Condensed milk flavor optimization method based on multilevel flavor characteristic analysis

Abstract

The invention provides a condensed milk flavor optimization method based on multilevel flavor feature analysis, and the method comprises the steps: carrying out the extraction of the contour of a fat globule through an edge detection algorithm for a segmented fat globule binary image, obtaining a contour curve of the fat globule, calculating the average curvature of the surface of the fat globule according to the curvature change of the contour curve, and obtaining the flavor of the condensed milk. Curvature data representing the shape of the fat globule is obtained; for the segmented fat globule binary image, extracting a membrane region of the fat globule by adopting a morphological processing algorithm, calculating the average membrane thickness of the fat globule according to the pixel number of the extracted membrane region, and obtaining membrane thickness data representing fat globule membrane characteristics; the method comprises the following steps: acquiring fat globule membrane component detection data of different condensed milk samples, including membrane protein content and phospholipid content, and calculating the proportion of protein and phospholipid in the fat globule membrane according to the membrane protein content and the phospholipid content to obtain protein-phospholipid proportion data representing the composition of the fat globule membrane.

Description

Technical Field

[0001] The invention relates to the field of information technology, and in particular to a condensed milk flavor optimization method for multi-level flavor characteristic analysis. Background Art

[0002] The flavor of condensed milk is not simply sweet, but a multi-layered and complex experience, which is closely related to the microstructure of its fat globules. However, there are many factors that affect the morphology of fat globules that affect the flavor of condensed milk, and they are interrelated. How to determine the best morphology has become a difficult problem. First, the size of the fat globule directly affects its surface area. Smaller fat globules have a larger specific surface area, which can theoretically release more encapsulated flavor substances, but at the same time, they are more prone to fat oxidation and produce bad flavors. Larger fat globules may lead to insufficient release of flavor substances and a heavy taste. Secondly, fat globules are not perfect spheres, and changes in their surface curvature will affect the structure and stability of the membrane. In areas with different curvatures, the thickness and composition of the membrane may be different, which will affect the adsorption and release of flavor substances. Thirdly, the thickness and composition of the fat globule membrane are also important influencing factors. The thickness of the membrane determines the diffusion rate of flavor substances, while the composition of the membrane, especially the ratio of protein and phospholipids, affects the permeability of the membrane and the adsorption capacity of different flavor substances. A higher protein ratio may improve the stability of the membrane, but it may also hinder the release of flavor substances. How to regulate the membrane thickness and protein-phospholipid ratio to achieve optimal flavor release remains a challenge. Summary of the invention

[0003] The present invention provides a condensed milk flavor optimization method for multi-level flavor characteristic analysis, which mainly comprises:

[0004] Obtain fat globule image data of different condensed milk samples, extract fat globule regions from fat globule images using image segmentation algorithms, obtain segmented fat globule binarization images, calculate equivalent diameters of fat globules based on pixel areas of fat globules in the binarization images, and obtain size distribution data of fat globules;

[0005] For the fat globule binary image after segmentation, the edge detection algorithm is used to extract the contour of the fat globule to obtain the contour curve of the fat globule. According to the curvature change of the contour curve, the average curvature of the fat globule surface is calculated to obtain the curvature data characterizing the shape of the fat globule;

[0006] The morphological processing algorithm is used to extract the membrane area of ​​the fat globule from the segmented fat globule binary image. The average membrane thickness of the fat globule is calculated based on the number of pixels in the extracted membrane area to obtain the membrane thickness data that characterizes the fat globule membrane characteristics.

[0007] Obtaining the fat globule membrane component detection data of different condensed milk samples, including membrane protein content and phospholipid content, calculating the ratio of protein to phospholipid in the fat globule membrane based on the membrane protein content and phospholipid content, and obtaining protein-phospholipid ratio data characterizing the composition of the fat globule membrane;

[0008] Obtain flavor characteristic evaluation data of different condensed milk samples, including flavor hierarchy scores, use the flavor hierarchy scores as sample label data, use fat globule size distribution, surface curvature, membrane thickness, and membrane component ratio as sample feature data, and construct a flavor hierarchy prediction model;

[0009] The flavor hierarchy prediction model is trained by using a support vector machine algorithm to obtain a trained flavor hierarchy prediction model, and the fat globule size distribution, surface curvature, membrane thickness, and membrane component ratio data of the condensed milk sample to be judged are input into the trained prediction model to obtain the flavor hierarchy prediction score of the sample;

[0010] According to the flavor hierarchy prediction scores, the flavor hierarchy of the samples is sorted to obtain the sample ranking results from high to low flavor hierarchy. According to the sorting results, the combination of fat globule size distribution, surface curvature, membrane thickness, and membrane component ratio corresponding to the sample with the highest flavor hierarchy is determined as the fat globule morphological feature combination with the highest flavor characteristic level.

[0011] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0012] The present invention discloses a method for optimizing the flavor of condensed milk by multi-level flavor feature analysis. The method obtains morphological feature data such as the size distribution, surface curvature and membrane thickness of fat globules in condensed milk samples through image analysis, and constructs a flavor hierarchy prediction model in combination with fat globule membrane component detection data. The specific steps include: segmenting and feature extraction of fat globule images, calculating the equivalent diameter, surface curvature and membrane thickness of fat globules; detecting the protein and phospholipid content ratio of fat globule membranes; using these feature data and flavor hierarchy scores as training data, and constructing a prediction model using a support vector machine algorithm. The model can predict the flavor hierarchy of the sample to be tested, and sort the samples to determine the combination of fat globule morphological features with the highest flavor hierarchy. The present invention realizes the objective quantitative evaluation of the flavor hierarchy of condensed milk, provides a scientific basis for the quality improvement of condensed milk and the development of new products, and helps to improve the flavor quality of condensed milk. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 The present invention is a flow chart of a condensed milk flavor optimization method for multi-level flavor characteristic analysis.

[0014] Figure 2The present invention is a schematic diagram of a condensed milk flavor optimization method for multi-level flavor characteristic analysis.

[0015] Figure 3 It is another schematic diagram of the condensed milk flavor optimization method of the present invention through multi-level flavor characteristic analysis. DETAILED DESCRIPTION

[0016] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0017] like Figure 1-3 The condensed milk flavor optimization method of the present embodiment with multi-level flavor characteristic analysis may specifically include:

[0018] S101, obtaining fat globule image data of different condensed milk samples, extracting fat globule regions from the fat globule images using an image segmentation algorithm, obtaining segmented fat globule binarization images, calculating equivalent diameters of the fat globules based on pixel areas of the fat globules in the binarization images, and obtaining size distribution data of the fat globules.

[0019] An image processor is used to obtain image data of condensed milk samples with calibration plate scales, and a conversion relationship between pixel points and actual sizes is established according to the calibration plate scales to obtain a calibration image; Gaussian low-pass filtering is performed on the calibration image to remove background noise, and a multi-level grayscale threshold is calculated using an OTSU algorithm, and an image contour boundary is extracted according to the grayscale threshold to obtain a boundary image; a seed point growth process is performed according to the boundary image, and if the grayscale value difference in the overlapping area exceeds the standard deviation of the neighborhood average, a watershed algorithm is used to divide the independent blocks to obtain a segmented image; fat globule contour area data is extracted from the segmented image, and the actual area value is calculated according to the conversion relationship between the pixel points and the actual size, and the particle size distribution result is obtained by calculating the equivalent diameter using a circular area formula.

[0020] Specifically, an image processor is used to obtain condensed milk sample image data by comparing with a standard calibration plate, a histogram equalization operation is performed on the image data, a Gaussian low-pass filter is performed on the central area of ​​the image to remove background noise points, a pixel point and actual size conversion relationship is established according to the scale size of the calibration plate, and enhanced processing data is obtained from the filtered image. The OTSU algorithm is used to iteratively calculate the multi-level grayscale threshold according to the enhanced data, and the contour boundary of the independent area after threshold segmentation is extracted. The boundary curvature value is calculated by the directed gradient function, and the edge integrity of the fat globule is judged according to the curvature value, and a binary image is generated in the edge complete area. The seed point growth process is performed on the binary image, and the regional grayscale value difference threshold is set as the neighborhood average standard deviation. The overlapping area is divided into independent blocks by the watershed algorithm, and the holes in the area are filled by the closing operation. The small area noise area is removed by the opening operation, and the pixel set of the independent area is marked. The contour area of ​​each fat globule was extracted from the set of marked areas. The actual area value was calculated using the size conversion relationship established by the calibration plate. The equivalent diameter value was obtained according to the circular area formula. The frequency distribution histogram was established for the obtained data, and the fat globule size distribution results were obtained from the histogram data. The calibration plate is a benchmark size conversion tool with a standard millimeter scale. The correspondence between the pixel points and the actual size was established by imaging the calibration plate and the condensed milk sample at the same time. Each 1 mm scale of the ruler corresponds to 25 pixels in the image. The surface coating of the calibration plate is matte black, which is in sharp contrast with the condensed milk sample. Histogram equalization adjusts the grayscale distribution of the image to make the grayscale value distribution more uniform and improve the image contrast. The grayscale values ​​of the original image are concentrated in the range of 120 to 180. After equalization, the grayscale values ​​are evenly distributed in the range of 0 to 255, which enhances the contrast between the edge of the fat globule and the background. The Gaussian low-pass filter uses a 5×5 template to remove high-frequency noise while maintaining the integrity of the edge contour of the fat globule. The OTSU algorithm calculates the optimal segmentation threshold through the maximum inter-class variance criterion, and performs multi-level iterative segmentation on the grayscale image. Each iteration calculates the average grayscale value and inter-class variance of the two subclasses. When the inter-class variance reaches the maximum value, the segmentation threshold is determined, and the threshold is 135. The directional gradient calculates the boundary curvature, and the curvature value greater than 0.85 is judged as a complete edge. The seed point growth selects the initial growth point from the central area of ​​the image, and the neighborhood grayscale difference is less than 15 as the growth criterion to achieve regional segmentation. The watershed algorithm processes overlapping fat globules and obtains independent blocks at the watershed line of the overlapping area. The closing operation fills the holes less than 50 square pixels in the area, and the opening operation removes the noise area less than 100 square pixels. The contour area of ​​the fat globule is obtained by counting the number of pixels in each marked area, and the pixel area is converted into the actual area value in combination with the calibration relationship. 25×25 pixels correspond to 1 square millimeter of actual area. The equivalent diameter is calculated by the circular area formula. Statistics show that the diameter of the fat globules is mainly distributed in the range of 2 to 8 microns, of which 4 to 6 microns accounts for 65%, constituting the main particle size distribution range.In practical applications, there are differences in the distribution of fat globules of different types of condensed milk samples. The fat globules of whole-fat condensed milk are larger in size and have a wide distribution range, while the fat globules of skim condensed milk are smaller in size and have a concentrated distribution. By quantitatively analyzing the size distribution characteristics of fat globules, the quality evaluation and classification of condensed milk can be achieved. The image processing parameter settings need to be adjusted according to the sample characteristics to ensure the accuracy of the segmentation results. A complete image acquisition and processing process plays an important role in obtaining reliable particle size distribution data.

[0021] S102, for the segmented fat globule binary image, use an edge detection algorithm to extract the contour of the fat globule to obtain the contour curve of the fat globule, calculate the average curvature of the fat globule surface according to the curvature change of the contour curve, and obtain curvature data characterizing the shape of the fat globule.

[0022] The Sobel operator is used to extract edge gradients in the horizontal and vertical directions of the binary image, and an edge intensity distribution map is obtained through iterative calculation based on the horizontal gradient and the vertical gradient; a refinement iterative operation is performed on the edge intensity distribution map, and the unclosed points are connected and compensated by a connectivity propagator to obtain a closed contour line; a point index sequence is constructed according to the closed contour line, and a three-point arc fitting function is used to calculate the curvature value of the point index sequence, and a smooth curvature value sequence is obtained through a Gaussian filter; if the curvature of a point in the smooth curvature value sequence is greater than zero, it is marked as a convex point, and if the curvature of the point is less than zero, it is marked as a concave point, and the average curvature result is obtained through a weighted operation of the extreme point curvature.

[0023] Specifically, the 3×3 Sobel edge detection operator is used to extract edge gradients in the horizontal and vertical directions of the fat globule binary image. The edge intensity distribution map is synthesized according to the horizontal and vertical gradients. The edge intensity threshold is determined by iteratively calculating the maximum inter-class variance in the map, and the points with edge intensity lower than the threshold are suppressed to generate an edge point map. The edge line width is compressed to a single pixel from the edge point map through a refinement iterative calculator. A 5×5 window scanner is used to detect the connectivity attribute value of each edge point for the refined edge map. The closure of the line segment where the point is located is judged according to the connectivity attribute value. The connectivity propagator is used to perform connection compensation on the unclosed points to generate a closed contour line. A point index sequence is constructed in a counterclockwise direction for the closed contour line. The curvature calculation is performed in the point sequence with 5 consecutive points as the window unit. The curvature value at the sequence point is obtained by a three-point arc fitting function. The curvature value sequence is smoothed by a 7-point Gaussian filter to obtain a smooth curvature value sequence of the contour curve. According to the sequence of smooth curvature values, the curvature of each point is marked as a convex point or a concave point. The contour curve is divided into sub-segments according to the marking results. The curvature values ​​in the sub-segments are supplemented with continuous curvature distribution by cubic spline interpolation. The coordinates of the local maximum and minimum points are extracted from the continuous curvature values. The average curvature result is obtained by weighted calculation of the curvature of the extreme points. The Sobel edge detection operator detects the horizontal and vertical gradients respectively through a 3×3 template. The horizontal template calculates the grayscale difference in the row direction of the image point by point, and the vertical template calculates the grayscale difference in the column direction. The square root of the sum of the squares of the gradient values ​​in the two directions is obtained. The intensity value at the edge of the fat globule is significantly higher than that in the background area. The edge intensity threshold is set between the background noise and the target edge. When iteratively calculating the maximum inter-class variance, the initial threshold is set to the edge intensity mean, and the image is divided into two categories: above the threshold and below the threshold. The average intensity value of the two categories of areas is calculated, and the new threshold is taken as the midpoint of the two average values. It is repeated until the threshold change is less than 0.1, and the final threshold is between 150 and 200 edge intensity values. During the refinement iteration process, the 8-neighborhood pixels of each edge point are scanned. If the number of non-zero pixels in the 8-neighborhood of the edge point is between 2 and 6, the edge points that meet the conditions are deleted, and the deletion operation is repeatedly performed on the edge map until the width of all edge lines is reduced to a single pixel. The connectivity of the edge point is determined by the pixel distribution in the 5×5 window. If there are more than two branch points in the neighborhood of the center point of the window, it means that the point is an endpoint. The connectivity propagation starts from the endpoint and extends to the surrounding areas. The search radius is 3 pixels. The nearest other endpoint is found to establish a connection and compensate for the disconnected contour segments. The closure of the contour line requires that all endpoints are paired and connected to form a closed curve. The point index sequence starts from any point on the contour line and records the coordinate values ​​in the counterclockwise direction. The serial number increases from 0. The curvature calculation window contains 5 consecutive points. The curvature of the center point is obtained by fitting a three-point arc. The arc is fitted by the three-point least squares method. The curvature value is the inverse of the arc radius.Gaussian filtering uses a 7-point weight template, with a central point weight of 0.4, and the weights of the two sides are 0.25, 0.15, and 0.05, respectively. The filter eliminates mutation values ​​in the curvature sequence. The curvature of the convex point is positive, and the curvature of the concave point is negative. The boundary between positive and negative curvatures is divided into sub-segments. The curvature within the sub-segment is continuously distributed through cubic spline interpolation, and the number of interpolation points is 3 times the number of original points. The local extreme point corresponds to the peak position of the absolute value of the curvature. The curvature of the maximum point exceeds 0.5, and the curvature of the minimum point is less than -0.5. The average curvature is calculated by weighted average of the extreme point curvatures. The weight is proportional to the absolute value of the curvature, reflecting the overall shape characteristics of the fat globule.

[0024] S103, extracting the membrane region of the fat globules using a morphological processing algorithm for the segmented fat globule binary image, calculating the average membrane thickness of the fat globules according to the number of pixels in the extracted membrane region, and obtaining membrane thickness data characterizing the membrane characteristics of the fat globules.

[0025] A circular structure corrosion operator is used to corrode the edges of the fat globule region in the binary image, and a membrane region map is obtained through a closing operation. The point with the maximum grayscale value in the membrane region map is used as a seed point, and a regional growth extractor is used to expand it in all directions, and a skeleton line of the membrane region is obtained through distance transformation. The skeleton line of the membrane region is scanned inwards and outwards to obtain the inner and outer boundary coordinate points of the membrane region contour. The inner and outer boundary coordinate points are measured along the reference direction to obtain a local membrane thickness sequence. If the sampling interval of the local membrane thickness sequence is a preset value, the average membrane thickness value is obtained based on Gaussian weighted averaging and Kalman filtering.

[0026] Specifically, the fat globule region is eroded by a 5×5 circular structure erosion operator on the binary image. The degree of erosion is determined by the ratio of the area of ​​the eroded region to the original region. The edge gap is filled by a 7×7 circular structure closing operation. The original region is subtracted from the eroded region to obtain a ring difference map. A 3×3 structure opening operation is performed on the difference map to remove independent noise points to obtain a membrane region map. The point with the maximum gray value is selected as the seed point in the membrane region map, and the region growth extractor is used to expand it around. The growth threshold is set as the gray standard deviation of the 8 neighborhoods of the seed point. The 5×5 structure closing operation is used to connect the breakpoints in the growth region. The skeleton line of the membrane region is extracted by distance transformation, and the complete membrane contour is obtained by scanning from the skeleton line to the inside and outside. The inner and outer boundary coordinate points are extracted according to the membrane region contour. The direction of the line between the corresponding points is used as the thickness measurement reference. The distance between the two points is measured at a 90-degree angle along the reference direction to obtain the local membrane thickness. The thickness value is filtered by a bilateral filter with a spatial scale of 5 pixels and a gray scale of 0.3. The jump points are removed by a 7-point sliding median filter to obtain a smooth thickness sequence. The smooth thickness sequence was sampled at 10-pixel intervals, and the local film thickness value was calculated by Gaussian weighted averaging within 5 pixels of the sampling point. The local film thickness sequence was filtered by a Kalman filter with a position noise variance of 2 and a measurement noise variance of 0.5. The average film thickness value was obtained by taking the mean of the filtered data sequence. The corrosion operation was to reduce the pixel points from the edge of the fat globule inward. The 5×5 circular structure corresponded to an actual spatial scale of about 0.5 microns, which matched the typical thickness of the fat globule membrane. The area of ​​the region after corrosion was reduced by 15% to 20% compared with the original area, indicating that the edge corrosion achieved the expected effect. The closing operation was used to connect the disconnected edge points, and the 7×7 structure filled the gap less than 0.7 microns to maintain edge continuity. The annular difference map reflects the distribution characteristics of the membrane area. The grayscale value of the membrane area in the difference map is between 180 and 220, which is significantly higher than the background area. The opening operation removes noise points with an area less than 9 pixels to highlight the real membrane structure. The seed point is selected in the area with a grayscale value of more than 215 to ensure that the growth starts from the real membrane area. During the region growing process, the grayscale standard deviation of the 8-neighborhood is about 12, which can effectively control the region expansion range as the growth threshold.

[0027] The 5×5 closing operation connects the break intervals of 2 to 3 pixels, and the distance transformation calculates the shortest distance from each pixel to the background. The maximum distance point constitutes the skeleton line. The width of the skeleton line in the membrane area is kept at 1 pixel, reflecting the center position of the membrane. The distance between the inner and outer boundary points reflects the local membrane thickness. The measurement reference direction is perpendicular to the membrane extension direction. The bilateral filter smoothes the membrane thickness change while maintaining the edge jump. The spatial scale of 5 pixels corresponds to the actual distance of 0.5 microns, and the grayscale scale of 0.3 retains significant thickness changes. The median filter window selects 7 points to eliminate isolated thickness outliers. The sampling interval of 10 pixels ensures that the thickness measurement points are independent of each other while maintaining sufficient spatial resolution. The Gaussian weighted average calculates the local mean within a range of 5 pixels, and the weight decays exponentially with increasing distance. The position noise variance of the Kalman filter reflects the natural fluctuation of the membrane thickness, and the measurement noise variance corresponds to the instrument measurement error. In practical applications, there are differences in the distribution characteristics of membrane thickness of different types of fat globules. The membrane thickness uniformity of fresh samples is good, and the membrane thickness of samples with a long storage time is uneven. The structural integrity of fat globules can be judged by quantitative analysis of membrane thickness data. Image processing parameters need to be calibrated according to the microscope resolution and magnification to ensure the accuracy of the measurement results. The complete membrane area extraction and thickness measurement process plays an important role in evaluating the stability of fat globules.

[0028] S104, obtaining fat globule membrane component detection data of different condensed milk samples, including membrane protein content and phospholipid content, and calculating the ratio of protein to phospholipid in the fat globule membrane based on the membrane protein content and phospholipid content, to obtain protein-phospholipid ratio data characterizing the composition of the fat globule membrane.

[0029] A methanol-water gradient eluent is used to separate components through a solid phase extraction column, and the component separation obtains a protein absorbance signal and a phospholipid fluorescence intensity signal; baseline smoothing is performed according to the protein absorbance signal and the phospholipid fluorescence intensity signal, and corrected signal data is obtained through internal standard correction; the corrected signal data is filtered by a wavelet filter and a low-pass filter, and component content values ​​are obtained through normal distribution function fitting; the molar concentration is calculated based on the component content values ​​using a protein and phospholipid molecular weight conversion coefficient, and the protein-phospholipid molar ratio is obtained through variance weighted averaging.

[0030] Specifically, the condensed milk sample was separated into components by an 18-carbon octyl silica gel solid phase extraction column, and a linear gradient elution of methanol-water concentration from 20% to 90% was used. The protein peak absorbance was measured at a wavelength of 280 nanometers, and the phospholipid fluorescence peak intensity was measured under the conditions of an excitation wavelength of 485 nanometers and an emission wavelength of 520 nanometers. The peak area was calculated by a trapezoidal integration method, and the five-point smoothing method was used to eliminate matrix interference from the integral baseline. A standard curve with a concentration of 0.1 to 1.0 mg / ml was prepared from a bovine serum albumin standard, and a standard curve with a concentration of 0.05 to 0.5 mg / ml was prepared from a fluorescently labeled phosphatidylcholine standard for phospholipids. The correlation coefficient of the standard curve was judged, and the judgment threshold was set to 0.995. The sample matrix effect was corrected by the internal standard method, and the concentration of each component in the sample was obtained from the calibration curve. A 4th-order wavelet filter was used to eliminate background noise for the detection signal, and a low-pass filter with a cutoff frequency of 20 Hz was used to filter out high-frequency interference. The filtered data was fitted with a normal distribution function, and the chi-square test was used to determine the consistency of the data distribution. The mean values ​​of the protein and phospholipid component contents were obtained from the data distribution. The protein-phospholipid mass ratio was converted according to the mean value of the component content, and the molar concentration was calculated using the conversion coefficient of 65,000 protein molecular weight and 7500 phospholipid molecular weight. The calculated results were weighted by variance, and the final protein-phospholipid molar ratio was obtained from the weighted results. Solid phase extraction uses an 18-carbon octyl silica gel column to separate proteins and phospholipids. The silica gel surface is hydrophobic and interacts with the hydrophobic region of the protein. The hydrophobic tail chains in the phospholipid molecules are also combined with the fixed structure. The hydrophilicity of the components is changed by adjusting the methanol concentration in the eluent to achieve selective elution. Gradient elution starts from 20% methanol, increases by 5% per minute, reaches 90% methanol concentration in 90 minutes, protein is eluted in the range of 45% to 65% methanol, phospholipids are eluted in the range of 70% to 85% methanol, and the elution intervals of the two components differ by more than 5% to avoid peak overlap. The wavelength of 280 nanometers corresponds to the characteristic absorption of tryptophan residues in proteins. During fluorescence detection, the 485-nanometer excitation light causes the phospholipid labeling group to produce electronic transitions, and the emission light is collected at 520 nanometers. In the preparation of the standard curve, 5 concentration points of 0.1, 0.3, 0.5, 0.7, and 1.0 mg / ml were selected for bovine serum albumin, and 5 concentration points of 0.05, 0.1, 0.2, 0.3, and 0.5 mg / ml were selected for phosphatidylcholine. Each concentration point was measured 3 times, and the peak area and concentration relationship curve was drawn. The correlation coefficient of 0.995 indicates that the concentration and response value have a good linear relationship. The amount of internal standard added is 20% of the sample concentration, which is used to correct the response deviation caused by the matrix effect. The 4th-order wavelet filter corresponds to 4 decomposition levels. The signal is decomposed into different frequency bands by high-pass and low-pass filters. After removing the high-frequency noise, the smooth signal is reconstructed. The 20 Hz low-pass filter removes the detector electronic noise and retains the component concentration change information. The kurtosis coefficient and skewness coefficient are obtained by normal distribution fitting. The p value in the chi-square test is greater than 0.05, which determines that the data conforms to the normal distribution.The ratio of protein molecular weight 65,000 to phospholipid molecular weight 7500 is about 87, which converts the mass ratio into a molar ratio. In variance weighted averaging, the weight is inversely proportional to the variance of the measured value, and data with small measurement errors receive a larger weight. In actual testing, the protein-phospholipid molar ratio of fresh samples is between 3.5 and 4.5, and the ratio decreases during storage due to protein degradation. The integrity of the fat globule membrane is closely related to the protein-phospholipid ratio, and a stable ratio indicates a stable membrane structure. Accurate quantification of membrane components is achieved by combining solid phase extraction with fluorescence detection.

[0031] S105. Obtain flavor characteristic evaluation data of different condensed milk samples, including flavor hierarchy scores, use the flavor hierarchy scores as label data of the samples, use fat globule size distribution, surface curvature, membrane thickness, and membrane component ratio as characteristic data of the samples, and construct a flavor hierarchy prediction model.

[0032] The characteristic data of the condensed milk sample are normalized by a normalizer to obtain normalized fat globule distribution values, surface curvature values, membrane thickness values ​​and membrane component ratio values; the Pearson correlation coefficient is calculated based on the normalized characteristic values, and the feature contribution is used to delete the feature combination with a correlation coefficient exceeding a threshold; a flavor hierarchy label vector is constructed for the data set after the features are deleted, and a stratified sampler is used to obtain training data and verification data; the feature importance value is calculated for the training data, and the features with a significance level lower than a preset threshold are selected by a forward stepwise regression method; a prediction model is established for the selected features, and the root mean square error and the determination coefficient are used to judge the accuracy of the prediction model on the verification data, and the prediction reliability level is determined according to the residual normality test result.

[0033] Specifically, the characteristic data of condensed milk samples were standardized from 0 to 1, and the fat globule size distribution value, surface curvature value, membrane thickness value, and membrane component ratio value were normalized by the normalizer. The Pearson correlation coefficient was used to calculate the correlation between features, and the features with low contribution in the feature combination with a correlation coefficient greater than 0.85 were deleted. Missing value and outlier detection were performed on the data set, and the data quality score was marked from the test results. The label vector was constructed according to the flavor hierarchy score, and the training data and validation data were divided into 7 to 3 ratios using a stratified sampler. The random forest parameters were optimized through 10-fold cross-validation, and the tree depth range was set to 5 to 15, the number of trees ranged from 50 to 200, and the feature sampling ratio was 0.6 to 0.9. The parameter combination was grid searched, and the accuracy of each parameter validation set was calculated. The Gini coefficient was used to calculate the feature importance for the training data, and the feature contribution was normalized. The features were selected by forward stepwise regression, and the inclusion criteria were set to a significance level of less than 0.05. The features with variance inflation factor greater than 5 were eliminated. The prediction model was built based on the selected features, and the model convergence was judged by the learning curve. The prediction operation was performed on the validation data, and the prediction accuracy was evaluated by the root mean square error and the coefficient of determination. The prediction confidence interval was calculated by the residual normality test, and the samples with prediction residuals exceeding 2 times the standard deviation were marked. The reliability level of the prediction results was determined in combination with the data quality score, and the sample data was recollected for the results with low reliability levels. The standardization process converted the features of different dimensions to the same scale. The original value of the fat globule size distribution was in the range of 2 to 8 microns, the curvature value was in the range of 0.5 to 1.5, the membrane thickness was in the range of 20 to 50 nanometers, and the protein-phospholipid ratio was between 3 and 5. After standardization, all features were mapped to the range of 0 to 1, which was convenient for the subsequent calculation of the correlation between features. The correlation coefficient reflects the degree of linear correlation between features, and its value ranges from -1 to 1. 0 indicates no correlation, and the larger the absolute value, the stronger the correlation. The correlation coefficient between fat globule size and membrane thickness is 0.87, which exceeds the threshold of 0.85. Through calculation, it is found that the membrane thickness feature contributes little to the prediction and is eliminated. The data quality score considers the proportion of missing values ​​and the number of outliers. The score ranges from 0 to 100. Data below 60 points need to be re-collected. Stratified sampling keeps the proportion of each type of sample in the training set and the validation set consistent. 70% of the samples are extracted from the original data set as the training set and 30% as the validation set. The 10-fold cross-validation divides the training data into 10 parts, and one of them is taken as validation, and the rest is used as training to evaluate the stability of the model. The tree depth controls the complexity of the random forest. Deeper trees are prone to overfitting, and shallower trees are underfitting. The optimal parameters are selected by the accuracy of the validation set. The Gini coefficient calculates the reduction in the impurity of the feature, reflecting the contribution of the feature to the classification. After normalization, the value is between 0 and 1. Forward stepwise regression starts with an empty feature set and adds the most significant features one at a time. The significance level of 0.05 is the standard threshold for statistical tests.A variance inflation factor greater than 5 indicates strong collinearity between features, which affects the stability of the model. The learning curve shows the trend of training error and validation error as the sample size increases, and convergence indicates that the sample size is sufficient. The root mean square error reflects the average deviation between the predicted value and the true value, and the coefficient of determination reflects the degree of variation explained by the model, with a value between 0 and 1. The normality test of the residuals determines whether the distribution of the prediction errors is symmetrically concentrated by calculating the skewness and kurtosis. The standard deviation is used to set the prediction interval. A residual exceeding 2 times the standard deviation indicates that the prediction deviates significantly from the true value. The reliability level of the prediction results is divided into three levels: high, medium, and low. The data quality score and the prediction residual are comprehensively evaluated. A low reliability level indicates that the prediction result is unreliable.

[0034] S106. Use a support vector machine algorithm to train the flavor hierarchy prediction model to obtain a trained flavor hierarchy prediction model, input the fat globule size distribution, surface curvature, membrane thickness, and membrane component ratio data of the condensed milk sample to be judged into the trained prediction model to obtain a flavor hierarchy prediction score for the sample.

[0035] The characteristic data of the condensed milk sample are subjected to standard deviation normalization to obtain normalized characteristic data, and a support vector machine is constructed by using a radial basis kernel function to obtain a training model from the normalized characteristic data; a characteristic contribution value is calculated by using the training model, and a characteristic weight screening threshold is obtained according to the characteristic contribution value, and features below the characteristic weight screening threshold are eliminated and then trained to obtain a new support vector machine; a Mahalanobis distance value of a sample to be judged is calculated according to the new support vector machine, and if the Mahalanobis distance value exceeds the mean threshold of the Mahalanobis distance of the training sample, the sample to be judged is marked as an abnormal sample; a kernel density estimator is used to calculate the predicted score probability for the abnormal sample, and if the predicted score probability is less than a preset probability threshold, it is determined that the abnormal sample does not meet the requirements.

[0036] Specifically, the standard deviation normalization is performed on the characteristic data of the condensed milk sample, and the values ​​are normalized to the interval of mean 0 and standard deviation 1. The normality of the characteristic distribution is verified by skewness test and kurtosis test, and the support vector machine is constructed by radial basis kernel function. The kernel function parameter is in the interval of 0.01 to 100 and the penalty coefficient is in the interval of 0.1 to 1000. The grid search is set to exponential growth, and the parameter combination is judged by the accuracy of the verification set. The verification accuracy and convergence speed are calculated for the training data. The support vector density statistics of the trained support vector machine are performed, and the support vector set is extracted by using a sequence minimum optimization searcher. The contribution value of each feature to the prediction result is obtained by a mutual information calculator, and the feature weight screening threshold is set to the maximum contribution value of 20%. The features below the threshold are eliminated, and the remaining features are used to train the new support vector machine. The convergence degree of the new support vector machine is judged from the accuracy curve. Standard deviation normalization is performed on the condensed milk sample data to be judged. The single sample Mahalanobis distance calculator is used to obtain the sample feature distance value. The discrimination threshold is set to 2 times the standard deviation of the mean Mahalanobis distance of the training sample. Samples with distance values ​​exceeding the threshold are marked as abnormal, and the prediction score of normal samples is obtained through support vector machine prediction operation. The confidence level is calculated according to the prediction error distribution of the training data, and the upper and lower bounds of the prediction score interval are set to 95%. The Shapiro normality tester is used to judge the normality of the prediction error. The probability of the prediction score falling into the distribution of historical data is calculated by kernel density estimation. The prediction score with a probability of less than 5% is marked, and the final prediction result is output. Standard deviation normalization converts different feature data to comparable scales. The mean of the fat globule size distribution in the original feature is 5 microns and the standard deviation is 1.5 microns. The mean of the surface curvature is 0.8 and the standard deviation is 0.2. The mean of the membrane thickness is 35 nanometers and the standard deviation is 8 nanometers. The mean of the protein-phospholipid ratio is 4.2 and the standard deviation is 0.6. After normalization, the mean of all features is 0 and the standard deviation is 1. The skewness coefficient is between -0.5 and 0.5, and the kurtosis coefficient is between 2.5 and 3.5, indicating that the data is close to normal distribution. The radial basis kernel function constructs nonlinear classification boundaries in the feature space. The kernel parameter controls the complexity of the mapping space. Too small a value leads to underfitting, and too large a value leads to overfitting. The grid search starts from 0.01 and increases by 10 times to 100. A total of 6 kernel parameter value points are selected. The penalty coefficient balances empirical risk and structural risk. The value starts from 0.1 and increases by 10 times to 1000. A total of 5 penalty coefficient value points are selected. The validation set accuracy is optimal when the kernel parameter is 1.0 and the penalty coefficient is 10. The support vector is the sample point located near the classification boundary. The number of support vectors reflects the complexity of the classification boundary. About 15% of the points in the training sample become support vectors. The mutual information calculation shows that the contribution value of fat globule size distribution is 0.45, the surface curvature is 0.25, the membrane thickness is 0.18, and the protein-phospholipid ratio is 0.12. The weight threshold was set to 0.09, and the protein-phospholipid ratio feature was eliminated, and the number of support vectors of the new model was reduced to 12%.The Mahalanobis distance measures the standardized distance from the sample point to the data center, taking into account the correlation between features. The mean Mahalanobis distance of the training sample is 2.8, the standard deviation is 0.9, and the abnormal threshold is set to 4.6. For the predicted score, the 95% confidence interval corresponds to 1.96 times the standard deviation range, and the prediction error Shapiro test p value is greater than 0.05, indicating that the error follows a normal distribution. The kernel density estimation constructs the distribution curve of historical data. The probability of the predicted score occurring is less than 5%, which usually indicates that the prediction result is unreliable. In practical applications, there are differences in the distribution characteristics of the flavor hierarchy scores of condensed milk samples of different qualities. The scores of fresh samples are stable, and the scores of samples with a long storage time fluctuate greatly. The mapping relationship between features and scores is established by support vector machine to realize the prediction and evaluation of flavor hierarchy. The model parameters need to be optimized according to the sample characteristics and scoring criteria to ensure the reliability of the prediction results. The complete feature extraction and prediction process plays an important role in the quality evaluation of condensed milk.

[0037] S107. Sort the flavor hierarchies of the samples according to the flavor hierarchy prediction scores to obtain a sample ranking result from high to low flavor hierarchy, and based on the ranking result, determine the combination of fat globule size distribution, surface curvature, membrane thickness, and membrane component ratio corresponding to the sample with the highest flavor hierarchy as the fat globule morphological feature combination with the highest flavor characteristic level.

[0038] The condensed milk samples are sorted in descending order according to their predicted scores for flavor hierarchy, and the Mann-Whitney test is used to obtain the significant difference results of the samples, which are used to mark the sample ranking levels; the Ward minimum variance clustering method is used to calculate the Euclidean distance for the top samples in the ranking levels, a distance matrix is ​​constructed by the Euclidean distance, a cluster pedigree diagram is obtained from the distance matrix, and a representative sample set is obtained by silhouette coefficient analysis of the cluster pedigree diagram; principal component analysis is performed on the representative sample set, and principal components are obtained by eigenvalue decomposition of the covariance matrix, and orthogonal basis vectors are obtained by Schmidt orthogonalization of the principal components, and the orthogonal basis vectors are used to construct a feature space; the Euclidean distance between sample points is calculated in the feature space, and a nearest neighbor sample set is constructed for the sample points whose Euclidean distance is less than a threshold value, and the most densely populated area sample points are obtained from the nearest neighbor sample set by density clustering, and the most densely populated area sample points are used to determine the optimal morphological feature combination value.

[0039] Specifically, the predicted scores of flavor hierarchy of condensed milk samples were sorted in descending order, and the Mann-Whitney test was used to determine the significance of the difference in scores of adjacent samples according to the 95% confidence interval of the predicted scores. The test level was set to 0.05, and the ranking level was marked for samples with significant differences. The ranking stability index was calculated by repeated sampling using the bootstrap method, and reliable ranking results were obtained from samples with a stability index greater than 0.9. Ward minimum variance clustering was used for the top ranked samples, and the Euclidean distance between samples was calculated to construct a distance matrix. The clustering pedigree diagram was drawn based on the distance matrix, and the optimal number of clusters was determined by the silhouette coefficient. The clustering results were tested for stability, and the silhouette coefficient threshold was set to 0.7. A representative sample set was extracted from the cluster center. Principal component analysis was performed on the representative sample set, and the principal components were extracted by eigenvalue decomposition of the covariance matrix. The cumulative contribution rate threshold was set to 85%, and the principal components that met the threshold were selected to construct the feature space. The orthogonal basis vectors were obtained by Schmidt orthogonalization, and the original features were projected on the orthogonal basis to obtain feature combinations. According to the feature combination, a four-dimensional feature vector space is constructed. The Euclidean distance between sample points is calculated in the feature space. The distance threshold is set to 20% of the maximum distance. Sample points with a distance less than the threshold are selected to form the nearest sample set. The sample points in the densest area are extracted by density clustering, and the optimal morphological feature combination value is obtained from the densest area. The ranking of predicted scores involves the judgment of the comparability of scores. The 95% confidence interval represents the fluctuation range of the predicted value. Taking the predicted score of 8.5 points as an example, its confidence interval is 8.2 to 8.8 points, which overlaps with the confidence interval of 8.7 to 9.3 points of the adjacent sample score of 9.0 points. The Mann-Whitney test is required to judge the significance of the difference. The test level of 0.05 corresponds to a critical value of 1.96, and a significant difference is determined to be greater than the critical value. The bootstrap method evaluates the stability of the ranking by repeated sampling. 1000 random samples are drawn from the original data, and the sample ranking position is calculated each time, and the frequency of each sample appearing in a specific ranking position is counted. The stability index of 0.9 means that 90% of the sampling results give the same ranking position, reflecting the reliability of the ranking results. The samples ranked at the top tend to have a higher stability index. Ward minimum variance clustering is based on the distance between samples. The Euclidean distance calculation takes into account all feature dimensions, and the eigenvalues ​​are standardized to ensure that the weights of each dimension are equal. The cluster pedigree diagram shows the hierarchical relationship between samples. The closer the distance, the earlier the samples are aggregated. The silhouette coefficient measures the tightness of clustering. The coefficient value of 0.7 indicates that the similarity of samples within the class is higher than the similarity between classes. Principal component analysis reduces the dimension of feature space. The covariance matrix reflects the correlation between features. The eigenvalue indicates the degree to which the principal component explains the variation of the original features. The cumulative contribution rate of 85% means that the retained principal component contains 85% of the information of the original data. Schmidt orthogonalization eliminates the correlation between principal components and constructs a standard orthogonal basis. The four-dimensional eigenvector contains information on fat globule size distribution, surface curvature, membrane thickness, and membrane component ratio. The Euclidean distance threshold selects 20% of the maximum distance to ensure that sufficiently similar sample points are screened out.Density clustering searches for dense areas of sample points in Euclidean space. The feature combination of dense areas is highly representative. In practical applications, samples with high flavor levels usually have stable feature combinations, such as concentrated distribution of fat globule size, uniform surface curvature, moderate membrane thickness, and coordinated protein-phospholipid ratio. The optimal feature combination is identified through multidimensional feature analysis, and the corresponding relationship between flavor levels and morphological characteristics is established. Statistical tests and stability assessments ensure the reliability of the results, and feature space dimensionality reduction and orthogonalization processing improve the representativeness of feature combinations.

[0040] What is disclosed above is only a preferred embodiment of the present invention, which certainly cannot be used to limit the scope of rights of the present invention. A person skilled in the art can understand that all or part of the processes of the above embodiments and equivalent changes made according to the claims of the present invention still fall within the scope of the invention.

Claims

1. A method for optimizing condensed milk flavor by multi-level flavor characteristic analysis, characterized in that: The method comprises: Obtain fat globule image data of different condensed milk samples, extract fat globule regions from fat globule images using image segmentation algorithms, obtain segmented fat globule binarization images, calculate equivalent diameters of fat globules based on pixel areas of fat globules in the binarization images, and obtain size distribution data of fat globules; For the fat globule binary image after segmentation, the edge detection algorithm is used to extract the contour of the fat globule to obtain the contour curve of the fat globule. According to the curvature change of the contour curve, the average curvature of the fat globule surface is calculated to obtain the curvature data characterizing the shape of the fat globule; The morphological processing algorithm is used to extract the membrane area of ​​the fat globule from the segmented fat globule binary image. The average membrane thickness of the fat globule is calculated based on the number of pixels in the extracted membrane area to obtain the membrane thickness data that characterizes the fat globule membrane characteristics. Obtaining the fat globule membrane component detection data of different condensed milk samples, including membrane protein content and phospholipid content, calculating the ratio of protein to phospholipid in the fat globule membrane based on the membrane protein content and phospholipid content, and obtaining protein-phospholipid ratio data characterizing the composition of the fat globule membrane; Obtain flavor characteristic evaluation data of different condensed milk samples, including flavor hierarchy scores, use the flavor hierarchy scores as sample label data, use fat globule size distribution, surface curvature, membrane thickness, and membrane component ratio as sample feature data, and construct a flavor hierarchy prediction model; The flavor hierarchy prediction model is trained by using a support vector machine algorithm to obtain a trained flavor hierarchy prediction model, and the fat globule size distribution, surface curvature, membrane thickness, and membrane component ratio data of the condensed milk sample to be judged are input into the trained prediction model to obtain the flavor hierarchy prediction score of the sample; According to the flavor hierarchy prediction scores, the flavor hierarchy of the samples is sorted to obtain the sample ranking results from high to low flavor hierarchy. According to the sorting results, the combination of fat globule size distribution, surface curvature, membrane thickness, and membrane component ratio corresponding to the sample with the highest flavor hierarchy is determined as the fat globule morphological feature combination with the highest flavor characteristic level.

2. The method according to claim 1, characterized in that The method comprises: obtaining fat globule image data of different condensed milk samples, extracting fat globule regions from the fat globule images using an image segmentation algorithm, obtaining segmented fat globule binarization images, and calculating equivalent diameters of the fat globules according to the pixel areas of the fat globules in the binarization images to obtain size distribution data of the fat globules, including: An image processor is used to obtain the image data of the condensed milk sample with the calibration plate scale, and a conversion relationship between pixel points and actual sizes is established according to the calibration plate scale to obtain a calibration image; Performing Gaussian low-pass filtering on the calibration image to remove background noise, using the OTSU algorithm to calculate multi-level grayscale thresholds, and extracting image contour boundaries according to the grayscale thresholds to obtain a boundary image; A seed point growth process is performed according to the boundary image, and if the gray value difference of the overlapping area exceeds the standard deviation of the neighborhood average value, a watershed algorithm is used to divide the independent blocks to obtain a segmented image; The fat globule contour area data is extracted from the segmented image, the actual area value is calculated according to the conversion relationship between the pixel point and the actual size, and the particle size distribution result is obtained by calculating the equivalent diameter through the circular area formula.

3. The method according to claim 1, characterized in that The method comprises: extracting the contour of the fat globule by using an edge detection algorithm from the segmented fat globule binary image to obtain the contour curve of the fat globule, calculating the average curvature of the fat globule surface according to the curvature change of the contour curve, and obtaining curvature data characterizing the shape of the fat globule, including: Using the Sobel operator to extract edge gradients in the horizontal and vertical directions of the binary image, and obtaining an edge intensity distribution map through iterative calculation according to the horizontal gradient and the vertical gradient; Performing a refinement iteration operation on the edge intensity distribution map, connecting and compensating the unclosed points through a connectivity propagator to obtain a closed contour line; Constructing a point index sequence according to the closed contour line, calculating the curvature value of the point index sequence using a three-point arc fitting function, and obtaining a smooth curvature value sequence through a Gaussian filter; If the curvature of a point in the smooth curvature value sequence is greater than zero, it is marked as a convex point. If the curvature of a point is less than zero, it is marked as a concave point. The average curvature result is obtained by weighted operation of the extreme point curvature.

4. The method according to claim 1, characterized in that: The method comprises: extracting the membrane region of the fat globules by using a morphological processing algorithm for the segmented fat globule binary image, calculating the average membrane thickness of the fat globules according to the number of pixels in the extracted membrane region, and obtaining membrane thickness data characterizing the membrane characteristics of the fat globules, including: A circular structure erosion operator is used to perform edge erosion on the fat globule region for the binary image, and a membrane region map is obtained by a closing operation; According to the maximum gray value point in the membrane region map as a seed point, a region growing extractor is used to expand to the surrounding areas, and a membrane region skeleton line is obtained by distance transformation; Scan from the skeleton line of the membrane area to the inner and outer sides to obtain the inner and outer boundary coordinate points of the membrane area contour; A local film thickness sequence is obtained by measuring the inner and outer boundary coordinate points along the reference direction. If the sampling interval of the local film thickness sequence is a preset value, an average film thickness value is obtained according to Gaussian weighted averaging and Kalman filtering.

5. The method according to claim 1, characterized in that The method of obtaining the fat globule membrane component detection data of different condensed milk samples, including membrane protein content and phospholipid content, and calculating the ratio of protein to phospholipid in the fat globule membrane according to the membrane protein content and phospholipid content, and obtaining protein phospholipid ratio data characterizing the composition of the fat globule membrane, includes: A methanol-water gradient eluent is used to separate components through a solid phase extraction column, and the component separation obtains a protein absorbance signal and a phospholipid fluorescence intensity signal; Performing baseline smoothing processing on the protein absorbance signal and the phospholipid fluorescence intensity signal, and obtaining corrected signal data by internal standard correction; The corrected signal data is filtered using a wavelet filter and a low-pass filter, and the component content value is obtained by fitting a normal distribution function; The molar concentration was calculated based on the component content using the protein and phospholipid molecular weight conversion coefficient, and the protein-phospholipid molar ratio was obtained by variance-weighted averaging.

6. The method according to claim 1, characterized in that The method comprises obtaining flavor characteristic evaluation data of different condensed milk samples, including flavor hierarchy scores, using the flavor hierarchy scores as label data of the samples, using fat globule size distribution, surface curvature, film thickness, and film component ratio as characteristic data of the samples, and constructing a flavor hierarchy prediction model, including: The characteristic data of condensed milk samples were normalized by a normalizer to obtain normalized fat globule distribution values, surface curvature values, membrane thickness values ​​and membrane component ratio values; The Pearson correlation coefficient is calculated based on the normalized eigenvalues, and the feature contribution is used to delete the feature combinations whose correlation coefficient exceeds the threshold; The flavor hierarchy label vector is constructed for the dataset after feature deletion, and the stratified sampler is used to obtain training data and validation data; Calculating feature importance values ​​for the training data, and selecting features whose significance levels are lower than a preset threshold by a forward stepwise regression method; A prediction model is established for the selected features, and the root mean square error and determination coefficient are used to judge the accuracy of the prediction model on the validation data. The prediction reliability level is determined according to the residual normality test results.

7. The method according to claim 1, characterized in that The flavor hierarchy prediction model is trained by using a support vector machine algorithm to obtain a trained flavor hierarchy prediction model, and the fat globule size distribution, surface curvature, membrane thickness, and membrane component ratio data of the condensed milk sample to be judged are input into the trained prediction model to obtain the flavor hierarchy prediction score of the sample, including: Performing standard deviation normalization on the characteristic data of the condensed milk sample to obtain normalized characteristic data, and constructing a support vector machine using a radial basis kernel function to obtain a training model from the normalized characteristic data; Calculate the feature contribution value through the training model, obtain the feature weight screening threshold according to the feature contribution value, remove the features below the feature weight screening threshold and then train to obtain a new support vector machine; Calculating the Mahalanobis distance value of the sample to be judged according to the new support vector machine, if the Mahalanobis distance value exceeds the mean threshold of the training sample Mahalanobis distance, marking the sample to be judged as an abnormal sample; A kernel density estimator is used to calculate the predicted score probability for the abnormal sample. If the predicted score probability is less than a preset probability threshold, it is determined that the abnormal sample does not meet the requirements.

8. The method according to claim 1, characterized in that The method comprises: ranking the flavor hierarchies of the samples according to the flavor hierarchies prediction scores to obtain a ranking result of the samples from high to low flavor hierarchies; and determining, according to the ranking result, a combination of fat globule size distribution, surface curvature, membrane thickness, and membrane component ratio corresponding to the sample with the highest flavor hierarchies as the fat globule morphological feature combination with the highest flavor characteristic level. According to the predicted scores of flavor levels of condensed milk samples, the samples are sorted in descending order, and the significant difference results of the samples are obtained by Mann-Whitney test, and the significant difference results are used to mark the ranking of the samples; For the top samples in the ranking level, the Ward minimum variance clustering method is used to calculate the Euclidean distance, a distance matrix is ​​constructed through the Euclidean distance, a cluster pedigree diagram is obtained from the distance matrix, and a representative sample set is obtained by analyzing the cluster pedigree diagram through the silhouette coefficient; Performing principal component analysis on the representative sample set, obtaining principal components by eigenvalue decomposition of the covariance matrix, obtaining orthogonal basis vectors by Schmidt orthogonalization of the principal components, and using the orthogonal basis vectors to construct a feature space; The Euclidean distance between sample points is calculated in the feature space, and a nearest neighbor sample set is constructed for the sample points whose Euclidean distance is less than a threshold value. The most densely populated area sample points are obtained from the nearest neighbor sample set through density clustering, and the most densely populated area sample points are used to determine the optimal morphological feature combination value.

Images