Machine learning based sweet potato amylase content analysis method and system

Abstract

The application provides a sweet potato amylase content analysis method and system based on machine learning, comprising: constructing a three-dimensional state space matrix according to a denoised time sequence, normalizing and mapping temperature gradient, pH fluctuation, time cumulative quantity and depth displacement quantity by using a fuzzy membership function to obtain a standardized state vector; determining amylase activity values under different state vectors by using a laboratory enzyme marker, establishing a sample mapping table of the state vector and the enzyme activity, removing outlier samples by using a quartile method to obtain cleaned training data matrix; if a combination of real-time collected key parameters exceeds a preset threshold range, triggering a loss function calculation module based on gradient descent, adjusting a parameter update step by using a weight decay coefficient, and outputting a process parameter correction amount; inputting an optimized population into an enzyme activity prediction model for iterative verification, and if a variance of a fitness function of three consecutive generations is less than a preset value, locking a final parameter combination and completing closed-loop regulation.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of information technology, and particularly to a sweet potato amylase content analysis method and system based on machine learning. BACKGROUND

[0002] As an important field of agricultural product deep processing, the quality of sweet potato starch products directly affects the development level of the food industry and the biological product industry. In the extraction process of sweet potato starch, soaking is a key link that determines the amylase content and activity in the final product. The content and activity of amylase not only affect the yield and quality of starch, but also are the core indicators for evaluating the stability of product quality.

[0003] Currently, in the control of the soaking process of sweet potato starch processing, it mainly relies on traditional manual experience or simple time-temperature program control. These methods simplify the complex soaking process into the control of several isolated parameters, and the operation is relatively extensive. However, the soaking process is a complex, multi-variable, non-linear dynamic system, in which there is a strong dynamic coupling relationship between the temperature gradient, pH fluctuation, material settlement (depth displacement) and soaking time and other process parameters. For example, the temperature difference at different depths in the soaking tank will affect the chemical reaction rate in the local area, and the change of pH will in turn affect the activity and stability of the enzyme, and these effects evolve over time. Although existing detection techniques can measure the amylase content, they lack the ability to analyze the dynamic change law in the soaking process, and cannot establish a quantitative relationship model between the soaking parameters and the amylase activity. At the same time, the traditional method is difficult to identify the key influencing factors among the numerous process parameters, resulting in a lack of pertinence and scientific basis for process optimization.

[0004] The change of amylase content in the soaking process of sweet potato presents a high degree of nonlinearity. The complexity of this nonlinear relationship makes it impossible for traditional linear analysis methods to accurately describe the internal relationship between the soaking degree and the amylase activity. The soaking degree, as a comprehensive indicator, involves parameters such as soaking depth and soaking completeness, and there is a mutual coupling and dynamic influence relationship among these parameters, which further aggravates the difficulty of modeling analysis. Due to the lack of effective quantitative evaluation system of soaking degree, the existing methods cannot accurately predict the change of amylase content in the soaking process, and it is also difficult to provide reliable theoretical guidance for the optimization and control of process parameters.

[0005] Therefore, how to construct a prediction model that can accurately describe the nonlinear relationship between the soaking degree and the amylase activity, and on this basis, establish a scientific quantitative evaluation system of soaking degree, and identify the key process parameters that are most sensitive to the amylase content, so as to realize the accurate prediction and intelligent control of amylase content in the soaking process of sweet potato, has become a key problem that needs to be solved in the development of sweet potato starch processing technology. SUMMARY

[0006] The purpose of the present application is to provide a sweet potato amylase content analysis method and system based on machine learning, aiming to solve the technical problems of inaccurate prediction and analysis of amylase content during the soaking process of sweet potatoes and low intelligence of control strategies in the prior art.

[0007] In one aspect, the present application provides a sweet potato amylase content analysis method based on machine learning, mainly comprising:

[0008] Temperature gradient data, pH dynamic fluctuation values, soaking time accumulation and depth displacement during the soaking process of sweet potatoes are collected by multiple sensors, high-frequency noise is eliminated by sliding window filtering, and denoised process parameter time series are obtained;

[0009] A three-dimensional state space matrix is constructed according to the denoised time series, and a fuzzy membership function is used to normalize and map the temperature gradient, pH fluctuation, time accumulation and depth displacement, to obtain a standardized state vector;

[0010] The amylase activity values under different state vectors are determined by a laboratory enzyme marker, a sample mapping table of state vectors and enzyme activity is established, outlier samples are removed by quartile method, and cleaned training data matrix is obtained;

[0011] A support vector regression algorithm is used to perform nonlinear fitting on the training data matrix, a high-dimensional mapping relationship between the state vector and the enzyme activity is processed by a radial basis kernel function transformation, a learning rate adjustment mechanism in gradient descent is introduced to dynamically optimize the kernel function parameters, and an enzyme activity prediction model is obtained;

[0012] The partial derivative matrix of each process parameter is calculated according to the output results of the prediction model, the importance of the temperature gradient, pH fluctuation, and time accumulation is sorted by the fitness function in genetic algorithm, and the key parameter combination is determined by the roulette selection operator;

[0013] If the real-time collected key parameter combination exceeds the preset threshold range, a loss function calculation module based on gradient descent is triggered, the parameter update step is adjusted by the weight decay coefficient, and the process parameter correction amount is output;

[0014] The current soaking state vector is recombined by chromosome coding according to the correction amount, a new parameter combination is generated by setting rules of the crossover probability of genetic algorithm, the optimal solution is selected by the elite reservation strategy, and the optimized process parameter population is obtained;

[0015] The optimized population is input into the enzyme activity prediction model for iterative verification, if the variance of the fitness function of the continuous three generations is less than the preset value, the final parameter combination is locked, and closed-loop regulation is completed.

[0016] In another aspect, the present application provides a sweet potato amylase content analysis system based on machine learning, mainly comprising:

[0017] A data acquisition and preprocessing module is configured to collect temperature gradient data, pH dynamic fluctuation values, soaking time accumulation and depth displacement during the soaking process of sweet potatoes by using multiple sensors, eliminate high-frequency noise by sliding window filtering, and obtain denoised process parameter time series.

[0018] A state space construction module is configured to construct a three-dimensional state space matrix according to the denoised time series, normalize and map the temperature gradient, pH fluctuation, time accumulation and depth displacement by using a fuzzy membership function, and obtain a standardized state vector.

[0019] A sample mapping and cleaning module is configured to determine the amylase activity value under different state vectors by using a laboratory enzyme marker, establish a sample mapping table of state vectors and enzyme activity, eliminate outliers by using a quartile method, and obtain cleaned training data matrix.

[0020] A model training module is configured to perform nonlinear fitting on the training data matrix by using a support vector regression algorithm, transform and process the high-dimensional mapping relationship between the state vector and the enzyme activity by using a radial basis kernel function, introduce a learning rate adjustment mechanism in gradient descent to dynamically optimize the kernel function parameters, and obtain an enzyme activity prediction model.

[0021] A parameter importance analysis module is configured to calculate the partial derivative matrix of each process parameter according to the output result of the prediction model, sort the importance of the temperature gradient, pH fluctuation, and time accumulation by using the fitness function in genetic algorithm, and determine the key parameter combination by using the roulette wheel selection operator.

[0022] A dynamic adjustment module is configured to trigger a loss function calculation module based on gradient descent if the real-time collected key parameter combination exceeds the preset threshold range, adjust the parameter update step by using a weight decay coefficient, and output the process parameter correction amount.

[0023] A genetic optimization module is configured to perform chromosome coding recombination on the current soaking state vector according to the correction amount, generate a new parameter combination by using the crossover probability setting rule of genetic algorithm, filter the optimal solution by using the elite reservation strategy, and obtain the optimized process parameter population.

[0024] An iterative convergence module is configured to input the optimized population into the enzyme activity prediction model for iterative verification, lock the final parameter combination if the variance of the fitness function is less than the preset value for three consecutive generations, and complete the closed-loop adjustment.

[0025] In summary, the present application discloses an intelligent optimization method for sweet potato soaking process, which is aimed at the problems of dynamic coupling of process parameters collected by multiple sensors and nonlinear response of enzyme activity. The standardized state vector is constructed by sliding window filtering and fuzzy membership normalization, and the parameter adaptive optimization is realized by combining support vector regression and genetic algorithm. Firstly, the present application uses the radial basis kernel function to fit the high-dimensional mapping relationship between the state vector and enzyme activity, dynamically adjusts the kernel parameter by gradient descent to improve the prediction accuracy, and then identifies the key parameter combination based on the partial derivative matrix and roulette selection. When the real-time data exceeds the threshold, the weight decay mechanism is triggered to calculate the correction amount. Finally, the optimized population is generated by chromosome coding recombination and elite reservation strategy, and the optimal process parameters are locked through iteration verification. This method realizes the closed-loop collaborative regulation of temperature gradient, pH fluctuation and time accumulation, significantly reduces the prediction error of sweet potato starch enzyme activity, and improves the stability of sweet potato soaking process. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 Flowchart of the sweet potato starch enzyme content analysis method based on machine learning of the present application;

[0027] Figure 2 Schematic diagram of state vector construction and model training of the present application;

[0028] Figure 3 Schematic diagram of parameter importance analysis and optimization of the present application;

[0029] Figure 4 Schematic diagram of closed-loop dynamic adjustment and optimization of the present application;

[0030] Figure 5 Structure schematic diagram of the sweet potato starch enzyme content analysis system based on machine learning of the present application;

[0031] Figure 6 Combined schematic diagram of SVR (Support Vector Regression) model performance verification according to an embodiment of the present application and comparison with prior art. DETAILED DESCRIPTION

[0032] The technical solutions in the embodiments of the present application will be described clearly and in detail below with reference to the drawings in the embodiments of the present application. The described embodiments are only a part of the embodiments of the present application. Embodiment one

[0033] As Figures 1-4 , the sweet potato starch enzyme content analysis method and system based on machine learning of the present embodiment can specifically include:

[0034] In step S101, the temperature gradient data, the dynamic fluctuation value of the pH value, the cumulative amount of soaking time and the depth displacement amount in the sweet potato soaking process are collected by using multiple sensors, and the high-frequency noise is eliminated by sliding window filtering to obtain the denoised process parameter time sequence.

[0035] As shown in Figure 2 The temperature gradient, the pH value, the soaking time and the depth displacement data are synchronously collected by using multiple sensors to generate the original process parameter data set. The original process parameter data set is processed by sliding window filtering to remove the high-frequency noise and obtain the denoised process parameter time sequence. According to the denoised process parameter time sequence, the dynamic correlation coefficient between the temperature gradient and the pH value is calculated to determine the coupling relationship therebetween. If the absolute value of the dynamic correlation coefficient is greater than a preset threshold value, the temperature gradient and the pH value are jointly analyzed to obtain the synergistic influence characteristics of both on the soaking time and the depth displacement. The synergistic influence characteristics are smoothed by using the Kalman filtering algorithm to generate a stable process parameter change trend. According to the stable process parameter change trend, the support vector regression algorithm is used to predict the dynamic evolution of the subsequent process parameters to obtain the predicted time sequence. If the deviation between the predicted time sequence and the actual collected data exceeds a preset threshold value, the sensor data collection frequency is adjusted to re-generate the original process parameter data set.

[0036] For example, during the sweet potato soaking process, the multi-sensor system first collects key process parameter data, such as temperature gradient data, which can be obtained by temperature sensors arranged at different depths (e.g., 0.5 meters, 1.0 meters, 1.5 meters) in the soaking tank. Assuming that the collected temperature values are 25.3°C, 24.8°C, and 24.2°C, respectively, a gradient distribution is formed, reflecting the heat transfer characteristics in the tank. The dynamic fluctuation value of the pH is collected every 5 minutes using a pH sensor, resulting in a set of data such as 6.5, 6.7, and 6.4, showing the slight fluctuation of the pH. The soaking time accumulation is recorded by a timer, assuming that the current soaking time is 2.5 hours. The depth displacement is measured by an ultrasonic sensor to measure the sinking depth of the sweet potato in the tank, assuming that the initial depth is 0.8 meters and the current depth is 0.85 meters, reflecting the sinking trend. Next, the collected time series data is denoised, and a sliding window filtering algorithm is used. For example, using a 5-point mean filter, for the temperature data sequence [25.3, 25.2, 25.4, 25.1, 25.3], the mean value of the first window is calculated as (25.3+25.2+25.4+25.1+25.3) / 5=25.26°C. The smoothed sequence is obtained by sliding the window, eliminating the influence of high-frequency noise. Similar methods are applied to the pH value and depth displacement data to ensure data stability. The denoised time series can be used for subsequent analysis, such as evaluating the uniformity of heat distribution through the temperature gradient change rate (ΔT / Δh= (25.3-24.2) / 1.0=1.1°C / meter), judging the stability of the chemical reaction by combining the pH fluctuation amplitude (maximum 6.7, minimum 6.4, fluctuation 0.3), and analyzing the sinking rate (0.05 / 2.5=0.02 meters / hour) by correlating the soaking time accumulation of 2.5 hours with the depth displacement increment of 0.05 meters, providing data support for optimizing the soaking process. The above processes are realized through automatic collection by sensors and algorithm processing, and the data analysis results can be stored in a database to form a dynamic monitoring closed loop of process parameters, ensuring the quality of sweet potato soaking.

[0037] In step S102, a three-dimensional state space matrix is constructed based on the denoised time series, and a fuzzy membership function is used to normalize and map the temperature gradient, pH fluctuation, time accumulation, and depth displacement, obtaining a standardized state vector.

[0038] The data after denoising is obtained from the time sequence, the main characteristics are extracted by using the principal component analysis method to generate an initial feature set. If the dimension of the initial feature set exceeds a preset threshold, the main components are retained through dimension reduction processing to obtain a compressed feature set. According to the compressed feature set, a three-dimensional state space matrix is constructed, and the matrix element distribution is determined by using the grid division method. For the three-dimensional state space matrix, the fuzzy membership function is applied to perform normalized mapping on the temperature gradient, the pH fluctuation, the time cumulative quantity and the depth displacement quantity to generate a normalized feature vector. If the variance of the normalized feature vector is lower than a preset threshold, the feature distribution is adjusted through kernel density estimation to obtain an optimized feature vector. According to the optimized feature vector, the k-means clustering algorithm is used to divide the state space to determine the state category. Through the state category, a standardized state vector is generated and stored as the basic data for subsequent analysis.

[0039] In the embodiment, the fuzzy membership function preferably adopts a trapezoidal or Gaussian membership function, because it can better reflect the smooth transition of the process parameters in the 'good / fair / poor' interval.

[0040] For example, in the process of constructing a three-dimensional state space matrix and subsequent processing, first, for the denoised time series data, assuming we have a set of temperature data sequences, after denoising processing, 100 time point data is obtained, each time point contains temperature value, pH value and depth value, the data range is respectively temperature 20 to 50 degrees Celsius, pH 4 to 9, depth 0 to 10 meters. When constructing a three-dimensional state space matrix, the data of the 100 time points is mapped into a three-dimensional matrix, with time, variable type and value as dimensions, where the time axis is 100, the variable type is 3 (temperature, pH, depth), and the value is the specific measured value, for example, the value of the first time point is [25.5, 6.2, 3.4]. Then, the fuzzy membership function is used to normalize and map the temperature gradient, pH fluctuation, time accumulation and depth displacement. The specific method is to calculate the gradient or fluctuation value of each variable, for example, the temperature gradient is calculated by the difference between adjacent time points, and the gradient sequence is obtained, for example, the temperature gradient from the first to the second time point is 0.3 degrees Celsius per unit time; the pH fluctuation is calculated by standard deviation, assuming that the standard deviation of the pH of the 100 time points is 0.8; the time accumulation is the serial number of the time point, the sum is 5050; the depth displacement is the change amplitude of the depth value, assuming that the maximum change is 2.5 meters. Then, the fuzzy membership function is designed to map these values to the range of 0 to 1, for example, the membership function of the temperature gradient is a linear function, when the gradient is less than 0.1, the membership degree is 0, and when it is greater than 1, the membership degree is 1, and the intermediate linear interpolation is calculated to obtain the temperature gradient membership degree of the first to the second time point, which is 0.3. Similarly, the pH fluctuation membership degree is designed based on the standard deviation, and the standard deviation 0.8 corresponds to the membership degree 0.6; the time accumulation and depth displacement also obtain the membership degree value through similar linear mapping, and finally form the standardized state vector, for example, the vector of a certain time point is [0.3, 0.6, 0.5, 0.4]. Through the above method, the system automatically completes data processing and normalization, and the generated vector can be used for subsequent state analysis or prediction model input, ensuring data consistency and comparability.

[0041] In step S103, the amylase activity value under different state vectors is measured by a laboratory enzyme marker, a sample mapping table of state vector and enzyme activity is established, and the quartile method is used to remove outliers to obtain the cleaned training data matrix.

[0042] As Figure 2As shown, the amylase activity values corresponding to different state vectors are collected by the enzyme marker to generate an initial data set. The quartile method is used to analyze the sample distribution of the initial data set to determine outliers. If the sample activity value exceeds the preset threshold of the interquartile range, the sample is removed to obtain the cleaned data set. According to the cleaned data set, a mapping table of state vectors and amylase activity is constructed. Features are extracted from the mapping table to generate a training data matrix. The principal component analysis method is used to reduce the dimension of the training data matrix to obtain an optimized feature matrix. The linear regression algorithm is used to model the optimized feature matrix to obtain the prediction model between the state vector and the enzyme activity.

[0043] For example, the amylase activity values under different state vectors are measured by the laboratory enzyme marker, a sample mapping table of state vectors and enzyme activity is established, and the quartile method is used to remove outliers, and the whole process of finally obtaining the cleaned training data matrix can be implemented as follows: First, the amylase activity under 10 different state vectors is measured by the enzyme marker, assuming that the state vector includes two dimensions of temperature and pH value, and the measured activity value data are as follows: state 1 (temperature 25°C, pH 6.5) activity value is 120.5 units / ml, state 2 (temperature 30°C, pH 7.0) activity value is 135.2 units / ml, and so on to state 10 (temperature 50°C, pH 8.5) activity value is 98.3 units / ml. These data are automatically recorded and stored as an initial sample mapping table by the instrument. Then, the activity value data is imported into the data processing system, and the quartiles are calculated to remove outliers, and the specific algorithm is as follows: First, sort the 10 activity values to get the median as 122.8 units / ml, the first quartile (Q1) as 110.4 units / ml, the third quartile (Q3) as 140.6 units / ml, the interquartile range (IQR) as 30.2 units / ml, the lower limit of the outlier range as Q1-1.5*IQR=65.1 units / ml, and the upper limit as Q3+1.5*IQR=185.9 units / ml. The system automatically filters and finds that the activity value of state 10 is 98.3 within the range, but if the sample activity value is lower than 65.1 or higher than 185.9, it is removed. Assuming there is no outlier, all 10 samples are retained. Finally, the cleaned data matrix is generated by the system, which includes a two-dimensional array of state vectors and corresponding activity values, such as the matrix row format [temperature, pH, activity value], for example, [25, 6.5, 120.5], and is stored as a training data file for subsequent modeling. The whole process forms a logical closed loop through data collection, algorithm calculation and matrix generation. If the sample size is insufficient, more state vector tests can be expanded to improve data diversity to ensure the accuracy of model training.

[0044] Step S104, a support vector regression algorithm is used to perform nonlinear fitting on the training data matrix, a high-dimensional mapping relationship between the state vector and the enzyme activity is processed through a radial basis kernel function transformation, a learning rate adjustment mechanism in gradient descent is introduced to dynamically optimize the kernel function parameters, and an enzyme activity prediction model is obtained.

[0045] The support vector regression algorithm is used to obtain the corresponding relationship between the state vector and the enzyme activity from the training data matrix, and a high-dimensional mapping is performed through a radial basis kernel function to obtain an initial mapping model. According to the initial mapping model, the prediction error between the state vector and the enzyme activity is calculated. If the prediction error exceeds the preset threshold, the width parameter of the radial basis kernel function is adjusted to obtain an updated mapping model. The gradient value of the kernel function parameter is obtained from the updated mapping model through the gradient descent algorithm to determine the parameter optimization direction. A dynamic learning rate adjustment mechanism is used to adaptively update the learning rate according to the trend of the gradient value to obtain the optimized kernel function parameter. The support vector regression algorithm is re-executed through the optimized kernel function parameter to update the high-dimensional mapping relationship to obtain a refined enzyme activity prediction model. If the prediction error of the refined enzyme activity prediction model is lower than the preset threshold, the model is saved to obtain the final enzyme activity prediction model.

[0046] For example, when constructing the enzyme activity prediction model, first, the training data matrix is standardized to scale the enzyme activity value to the interval [0, 1], for example, using MinMaxScaler to normalize the original data such as pH value 7.2, temperature 35℃, etc. Then, the support vector regression (SVR) algorithm is used, and the initial kernel function is set to radial basis kernel (RBF), the parameters include penalty coefficient C=1.0, kernel coefficient gamma=0.01, and the parameters are optimized in 10-fold cross-validation through grid search, for example, when gamma is searched from 0.001 to 0.1 with logarithmic step, it is found that when gamma=0.05, the model has the lowest mean square error (MSE) of 0.12 on the validation set. To handle the nonlinear relationship between the state vector and the enzyme activity, the kernel trick is used to map the original 4-dimensional features to a high-dimensional space, for example, the RBF kernel is used to calculate the Euclidean distance between two samples x1=[0.3, 0.6] and x2=[0.4, 0.5], and the kernel function value is exp(-0.02 / 0.05)=0.67. A dynamic learning rate mechanism is introduced to optimize the kernel parameters, and the Adam optimizer is used with an initial learning rate of 0.001, which is attenuated by 10% every 100 iterations. After 500 iterations, gamma converges to 0.048, and the R² of the test set reaches 0.89 at this time. The final model is verified by the leave-one-out method, and the predicted enzyme activity of a new sample such as substrate concentration 0.8mM is 1.2U / mL, with an error controlled within ±0.15U / mL. The whole process is realized through Scikit-learn to automate training and evaluation, ensuring closed-loop execution of parameter optimization and prediction.

[0047] Step S105, the partial derivative matrix of each process parameter is calculated according to the output result of the prediction model, the importance of the temperature gradient, the pH fluctuation and the time accumulation is sorted by using the fitness function in the genetic algorithm, and the key parameter combination is determined by using the roulette wheel selection operator.

[0048] As shown in Figure 3 The output result is obtained from the prediction model, the partial derivative matrix of each process parameter is calculated, and the process parameter matrix is obtained. According to the process parameter matrix, the influence degree of each parameter on the model output is quantified by using the partial derivative calculation method, and the parameter influence matrix is obtained. The partial derivative values of the temperature gradient, the pH fluctuation and the time accumulation are extracted from the parameter influence matrix, and the parameter influence vector is obtained. The parameter influence vector is evaluated by using the fitness function of the genetic algorithm, and the importance of each parameter is sorted. The key parameter combination is selected from the importance sorting by using the roulette wheel selection operator, and the optimal parameter set is obtained. According to the optimal parameter set, the input configuration of the prediction model is adjusted, and the updated model output is obtained. If the deviation of the updated model output from the expected value exceeds the preset threshold, the optimal parameter set is iteratively adjusted, and the final key parameter combination is obtained.

[0049] For example, the partial derivative matrix of each process parameter is calculated based on the output result of the prediction model, and the importance of the temperature gradient, the pH fluctuation, and the time accumulation is sorted by using the fitness function in the genetic algorithm. The specific implementation method of determining the key parameter combination by using the roulette selection operator is as follows: first, the output data of the prediction model is used. It is assumed that the temperature gradient is 2.5 degrees per minute, the pH fluctuation is 0.3 units, and the time accumulation is 120 minutes. The partial derivative matrix is calculated by using the numerical differentiation method. The specific algorithm is that a small perturbation (such as a temperature gradient perturbation of 0.01 degrees) is performed on each parameter. The change in the model output is recorded, and the partial derivative value is calculated. The result matrix is that the temperature gradient partial derivative is 1.2, the pH fluctuation is 0.8, and the time accumulation is 0.5, which indicates that the temperature gradient has the greatest influence on the model output. Second, the fitness function of the genetic algorithm is constructed. The partial derivative value is used as the weight, and the fitness score is calculated. The formula is fitness = 1.2 x temperature gradient + 0.8 x pH fluctuation + 0.5 x time accumulation. It is assumed that the score of a certain parameter combination is 3.9. Through 100 iterations, the combination with the highest score of 4.2 is selected. The analysis shows that the temperature gradient has the highest weight, and the importance is sorted as temperature gradient, pH fluctuation, and time accumulation. Finally, the roulette selection operator is used. The probability is allocated according to the proportion of the fitness. The probability of the temperature gradient is 50%, the probability of the pH fluctuation is 30%, and the probability of the time accumulation is 20%. The key parameter combination is obtained by using the random number generator to simulate the selection. It is obtained that the key parameter combination is the temperature gradient and the pH fluctuation, and the contribution rate of the process stability is 80%. Therefore, these two parameters are preferentially optimized. The above process is realized by using the Python script to automatically calculate and select. The partial derivative calculation module calls the NumPy library. The genetic algorithm iteration uses the DEAP library. The roulette selection is completed by using the random number module. The logic is rigorous. The parameter values and the analysis results are consistent. A complete technical chain is formed.

[0050] In step S106, if the real-time collected key parameter combination exceeds the preset threshold range, a loss function calculation module based on gradient descent is triggered. The parameter update step is adjusted by using the weight decay coefficient, and the process parameter correction amount is output.

[0051] As Figure 4As shown, the data stream of the key parameters is obtained through the real-time acquisition system, the parameter combination is continuously monitored during the acquisition process, it is judged whether it exceeds the preset threshold range, and a preliminary abnormality determination result is obtained. If the determination result shows that the parameter combination exceeds the preset threshold range, the abnormal data is transmitted to the loss calculation module, the gradient descent method is used to iteratively calculate the loss function, and the minimum direction of the loss value is determined. According to the result of loss calculation, the weight attenuation coefficient is obtained, the parameter update step is dynamically adjusted, and the adjusted step value is output. The process parameters are preliminarily corrected through the adjusted step value, a first corrected data set is generated, and it is stored in the temporary database. If the parameter values in the first corrected data set still exceed the preset threshold range, the gradient descent module is called again to optimize the corrected data set twice, and a second corrected data set is obtained. According to the second corrected data set, it is judged whether the process parameters meet the preset range, if not, the data is transmitted to the standby processing module, the parameter fine tuning is performed through the pre-established rule base, and the final correction amount is output. The process parameters in the real-time system are updated through the final correction amount, the adjustment log is recorded in the database at the same time, and the parameter optimization process is completed.

[0052] In step S107, the current soaking state vector is recombined by chromosome coding according to the correction amount, a new parameter combination is generated by using the crossover probability setting rule of the genetic algorithm, the optimal solution is selected by the elite reservation strategy, and an optimized process parameter population is obtained.

[0053] The current soaking state vector is obtained, real-time data is collected from the production equipment sensor, and the initial value of the state vector is determined. The initial value of the state vector is mapped to a chromosome sequence by a chromosome coding rule, and an initial parameter population is generated. The chromosome sequence is operated by crossover by using the crossover probability rule of the genetic algorithm, and the recombined chromosome sequence is obtained. If the recombined chromosome sequence meets the preset fitness threshold, the sequence is retained; if not, a new chromosome sequence is generated by mutation operation. According to the elite reservation strategy, the sequence with the highest fitness is selected from the recombined chromosome sequence, and the candidate process parameters are determined. The candidate process parameters are restored to the optimized process parameters by decoding operation, and an optimized parameter population is generated. For the optimized parameter population, simulation verification is performed, whether the population meets the production constraint condition is judged, and a final process parameter population is obtained.

[0054] For example, the chromosome coding recombination and process parameter optimization process for the soaking state vector assumes that the current soaking state vector is an array containing temperature, humidity, and time parameters, such as [25.5, 60.0, 120.0], which represents a temperature of 25.5 degrees Celsius, a humidity of 60.0%, and a soaking time of 120 minutes. First, the vector is converted to binary form through chromosome coding, with the temperature range set to 20-30 degrees Celsius and the accuracy set to 0.1, which requires 7 bits of binary representation. Similarly, the humidity range of 50-70% requires 6 bits, and the time range of 100-140 minutes requires 6 bits, resulting in a final encoding of a 19-bit binary string, such as 0110010_011110_0111100. Next, the crossover probability setting rule of the genetic algorithm is used, with the crossover probability set to 0.7. A random value between 0 and 1 is generated by a random number generator, such as 0.65, which is less than 0.7, so the crossover operation is performed. Select another chromosome, such as [26.0, 62.0, 125.0], and perform single-point crossover at a random position, such as the 10th position, to generate two new individuals. Subsequently, the elitist strategy is used to directly retain the individual with the highest fitness in the population to the next generation. Assuming that the fitness function is the process effect score, the new individual with the highest score of 85.3, [25.8, 61.5, 122.5], is retained, and the other individuals are selected through roulette wheel selection probability. The probability calculation is based on the proportion of fitness, such as a score of 85.3 accounting for 17% of the total fitness of 500, so there is a 17% probability of being selected. Finally, after multiple iterations, the optimized process parameter population converges to the optimal solution, such as [25.7, 61.2, 123.0], with a fitness score of 88.9, which is significantly improved compared to the initial value. The entire process is automatically implemented through the algorithm, and data analysis shows that temperature and humidity fine-tuning have a significant impact on process effectiveness, while time parameters need to be further verified in combination with device response speed to form a complete logic chain from coding to optimization, ensuring the scientificity and feasibility of parameter adjustment.

[0055] Step S108, input the optimized population into the enzyme activity prediction model for iterative verification. If the variance of the fitness function is less than the preset value for three consecutive generations, the final parameter combination is locked, and the closed-loop regulation is completed.

[0056] As Figure 4As shown, by transmitting the data of the optimized population to the enzyme activity prediction model, the initial iteration verification process is started, and the fitness value calculation result of each generation is obtained. According to the fitness value of each generation, the function variance data of three consecutive generations is calculated, and it is judged whether it is lower than the preset threshold value. If it is lower than the preset threshold value, the current parameter combination information is recorded. If the function variance data of three consecutive generations is lower than the preset threshold value, the final locking mechanism is triggered, and the current parameter combination is determined as the optimal solution. For the locked parameter combination, the corresponding enzyme activity prediction result is obtained, and the intermediate data record of the closed-loop regulation is generated. The pre-established verification process is used to check the intermediate data record for multiple rounds, judge the stability of the prediction result, and obtain the final verification conclusion. Through the analysis of the verification conclusion, the feedback mechanism of the closed-loop regulation is adjusted, the convergence state of the regulation process is determined, and the closed-loop processing of the entire iteration verification is completed.

[0057] For example, in the process of inputting the optimized population into the enzyme activity prediction model for iteration verification, first, an initial population is generated by a genetic algorithm, assuming that the population size is 100 individuals, each individual represents a set of enzyme parameter combinations, including temperature, pH value and substrate concentration, for example, individual 1 has parameters of temperature 45°C, pH value 6.5, and substrate concentration 0.1 mol / L. Then, these parameters are input into the enzyme activity prediction model based on random forest, and the fitness value of each individual is calculated, assuming that the fitness function is the enzyme activity value, and the predicted activity value of individual 1 is 85.3 units. Subsequently, iteration verification is performed, and each generation is updated by selection, crossover and mutation operations to update the population, and the fitness value of the new generation is calculated, for example, the activity value of individual 1 in the second generation is increased to 87.2 units. At the same time, the variance of the fitness function of each generation is recorded, assuming that the variance of the first generation is 12.5, the variance of the second generation is 10.8, and the variance of the third generation is 9.2, and if the preset variance threshold is 10.0, the variance of three consecutive generations is less than the threshold, and the system automatically triggers the locking mechanism to determine the final parameter combination as temperature 46°C, pH value 6.7, and substrate concentration 0.12 mol / L. In this process, the system judges the stability of the population by variance analysis, and makes trend prediction combined with historical data. If the variance does not meet the standard, the genetic algorithm parameters such as the crossover rate are automatically adjusted from 0.7 to 0.8 to speed up the convergence. The final parameter combination is stored in the database and connected to the downstream production system to automatically generate an optimized enzyme reaction condition configuration file, ensuring the continuity of closed-loop regulation. Through the above process, the system realizes the fully automated process from population optimization to parameter locking, logically forming a complete chain from data input, model prediction to result output.

[0058] To sum up, the present application is not a simple superposition of support vector regression (SVR) and genetic algorithm GA, but through the bridge of "partial derivative + fitness function", the accurate local prediction ability of SVR and the strong global search ability of GA are organically combined; and through the double-layer control architecture of "SVR + GA", the real-time responsiveness and global optimization ability of the analysis method are realized, which is a synergistic technical effect that a single algorithm cannot achieve. Example two

[0059] As Figure 5 shown, the present application also provides a sweet potato amylase content analysis system based on machine learning, mainly comprising:

[0060] A data acquisition and preprocessing module is used to collect temperature gradient data, pH dynamic fluctuation values, soaking time accumulation and depth displacement during the sweet potato soaking process by using multiple sensors, eliminate high-frequency noise through sliding window filtering, and obtain denoised process parameter time series;

[0061] A state space construction module is used to construct a three-dimensional state space matrix according to the denoised time series, normalize map the temperature gradient, pH fluctuation, time accumulation and depth displacement by using a fuzzy membership function, and obtain a standardized state vector;

[0062] A sample mapping and cleaning module is used to determine the amylase activity value under different state vectors by using a laboratory enzyme marker, establish a sample mapping table of state vectors and enzyme activity, remove outliers by using the quartile method, and obtain a cleaned training data matrix;

[0063] A model training module is used to perform nonlinear fitting on the training data matrix by using a support vector regression algorithm, process the high-dimensional mapping relationship between the state vector and the enzyme activity by using a radial basis kernel function transformation, introduce a learning rate adjustment mechanism in gradient descent to dynamically optimize the kernel function parameters, and obtain an enzyme activity prediction model;

[0064] A parameter importance analysis module is used to calculate the partial derivative matrix of each process parameter according to the prediction model output result, sort the temperature gradient, pH fluctuation, time accumulation by importance by using the fitness function in genetic algorithm, and determine the key parameter combination by using the roulette selection operator;

[0065] A dynamic adjustment module is used to trigger the loss function calculation module based on gradient descent if the real-time collected key parameter combination exceeds the preset threshold range, adjust the parameter update step by using the weight decay coefficient, and output the process parameter correction amount;

[0066] a genetic optimization module, configured to perform chromosome coding recombination on the current soaking state vector according to the correction amount, generate a new parameter combination by using a cross probability setting rule of a genetic algorithm, screen an optimal solution by using an elite reservation strategy, and obtain a population of optimized process parameters;

[0067] an iterative convergence module, configured to input the optimized population into an enzyme activity prediction model for iterative verification, lock a final parameter combination if a variance of an adaptive function of three consecutive generations is less than a preset value, and complete closed-loop adjustment. Embodiment Three

[0068] This embodiment describes the specific process of applying the method and system of the present application to the soaking process in the actual production of sweet potato starch, and verifies the technical effect.

[0069] The system of the present application is deployed in the middle of a soaking tank in a sweet potato processing workshop. Specifically, high-precision temperature sensors are installed at different depths (for example, 0.3 meters, 0.8 meters and 1.3 meters) in the soaking tank to obtain temperature gradient data; an online pH meter is installed to monitor the dynamic fluctuation of the acid-base degree; and an ultrasonic liquid level meter is used to monitor the settlement displacement of the sweet potato accumulation. The system records the soaking time accumulation through a timer. The data acquisition frequency is set to once every 2 minutes.

[0070] 1. Model training phase (S101-S104)

[0071] First, 60 batches of soaking experiments were conducted. In each batch of experiment, different initial water temperature, pH adjuster addition amount and other parameters were artificially set, and the time sequence of the above-mentioned multi-sensor data in the entire soaking process was recorded. After each batch, samples were taken and the final amylase activity value was accurately measured by a laboratory enzyme marker.

[0072] The system first performs sliding window filtering and denoising on the 60 groups of time sequence data collected. Then, the fuzzy membership function is applied to normalize the temperature gradient, acid-base fluctuation, time accumulation and depth displacement after denoising, and 60 standardized state vectors are constructed. The 60 state vectors are paired with their corresponding measured enzyme activity values to form a sample mapping table. Through quartile analysis, it is found that the enzyme activity values of 2 samples deviate too much from the normal range, which are judged as outliers and are excluded, and finally a training data matrix containing 58 effective samples is obtained. Using the matrix, the system trains a support vector regression (SVR) enzyme activity prediction model.

[0073] 2. Closed-loop control and effect verification phase (S105-S108)

[0074] The trained model is deployed in an online control system for intelligent regulation of new production batches. The system runs step S105 to determine that, under the current process conditions, the "temperature gradient" and the "cumulative amount of soaking time" are the two key parameter combinations that have the greatest impact on the final enzyme activity by analyzing the partial derivative matrix of the model and the genetic algorithm.

[0075] In a new soaking process, the system real-time monitors that due to changes in the external environment temperature, the water temperature at the upper layer of the soaking tank drops too quickly, causing the "temperature gradient" to exceed the preset healthy threshold. The system immediately triggers the dynamic adjustment module S106 to calculate a "correction amount" of heating power based on the gradient descent algorithm. The correction amount is sent to the genetic optimization module S107 to recombine the "chromosome" of the current control strategy, and through several generations of iteration, an optimal heating power and soaking time collaborative control strategy for the next hour is quickly generated. The strategy is verified by iteration S108, and its fitness function variance converges rapidly within three generations. The system then locks the optimization strategy and issues it to the actuator.

[0076] 3. Technical effect verification

[0077] To verify the technical effect of the method and system of the present application, the trained SVR model is applied to 58 training samples and 10 test samples not involved in training for verification. The predicted amylase activity value is compared with the actual value determined by the laboratory, and compared with the production process using traditional fixed parameter control. Please refer to the attached Figure 6 , which intuitively demonstrates the technical advantages of the present application through a combination of graphs.

[0078] Appendix Figure 6 (a) Precision verification scatter plot of the prediction model of the present application. In the figure, the horizontal axis is the actual amylase activity value measured by the enzyme label instrument in the laboratory, and the vertical axis is the corresponding activity value predicted by the system of the present application. Ideally, all data points should completely fall on the y=x diagonal line (shown by the dashed line in the figure) representing the predicted value equal to the true value. As can be seen from the figure, all data points (hollow circles) are closely around the y=x diagonal line. Calculation shows that the determination coefficient (R 2 ) of the predicted value reaches 0.92, indicating that the prediction model of the present application has high linear fitting degree and prediction accuracy.

[0079] Appendix Figure 6(b) compares the process stability of the method of the present invention with that of the prior art. The bar chart compares the standard deviation of the final enzyme activity between different batches in production processes using both prior art (e.g., fixed time and temperature control) and the method of the present invention (closed-loop intelligent regulation based on SVR model and genetic algorithm). As shown in the figure, the production line using the prior art, due to its inability to respond to dynamic changes in the process in real time, exhibits a batch-to-batch standard deviation of product activity exceeding 18%, resulting in poor product quality consistency. In stark contrast, the method of the present invention, through precise online prediction and dynamic closed-loop regulation, significantly reduces the batch-to-batch standard deviation of product activity to 3.1%.

[0080] Comprehensive Appendix Figure 6 The results analysis of (a) and 6(b) show that the present invention, through precise machine learning model prediction and dynamic closed-loop control strategy, not only achieves high-precision prediction of key process indicators (amylase activity), but also greatly improves the stability of the production process and the predictability of the final product quality. It solves the technical problems of inaccurate prediction, extensive control and large fluctuations in product quality in the prior art, and has significant technological progress and industrial application value.

[0081] It should be noted that the above examples are merely some specific embodiments of the present invention. Obviously, the present invention is not limited to the above embodiments and many variations are possible. All variations that can be directly derived or conceived by those skilled in the art from the content disclosed in this invention should be considered within the scope of protection of this invention.

Claims

1. A machine learning-based method for analyzing sweet potato amylase content, characterized in that, The method includes: Multiple sensors were used to collect temperature gradient data, pH dynamic fluctuation values, cumulative soaking time and depth displacement during the sweet potato soaking process. High-frequency noise was eliminated by sliding window filtering to obtain the time sequence of denoised process parameters. A three-dimensional state space matrix is ​​constructed based on the denoised time series. The temperature gradient, pH fluctuation, time accumulation and depth displacement are normalized and mapped using fuzzy membership functions to obtain a standardized state vector. The amylase activity values ​​under different state vectors were measured by a laboratory microplate reader. A sample mapping table between state vectors and enzyme activities was established. Outliers were removed by the quartile method to obtain the cleaned training data matrix. The support vector regression algorithm is used to perform nonlinear fitting on the training data matrix. The high-dimensional mapping relationship between the state vector and enzyme activity is handled by radial basis kernel function transformation. The learning rate adjustment mechanism in gradient descent is introduced to dynamically optimize the kernel function parameters, and an enzyme activity prediction model is obtained. The partial derivative matrix of each process parameter is calculated based on the output of the prediction model. The fitness function in the genetic algorithm is used to rank the importance of temperature gradient, pH fluctuation and time accumulation. The combination of key parameters is determined by roulette wheel selection operator. If the combination of key parameters collected in real time exceeds the preset threshold range, the gradient descent-based loss function calculation module is triggered, and the parameter update step size is adjusted by the weight decay coefficient, and the process parameter correction amount is output. Based on the correction amount, the current soaking state vector is recombined with chromosome encoding. A new parameter combination is generated by setting the crossover probability using a genetic algorithm. The optimal solution is selected by an elite retention strategy to obtain the optimized process parameter population. The optimized population input enzyme activity prediction model is iteratively validated. If the variance of the fitness function is less than the preset value for three consecutive generations, the final parameter combination is locked to complete the closed-loop regulation.

2. The method for analyzing sweet potato amylase content based on machine learning according to claim 1, characterized in that, The process involves using multiple sensors to collect temperature gradient data, dynamic pH fluctuations, cumulative soaking time, and depth displacement during the sweet potato soaking process. High-frequency noise is eliminated through sliding window filtering to obtain a denoised time-series sequence of process parameters, including: Multiple sensors are used to synchronously collect temperature gradient, pH value, immersion time and depth displacement data to generate raw process parameter datasets. The original process parameter dataset is processed by sliding window filtering to remove high-frequency noise and obtain the denoised process parameter time series. Based on the time sequence of the denoised process parameters, the dynamic correlation coefficient between the temperature gradient and the pH value is calculated to determine the coupling relationship between the two. If the absolute value of the dynamic correlation coefficient is greater than the preset threshold, the temperature gradient and pH value are jointly analyzed to obtain the synergistic influence characteristics of the two on immersion time and depth displacement. The synergistic influence characteristics are smoothed by using the Kalman filter algorithm to generate stable trends in process parameter changes; Based on the stable trend of process parameter changes, the support vector regression algorithm is used to predict the dynamic evolution of subsequent process parameters, and the predicted time series is obtained. If the deviation between the predicted time series and the actual collected data exceeds a preset threshold, the sensor data acquisition frequency is adjusted, and the original process parameter dataset is regenerated.

3. The method for analyzing sweet potato amylase content based on machine learning according to claim 1, characterized in that, The process involves constructing a three-dimensional state space matrix based on the denoised time sequence, and then normalizing the temperature gradient, pH fluctuations, time accumulation, and depth displacement using a fuzzy membership function to obtain a standardized state vector, including: Denoising data is obtained from time series sequences, and principal component analysis is used to extract the main features to generate an initial feature set. If the dimension of the initial feature set exceeds a preset threshold, the main components are retained through dimensionality reduction to obtain a compressed feature set. Based on the compressed feature set, a three-dimensional state space matrix is ​​constructed, and the distribution of matrix elements is determined by the grid partitioning method. For the three-dimensional state space matrix, a fuzzy membership function is applied to perform normalization mapping on the temperature gradient, pH fluctuation, time accumulation, and depth displacement to generate normalized feature vectors. If the variance of the normalized feature vector is lower than a preset threshold, the feature distribution is adjusted by kernel density estimation to obtain an optimized feature vector. Based on the optimized feature vectors, the k-means clustering algorithm is used to divide the state space and determine the state categories. Standardized state vectors are generated based on state categories and stored as the foundational data for subsequent analysis.

4. The method for analyzing sweet potato amylase content based on machine learning according to claim 1, characterized in that, The process involves measuring amylase activity values ​​under different state vectors using a laboratory ELISA reader, establishing a sample mapping table between state vectors and enzyme activities, and using the quartile method to remove outliers, resulting in a cleaned training data matrix, including: An initial dataset was generated by collecting amylase activity values ​​corresponding to different state vectors using an enzyme-linked immunosorbent assay (ELISA) reader. The quartile method is used to analyze the sample distribution of the initial dataset and identify outliers; If the activity value of a sample exceeds the preset threshold of the interquartile range, the sample is removed, and the cleaned dataset is obtained. Based on the cleaned dataset, construct a mapping table between state vectors and amylase activity; Extract features from the mapping table to generate a training data matrix; Principal component analysis is used to reduce the dimensionality of the training data matrix, resulting in an optimized feature matrix. A linear regression algorithm was used to model the optimized feature matrix to obtain a predictive model between the state vector and enzyme activity.

5. The method for analyzing sweet potato amylase content based on machine learning according to claim 1, characterized in that, The method employs a support vector regression algorithm to perform nonlinear fitting on the training data matrix, uses radial basis function transformation to handle the high-dimensional mapping relationship between the state vector and enzyme activity, and introduces a learning rate adjustment mechanism from gradient descent to dynamically optimize the kernel function parameters, thereby obtaining an enzyme activity prediction model, including: The support vector regression algorithm is used to obtain the correspondence between state vectors and enzyme activities from the training data matrix. The initial mapping model is obtained by performing high-dimensional mapping through the radial basis kernel function. Based on the initial mapping model, the prediction error between the state vector and enzyme activity is calculated. If the prediction error exceeds the preset threshold, the width parameter of the radial basis kernel function is adjusted to obtain the updated mapping model. The gradient descent algorithm is used to obtain the gradient values ​​of the kernel function parameters from the updated mapping model, thereby determining the direction of parameter optimization. A dynamic learning rate adjustment mechanism is adopted to adaptively update the learning rate according to the changing trend of the gradient value, thereby obtaining optimized kernel function parameters; By optimizing the kernel function parameters, the support vector regression algorithm is re-executed to update the high-dimensional mapping relationship, resulting in a refined enzyme activity prediction model. If the prediction error of the refined enzyme activity prediction model is lower than the preset threshold, the model is saved to obtain the final enzyme activity prediction model.

6. The method for analyzing sweet potato amylase content based on machine learning according to claim 1, characterized in that, The process involves calculating the partial derivative matrices of each process parameter based on the prediction model output, ranking the importance of temperature gradient, pH fluctuation, and time accumulation using the fitness function in a genetic algorithm, and determining the combination of key parameters using a roulette wheel selection operator. This includes: The output results are obtained from the prediction model, and the partial derivative matrices of each process parameter are calculated to obtain the process parameter matrix. Based on the process parameter matrix, the influence of each parameter on the model output is quantified by using the partial derivative calculation method, and the parameter influence matrix is ​​obtained. The partial derivatives of temperature gradient, pH fluctuation and time accumulation are extracted from the parameter influence matrix to obtain the parameter influence vector. The fitness function of the genetic algorithm is used to evaluate the parameter influence vector and obtain the importance ranking of each parameter; By using a roulette wheel selection operator, key parameter combinations are selected from the importance ranking to obtain the optimal parameter set; Based on the optimized parameter set, the input configuration of the prediction model is adjusted to obtain the updated model output; If the deviation between the updated model output and the expected value exceeds a preset threshold, the optimal parameter set will be iteratively adjusted to obtain the final key parameter combination.

7. The method for analyzing sweet potato amylase content based on machine learning according to claim 1, characterized in that, If the combination of key parameters collected in real time exceeds a preset threshold range, the gradient descent-based loss function calculation module is triggered. This module adjusts the parameter update step size using a weight decay coefficient and outputs the process parameter correction amount, including: The system acquires data streams of key parameters in real time, continuously monitors parameter combinations during the acquisition process, determines whether they exceed preset threshold ranges, and obtains preliminary anomaly judgment results. If the judgment result shows that the parameter combination exceeds the preset threshold range, the abnormal data is transmitted to the loss calculation module, and the gradient descent method is used to iteratively calculate the loss function to determine the direction of minimizing the loss value. Based on the loss calculation results, obtain the weight decay coefficient, dynamically adjust the parameter update step size, and output the adjusted step size value. The process parameters are initially corrected by adjusting the step size value, the first corrected dataset is generated, and it is stored in a temporary database. If the parameter values ​​in the first corrected dataset still exceed the preset threshold range, the gradient descent module is called again to perform a second optimization on the corrected dataset to obtain the second corrected dataset. Based on the second corrected dataset, it is determined whether the process parameters meet the preset range. If they still do not meet the preset range, the data is transmitted to the backup processing module, where the parameters are fine-tuned using a pre-established rule base, and the final corrected amount is output. The process parameters in the real-time system are updated by the final correction amount, and the adjustment log is synchronously recorded in the database to complete the parameter optimization process.

8. The method for analyzing sweet potato amylase content based on machine learning according to claim 1, characterized in that, The process involves chromosome encoding and recombination of the current soaking state vector based on the correction amount, generating new parameter combinations using crossover probability setting rules in a genetic algorithm, and selecting the optimal solution through an elite retention strategy to obtain an optimized process parameter population, including: Obtain the current soaking state vector by collecting real-time data from the sensors of the production equipment and determining the initial value of the state vector. By using chromosome encoding rules, the initial values ​​of the state vector are mapped to chromosome sequences to generate an initial parameter population; The crossover probability rules of a genetic algorithm are used to perform a crossover operation on the chromosome sequence to obtain a recombinant chromosome sequence; If the recombined chromosome sequence meets the preset fitness threshold, then the sequence is retained; If the conditions are not met, a new chromosome sequence is generated through mutation operations; Based on the elite retention strategy, the sequence with the highest fitness is screened from the recombinant chromosome sequence to determine candidate process parameters; Through decoding, the candidate process parameters are restored to the optimized process parameters, and an optimized parameter population is generated. For the optimized parameter population, a simulation verification is performed to determine whether the population meets the production constraints, and the final process parameter population is obtained.

9. The method for analyzing sweet potato amylase content based on machine learning according to claim 1, characterized in that, The optimized population input enzyme activity prediction model is iteratively validated. If the variance of the fitness function is less than a preset value for three consecutive generations, the final parameter combination is locked to complete closed-loop regulation, including: By transmitting the optimized population data to the enzyme activity prediction model, the initial iterative verification process is initiated to obtain the fitness value calculation results for each generation. Based on the fitness value of each generation, calculate the function variance data for three consecutive generations, determine whether it is lower than the preset threshold, and if it is lower than the preset threshold, record the current parameter combination information. If the variance data of the function for three consecutive generations is lower than the preset threshold, the final locking mechanism is triggered to determine that the current combination of parameters is the optimal solution. For the locked parameter combination, obtain the corresponding enzyme activity prediction results and generate intermediate data records for closed-loop regulation; A pre-established verification process is used to perform multiple rounds of verification on intermediate data records to determine the stability of the prediction results and obtain the final verification conclusion. By analyzing the verification results, the feedback mechanism of the closed-loop regulation is adjusted, the convergence state of the regulation process is determined, and the closed-loop processing of the entire iterative verification is completed.

10. A machine learning-based sweet potato amylase content analysis system, characterized in that, The system includes: The data acquisition and preprocessing module is used to collect temperature gradient data, pH dynamic fluctuation value, cumulative soaking time and depth displacement during the sweet potato soaking process using multiple sensors. High-frequency noise is eliminated by sliding window filtering to obtain the noise-reduced process parameter time sequence. The state space construction module is used to construct a three-dimensional state space matrix based on the denoised time series. It uses a fuzzy membership function to normalize and map the temperature gradient, pH fluctuation, time accumulation and depth displacement to obtain a standardized state vector. The sample mapping and cleaning module is used to measure the amylase activity value under different state vectors using a laboratory microplate reader, establish a sample mapping table between state vectors and enzyme activities, remove outlier samples using the quartile method, and obtain a cleaned training data matrix. The model training module is used to perform nonlinear fitting of the training data matrix using the support vector regression algorithm. It processes the high-dimensional mapping relationship between the state vector and enzyme activity through radial basis kernel function transformation, and introduces the learning rate adjustment mechanism in gradient descent to dynamically optimize the kernel function parameters, thereby obtaining the enzyme activity prediction model. The parameter importance analysis module is used to calculate the partial derivative matrix of each process parameter based on the output of the prediction model. It uses the fitness function in the genetic algorithm to rank the importance of temperature gradient, pH fluctuation and time accumulation, and determines the combination of key parameters through roulette wheel selection operator. The dynamic adjustment module is used to trigger the gradient descent-based loss function calculation module if the combination of key parameters collected in real time exceeds the preset threshold range. The module adjusts the parameter update step size through the weight decay coefficient and outputs the process parameter correction amount. The genetic optimization module is used to perform chromosome encoding and recombination on the current soaking state vector according to the correction amount, generate new parameter combinations by setting the crossover probability rules of the genetic algorithm, and select the optimal solution through the elite retention strategy to obtain the optimized process parameter population. The iterative convergence module is used to iteratively verify the optimized population input enzyme activity prediction model. If the variance of the fitness function is less than the preset value for three consecutive generations, the final parameter combination is locked to complete the closed-loop adjustment.

Images