Method for optimizing plant extract preparation based on machine learning and multi-objective algorithm

Abstract

The invention discloses a method for optimizing plant extract preparation based on machine learning and a multi-objective algorithm, and belongs to the technical field of plant extraction. The method comprises the steps that firstly, data in the plant extract preparation process are collected, then a non-linear mapping relation between process parameters and product quality is established through a back propagation neural network model, prediction of multiple quality indexes of an extraction process and a product is achieved, and finally the process parameters are optimized in combination with a multi-objective optimization algorithm. The optimized process parameters are adopted to prepare the plant extract, so that the nutrition and sensory quality of the extracted product can be remarkably improved. The method meets multi-target, complex and high-dimensional optimization scenes, is suitable for preparation of any plant extract, and is wide in application range.

Description

Technical Field

[0001] The present invention relates to the technical field of plant extraction, and particularly relates to a method for optimizing the preparation of plant extracts based on machine learning and multi-objective algorithms. Background Art

[0002] Plant extracts are natural substances separated and extracted from plant tissues by physical or chemical methods, and are usually used in industries such as medicine, food, and cosmetics. Plant extracts have many health-promoting functions, including antibacterial, anti-cancer, anti-diabetic, anti-inflammatory, anti-obesity, antioxidant, antiviral, and neuroprotective effects, etc. These functions are mainly attributed to chemical components in plants such as flavonoids, polysaccharides, phenols, terpenoids, etc.

[0003] The technological process of plant extracts mainly includes raw material collection, raw material preparation, leaching extraction, concentration, drying, and pulverization. Among them, leaching extraction is the key link in the preparation of plant extracts. Common extraction methods include water extraction, alcohol extraction, supercritical fluid extraction, etc. Process conditions during the extraction process, such as solvent type, temperature, time, etc., will all affect the extraction efficiency of target components. Many studies have reported the influence of extraction conditions on the chemical composition or sensory quality of extracts. Therefore, it is necessary to experiment and explore to optimize the extraction process conditions.

[0004] Traditional methods for optimizing the preparation process of plant extracts include the trial-and-error method and the response surface method (Response surface methodology as a tool to optimize the extraction of bioactive compounds from plant sources. Journal of the Science of Food and Agriculture, 2022, 103(1): 26 - 36.), etc. However, the traditional trial-and-error method has high costs and a time-consuming process; the response surface has limitations in dealing with high-order non-linear relationships, which limits its fitting ability. Moreover, most current studies are based on a single optimization goal, and traditional optimization models are difficult to meet multi-objective, complex, and high-dimensional optimization scenarios.

[0005] With the rapid development of artificial intelligence, the performance of machine learning has gradually surpassed traditional statistical methods. Machine learning can identify complex patterns in analyzing large datasets and high-dimensional spaces and make accurate predictions from new data. Multi-objective optimization algorithms are used to solve optimization problems with multiple conflicting objectives. These algorithms aim to find a set of solutions to achieve an optimal balance among multiple objectives. There has been no research report on applying multi-objective optimization algorithms to the optimization of plant extract processes. Summary of the Invention

[0006] The purpose of the present invention is to provide a method for optimizing the extraction process based on machine learning and multi-objective algorithms, so as to obtain plant extracts with excellent nutritional and sensory qualities, so as to enhance the application of the extracts in the fields of food, cosmetics and medicine.

[0007] To achieve the above object, the present invention adopts the following technical solution: The present invention provides a method for optimizing the preparation of plant extracts based on machine learning and multi-objective algorithms, comprising the following steps: (1) Obtaining process parameters and corresponding quality indicators by preparing plant extracts to form a data set; the number of process parameters corresponding to a plant extract is ≥4, and the number of quality indicators is ≥2; (2) Screening the sample data set, removing outliers and normalizing it, and then dividing the optimized sample data set into a training set and a test set; (3) Using the process parameters in the training set as input and the quality indicators as output, the back propagation neural network model is used for training to establish a nonlinear mapping relationship between the process parameters and the quality indicators. The test set is then used to test and optimize the model parameters to obtain a prediction model. (4) The NSGA-II algorithm is used to take the prediction model as a proxy model, maximize or minimize the quality index as the optimization target, and take the applicable process parameter range as the constraint condition of the optimization target, and output the Pareto optimal solution set of the process parameters.

[0008] The present invention collects data during the preparation of plant extracts, uses a back-propagation neural network model to establish the relationship between process parameters and product quality, and then optimizes the process parameters in combination with a multi-objective optimization algorithm for the preparation of plant extracts to improve the nutritional and sensory quality of the extracts.

[0009] The plant extract refers to a natural substance separated and extracted from plant tissues, and the plant tissues may be, but are not limited to, roots, stems, leaves, and flowers of plants.

[0010] In step (1), the process parameters in the preparation process of the plant extract and the quality indicators of the extracted product are collected to obtain sample data.

[0011] Furthermore, the preparation process of the plant extract includes: raw material crushing, ultrasound-assisted extraction, filtration, concentration, and drying.

[0012] Furthermore, the process parameters include at least four of: raw material particle size, material-liquid ratio, pH value, leaching temperature, leaching time, and ultrasonic power.

[0013] Furthermore, the quality indicators include: at least two of: yield, active ingredient content, color, smell, and taste.

[0014] As a specific embodiment of the present invention, a golden yellow extract rich in phenylethanoid glycosides is extracted from the flowers of Osmanthus fragrans using a water extraction method.

[0015] The process parameters include: solid-liquid ratio, pH value, extraction temperature, extraction time, and ultrasonic power.

[0016] The quality indicators include: yield, total phenylethanoid glycoside content, verbascoside content, salidroside content and CIELAB color parameters L*, a*, b*.

[0017] As another specific embodiment of the present invention, a water extraction method is used to extract an extract rich in nutrients such as chlorogenic acid and having a strong caramel flavor and a fresh aroma from coffee beans.

[0018] The process parameters include: particle size, material-liquid ratio, extraction temperature, extraction time, and ultrasonic power.

[0019] Quality indicators include: yield, caffeine content, chlorogenic acid content, trigonelline, aroma and taste.

[0020] Furthermore, the number of samples should be able to cover the range of different process parameters and quality indicators so that the model can learn the inherent laws and patterns of the data.

[0021] In step (2), the collected data is preprocessed to remove abnormal data and normalize it to remove the influence of dimension. Then, the optimized sample data set is divided into training set and test set.

[0022] Furthermore, the Z-core standardization method was used to process the quality index data, and the calculation formula was as follows: Formula (1).

[0023] Formula (1): ; Where Z represents the standard score, X represents the value of the original data point, i represents the index of the data point, μ represents the mean value of the feature, and σ represents the standard deviation of the feature.

[0024] Furthermore, the dataset is divided into a training set and a test set at a ratio of 8:2. The training set is used for hyperparameter adjustment and model construction, and the test set is used to evaluate model performance.

[0025] In step (3), the preprocessed data set is input into the back propagation neural network model for training, and the information of process parameters and corresponding quality indicators in the training data samples is used to establish a nonlinear mapping relationship between process parameters and quality indicators using back propagation neural network learning.

[0026] The research of the present invention shows that, compared with models such as support vector machine (SVM), random forest (RF) and response surface algorithm, the prediction model constructed by using the backpropagation neural network model has higher accuracy.

[0027] Further, the backpropagation neural network model includes 1 input layer, 1 hidden layer and 1 output layer. The number of neurons in the input layer is m, the number of neurons in the output layer is n, and the number of neurons in the hidden layer is , where m is the number of independent variables, n is the number of optimization objectives, and a is a constant from 1 to 10. The independent variables are process parameters, and the optimization objectives are quality indicators.

[0028] Further, the parameters set for model training are as follows: the learning rate is 0.001 - 0.01, the allowable error for training is less than 1×10 -6 -1×10 -8 , and the number of iterations is 1000 - 10000.

[0029] In step (4), the backpropagation neural prediction model trained in step (3) is used as the objective function, and the NSGA-II algorithm is used. With maximizing or minimizing the quality indicator as the optimization objective and the parameter range of the equipment as the constraint condition of the optimization objective, the Pareto optimal solution set of the processing parameters is output. Maximizing the quality indicator means maximizing the dominant indicator, and minimizing the quality indicator means minimizing the inferior indicator.

[0030] The research of the present invention shows that, compared with the particle swarm optimization algorithm (PSO) and genetic algorithm (GA), the optimization model using the NSGA-II algorithm has higher accuracy.

[0031] Further, the processing method of the NSGA-II algorithm includes the following steps: S1. Coupling the trained backpropagation neural network into the NSGA-II algorithm to replace the fitness function; S2. Defining the optimization problem, including the number of input variables, the number of objective functions, the population size, the number of iterations, the crossover probability, the mutation probability, and the range of constraint variables.

[0032] Further, the population size of the NSGA-II algorithm is set to 100 - 200, the maximum number of iterations is 100 - 500 times, the crossover probability is set to 0.5 - 1, and the mutation probability is set to 0.001 - 0.1.

[0033] Furthermore, the population size of the NSGA-II algorithm is set to 100, the maximum number of iterations is 200 times, the crossover probability is set to 0.8, and the mutation probability is set to 0.05.

[0034] The beneficial effects of the present invention are: (1) The present invention first proposes to combine machine learning and multi-objective optimization algorithms to optimize the process conditions for the preparation of plant extracts. Specifically, first, data in the process of plant extract preparation is collected. Then, a backpropagation neural network model is used to establish a non-linear mapping relationship between process parameters and product quality, realizing the prediction of extraction processes and multiple quality indicators of products. Finally, the process parameters are optimized by combining multi-objective optimization algorithms. Using the optimized process parameters for the preparation of plant extracts can significantly improve the nutritional and sensory qualities of the extraction products.

[0035] (2) The present invention meets the optimization scenarios of multi-objectives, complexity, and high dimensions, and is applicable to the preparation of any plant extracts, with a wide application range. Different from traditional experience-based optimization and traditional model optimization, the present invention can achieve fast, accurate, multi-objective, and high-efficiency process optimization, reducing the trial-and-error cost. Brief Description of the Drawings

[0036] Figure 1 It is a flow chart of the optimization method for preparing high-quality plant extracts based on machine learning and multi-objective algorithms of the present invention.

[0037] Figure 2 It is the performance characterization of the BP-ANN model of the present invention.

[0038] Figure 3 It is the principal component analysis diagram of the output results of the present invention by BP-ANN-NSGA-II. Detailed Embodiments

[0039] The present invention will be further described below in conjunction with specific embodiments. The following embodiments are only used to illustrate the present invention and are not used to limit the scope of application of the present invention. Without departing from the spirit and essence of the present invention, any modification or replacement of the methods, steps, or conditions of the present invention shall fall within the scope of the present invention.

[0040] The test methods used in the following embodiments are all conventional methods unless otherwise specified; the materials, reagents, etc. used are, unless otherwise specified, reagents and materials that can be obtained from commercial channels.

[0041] The software used in the following embodiments is as follows: Matlab tool is used to construct a multi-objective optimization model, and the download address is HYPERLINK "http: / / matlab.p2hp.com" http: / / matlab.p2hp.com, with the version being 9.14.0.

[0042] Design expert is used for the design of orthogonal arrays. The download address is https: / / http: / / www.statease.com / , and the version is 13.0.5.

[0043] Python machine learning library scikit-learn: The download address for Python is HYPERLINK "https: / / www.python.org / downloads / windows / " https: / / www.python.org / downloads / windows / , the version is 3.11.4. Use the pip install scikit-learn command in the command prompt window to download the scikit-learn package, and the version is 1.3.0.

[0044] Example 1 This example provides a method for preparing high-quality osmanthus extract based on machine learning and multi-objective optimization algorithms. The technical route is as Figure 1 shown, and it includes the following steps: 1. Preparation of osmanthus extract Raw materials: The variety is Jin Gui. The flowers of osmanthus trees are collected and made into dried osmanthus after vacuum freeze-drying for more than 48 hours. Grind the dried osmanthus using a blender, and then sieve it (20 - 40 mesh).

[0045] Ultrasonic extraction: Weigh 10 g of osmanthus powder, add ultrapure water according to the corresponding solid-liquid ratio for extraction, and set the parameters of the ultrasonic extractor according to Table 1.

[0046] Concentration and drying: Filter the suspension after ultrasonic extraction under a vacuum of 0.01 MPa, collect the filtrate, rotary evaporate and concentrate it under a vacuum of 0.01 Mpa. The concentrated solution is pre-frozen at -80 °C for 12 h and then transferred to a vacuum freeze-dryer for drying for 24 hours until a light yellow osmanthus extract is obtained.

[0047] Finally, store the product at -80 °C for future use.

[0048] Table 1. Actual factors and coded factor levels used in the Box-Behnken method (osmanthus extract) Run <![CDATA[Liquid-to-solid ratio (mL / g) X 1 > <![CDATA[Time (min) X 2 > <![CDATA[Temperature (°C) X 3 > <![CDATA[Ultrasonic power (W) X 4 > <![CDATA[pH X 5 > 1 15（-1） 50（0） 60（0） 300（0） 5（0） 2 25（1） 50（0） 50（-1） 300（0） 5（0） 3 20（0） 50（0） 60（0） 250（-1） 5（0） 4 20（0） 60（1） 60（0） 300（0） 3（-1） 5 20（0） 50（0） 60（0） 250（-1） 5（0） 6 25（1） 50（0） 70（1） 350（1） 5（0） 7 20（0） 50（0） 70（1） 250（-1） 7（1） 8 25（1） 50（0） 50（-1） 300（0） 5（0） 9 20（0） 50（0） 60（0） 350（1） 3（-1） 10 15（-1） 50（0） 70（1） 250（-1） 5（0） 11 20（0） 40（-1） 60（0） 350（1） 5（0） 12 20（0） 50（0） 60（0） 300（0） 5（0） 13 20（0） 40（-1） 50（-1） 300（0） 5（0） 14 15（-1） 60（1） 60（0） 300（0） 5（0） 15 15（-1） 40（-1） 60（0） 300（0） 5（0） 16 20（0） 60（1） 60（0） 300（0） 5（0） 17 20（0） 40（-1） 60（0） 250（-1） 5（0） 18 20（0） 40（-1） 50（-1） 300（0） 7（1） 19 20（0） 50（0） 60（0） 250（-1） 3（-1） 20 25（1） 50（0） 60（0） 300（0） 7（1） 21 20（0） 40（-1） 70（1） 300（0） 3（-1） 22 20（0） 40（-1） 60（0） 300（0） 5（0） 23 20（0） 50（0） 60（0） 350（1） 5（0） 24 25（1） 60（1） 60（0） 300（0） 5（0） 25 20（0） 50（0） 60（0） 300（0） 5（0） 26 15（-1） 50（0） 60（0） 300（0） 3（-1） 27 20（0） 50（0） 70（1） 350（1） 5（0） 28 25（1） 50（0） 70（1） 250（-1） 5（0） 29 20（0） 50（0） 60（0） 300（0） 7（1） 30 20（0） 50（0） 60（0） 300（0） 7（1） 31 15（-1） 50（0） 60（0） 300（0） 3（-1） 32 20（0） 50（0） 50（-1） 350（1） 5（0） 33 25（1） 50（0） 60（0） 250（-1） 5（0） 34 20（0） 50（0） 70（1） 300（0） 7（1） 35 20（0） 50（0） 50（-1） 300（0） 7（1） 36 20（0） 60（1） 60（0） 250（-1） 5（0） 37 20（0） 50（0） 70（1） 300（0） 3（-1） 38 15（-1） 50（0） 60（0） 300（0） 7（1） 39 20（0） 50（0） 60（0） 350（1） 7（1） 40 15（-1） 50（0） 50（-1） 300（0） 5（0） 41 15（-1） 50（0） 60（0） 350（1） 5（0） 42 25（1） 50（0） 60（0） 300（0） 3（-1） 43 20（0） 50（0） 60（0） 300（0） 5（0） 44 20（0） 60（1） 60（0） 350（1） 5（0） 45 20（0） 50（0） 60（0） 300（0） 5（0） 46 20（0） 60（1） 50（-1） 300（0） 5（0） Note: Actual value (coded value).

[0049] 2. Formation of a dataset for detecting the quality of plant extracts 2.1 Determine the content of total phenyl ethanol glycosides using ultraviolet absorption spectrophotometry Prepare a stock solution of acteoside standard (1 mg / mL) and dilute it to concentrations of 10, 20, 50, 100, and 200 μg / mL. Measure the absorbance at a wavelength of 334 nm to construct the standard curve of acteoside.

[0050] Dilute the osmanthus extract with water to an appropriate concentration, and then add 200 μL to each well of a 96-well plate. Use a microplate reader to measure the absorbance of total phenylethanoid glycosides at 334 nm. Calculate the content of total phenylethanoid glycosides according to the acteoside standard curve, expressed as mg acteoside / g.

[0051] 2.2 Determine the contents of acteoside and salidroside by high performance liquid chromatography Accurately weigh 10 mg of acteoside and salidroside standards respectively, dissolve them in ultrapure water and make up to 10 mL, shake well, and prepare stock solutions with a concentration of 1 mg / mL for standby. Take an appropriate amount from the prepared stock solution and perform gradient dilution to make the concentrations 2.5, 5, 10, 30, 50, 100, 200 μg / mL. Filter the diluted solution through a 0.22 μm aqueous filter membrane and inject it into the high performance liquid chromatography system for analysis. Use the high performance liquid chromatography system to detect standard solutions of different concentrations, and create a standard curve based on the correlation between peak area and concentration.

[0052] Dissolve and dilute the osmanthus extract with water to an appropriate concentration, then filter the diluted solution through a 0.22 μm aqueous filter membrane and inject it into a liquid phase injection vial for analysis by the high performance liquid chromatography system. Compare the peak elution time and peak area of the standard, and calculate the contents of acteoside and salidroside in the osmanthus extract sample according to the standard curve.

[0053] Liquid chromatography conditions are as follows: Chromatographic column: C18 (4.6×250 mm, 5 μm) or equivalent specification; Detector: UV detector or diode array detector; Detection wavelength: 280 / 285 nm; Column temperature: 35 °C; Injection volume: 10 μL; Flow rate: 1.0 mL / min. Mobile phase: A is 0.1% acetic acid aqueous solution, B is acetonitrile. Gradient elution conditions are: 0 min, mobile phase A 90%, mobile phase B 10%; 10 min, mobile phase A 80%, mobile phase B 20%; 30 min, mobile phase A 75%, mobile phase B 25%; 31 min, mobile phase A 20%, mobile phase B 80%; 32 min, mobile phase A 80%, mobile phase B 20%; 33 min, mobile phase A 90%, mobile phase B 10%; 35 min, mobile phase A 90%, mobile phase B 10%.

[0054] 2.3 Detection of the yield of osmanthus extract The extraction of osmanthus extract was repeated three times for each set of parameters. The mass of the osmanthus extract was accurately weighed and divided by the mass of the raw material to obtain the yield of the osmanthus extract, and the results were presented as an average value.

[0055] 2.4 Determination of the color sensory quality of osmanthus extract using a color difference meter The osmanthus extract was dissolved in water to obtain a uniform color, and the color parameters of the osmanthus extract were measured using a portable color spectrophotometer. The measurements were performed in triplicate to obtain the average value and standard deviation. The results were expressed using CIELAB parameters (L*, a*, b*). L*, a*, and b* represent lightness, red-green, and yellow-blue attributes, respectively. A higher L* value indicates a lighter color of the OFE, while a lower L* value indicates a darker hue. A negative a* value corresponds to a green tone, while a positive a* value indicates red. Additionally, a lower negative b* value indicates an increased blue shade, while a higher positive b* value indicates a stronger yellow tone.

[0056] The quality data obtained using the above experimental methods and the corresponding process parameters formed a dataset for the next step of machine learning modeling. Specifically, in the machine learning modeling, the process parameters were used as the input variables of the machine learning model, and the quality indicators were used as the output variables of the machine learning model.

[0057] 3. Data preprocessing and dataset partitioning Since the obtained quality indicators were not on the same order of magnitude, the Z-core normalization method was used for processing. The calculation formula was provided in Equation (1), and this normalization process was implemented using the StandardScaler function in the scikit-learn (version 1.3.0) library in Python 3.11.4.

[0058] Equation (1): ; where Z represents the standard score, X represents the value of the original data point, i represents the index of the data point, μ represents the mean of the feature, and σ represents the standard deviation of the feature.

[0059] Dataset partitioning: The dataset was randomly partitioned into a training set that contained 80% of the samples for hyperparameter tuning and model construction, while the remaining 20% was used as a test set to evaluate the performance of the model.

[0060] 4. Construction of a backpropagation neural network model The above dataset was used to establish a non-linear mapping relationship between the process parameters and the quality indicators using backpropagation neural network learning. Then, seven quality indicators (yield, total acteoside content, verbascoside content, salidroside content, and CIELAB color parameters L*, a*, b*) were predicted respectively, and the prediction results were evaluated as follows: The backpropagation neural network model includes 1 input layer, 1 hidden layer, and 1 output layer. The number of neurons in the input layer is m, the number of neurons in the output layer is n, and the number of neurons in the hidden layer is , where m is the number of independent variables, n is the number of optimization objectives, and a is a constant between 1 and 10. The learning rate is 0.01, and the training allowable error is less than 1×10 -8 . The number of iterations is 1000.

[0061] 5. Model Performance Evaluation To evaluate the effectiveness of the machine learning algorithm on the training and test datasets, two commonly used key evaluation metrics were calculated: R² and RMSE. R² was calculated using Equation (2), and RMSE was calculated using Equation (3).

[0062] Equation (2): ; Equation (3): ; In the equations, SS res is the sum of the squares of the residuals, that is, the sum of the squared differences between the actual observed values and the model predicted values; SS tot is the total sum of squares, that is, the sum of the squared differences between the actual observed values and the average of the observed values; y, and represent the true value, predicted value, and average value respectively, N is the number of samples, and i represents the index of the data point. RMSE represents the standard deviation of the prediction error.

[0063] The model was tested using the test set partitioned in Step 3, and the parameters were adjusted using the grid search method. Finally, the hyperparameter combination with the highest R² and the lowest RMSE as shown Figure 2 is the optimal extract quality prediction model.

[0064] 6. NSGA-II Model Construction and Experimental Verification Using the NSGA-II algorithm, the trained backpropagation neural prediction model was used as the surrogate model, with maximizing seven quality indicators as the optimization objective and the parameter range of the equipment as the constraint condition of the optimization objective, and the Pareto optimal solution set of the processing parameters was output.

[0065] The population size of the NSGA-II algorithm was set to 100, the maximum number of iterations was set to 200 times, the crossover probability was set to 0.8, and the mutation probability was set to 0.05.

[0066] Principal component analysis was performed based on the results of NSGA-II as shown Figure 3 , and the corresponding specific process parameters are shown in Table 2.

[0067] Table 2. Principal Component Analysis of NSGA-II Operation Results Category Liquid ratio mL / g Time min Temperature °C Ultrasonic power W pH 1 23.4552（23） 40.2215（40） 59.2254（59） 339.0291（340） 6.2431（6.2） 2 15.7875（16） 52.0425（52） 53.7063（54） 343.3803（345） 5.2403（5.2） 3 15.0586（15） 52.4902（52） 50.0710（50） 256.0671（255） 6.2351（6.2） Note: The values outside the brackets are theoretical values, and the values inside the brackets are the accurate values of the equipment.

[0068] Experimental verification was carried out according to the process parameters in Table 2, and the results are shown in Table 3.

[0069] Table 3. Experimental verification of three types of multi-objective optimization results (BP-ANN-NSGA-II)

[0070] It can be seen from Table 3 that the relative error is less than 10%, indicating the reliability of the model.

[0071] Example 2 This example provides a method for preparing high-quality coffee extract based on machine learning and multi-objective optimization algorithms. The technical route is as Figure 1 shown, including the following steps: 1. Preparation of coffee extract Raw materials: The coffee bean variety is Catimor, and the main production area is Yunnan. Use a blender to grind the dried coffee beans, and then sieve to select coffee powders with particle sizes of 0.5 mm, 0.75 mm, and 1 mm respectively.

[0072] Ultrasonic extraction: Weigh 10 g of coffee powder, add ultrapure water according to the corresponding solid-liquid ratio for extraction, and set the parameters of the ultrasonic extractor according to Table 4.

[0073] Concentration and drying: Filter the leaching solution, concentrate the filtrate under reduced pressure, and pre-freeze the sample in a -80°C refrigerator for 12 h. Transfer the pre-frozen sample to a vacuum freeze dryer and dry for 24 hours until the coffee extract is obtained.

[0074] Finally, store the product at -80°C for future use.

[0075] Table 4. Actual factors and coded factor levels used in the Box-Behnken method (coffee extract) Run <![CDATA[Liquid-to-solid ratio (mL / g) X 1 > <![CDATA[Time (min) X 2 > <![CDATA[Temperature (°C) X 3 > <![CDATA[Ultrasonic power (W) X 4 > <![CDATA[Particle size (mm) X 5 > 1 15（-1） 50（0） 60（0） 300（0） 0.75（0） 2 25（1） 50（0） 50（-1） 300（0） 0.75（0） 3 20（0） 50（0） 60（0） 250（-1） 0.75（0） 4 20（0） 60（1） 60（0） 300（0） 0.5（-1） 5 20（0） 50（0） 60（0） 250（-1） 0.75（0） 6 25（1） 50（0） 70（1） 350（1） 0.75（0） 7 20（0） 50（0） 70（1） 250（-1） 1（1） 8 25（1） 50（0） 50（-1） 300（0） 0.75（0） 9 20（0） 50（0） 60（0） 350（1） 0.5（-1） 10 15（-1） 50（0） 70（1） 250（-1） 0.75（0） 11 20（0） 40（-1） 60（0） 350（1） 0.75（0） 12 20（0） 50（0） 60（0） 300（0） 0.75（0） 13 20（0） 40（-1） 50（-1） 300（0） 0.75（0） 14 15（-1） 60（1） 60（0） 300（0） 0.75（0） 15 15（-1） 40（-1） 60（0） 300（0） 0.75（0） 16 20（0） 60（1） 60（0） 300（0） 0.75（0） 17 20（0） 40（-1） 60（0） 250（-1） 0.75（0） 18 20（0） 40（-1） 50（-1） 300（0） 1（1） 19 20（0） 50（0） 60（0） 250（-1） 0.5（-1） 20 25（1） 50（0） 60（0） 300（0） 1（1） 21 20（0） 40（-1） 70（1） 300（0） 0.5（-1） 22 20（0） 40（-1） 60（0） 300（0） 0.75（0） 23 20（0） 50（0） 60（0） 350（1） 0.75（0） 24 25（1） 60（1） 60（0） 300（0） 0.75（0） 25 20（0） 50（0） 60（0） 300（0） 0.75（0） 26 15（-1） 50（0） 60（0） 300（0） 0.5（-1） 27 20（0） 50（0） 70（1） 350（1） 0.75（0） 28 25（1） 50（0） 70（1） 250（-1） 0.75（0） 29 20（0） 50（0） 60（0） 300（0） 0.5（1） 30 20（0） 50（0） 60（0） 300（0） 1（1） 31 15（-1） 50（0） 60（0） 300（0） 0.5（-1） 32 20（0） 50（0） 50（-1） 350（1） 0.75（0） 33 25（1） 50（0） 60（0） 250（-1） 0.75（0） 34 20（0） 50（0） 70（1） 300（0） 1（1） 35 20（0） 50（0） 50（-1） 300（0） 1（1） 36 20（0） 60（1） 60（0） 250（-1） 0.75（0） 37 20（0） 50（0） 70（1） 300（0） 0.5（-1） 38 15（-1） 50（0） 60（0） 300（0） 1（1） 39 20（0） 50（0） 60（0） 350（1） 1（1） 40 15（-1） 50（0） 50（-1） 300（0） 0.75（0） 41 15（-1） 50（0） 60（0） 350（1） 0.75（0） 42 25（1） 50（0） 60（0） 300（0） 0.5（-1） 43 20（0） 50（0） 60（0） 300（0） 0.75（0） 44 20（0） 60（1） 60（0） 350（1） 0.75（0） 45 20（0） 50（0） 60（0） 300（0） 0.75（0） 46 20（0） 60（1） 50（-1） 300（0） 0.75（0） Note: Actual value (coded value).

[0076] 2. Quality detection of plant extract 2.1 Caffeine content Weigh a certain amount of caffeine standard, dissolve it with methanol and make up the volume to a certain value to prepare a standard solution with a concentration of 2 mg / mL. Dilute the caffeine standard solution with methanol to different concentrations (such as 10 μg / mL, 20 μg / mL, 40 μg / mL, etc.). Filter the diluted solution through a 0.45 μm filter membrane, and the filtrate is used for liquid chromatography analysis. Use a high-performance liquid chromatography system to detect standard solutions with different concentrations, and create a standard curve based on the correlation between the peak area and the concentration.

[0077] The liquid chromatography conditions are as follows: Chromatographic column: C18 (4.6×250 mm, 5 μm) or equivalent specification; Detector: UV detector or diode array detector; Detection wavelength: 272 nm; Column temperature: 30 °C; Injection volume: 10 μL; Flow rate: 1.0 mL / min. Mobile phase: A is methanol, B is water.

[0078] Weigh 1 g of coffee extract, dissolve it with water, shake well, let it stand, and take the supernatant for filtration through a filter membrane for liquid chromatography detection.

[0079] 2.2 Chlorogenic acid content Prepare a stock solution of chlorogenic acid (2 mg / mL) and dilute it to concentrations of 2, 10, 20, 40, and 80 μg / mL. Filter the diluted solution through a 0.45 μm filter membrane, and the filtrate is used for liquid chromatography analysis. Use a high-performance liquid chromatography system to detect standard solutions with different concentrations, and create a standard curve based on the correlation between the peak area and the concentration.

[0080] The liquid chromatography conditions are as follows: Chromatographic column: C18 (4.6×250 mm, 5 μm) or equivalent specification; Detector: UV detector or diode array detector; Detection wavelength: 327 nm; Column temperature: 35 °C; Injection volume: 10 μL; Flow rate: 1.0 mL / min. Mobile phase: A is 0.5% acetic acid solution, B is acetonitrile.

[0081] Weigh 1 g of coffee extract, dissolve it with water, shake well, let it stand, and take the supernatant for filtration through a filter membrane for liquid chromatography detection.

[0082] 2.3 Trigonelline Prepare a stock solution of trigonelline (2 g / mL) and dilute it to concentrations of 1, 2, 5, 10, 20, and 50 mg / L. Filter the diluted solution through a 0.45 μm aqueous filter membrane, and the filtrate is used for liquid chromatography analysis. Use a high-performance liquid chromatography system to detect standard solutions with different concentrations, and create a standard curve based on the correlation between the peak area and the concentration.

[0083] The liquid chromatography conditions are as follows: amino column: C18 (4.6×250 mm, 5 μm) or equivalent; detector: ultraviolet detector or diode array detector; detection wavelength: 260 nm; column temperature: 35 °C; injection volume: 10 μL; flow rate: 1.0 mL / min. Mobile phase: A is methanol and B is water.

[0084] Weigh 1 g of coffee extract, dissolve it in water, shake well, let it stand, and take the supernatant for filtration through a filter membrane for liquid phase detection.

[0085] 2.4 Aroma and taste The aroma and taste scores of the coffee extract are determined by professional coffee tasters, and the aroma and taste of the coffee extract are quantified according to the coffee SCA scoring system.

[0086] 3. Data preprocessing Since the obtained quality indicators are not in the same order of magnitude, the Z-core standardization method is used for processing, and the method is the same as that in Example 1.

[0087] Dataset division: The dataset is randomly divided into a training set, which contains 80% of the samples for hyperparameter tuning and model construction, and the remaining 20% is used as a test set to evaluate the model performance.

[0088] 4. Construction of backpropagation neural network model The above dataset uses a backpropagation neural network to learn the non-linear mapping relationship between the features automatically extracted from the processing parameters and the quality indicators, and then predicts the caffeine content, chlorogenic acid content, trigonelline, aroma and taste respectively and evaluates according to the prediction results; the parameter settings for constructing the backpropagation neural network model refer to Example 1.

[0089] 5. Model performance evaluation To evaluate the effectiveness of the machine learning algorithm on the training and test datasets, two commonly used key evaluation indicators are calculated: R² and RMSE. The method is the same as that in Example 1.

[0090] Use the divided test set in Step 3 for model testing, and use the grid search method for parameter tuning. Finally, the hyperparameter combination with the highest R² and the lowest RMSE is the optimal extract quality prediction model.

[0091] 6. Construction of NSGA-II model and experimental verification The NSGA-II algorithm is adopted, and the trained backpropagation neural prediction model is used as the surrogate model. Taking the maximization or minimization of the required quality indicators and the parameter range of the equipment as the optimization objective and the constraint conditions of the optimization objective, the Pareto optimal solution set of the processing parameters is output.

[0092] The population size of the NSGA-II algorithm is set to 100, the maximum number of iterations is 200, the crossover probability is set to 0.8, and the mutation probability is set to 0.05.

[0093] Principal component analysis dimensionality reduction is performed based on the results of NSGA-II, and the corresponding specific process parameters are shown in Table 5.

[0094] Table 5. Principal component analysis of the NSGA-II operation results

[0095] Note: The values outside the brackets are theoretical values, and the values inside the brackets are the accurate values of the equipment. Experimental verification is carried out according to the process parameters in Table 5, and the results are shown in Table 6.

[0096] Table 6. Experimental verification of the multi-objective optimization results of three categories (BP-ANN-NSGA-II)

[0097] It can be seen from Table 6 that the relative error is less than 10%, indicating the reliability of the model.

[0098] Comparative Example 1 This comparative example uses the response surface method, support vector machine (SVM), and random forest (RF) to optimize the preparation method of osmanthus extract. The method is as follows: The process parameters and corresponding quality indicators are obtained by preparing osmanthus extract to form a data set, the same as in Example 1. Weights are assigned to multiple quality indicators and converted into a comprehensive objective function. A model is constructed using the response surface and other machine learning algorithms, and performance evaluation is carried out. Experimental verification is carried out according to the model optimization results.

[0099] Specifically, since the model output of the response surface optimization method or other machine learning algorithms is a single dependent variable and cannot directly adapt to multi-objective optimization, it is necessary to rely on weight assignment to convert the multi-objective problem into a single-objective problem, and the weight reflects the importance of each objective.

[0100] In python, the quality indicators of the extract are converted into a comprehensive objective function using formula (4).

[0101] Formula (4):

[0102] In the formula: is the value of the comprehensive objective function to be optimized: is the actual value of each optimization objective, which can be obtained through experiments, is the weight factor of the i-th optimization objective, mainly representing the importance degree of each sub-objective function.

[0103] When conducting research in the article, each sub-objective function is regarded as equally important, so there is = ±1 (i = 1, 2, 3, 4, ……). According to the experimental results, the system response values under different combinations of variable level values are obtained and are transformed into a function related to according to formula (4).

[0104] The polyfit function in Matlab is used to implement the polynomial fitting of the response surface algorithm, and R² and RMSE are used to evaluate the prediction performance of the algorithm. The results are shown in Table 7.

[0105] The machine learning algorithm modeling is implemented using the Python machine learning library scikit-learn to select the model with the best performance, and R² and RMSE are used to evaluate the prediction performance of different algorithms to select the algorithm with the best prediction performance. The specific algorithms are as follows: The support vector regression algorithm is used for training on the training set, and the optimal parameters are obtained using the grid search method. The parameters considered are: kernel function, regularization coefficient, Gamma parameter, etc.

[0106] The random forest regression algorithm is used for training on the training set, and the optimal parameters are obtained using the grid search method. The parameters considered are: the number of trees (n_estimators), the maximum depth of the tree (max_depth), the minimum number of samples required for node splitting (min_samples_split), the minimum number of samples required for leaf nodes (min_samples_leaf), the number of features considered for the best split (max_features), resampling (bootstrap = True).

[0107] Table 7. Performance characterization of the response surface algorithm and machine learning algorithm models Model <![CDATA[Training set R 2 > RMSE of training set R² of test set RMSE of training set Support vector machine 0.67 1.42 0.21 0.69 Random forest 0.70 1.56 0.17 0.58 Response surface algorithm 0.55 2.45 0.49 0.77 The model performance evaluations of SVM, RF, and RSM are shown in Table 7, and the corresponding optimization results are shown in Table 8.

[0108] Table 8. Single-objective optimization results of three machine learning algorithms

[0109] Experimental verification is carried out according to the experimental parameters in Table 8, and the results are shown in Table 9.

[0110] Table 9. Experimental verification of single-objective optimization results of three machine learning algorithms

[0111] As shown in Table 9, the relative error is greater than 10%, indicating the unreliability of the model. Therefore, the use of the weight allocation method often over-simplifies the multi-objective optimization problem and may not produce accurate optimization results.

[0112] Comparative Example 2 Taking the process optimization of the preparation process of osmanthus extract in Example 1 as an example, an artificial neural network (ANN) model of processing parameters and optimization objectives was established. Based on the constructed ANN model as the objective function, PSO and GA algorithms were used, and multiple process parameters were optimized according to the algorithm results for experimental verification. The specific steps are as follows: 1. Obtain process parameters and corresponding quality indicators by preparing osmanthus extract to form a data set, the same as steps 1-3 of Example 1.

[0113] 2. Construction of artificial neural network model The above data set uses artificial neural network learning to establish a non-linear mapping relationship between process parameters and quality indicators, and then predicts seven quality indicators (yield, total phenylethanoid glycoside content, acteoside content, salidroside content, and CIELAB color parameters L*, a*, b*) respectively and evaluates according to the prediction results, as follows: The artificial neural network model includes 1 input layer, 1 hidden layer and 1 output layer. The number of neurons in the input layer is m, the number of neurons in the output layer is n, and the number of neurons in the hidden layer is , where m is the number of independent variables, n is the number of optimization objectives, and a is a constant from 1 to 10. The learning rate is 0.01, and the training allowable error is less than 1×10 -8 , and the number of iterations is 1000.

[0114] 3. Model performance evaluation, the method is the same as step 5 of Example 1.

[0115] 4. Construction and experimental verification of PSO and GA models Using PSO and GA algorithms, the trained artificial neural prediction model is used as the surrogate model, with maximizing seven quality indicators as the optimization objective and the parameter range of the equipment as the constraint condition of the optimization objective, and the Pareto optimal solution set of processing parameters is output.

[0116] The number of particles of the PSO algorithm is set to 40, the inertia weight is set to 0.7, the learning factor is set to 2.0, the maximum speed is set to 0.1, and the number of iterations is set to 1000.

[0117] The population size of the GA algorithm is set to 100, the maximum number of iterations is 200 times, the crossover probability is set to 0.8, and the mutation probability is set to 0.05.

[0118] Perform principal component analysis based on the results of PSO and GA, and the corresponding experimental verification results are shown in Table 10.

[0119] Table 10. Experimental verification of ANN model and different multi-objective optimization results

[0120] The results show that taking ANN as the objective function, the relative error between the theoretical values and the actual values obtained by using PSO and GA algorithms respectively is greater than that of BP-ANN-NSGA-II, indicating the relative inaccuracy of the model.

Claims

1. A method for optimizing the preparation of plant extracts based on machine learning and multi-objective algorithms, characterized in that: The following steps are involved: (1) Obtaining process parameters and corresponding quality indicators by preparing plant extracts to form a data set; the number of process parameters corresponding to a plant extract is ≥4, and the number of quality indicators is ≥2; (2) Screening the sample data set, removing outliers and normalizing it, and then dividing the optimized sample data set into a training set and a test set; (3) Using the process parameters in the training set as input and the quality indicators as output, the back propagation neural network model is used for training to establish a nonlinear mapping relationship between the process parameters and the quality indicators. The test set is then used to test and optimize the model parameters to obtain a prediction model. (4) The NSGA-II algorithm is used to take the prediction model as a proxy model, maximize or minimize the quality index as the optimization target, and take the applicable process parameter range as the constraint condition of the optimization target, and output the Pareto optimal solution set of the process parameters.

2. The method for optimizing the preparation of plant extracts based on machine learning and multi-objective algorithm according to claim 1, characterized in that: The preparation process of the plant extract comprises: raw material crushing, ultrasonic-assisted extraction, filtration, concentration and drying.

3. The method for optimizing the preparation of plant extracts based on machine learning and multi-objective algorithm according to claim 1 or 2, characterized in that: The process parameters include: at least four of raw material particle size, material-liquid ratio, pH value, leaching temperature, leaching time, and ultrasonic power.

4. The method for optimizing the preparation of plant extracts based on machine learning and multi-objective algorithm according to claim 1, characterized in that: The quality indicators include: at least two of yield, active ingredient content, color, smell, and taste.

5. The method for optimizing the preparation of plant extracts based on machine learning and multi-objective algorithm according to claim 1, characterized in that: In step (2), the quality index data is processed using the Z-core standardization method.

6. The method for optimizing the preparation of plant extracts based on machine learning and multi-objective algorithm according to claim 1, characterized in that: In step (2), the data set is divided into a training set and a test set in a ratio of 8:

2.

7. The method for optimizing the preparation of plant extracts based on machine learning and multi-objective algorithm according to claim 1, characterized in that: In step (3), the back propagation neural network model includes 1 input layer, 1 hidden layer and 1 output layer, the number of neurons in the input layer is m, the number of neurons in the output layer is n, and the number of neurons in the hidden layer is , where m is the number of independent variables, n is the number of optimization targets, and a is a constant between 1 and 10; the learning rate is 0.001-0.01, and the training error is allowed to be less than 1×10 -6 -1×10 -8 , the number of iterations is 1000-10000.

8. The method for optimizing the preparation of plant extracts based on machine learning and multi-objective algorithm according to claim 1, characterized in that: In step (3), the model is tested using the test set, the determination coefficient R² and the root mean square error RMSE are calculated, and the model parameters are adjusted using the grid search method to obtain the hyperparameter combination with the highest R² and the lowest RMSE, thereby obtaining the prediction model.

9. The method for optimizing the preparation of plant extracts based on machine learning and multi-objective algorithm according to claim 1, characterized in that: In step (4), the processing method of the NSGA-II algorithm includes the following steps: S1. The trained back-propagation neural network is used as a proxy model and coupled to the NSGA-II algorithm to replace the fitness function. S2. Define the optimization problem, including the number of input variables, the number of objective functions, the population size, the number of iterations, the crossover probability, the mutation probability, and the range of the constraint variables.

10. The method for optimizing the preparation of plant extracts based on machine learning and multi-objective algorithm according to claim 9, characterized in that: The population size of the NSGA-II algorithm is set to 100-200, the maximum number of iterations is 100-500, the crossover probability is set to 0.5-1, and the mutation probability is set to 0.001-0.1.

Images