due to the relatively high linearity of the platform measured values, it is expected to calibrate them through polynomial curve fitting to improve the accuracy. Calibrate μ a and μ’ s respectively through cubic and first-order (linear) polynomial curve fittings. The calibration formulas are as follows:

[0082] μ a,ref = a 1 ·μ a,mea 3 + a 2 ·μ a,mea 2 + a 3 ·μ a,mea + b 1

[0083] μ' s,ref = k·μ' s,mea + b 2

[0084] where μ a,ref and μa,mea are μ respectively a reference value and measured value, μ’ s,ref and μ’ s,mea are μ’ respectively s reference value and measured value, a 1 、a 2 、a 3 、b 1 and k, b 2 are μ respectively a and μ’ s correction parameters. Further, μ a and μ’ s The correction value and the reference value are compared, and the relative error at each wavelength is calculated as a parameter to finally verify the accuracy of the platform. The calculation results are as Figure 13 shown. It can be found from the figure that μ a and μ’ s After correction, the relative error with the reference value is greatly reduced. In the non-noise area, the highest relative error is less than 15%, and the highest relative error all appears in the phantom with the lowest concentration of India Ink and TiO 2 . This is because when the concentrations of the two components are the lowest, their μ a and μ’ s reference values are also the smallest. Therefore, when the absolute error is small, the calculated relative error is larger.

[0085] (4) Optical feature acquisition

[0086] Extract μ a and μ’ s from the spectra of the vegetable oil samples at the excitation wavelength (375 nm) and fluorescence emission wavelength (424 nm) of AFB 1 . According to the correction parameters obtained in the previous phantom correction process, correct and calculate the extracted μ a and μ’ s values, and further calculate the diffusion coefficients (D a , D s ) at the two wavelengths according to the following formula: ex 、D em ):

[0087]

[0088] where μ’ t,ex and μ’ t,em are the corrected values of the transport coefficients at 375 and 424 nm respectively, μ’ s,ex and μ’ s,em are the μ’ s corrected values at 375 and 424 nm respectively, μ a,ex and μa,em μ at 375 and 424 nm respectively a Correction value.

[0089] Extract the fluorescence intensity value (F) at 424 nm from the fluorescence intensity spectrum of the vegetable oil sample. mea ).

[0090] The 128 vegetable oil samples were divided into a calibration group (96) and a validation group (32) in a ratio of 3:1. ex , D em , μ a,ex , μ' s,ex , μ a,em 、F mea As the model input value, each sample is added with AFB 1 The concentration of F int,AFB1 As the output value, the input value and output value are normalized respectively, and the one-dimensional convolutional neural network is used to calibrate the model. 1 The convolutional neural network structure for intrinsic fluorescence recovery is as follows Figure 14 As shown in the figure, it includes: 1 input layer, 2 convolutional layers, 2 batch normalization layers, 2 ReLU layers, 1 pooling layer, 1 dropout layer, 1 fully connected layer, and 1 regression output layer. The convolution kernel size is 3*1, and the number of convolution kernels is 16 and 100 respectively. The pooling layer is the maximum pooling. The optimization algorithm is the ADAM algorithm.

[0091] The D of the 32 validation group samples ex , D em , μ a,ex , μ' s,ex , μ a,em 、F mea After normalization, the AFB in the validation group samples was calculated. 1 Intrinsic fluorescence recovery value (F' int,AFB1 ). AFB measured by the calibration group and validation group phantom 1 Unrecovered fluorescence vs. theoretical intrinsic fluorescence value, intrinsic fluorescence recovery value vs. theoretical intrinsic fluorescence value Figure 15 As shown, from the comparison of the left and right sub-images, it can be found that before the intrinsic fluorescence recovery, the same AFB 1 The fluorescence intensity of vegetable oil samples with different concentrations and matrices varies greatly, which results in the lack of fluorescence intensity when using only AFB. 1 AFB 1 Concentration prediction is simply not feasible. After fluorescence recovery, the intrinsic fluorescence recovery value is basically consistent with the theoretical value, indicating that the proposed intrinsic fluorescence recovery scheme has an excellent recovery effect.

[0092] According to the previous AFB1 The measurement results of the working fluid are based on AFB 1 The theoretical value of intrinsic fluorescence is the independent variable, with AFB 1 The actual concentration is the dependent variable, and the AFB 1 Intrinsic fluorescence AFB 1 The concentration calculation formula is:

[0093] c AFB1 =0.0467·F' int,AFB1 -58.925

[0094] When this formula is applied in practice, F' int,AFB1 is the intrinsic fluorescence recovery value, c' AFB1 For AFB 1 The concentration was calculated. The F' of the calibration group and the validation group int,AFB1 Substituting the above formula, we can calculate AFB 1 The AFB of the validation group was calculated based on the intrinsic fluorescence recovery value. 1 Content vs AFB 1 Reference value of content Figure 16 shown.

[0095] Correction Group AFB 1 The correlation coefficient (R c ) is 0.978, the root mean square error (RMSEC) is 4.609, and the validation group AFB 1 The correlation coefficient (R v ) was as high as 0.965, the root mean square error (RMSEV) was as low as 4.611, and the relative analytical deviation (RPD) was 3.780, indicating that the one-dimensional convolutional neural network proposed in the present invention can be used to identify AFB in complex vegetable oil matrices. 1 The intrinsic fluorescence recovery method can effectively recover the distorted fluorescence from the continuously changing complex matrix. The recovered fluorescence can be used for AFB in plant oil. 1 The model is simple and reliable for accurate prediction of content.

[0096] The above examples are only specific embodiments of the present invention. Obviously, the present invention is not limited to the above examples, and many variations are possible. All variations that can be directly derived or associated with the contents disclosed by a person skilled in the art should be considered as the protection scope of the present invention.

Claims

1. A method for restoring the intrinsic fluorescence of AFB1 in a complex matrix of vegetable oil based on a one-dimensional convolutional neural network, characterized in that: include: Obtain plant oil samples with different matrix gradients and containing different concentration gradients of AFB1 contamination; The absorption, scattering and fluorescence spectra of the vegetable oil samples were obtained through the optical platform, and the fluorescence intensity of the AFB1 standard solution at different concentrations was measured as the intrinsic fluorescence; The repeatability, linearity and accuracy of the optical platform are verified and calibrated through the phantom to obtain the correction parameters of the optical platform absorption and scattering; The absorption coefficient, reduced scattering coefficient, diffusion coefficient of AFB1 at the excitation wavelength and the absorption coefficient and diffusion coefficient at the fluorescence emission wavelength were extracted, corrected and calculated; According to the above five parameters and the measured distorted fluorescence intensity, an intrinsic fluorescence regression model was established using a one-dimensional convolutional neural network to calculate the restored AFB1 characteristic fluorescence; Based on the restored AFB1 characteristic fluorescence, the quantitative calculation formula of AFB1 content was obtained by linear regression analysis.

2. The method for restoring the intrinsic fluorescence of AFB1 in a complex matrix of vegetable oil based on a one-dimensional convolutional neural network according to claim 1, characterized in that: The vegetable oil sample acquisition process includes: 1.1) Four vegetable oils with different intrinsic fluorescent components and pigments (vitamin E, polyphenols, carotenoids, chlorophyll) in the matrix were selected from the market as crude oils; 1.2) Mix the four crude oils separately to obtain oil samples with more gradient matrix; 1.3) Prepare AFB1 stock solution (10000 μg kg -1 ); 1.4) Add 0, 9, 18, 36, 72, 108, 144, and 180 μL of AFB1 stock solution to each oil sample by pipette and stir evenly to prepare AFB1 contamination levels of 0, 2.5, 5, 10, 20, 30, 40, and 50 μg kg -1 of vegetable oil.

3. The method for restoring the intrinsic fluorescence of AFB1 in a complex matrix of vegetable oil based on a one-dimensional convolutional neural network according to claim 1, characterized in that: The absorption, scattering and fluorescence spectrum acquisition, AFB1 intrinsic fluorescence acquisition process includes: 2.1) The absorption coefficient (μ) is calculated based on the transmittance and reflectivity of the oil sample by combining the double integrating sphere technique with the reverse multiplication algorithm. a ) and the reduced scattering coefficient (μ' s ) spectrum, obtaining the fluorescence intensity spectrum of the oil sample by laser induced fluorescence spectroscopy; 2.2) Pipette 1 mL of AFB1 mother solution into a 100 mL volumetric flask, add acetonitrile to make up to volume, and obtain 100 μg kg -1 AFB1 solution was taken in different volumes of 100 μg kg -1 AFB1 solution was added to a 25 mL volumetric flask and fixed to volume with acetonitrile to obtain 0, 0.1, 1, 10, 20, 30, 40, and 50 μg kg -1 AFB1 working solution, and the fluorescence intensity of AFB1 working solution was obtained by laser induced fluorescence spectroscopy as the intrinsic fluorescence of different concentrations. Through linear fitting, the AFB1 concentration (c' AFB1 ) is the independent variable, and the fluorescence intensity at 424 nm (F int,AFB1 ) is the dependent variable, and the calculation formula of AFB1 intrinsic fluorescence is obtained: F int,AFB1 =21.405·c' AFB1 +1261.3 c' AFB1 =0, 2.5, 5, 10, 20, 30, 40, 50μg·kg -1 Substituting into the formula, the theoretical value of intrinsic fluorescence of AFB1 in oil sample (F int,AFB1 ).

4. The method for restoring the intrinsic fluorescence of AFB1 in a complex matrix of vegetable oil based on a one-dimensional convolutional neural network according to claim 1, characterized in that: The optical platform repeatability, linearity and accuracy verification and calibration process includes: 4.1) preparing 2 groups of liquid phantoms comprising India Ink and TiO2, wherein the volume concentrations of India Ink in group 1 were 0.005%, 0.010%, 0.020%, 0.040%, and 0.080%, and the volume concentration of TiO2 was 0.020%, and the volume concentration of TiO2 in group 2 were 0.004%, 0.008%, 0.012%, 0.016%, 0.020%, 0.024%, 0.028%, 0.032%, 0.036%, and 0.040%, and the volume concentration of India Ink was 0.020%, and a total of 15 phantoms were prepared; 4.2) The collimated transmittance of pure India Ink solution at 5 concentrations was measured, and the μ of India Ink at different wavelengths and concentrations was calculated according to the Lambert-Beer law (as shown below): a Reference value, μ a,ref =lg(1 / T) / b Where T is the measured collimated transmittance, b is the optical path length of the cuvette; Using Mie algorithm, we calculated the μ' of TiO2 at different wavelengths and concentrations. s Reference value; 4.3) The transmission intensity and reflection intensity of the standard diffuse reflector and one of the phantoms (IndiaInk: 0.010%, TiO2: 0.020%) were tested at different time points using the double integrating sphere technology, and the coefficient of variation of the transmission intensity and reflection intensity at each wavelength was calculated to verify the stability of the platform; 4.4) The double integrating sphere technique combined with the reverse multiplication algorithm is used to calculate μ according to the transmittance and reflectivity of the phantom. a With μ' s The measured values ​​are India Ink and TiO2 concentration as independent variables, μ a With μ' s The measured value is the dependent variable, and the correlation coefficient at each wavelength is calculated to verify the linearity of the platform; 4.5) The phantom μ a , μ' s The measured value was compared with the reference value, and the relative error at each wavelength was calculated. The μ a With μ' s Correction is performed to obtain μ a With μ' s Correction parameters and μ a With μ' s The correction value is compared with the reference value, and the relative error at each wavelength is calculated as the final parameter for verifying the accuracy of the platform.

5. The method for restoring the intrinsic fluorescence of AFB1 in a complex matrix of vegetable oil based on a one-dimensional convolutional neural network according to claim 1, characterized in that: The excitation wavelength and fluorescence emission wavelength of AFB1 are 375 and 424 nm, respectively. ex ) and 424nm(D em ) are calculated as follows: where μ' t,ex With μ' t,em are the transmission coefficient correction values ​​at 375 and 424 nm, μ' s,ex With μ' s,em are the reduced scattering coefficient correction values ​​at 375 and 424 nm, μ a,ex With μ a,em These are the absorption coefficient correction values ​​at 375 and 424 nm, respectively.

6. The method for restoring the intrinsic fluorescence of AFB1 in a complex matrix of vegetable oil based on a one-dimensional convolutional neural network according to claim 1, characterized in that: The one-dimensional convolutional neural network includes: 1 input layer, 2 convolutional layers, 2 batch normalization layers, 2 ReLU layers, 1 pooling layer, 1 discard layer, 1 fully connected layer, and 1 regression output layer; the input layer has 6 input parameters, namely: absorption coefficient correction value at 375nm, reduced scattering coefficient correction value, diffusion coefficient, absorption coefficient correction value at 424nm, diffusion coefficient and fluorescence intensity; the convolution kernel size is 3*1, and the number of convolution kernels is 16 and 100 respectively; the pooling layers are all maximum pooling; and the optimization algorithm is the ADAM algorithm.

7. The one-dimensional convolutional neural network according to claim 5, characterized in that: Before model training, all oil samples were divided into a calibration group and a verification group in a ratio of 3:

1. The calibration group was used to train the one-dimensional convolutional neural network model, and the verification group was used to verify the model effect. The input parameters and the true value of the intrinsic fluorescence of the oil sample were normalized before being used for model training. When the model training was completed and the calculated value was output, the calculated value of the intrinsic fluorescence was denormalized.

8. The method for restoring the intrinsic fluorescence of AFB1 in a complex matrix of vegetable oil based on a one-dimensional convolutional neural network according to claim 1, characterized in that: The process of obtaining the simple linear calculation formula for AFB1 concentration is as follows: through linear fitting, the theoretical value of AFB1 intrinsic fluorescence is used as the independent variable, and the actual concentration of AFB1 is used as the dependent variable, and the calculation formula for AFB1 concentration based on AFB1 intrinsic fluorescence is obtained as follows: c AFB1 =0.0467·F' int,AFB1 -58.925 When this formula is applied in practice, F' int,AFB1 is the intrinsic fluorescence recovery value, c' AFB1 Calculated value for AFB1 concentration.

Images