Grease crystallization characteristic prediction method and system based on deep learning model

Abstract

The invention discloses a deep learning model-based grease crystallization characteristic prediction method and a deep learning model-based grease crystallization characteristic prediction system. The prediction method comprises the following steps: establishing a grease sample database comprising single oil, blend oil and modified oil: constructing a multi-layer feedforward neural network model, setting an input layer as an original index of grease, and setting an output layer as a crystallization characteristic index of the grease; training and fitting the constructed neural network model; global optimization is carried out on weight parameters or model structures of the neural network; the optimized neural network model is verified; predicting the grease sample to be detected; and updating the grease sample database. The neural network model constructed by the invention can accurately and quickly predict the crystallization characteristics of the grease, has universality, greatly saves the detection cost, accelerates the research and development period of related products, and reduces a large amount of repetitive characterization test work.

Description

Technical Field

[0001] The present invention relates to a method for predicting the crystallization characteristics of oils and fats, and particularly to a method and system for predicting the crystallization characteristics of oils and fats based on a deep learning model. Background Art

[0002] In the food industry, the crystallization characteristics of oils and fats directly affect the quality and performance of products. The crystallization of oils and fats is a complex physico-chemical process affected by multiple factors. The interaction of multiple factors makes it difficult to accurately predict the crystallization process. Traditional physical and chemical property determination methods mainly conduct actual measurements and characterizations on oil and fat samples. Different samples are obtained through multiple repetitive operations and then subjected to combined analysis. Although various chromatographic methods and electrochemical analysis methods are precise in structure, they are time-consuming and require a large number of repeated experiments. The above problems result in a long R & D cycle and high detection costs for oil and fat products, especially for enzymatically modified oils and fats.

[0003] Enzymatic modification of animal and vegetable oils is a green chemical method that uses lipase as a biocatalyst to rearrange and optimize the fatty acyl groups in animal and vegetable oils. It has the advantages of high efficiency, mildness, and environmental friendliness, and has broad application prospects in the fields of food, medicine, cosmetics, and industrial materials. Due to the diversity of the fatty acyl group composition and arrangement on triglycerides, as well as the diversity of triglyceride content, the mechanism of fatty acyl group change after transesterification is complex and is non-linearly affected by many factors such as the type of substrate oil, reaction temperature, reaction time, catalyst addition amount, and type of lipase. At the same time, it is difficult to accurately identify the isomers of triglyceride molecules, resulting in the inability of existing models to dynamically invert the reaction conditions and accurately predict their crystallization characteristics. Summary of the Invention

[0004] In order to overcome the above-mentioned shortcomings and deficiencies of the prior art, the purpose of the present invention is to provide a method for predicting the crystallization characteristics of oils and fats based on a deep learning model. Using the original indexes of oils and fats, namely total fatty acid composition, sn-2 fatty acid composition, and triglyceride composition, as the input of the neural network model, the constructed neural network model can accurately and quickly predict the crystallization characteristics of oils and fats and has universality.

[0005] Another purpose of the present invention is to provide a system for predicting the crystallization characteristics of oils and fats based on a deep learning model.

[0006] The purpose of the present invention is achieved through the following technical solutions:

[0007] The present invention provides a method for predicting the crystallization characteristics of oils and fats based on a deep learning model, including the following steps:

[0008] S1 Establish an oil and fat sample database:

[0009] S11 Select multiple single oils; compound the single oils to prepare multiple blended oils; use lipase to perform enzymatic modification on the single oils and blended oils at different temperatures and different reaction times to obtain modified oils;

[0010] S12 Test and record the original indexes and crystallization characteristic indexes of the single oils, blended oils, and enzymatically modified oils, and perform normalization and standardization processing on the original index data and crystallization characteristic index data;

[0011] The original indexes include total fatty acid composition, sn-2 position fatty acid composition, and triglyceride composition;

[0012] The crystallization characteristic indexes include solid fat content and crystallization rate at different temperatures within the range of 0 to 40°C;

[0013] S2 Construct a multi-layer feedforward neural network model, and set the input layer as the original indexes of the oil and fat, and the output layer as the crystallization characteristic indexes of the oil and fat;

[0014] S3 Train and fit the neural network model constructed in step S2;

[0015] S4 Perform global optimization on the weight parameters or model structure of the neural network;

[0016] S5 Verify the neural network model optimized in step S4, and the verification includes forward prediction and reverse deduction;

[0017] S6 Use the neural network model verified in step S5 to predict the oil and fat sample to be tested.

[0018] Preferably, the prediction method for the crystallization characteristics of oil and fat based on the deep learning model further includes the following steps:

[0019] S7 Record the original indexes and predicted crystallization characteristic indexes of the oil and fat sample to be tested in step S6 into the sample database, and update the oil and fat sample database.

[0020] Preferably, the single oils in step (1) include palm oil, soybean oil, rapeseed oil, sunflower oil, peanut oil, palm kernel oil, coconut oil, olive oil, corn oil, and linseed oil.

[0021] Preferably, in step S2, the neural network model has three layers, including an input layer, a hidden layer, and an output layer, and uses the Fletcher-Reeves variable gradient correction algorithm for the regression task; the hidden layer uses the non-linear Sigmoid function as the transfer function, and the linear function is used as the transfer function in the output layer; the number of nodes in the hidden layer is experimentally selected from 3 to 15, and the optimal number of nodes is judged by the mean square error of the prediction model; the learning rate is selected between 0.01 and 0.8; a momentum coefficient is introduced, and its selection range is 0 to 1.

[0022] Preferably, during the training and fitting process in step S3, the error is minimized through the backpropagation algorithm to realize the fitting of the crystallization characteristic indexes of the modified oil.

[0023] Preferably, during the training and fitting process in step S3, a penalty term is added to the loss function to realize the regularization of the loss function:

[0024]

[0025] where: J(θ) is the regularized loss function; loss(θ) is the original loss function; θ is the model parameter; λ is the penalty coefficient; n is the number of parameters; θ i is the i-th model parameter.

[0026] Preferably, the global optimization of the weight parameters or model structure of the neural network in step S4 is specifically carried out by using a genetic algorithm. Through multiple iterative optimizations, the generalization ability of the neural network on the data set reaches the optimal state; the genetic algorithm includes steps of individual coding, fitness function design, selection, crossover, and mutation.

[0027] Preferably, the forward prediction in step S5 uses the trained neural network model. After inputting the given original indexes, the solid fat content and crystallization rate are predicted forward and compared with the actual measurement results to verify the accuracy of the forward prediction of the model; the reverse deduction sets the target oil crystallization characteristics, inputs them reversely into the neural network model, and uses the genetic algorithm to find the best original indexes of the oil.

[0028] The present invention provides a prediction system for the crystallization characteristics of oil based on a deep learning model, including a memory and a processor;

[0029] The memory is used to store non-temporary computer instructions;

[0030] When the non-temporary computer instructions are executed by the processor, the processor implements the prediction method for the crystallization characteristics of oil based on the deep learning model.

[0031] The present invention provides a storage medium for storing non-temporary computer instructions, which, when run, execute the prediction method for the crystallization characteristics of oil based on the deep learning model.

[0032] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0033] (1) The prediction method of the crystallization characteristics of fats and oils based on a deep learning model of the present invention overcomes the prediction errors caused by the difficult accurate identification of the isomers of triglyceride molecules by introducing the fatty acid composition at the n-2 position and using the total fatty acid composition and triglyceride composition together as the input of the neural network model, and can quickly and efficiently predict the crystallization characteristics (solid fat content, crystallization rate), greatly saving the detection cost.

[0034] (2) The prediction method of the crystallization characteristics of fats and oils based on a deep learning model of the present invention can not only predict the physical and chemical characteristics of fat and oil products under given original indexes in a forward direction, but also inversely deduce the original indexes required for the target fat and oil crystallization characteristic indexes, realizing precise process design and parameter optimization, accelerating the development cycle of related products, and reducing a large amount of repetitive characterization and testing work.

[0035] (3) The prediction method of the crystallization characteristics of fats and oils based on a deep learning model of the present invention uses lipase to catalyze the reaction of single oil and blended oil, and conducts different degrees of transesterification reactions on the single oil and blended oil by selecting the reaction time and reaction temperature, deriving a large number of modified oils with physical and chemical characteristics different from those of the single oil and blended oil. The database established by using the large amount of experimental characterization data of these oils (single oil, blended oil, modified oil) comprehensively reflects the differences in the physical and chemical characteristics of different oils, and has high representativeness and reliability.

[0036] (4) The prediction method of the crystallization characteristics of fats and oils based on a deep learning model of the present invention uses a neural network to fit complex non-linear relationships and combines genetic algorithm global optimization to further improve the prediction accuracy.

[0037] (5) The prediction method of the crystallization characteristics of fats and oils based on a deep learning model of the present invention can continuously incorporate new data into the fat and oil sample database, and use incremental learning to retrain and optimize the neural network model, so that the model can adapt to a wider variety of fat and oil types and process parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 is a flowchart of the prediction method of the crystallization characteristics of fats and oils based on a deep learning model of the present invention.

[0039] Figure 2 is a comparison result of the predicted values and measured values of the solid fat content by different models in Examples and Comparative Examples 1-5. DETAILED DESCRIPTION OF THE INVENTION

[0040] The present invention will be further described in detail below with reference to the examples, but the implementation manners of the present invention are not limited thereto.

[0041] EXAMPLES

[0042] As Figure 1As shown in the figure, an embodiment of the present invention provides a method for predicting the crystallization characteristics of oils and fats based on a deep learning model, including the following steps:

[0043] S1 Establish an oil and fat sample database:

[0044] S11 Select a variety of single oils; prepare a variety of blended oils by compounding single oils; use lipase to perform enzymatic modification on single oils and blended oils at different temperatures and different reaction times to obtain modified oils;

[0045] S12 Test and record the original indexes and crystallization characteristic indexes of single oils, blended oils, and enzymatically modified oils, and perform normalization and standardization processing on the original index data and crystallization characteristic index data to prevent reaching a supersaturated state during the training process and prepare for subsequent model training;

[0046] The original indexes include total fatty acid composition, sn-2 fatty acid composition, and triglyceride composition;

[0047] The crystallization characteristic indexes include solid fat content and crystallization rate at different temperatures within the range of 0 to 40 °C;

[0048] S2 Construct a multi-layer feedforward neural network model, set the input layer as the original indexes of oils and fats, and the output layer as the crystallization characteristic indexes of oils and fats; the structure design and operation of the network model are carried out in MATLAB R2024a;

[0049] S3 Train and fit the neural network model constructed in step S2;

[0050] S4 Perform global optimization on the weight parameters or model structure of the neural network;

[0051] S5 Verify the optimized neural network model in step S4, and the verification includes forward prediction and reverse deduction;

[0052] S6 Use the neural network model verified in step S5 to predict the oil and fat sample to be tested.

[0053] In the embodiment, after step S6, the following steps are also carried out:

[0054] S7 Record the original indexes of the oil and fat sample to be tested in step S6 and the predicted crystallization characteristic indexes into the sample database, and update the sample database. With the continuous accumulation of new data, new data can be continuously incorporated into the database, and the neural network model can be retrained and optimized using incremental learning to make the model adapt to a wider variety of oil and fat types and process parameters.

[0055] In this embodiment, the single oil described in step S11 includes palm oil, soybean oil, rapeseed oil, sunflower seed oil, peanut oil, palm kernel oil, coconut oil, olive oil, corn oil, and linseed oil.

[0056] The blended oil of this embodiment is compounded from the above single oils, including pairwise compounding and multi-component compounding. Among them, the pairwise compounding is: randomly pairwise compounding from the above vegetable oils and mixing them according to different mass ratios to obtain the blended oil, where the mass ratios are set to 25:75, 50:50, and 75:25; for the random compounding of more than two single oils, at least three or more mass ratio parameters are set for mixing.

[0057] In the embodiment, the enzymatic modification in step S11 is specifically: using the single vegetable oil and the blended oil in step S11 to carry out transesterification reaction under controlled reaction conditions (such as different reaction temperatures, reaction times, enzyme dosages, etc.) to obtain the modified oil. Among them, the reaction temperature is preferably 50-80 °C, the reaction time is preferably 0-50 min, and the lipase is preferably Lipozyme TL IM. The addition amount is preferably 0-10% wt of the mass of the reaction substrate.

[0058] In this embodiment, the original indexes of total fatty acid composition, sn-2 position fatty acid composition, and triglyceride composition in step S12 are detected by the following methods respectively:

[0059] Among them, the total fatty acid composition analysis: The total fatty acid composition analysis was carried out using gas chromatography (GC). After converting the sample into fatty acid methyl esters, gas phase analysis was performed. The lipid methylation referred to the AOCS method with slight modification. 2 mL of a 0.5 mol / L potassium hydroxide-methanol solution was added to the round-bottom flask for extracting lipids, and saponification reaction was carried out by shaking in a 70 °C water bath for 10 min. After cooling to room temperature, 2 mL of boron trifluoride-methanol solution was added, and methylation reaction was carried out in a 70 °C water bath for 5 min. After the reaction was complete, it was cooled to room temperature, 3 mL of chromatographic grade n-hexane reagent was added, shaken well, then 2 mL of saturated brine was added, shaken again, and after standing for layering, the upper layer liquid was aspirated into a 5 mL centrifuge tube, a little anhydrous sodium sulfate solid powder was added, shaken for 1 min and then centrifuged, the supernatant was taken out, passed through a 0.22 μm organic filter membrane into a gas chromatography injection vial, and was ready for analysis on the machine. Standard samples were used for qualitative analysis of fatty acids to determine the fatty acid composition, and area normalization method was used to analyze the composition content of fatty acids. Analysis was carried out using GC (7820-A, Agilent, Wilmington, DE, USA), equipped with a flame ionization detector and a CP-sil88 chromatographic column (100 m × 0.25 mm × 0.2 mm, Agilent, Wilmington, DE, USA). Nitrogen was used as the carrier gas, the column pressure was 30.8 psi, the injection volume was 1 μL, the split ratio was 40:1, the initial oven temperature was set at 120 °C and held for 3 min, then it was heated to 175 °C at a rate of 8 °C / min and held for 18 min. The detector and injector temperatures were both 260 °C. This method was used to analyze the fatty acid composition of the sample with reference to the AOCS method.

[0060] Analysis of the fatty acid composition at the sn-2 position: After dissolving the sample in n-hexane, the solution passing through the alumina chromatography column was rotary evaporated to obtain an oil sample. Then, 0.1 g of this oil sample (pre-prepared warm water bath preheated to 40 °C), 20 mg of porcine pancreatic lipase, and 2 mL of Tris buffer solution with pH = 8.0 were added to a centrifuge tube. After mixing, 0.2 mL of 220 g / L calcium chloride solution and 0.5 mL of 1 g / L sodium cholate solution were added. After the above operations were completed within 30 s, the above centrifuge tube was placed in a 40 °C water bath and shaken well for 4 min. After taking it out of the water bath, it was shaken with a vortex mixer for 1 min, then 1 mL of ether and 1 mL of 6 mol / L hydrochloric acid solution were added, and it was centrifuged in a centrifuge for layer separation (3000 r / min, 5 min). The upper ether layer was taken for thin-layer chromatography analysis (the developing agent ratio was n-hexane: anhydrous ether: formic acid = 70:30:1.5). After development, it was taken out of the developing tank and dried for 5 - 10 min, then the thin-layer plate was placed in an iodine tank for color development. The sn-2 MAG band was scraped, extracted with ether, and then converted into fatty acid methyl ester for gas phase analysis. The methylation conditions and gas phase conditions were the same as those in the above total fatty acid composition analysis method.

[0061] Analysis of the triglyceride composition: A flame ionization detector and an Rtx-65TG capillary column (30 m × 0.25 mm × 0.1 um, RESTEK, USA) were equipped. Hydrogen was used as the carrier gas with a flow rate of 40 mL / min, the column pressure was 12.38 psi, the injection volume was 1 μL, the split ratio was 25:1, the initial oven temperature was set at 280 °C and maintained for 1.5 min, then it was heated at 10 °C / min to 340 °C and maintained for 9.5 min, and then it was heated at 1 °C / min to 350 °C and maintained for 12 min. This method was determined with reference to the AOCS method.

[0062] In this embodiment, the solid fat content and crystallization rate, the crystallization characteristic indexes in step S12, were detected by the following methods respectively:

[0063] Method for measuring the solid fat content: Using a pulsed Nuclear Magnetic Resonance (p-NMR) spectrometer (Rruker PC / 20 series micro-optics, Bruker Optics Co., Ltd., Milton, Ontario, Canada), according to the AOCS official method Cd 16-81, the p-NMR tube was filled with about 2 mL of molten sample, and then it was placed in a 60 °C water bath for 30 min to melt and eliminate the crystallization memory. Then, all the p-NMR tubes filled with samples were transferred to a thermostat at 0 °C and kept for 60 min. Before measurement, the P-NMR tubes were successively kept at 0, 10, 15, 20, 25, 30, 35, 40 °C. Except for keeping at 0 °C for 60 min, they were kept at other temperatures for 30 min.

[0064] Among them, the method for measuring the crystallization rate: First, the sample is kept at 60 °C for 30 min to melt and eliminate the crystallization memory, and then placed at the selected temperature to measure the change in the solid fat content at different crystallization times under this temperature condition. The time range is 0 - 100 min. Every 2 min within 0 - 30 min, every 10 min within 30 - 60 min, and every 20 min within 60 - 100 min, measure and record the SFC at the current crystallization time.

[0065] The curve of the measured SFC versus time is fitted using the Avrami equation,

[0066]

[0067] where SFC(t) is the SFC value at crystallization time t, SFC(∞) is the SFC value at complete crystallization, K is the crystallization rate constant, and n is the Avrami exponent.

[0068] In this embodiment, in step S12, the data is normalized and standardized. Specifically:

[0069] Among them, the sample data of the input and output layers is normalized and restricted within the interval [0, 1].

[0070]

[0071] x i represents the i-th input or output data, x min represents the minimum value of the data change, x max represents the maximum value of the data change;

[0072] Among them, Z-score standardization is used to improve the performance and convergence speed of the model, changing the mean of the data to 0 and the standard deviation to 1.

[0073]

[0074] μ represents the mean, and σ represents the standard deviation.

[0075] In this embodiment, the neural network model described in step S2 has three layers, including an input layer, a hidden layer, and an output layer. Among the five improved Back-Propagation learning algorithms for regression tasks (steepest descent algorithm, momentum algorithm, Fletcher-Reeves variable gradient correction algorithm, quasi-Newton OSS algorithm, and L-M algorithm), the Fletcher-Reeves variable gradient correction algorithm is preferably used for the regression task; the hidden layer uses the non-linear Sigmoid function as the transfer function, and the output layer uses the linear function as the transfer function. At the same time, according to the empirical formula for selecting the number of hidden layer nodes, the number of hidden layer nodes is experimentally tested from 3 to 15 respectively, and the optimal number of nodes is judged by the mean square error of the prediction model. The learning rate is selected between 0.01 and 0.8. A momentum coefficient is introduced, and the selection range is 0-1, which reduces the oscillation trend during the training process of the neural network algorithm, improves the convergence, and increases the reliability of the algorithm.

[0076] In this embodiment, during the training and fitting process described in step S3, the error is minimized through the backpropagation algorithm to achieve the fitting of the crystallization characteristic index of the modified oil. A machine learning prediction model for regression tasks based on error backpropagation is established using the neural network toolbox V 8.3 provided by MATLAB R2024a. The collected data set is randomly divided into three groups: 85% for the training set and 15% for the validation set.

[0077] In the embodiment, during the training and fitting process described in step S3, to solve the overfitting problem and improve the generalization ability of the model, a penalty term is added to the loss function to regularize the loss function, which is related to the complexity of the model parameters. Usually, L1 regularization is used to achieve this. The sum of the absolute values of the model weights is added to the loss function, which tends to produce a sparse weight matrix, that is, many weights become zero, thus achieving the effect of feature selection. Specifically as follows:

[0078]

[0079] Where: J(θ) is the regularized loss function; loss(θ) is the original loss function; θ is the model parameter;

[0080] λ is the regularization parameter (also known as the penalty coefficient); n is the number of parameters; θ i is the i-th model parameter.

[0081] In this embodiment, the global optimization of the weight parameters or model structure of the neural network described in step S4 is specifically carried out using the genetic algorithm for global optimization. Through multiple iterative optimizations, the generalization ability of the neural network on the data set reaches the optimal state; the genetic algorithm includes steps of individual coding, fitness function design, selection, crossover, and mutation.

[0082] In this embodiment, in the verification step described in step S5, forward prediction means using the trained neural network model to perform forward prediction of the solid fat content and crystallization rate after inputting the given original indexes, and comparing with the actual measurement results to verify the forward prediction accuracy of the model; reverse deduction means setting the target oil crystallization characteristics (such as the desired solid fat content and crystallization rate), inputting them reversely into the neural network model, and using the genetic algorithm to find the optimal original indexes of the oil (total fatty acid composition, sn-2 fatty acid composition, triglyceride composition).

[0083] During the verification process, statistical parameter combinations such as the mean square error (MSE) and the correlation coefficient (R 2 , correlation coefficient) are used to judge the prediction performance of the neural network.

[0084]

[0085] Among them, Y i represents the actual value of the i-th sample, and y i represents the predicted value corresponding to the i-th sample; y m represents the average value of the actual values of all samples.

[0086] An embodiment of the present invention also provides a prediction system for oil crystallization characteristics based on a deep learning model, including a memory and a processor;

[0087] The memory is used to store non-temporary computer instructions;

[0088] When the non-temporary computer instructions are executed by the processor, the processor implements the prediction method for oil crystallization characteristics based on the deep learning model.

[0089] An embodiment of the present invention also provides a storage medium for storing non-temporary computer instructions, which, when run, execute the prediction method for oil crystallization characteristics based on the deep learning model.

[0090] The method of the embodiment of the present invention can provide fast and efficient prediction for the determination of oil crystallization characteristics behavior, and greatly save the detection cost. For the known original indexes of oil (total fatty acid composition, sn-2 fatty acid composition, triglyceride composition), its crystallization characteristics (solid fat content, crystallization rate) can be quickly predicted. At the same time, according to the specific crystallization characteristics of the oil required by the product, its original indexes can be inversely deduced, accelerating the research and development cycle of related products and reducing a large amount of repetitive characterization and testing work.

[0091] Comparative Example 1

[0092] In this comparative example, only the total fatty acid composition data is used as the input layer to construct a BP neural network prediction model; except for this, the method of establishing the neural network model is the same as that in the embodiment.

[0093] Comparative Example 2

[0094] In this comparative example, only the triglyceride composition data is used as the input layer to construct a BP neural network prediction model. Except for this, the method of establishing the neural network model is the same as that in the embodiment.

[0095] Comparative Example 3

[0096] In this comparative example, only the total fatty acid composition data and the triglyceride composition data are used as the input layer to construct a BP neural network prediction model. Except for this, the method of establishing the neural network model is the same as that in the embodiment.

[0097] Comparative Example 4

[0098] In this comparative example, the data of the modified oil is not included in the database, that is, only the original indexes (total fatty acid composition, sn-2 fatty acid composition, triglyceride composition) of the single oil and the blended oil and the crystallization characteristic indexes (solid fat content, crystallization rate) are systematically recorded. Except for this, the method of establishing the neural network model is the same as that in the embodiment.

[0099] Comparative Example 5

[0100] The difference is that the BP neural network model established in this comparative example is not optimized by the genetic algorithm. Except for this, the method of establishing the neural network model is the same as that in the embodiment.

[0101] Figure 2 Figure 2 shows the comparison results of the predicted values and the measured values of the solid fat content by different models in the embodiment and Comparative Examples 1-5; Table 2 shows the predicted values and the measured values of the crystallization rate (K value) at 10 °C by different models in the embodiment and Comparative Examples 1-5 as shown in Table 1. The embodiment shows the best prediction accuracy, is closest to the measured value, and the result is relatively accurate. At the same time, it is evaluated by four verification indexes of the determination coefficient (R 2 ) of the modeling set, the mean square error (MSE), the mean absolute error (MAE), and the mean relative error (MRE) of the verification data set (as shown in Table 2).

[0102] Table 1

[0103]

[0104] According to the data in Table 1, it can be seen that the prediction model based on the total fatty acid composition, the sn-2 fatty acid composition, and the triglyceride composition has higher prediction accuracy than the prediction model established simply using the total fatty acid composition and the triglyceride composition data, and its R2 The closest to 1, with the smallest difference from the measured value.

[0105] After training is completed, the model is tested using the test set data. It is stipulated that if the error between the predicted value and the measured value is less than 10%, it is considered excellent; if the error is between 10% - 20%, it is considered good; if the error range is between 20% - 30%, it is considered average; if the error range is greater than 30%, it is considered poor. Table 3 shows the predicted value distribution and accuracy of each example and comparative example.

[0106] Table 3

[0107]

[0108]

[0109] It can be seen from the above examples and comparative examples that the BP neural network model combining the total fatty acid composition data, triglyceride composition, and sn-2 position fatty acid composition data provided by the examples of the present invention can quickly predict the crystallization characteristics of oils and fats with high precision through genetic algorithm inversion.

[0110] The above embodiments are the preferred embodiments of the present invention, but the embodiments of the present invention are not limited by the described embodiments. Any other changes, modifications, substitutions, combinations, and simplifications made without departing from the spirit and principle of the present invention shall be equivalent replacement methods and are all included in the protection scope of the present invention.

Claims

1. A prediction method for the crystallization characteristics of oils and fats based on a deep learning model, characterized in that, It includes the following steps: S1 Establish a grease sample database: S11 Select multiple single oils; compound the single oils to prepare multiple blended oils; use lipase to carry out enzymatic modification on the single oils and blended oils at different temperatures and different reaction times to prepare modified oils; S12 Test and record the original indexes and crystallization characteristic indexes of the single oils, blended oils, and enzymatically modified oils, and perform normalization and standardization processing on the original index data and crystallization characteristic index data; The original indexes include total fatty acid composition, sn-2 fatty acid composition, and triglyceride composition; The crystallization characteristic indexes include solid fat content and crystallization rate at different temperatures within the range of 0 to 40°C; S2 Construct a multi-layer feedforward neural network model, and set the input layer as the original indexes of the grease and the output layer as the crystallization characteristic indexes of the grease; S3 Train and fit the neural network model constructed in step S2; S4 Perform global optimization on the weight parameters or model structure of the neural network; S5 Verify the neural network model optimized in step S4, and the verification includes forward prediction and reverse deduction; S6 Use the neural network model verified in step S5 to predict the grease sample to be tested.

2. The prediction method for the characteristics of oil crystallization based on a deep learning model according to claim 1, wherein It also includes the following steps: S7 Record the original indexes of the grease sample to be tested in step S6 and the predicted crystallization characteristic indexes into the grease sample database, and update the grease sample database.

3. The prediction method for the crystallization characteristics of grease based on a deep learning model according to claim 1, wherein The single oils in step (1) include palm oil, soybean oil, rapeseed oil, sunflower oil, peanut oil, palm kernel oil, coconut oil, olive oil, corn oil, and linseed oil.

4. The prediction method for the crystallization characteristics of fats and oils based on a deep learning model according to claim 1, characterized in that In step S2, the neural network model has three layers, including an input layer, a hidden layer, and an output layer, and uses the Fletcher-Reeves variable gradient correction algorithm for the regression task; the hidden layer uses the non-linear Sigmoid function as the transfer function, and the linear function is used as the transfer function in the output layer; the number of nodes in the hidden layer is experimentally selected from 3 to 15, and the optimal number of nodes is judged by the mean square error of the prediction model; the learning rate is selected between 0.01 and 0.8; a momentum coefficient is introduced, and its selection range is 0 to 1.

5. The prediction method for the crystallization characteristics of fats and oils based on a deep learning model according to claim 1, wherein In the training and fitting process in step S3, the error is minimized through the backpropagation algorithm to realize the fitting of the crystallization characteristic indexes of the modified grease.

6. The method for predicting the crystallization characteristics of grease based on a deep learning model according to claim 1 or 5, characterized in that In the training and fitting process in step S3, a penalty term is added to the loss function to realize the regularization of the loss function: Among them: J(θ) is the regularized loss function; loss(θ) is the original loss function; θ is the model parameter; λ is the penalty coefficient; n is the number of parameters; θ i is the i-th model parameter.

7. The prediction method for the crystallization characteristics of grease based on a deep learning model according to claim 1, wherein The global optimization of the weight parameters or model structure of the neural network in step S4 is specifically carried out by using a genetic algorithm for global optimization. Through multiple iterations of optimization, the generalization ability of the neural network on the data set reaches the optimal state; the genetic algorithm includes steps of individual coding, fitness function design, selection, crossover, and mutation.

8. The prediction method for the crystallization characteristics of grease based on a deep learning model according to claim 1, wherein The forward prediction described in step S5 uses the trained neural network model. After inputting the given original metrics, it predicts the solid fat content and crystallization rate in the forward direction and compares with the actual measurement results to verify the accuracy of the forward prediction of the model. The reverse deduction is to set the target oil crystallization characteristics, input them reversely into the neural network model, and use the genetic algorithm to find the optimal original oil metrics.

9. A prediction system for the crystallization characteristics of grease based on a deep learning model, characterized in that, It includes a memory and a processor; The memory is used to store non-temporary computer instructions; When the non-temporary computer instructions are executed by the processor, the processor implements the prediction method for oil crystallization characteristics based on the deep learning model according to any one of claims 1 to 8.

10. A storage medium, characterized in that, It is used to store non-temporary computer instructions, and when the non-temporary computer instructions are run, it executes the prediction method for oil crystallization characteristics based on the deep learning model according to any one of claims 1 to 8.

Images