Yellow rice wine fermentation prediction method and system based on anfis and random fractal search algorithm

Abstract

The application provides a yellow rice wine fermentation prediction method and system based on an ANFIS and a random fractal search algorithm, the method comprising collecting data samples of a pre-fermentation process of different production batches of yellow rice wine; dividing the data samples into a training set and a test set, and performing normalization processing on the data samples; inputting the processed data samples into a multi-output adaptive neural fuzzy inference system model constructed in advance, identifying and optimizing model parameters of the multi-output adaptive neural fuzzy inference system model by using a hierarchical learning random fractal search algorithm, obtaining an optimized multi-output adaptive neural fuzzy inference system model, and predicting a yellow rice wine fermentation state. The application improves the precision and generalization ability of the model, and can achieve good prediction of the fermentation state of different production batches of yellow rice wine.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of yellow rice wine fermentation, and particularly relates to a yellow rice wine fermentation prediction method and system based on an ANFIS (Adaptive Network-based Fuzzy Inference System) and a random fractal search algorithm. BACKGROUND

[0002] Yellow rice wine is a unique wine in China, and has both drinking and medicinal functions. The essence of yellow rice wine fermentation is a biochemical reaction process with multiple variable inputs and outputs. Under different environmental conditions, malt, yeast and rice interact with each other to generate various metabolites such as sugars, organic acids, ethanol, esters, aldehydes and amino acids. The yellow rice wine fermentation process mainly includes traditional manual brewing and mechanized brewing. On the one hand, manual brewing requires the manual experience of a brewer, and has gradually been unable to meet the needs of modern yellow rice wine fermentation process production and control; on the other hand, due to the detection equipment and analysis instruments, the yellow rice wine fermentation process using the new mechanized technology is difficult to be monitored and controlled in real time, which is prone to cause the problem of poor quality stability of yellow rice wine, and therefore, the automation level thereof is far behind other industrial production processes. How to realize the automatic control of yellow rice wine fermentation is a key problem faced by the yellow rice wine industry.

[0003] At present, the research on yellow rice wine fermentation automation mainly includes two aspects.

[0004] One is to monitor and control the environmental variables in the fermentation process. For example, Liu Zhu et al. realized the analog collection of temperature and dissolved oxygen by using a LabVIEW software platform, and realized the monitoring of the yellow rice wine fermentation state by controlling the motor and valve of the fermentation tank equipment; Zhou Yun et al. designed a dissolved oxygen control system based on fuzzy PID in view of the nonlinearity of dissolved oxygen in the fermentation process, which has a certain practicability; Gao Donglei et al. realized the temperature control in the fermentation process by using an iterative learning controller, and optimized the parameters of the iterative learning controller by using a BP neural network (Back Propagation Neural Network) to improve the control precision; Xu Ling et al. designed a digital Smith predictive incremental PID controller in view of the delay problem existing in the temperature control in the yellow rice wine fermentation process, and the simulation experimental results show that the method can solve the adverse effects of the large time lag of the temperature control object in the yellow rice wine fermentation on the dynamic performance of the system, and improve the dynamic performance of the system; Huang Suixi used the powerful fitting capability of a BP neural network to establish a temperature prediction model of the pre-fermentation process of yellow rice wine, by taking the temperature data values collected at the first five time points in the fermentation process as the input of the BP neural network model, and taking the temperature value at the next time point as the output. The above research work can better realize the monitoring of the temperature or dissolved oxygen in the fermentation process, but more important data such as the key biological variables in the fermentation process are rarely involved.

[0005] Second, to establish a reasonable mathematical model of the fermentation process of yellow rice wine, to quantitatively and dynamically describe the fermentation process, and to realize the automation and optimal control of the fermentation process of yellow rice wine. Liu Dengfeng et al. first discussed the effects of four enzymes, koji and three fermentation temperatures on the key biological variables such as sugars, organic acids, glycerol and ethanol in the fermentation process of yellow rice wine, providing a theoretical basis for the modeling of yellow rice wine fermentation. Liu Dengfeng et al. conducted yellow rice wine fermentation experiments under five different temperature conditions, and proposed a simultaneous saccharification and fermentation (SSF) kinetic model for the biochemical reactions and main products in the fermentation process. The model parameters were optimized by the least squares method. The model structure well reflects the kinetic changes measured under different fermentation temperatures. Liu Dengfeng et al. first monitored the whole process of the pre-fermentation stage in the actual production process of Nv'erhong yellow rice wine, and measured the total sugar, alcohol, total acid and pH in the fermentation process offline. Based on the fermentation kinetics, a simultaneous saccharification and fermentation model SSF for the pre-fermentation process of yellow rice wine in industrial production was established, and the model parameters were optimized and identified by the Marquardt method based on the least squares method. Zong Yuan et al. proposed an ant lion optimization algorithm with Levy flight and Cauchy mutation (LCALO) to solve the problem of easy falling into local optimum and slow convergence speed when identifying the model parameters of yellow rice wine fermentation process based on the Levenberg-Marquardt method. The LCALO algorithm improves the accuracy of the SSF model under fixed temperature conditions. The main shortcomings of the above work are that the robustness and accuracy of the model need to be improved. On the one hand, due to the complexity of the internal mechanism of fermentation, the model based on fermentation kinetics is established under specified fermentation environment conditions, and can only reflect the main chemical reactions and main products. Under different production environment conditions, there is a certain degree of difference in the data collected in the fermentation process, and the generalization ability of the model on new data is poor. On the other hand, a large amount of available and useful data is difficult to obtain due to the limitations of collection equipment, analysis instruments, time and other reasons, and there is a problem of information missing. This leads to the fact that the amount of collected data cannot be sufficient for the training of the model, and the prediction accuracy of the model is not high. SUMMARY

[0006] The embodiment of the present application provides a yellow rice wine fermentation prediction method and system based on ANFIS and random fractal search algorithm, which is used to solve the problem that the accuracy and generalization ability of the modeling method of the yellow rice wine fermentation process in the prior art cannot meet the industrial requirements.

[0007] To address the aforementioned technical problems, embodiments of the present invention provide a method for predicting the fermentation of rice wine based on ANFIS and a random fractal search algorithm. This method includes:

[0008] S1: Collect data samples of the pre-fermentation process in the production of rice wine from different batches;

[0009] S2: Divide the data samples into training set and test set, and normalize the data samples.

[0010] S3: Input the processed data sample into the pre-constructed multi-output adaptive neural fuzzy inference system model, and use the hierarchical learning stochastic fractal search algorithm to identify and optimize the model parameters of the multi-output adaptive neural fuzzy inference system model to obtain the optimized multi-output adaptive neural fuzzy inference system model.

[0011] S4: Predict the fermentation status of rice wine;

[0012] The hierarchical learning stochastic fractal search algorithm includes:

[0013] By incorporating Levi flight into a hierarchical learning strategy, a hierarchical learning strategy with Levi flight is obtained.

[0014] By incorporating the hierarchical learning strategy with Levy flight into the stochastic fractal search algorithm, a hierarchical learning stochastic fractal search algorithm is obtained.

[0015] Preferably, the input data of the multi-output adaptive neural fuzzy inference system model includes sampling time, temperature, pH, total sugar concentration, total acid concentration, and alcohol content, and the output data includes the total sugar concentration, total acid concentration, and alcohol content at the next time step.

[0016] Preferably, the training set of the multi-output adaptive neural fuzzy inference system model is set to D = (X, Y).

[0017] Where X = (x1, x2, ..., x N )∈R n×N The input data is Y = (y1, y2, ..., y...). k )∈R k×N These are the actual output data, with a training set size of N. It is the n-dimensional feature vector of the i-th instance, where i = 1, ..., N, indicating that the input data is n-dimensional. The i-th instance is a k-dimensional label vector, i = 1, ..., N, indicating that the output data is k-dimensional. The output of the j-th layer node is denoted as O. j , j = 1, 2, ..., 5.

[0018] Preferably, the multi-output adaptive neural fuzzy inference system model includes:

[0019] The first layer: fuzzy layer, for fuzzy operation on input data, calculate the membership of each dimension feature, the output is:

[0020]

[0021] Where, A i,j (i = 1, …, n; j = 1, 2) is the membership of each feature corresponding to the set, x i is the i-th dimension feature, is the membership function, exp[] is the Gaussian function, c ij , σ ij is the mean and standard deviation of Gaussian function respectively;

[0022] The second layer: weight layer of rules, for realizing fuzzy reasoning process, the output of each node represents the credibility of a rule, the output is:

[0023]

[0024] Where, w l is the credibility of the l-th rule, is the membership function, ∏ is the product operation;

[0025] The third layer: weight normalization layer of rules, for making the numerical analysis range of rule weight concentrated in 0 ~ 1, and the sum of all weights is 1, the output is:

[0026]

[0027] Where, is the normalization of the weight of the l-th rule, ∑ is the sum operation;

[0028] The fourth layer: fuzzy rule output layer, for outputting the result after operation of rules and weights, the output is:

[0029]

[0030] Where, is the normalization of the weight of the l-th rule, is the l-th rule;

[0031] The fifth layer: output layer, for calculating the error of the whole model, the output is:

[0032]

[0033] Where, x i is the i-th dimension feature, is the consequent parameter vector of the model, ∑ is the sum operation Finally, the output of the multi-output adaptive neural fuzzy reasoning system model is

[0034] Preferably, the random fractal search algorithm comprises the following steps:

[0035] Setting the optimization object is a global optimization function f(x1,x2,…,x d , and the optimization goal is to minimize the fitness value of function f;

[0036] First, the particle swarm P∈R d×N is randomly initialized, and the initialization formula is as follows:

[0037] P i = LB+ε×(UB-LB),i=1,2,…,N (6)

[0038] Where LB and UB are the lower and upper bounds of the particle position distribution, respectively, and ε is a uniformly distributed random number in the range of [0,1];

[0039] After initializing all particles, the fitness value corresponding to each particle is calculated, and the particle with the minimum value is recorded as the optimal particle BP;

[0040] Then, in the fractal stage, a new particle is generated by using Gaussian distribution as the random walk method, and the formula is as follows:

[0041]

[0042] Where ε and ε′ are uniformly distributed random numbers between 0 and 1, BP and P i are the optimal particle and the i-th particle, respectively, and the values of μ BP and μ P are |μ BP | and |μ i |, respectively. The standard deviation in the Gaussian distribution is calculated as follows:

[0043]

[0044] Where g represents the number of iterations, and represents that the step of the Gaussian function is decreasing as the number of iterations increases.

[0045] Finally, in the update stage, the first step is to sort all particles according to the fitness value, and give each particle a probability value that follows a uniform distribution, and the formula is as follows:

[0046]

[0047] When Pa i <ε, the j-th dimensional value in the particle is updated as follows:

[0048] P' i (j) = P r (j) - ε x (P t (j) - P i (j)), j = 1, 2,..., d (10)

[0049] where P i ' is the particle P i updated position, P r and P t are two particles randomly selected from the particle swarm;

[0050] Before starting the second step, the particles obtained in the first step are sorted again according to the formula P i < ε, the particles are updated according to the following formula, otherwise they are not updated, the formula is as follows:

[0051]

[0052] where P r and P t are two particles randomly selected from the particle swarm obtained in the first step, is a random number generated by a Gaussian normal distribution.

[0053] Preferably, the hierarchical learning strategy comprises the following steps:

[0054] According to the function fitness value corresponding to the particle, it is divided into different levels, and the particle swarm with a set number N is divided into NL levels, each level is represented as L i (1≤i≤NL), each level has the same number of particles;

[0055] First, the particles in the particle swarm are arranged in ascending order according to their corresponding function fitness values, and are divided into four levels (L1-L4);

[0056] Then, the particles in L4 learn from the particles in L1-L3, the particles in L3 learn from the particles in L1-L2, the particles in L2 learn from the particles in L1, and the particles in L1 are not updated and directly enter the next cycle, the updating method of the particles is as follows:

[0057]

[0058] where P i,j is the jth particle in the L i th level, V i,j is its speed; and are two particles randomly selected from the two higher layers rl1 and rl2, k1 and k2 represent the index values of the particles in the respective layers; higher than higher, i.e. better than particles; r1, r2 and a3 are random numbers between [0, 1], are random numbers between [0, 1].

[0059] Preferably, the Levy flight function expression is as follows:

[0060]

[0061] wherein r a and r b are random numbers between [0, 1], and β is a constant with a value of 1.5, and the standard deviation of the Levy flight is calculated as follows:

[0062]

[0063] wherein Γ(x) = (x-1)!.

[0064] Preferably, the Levy flight is integrated into the hierarchical learning strategy to obtain a hierarchical learning strategy with Levy flight, which comprises:

[0065] The particles in L3 and L4 are further optimized by using the Levy flight, and the particle updating mode is as follows:

[0066]

[0067] wherein P i,j is the jth particle in the L i th layer, and V i,j is the velocity thereof; and are two particles randomly selected from the two higher layers rl1 and rl2, k1 and k2 represent the index values of the particles in the respective layers; higher than higher, i.e. better than particles; r1, r2 and r3 are random numbers between [0, 1], are random numbers between [0, 1], LB and UB are the lower bound and the upper bound of the particle position distribution respectively, and levy() is the Levy flight function.

[0068] Preferably, the hierarchical learning strategy with Levy flight is integrated into the random fractal search algorithm to obtain a hierarchical learning random fractal search algorithm, which comprises the following steps:

[0069] step1: initialize particle swarm P, calculate the function fitness value F corresponding to each particle;

[0070] step2: enter the iteration loop;

[0071] step3: according to formula (7), each particle generates a new particle, and updates the function fitness value corresponding to the new particle;

[0072] step4: sort the particle swarm according to formula (9), if Pa i <ε, update the particle P i according to formula (10), otherwise, keep the particle P i unchanged;

[0073] step5: according to the function fitness value, divide the particle swarm P into four levels L1-L4, wherein the particles in L1 are not updated, the particles in L2 are updated according to formula (12), the particles in L3 and L4 are updated according to formula (15), and a new particle swarm P new is obtained;

[0074] step6: calculate the function fitness value F new corresponding to each particle in the particle swarm P new ;

[0075] step7: if F new <F, continue iteration, otherwise, end iteration.

[0076] The application provides a yellow rice wine fermentation prediction system based on an ANFIS and a random fractal search algorithm, which comprises:

[0077] A data acquisition module is configured to acquire data samples of a pre-fermentation process of different production batches of yellow rice wine.

[0078] A data preprocessing module is configured to divide the data samples into a training set and a test set, and perform normalization processing on the data samples.

[0079] A model optimization module is configured to input the processed data samples into a pre-constructed multi-output adaptive neural fuzzy inference system model, identify and optimize model parameters of the multi-output adaptive neural fuzzy inference system model by using a hierarchical learning random fractal search algorithm, and obtain an optimized multi-output adaptive neural fuzzy inference system model.

[0080] A model prediction module is configured to predict a fermentation state of yellow rice wine.

[0081] The hierarchical learning random fractal search algorithm comprises:

[0082] The level learning strategy is combined with the Levy flight to obtain a level learning strategy with the Levy flight.

[0083] The level learning strategy with the Levy flight is combined into a random fractal search algorithm to obtain a level learning random fractal search algorithm.

[0084] From the above technical solutions, the present application has the following advantages:

[0085] The present application provides a yellow rice fermentation prediction method and system based on an ANFIS and a random fractal search algorithm. The present application proposes a multi-output adaptive network-based fuzzy inference system (MOANFIS) in view of the characteristics of the diversity of yellow rice fermentation products. The MOANFIS combines fuzzy logic and neural networks and has a very efficient ability to process nonlinear problems. The level learning strategy in the level learning particle swarm optimization algorithm (LLSO) is combined with the Levy flight to propose a level learning stochastic fractal search algorithm (LLSFS). The LLSFS has excellent global search ability when facing high-dimensional optimization problems. The LLSFS algorithm is used for the optimization of the MOANFIS model. The experimental results show that the algorithm improves the precision and generalization ability of the model and can realize good prediction of the fermentation state of different production batches of yellow rice. BRIEF DESCRIPTION OF DRAWINGS

[0086] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly described below. The features and advantages of the present application can be more clearly understood by referring to the drawings. The drawings are schematic and should not be understood as any limitation on the present application. Those skilled in the art can obtain other drawings according to these drawings without creative effort. Among them:

[0087] Figure 1 A flowchart of a yellow rice fermentation prediction method based on an ANFIS and a random fractal search algorithm according to the embodiments is provided.

[0088] Figure 2 A structure diagram of a MOANFIS model according to the embodiments is provided.

[0089] Figure 3 A particle fractal process schematic diagram according to the embodiments is provided.

[0090] Figure 4 A structural diagram of hierarchical learning provided in the embodiments;

[0091] Figure 5 A convergence curve diagram of the MOANFIS according to each algorithm provided in the embodiments;

[0092] Figure 6 Prediction effects of the LLSFS-MOANFIS model on the test set according to the embodiments, wherein, (a) is a total sugar concentration prediction diagram of the next moment, (b) is a total acid concentration prediction diagram of the next moment, and (c) is an alcohol degree prediction diagram of the next moment;

[0093] Figure 7 A block diagram of a yellow rice fermentation prediction system based on an ANFIS and a random fractal search algorithm according to the embodiments. DETAILED DESCRIPTION

[0094] To make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0095] As shown in Figure 1 the embodiments of the present application propose a yellow rice fermentation prediction method based on an ANFIS and a random fractal search algorithm, which comprises:

[0096] S1: collecting data samples of a pre-fermentation process in the production of yellow rice of different production batches;

[0097] S2: dividing the data samples into a training set and a test set, and performing normalization processing on the data samples;

[0098] S3: inputting the processed data samples into a multi-output adaptive neural fuzzy inference system model constructed in advance, identifying and optimizing the model parameters of the multi-output adaptive neural fuzzy inference system model by using a hierarchical learning random fractal search algorithm, and obtaining an optimized multi-output adaptive neural fuzzy inference system model;

[0099] S4: predicting the fermentation state of the yellow rice.

[0100] The application provides a yellow rice wine fermentation prediction method based on an ANFIS and a random fractal search algorithm.

[0101] Further, in step S1:

[0102] Data samples of a pre-fermentation process in the generation of yellow rice wine of different production batches are collected;

[0103] Further, in step S2:

[0104] The data samples are divided into a training set and a test set, and the data samples are normalized;

[0105] Further, in step S3:

[0106] The input data are: sampling time (h), temperature (T), PH, total sugar concentration, total acid concentration and alcohol content in the fermentation process; and the output data are: total sugar concentration (h+2), total acid concentration (h+2) and alcohol content (h+2) at the next moment.

[0107] The traditional ANFIS is established based on a T-S fuzzy system, can automatically extract rules from input and output data by using the learning mechanism of a neural network, has clear language expression ability of a fuzzy system, but only has one output node. The yellow rice wine fermentation process can be regarded as a system with multiple variable inputs and multiple variable outputs, and fermentation products include sugar, organic acid, alcohol and alcohol. For this problem, the single-dimensional output of the ANFIS is expanded to a multi-dimensional output, i.e. a MOANFIS, and the structure is as shown in Figure 2 The rectangular node represents an adaptive node containing system parameters, and the circular node represents a fixed node without system parameters.

[0108] For a k-dimensional output model, the training data set is set as D=(X, Y), wherein X=(x1, x2, …, xk) ∈ Rn is input data, Y=(y1, y2, …, yk) ∈ Rn is output data, and n is the number of input variables. N n×N k k×N ​​​is the real output data, the training set size is N, is the n-dimensional feature vector of the i-th instance, i = 1, …, N, indicating that the input data is n-dimensional, is the k-dimensional label vector of the i-th instance, i = 1, …, N, indicating that the output data is k-dimensional, and the output of the j-th layer node is denoted as O j , j = 1, 2, …, 5.

[0109] The multi-output adaptive neural fuzzy inference system (MOANFIS) model comprises:

[0110] The first layer is a fuzzy layer, which is used for fuzzy operation on input data and calculates the membership degree of each dimension feature. In the embodiment of the present application, a Gaussian function is selected as the membership function, A i,j (i = 1, …, n; j = 1, 2) is the fuzzy set corresponding to each feature. The output of this layer is:

[0111]

[0112] wherein x i is the i-th dimension feature, is the membership function, exp[] is the Gaussian function, c ij and σ ij are the mean and standard deviation of the Gaussian function, respectively, and are the antecedent parameters of the fuzzy rule;

[0113] The second layer is a rule weight layer, which is used for implementing fuzzy inference process, and the output of each node represents the credibility of a certain rule, and the output is:

[0114]

[0115] wherein w l is the credibility of the l-th rule, is the membership function, and ∏ is the product operation;

[0116] The third layer is a rule weight normalization layer, which is used for concentrating the numerical analysis range of the rule weight in 0-1 and making the sum of all weights equal to 1, and the output is:

[0117]

[0118] wherein, is the normalization of the weight of the l-th rule, and ∑ is the summation operation;

[0119] The fourth layer is a fuzzy rule output layer, which is used for outputting the result after the operation of the rule and each weight. For an n-dimensional input vector x = (x1, x2, …, x n , the k-dimensional output vector of the MOANFIS is represented as follows:

[0120] R l : if x1∈A 1,j , x2∈A 2,j , …, x n ∈A n,j , then

[0121]

[0122] where A 1,j , A 2,j , …, A n,j (j = 1, 2) denote fuzzy subsets, θ[l]∈R (n+1)×k denotes the consequent parameter matrix of the lth fuzzy rule corresponding to the k-dimensional output vector, and the consequent parameter matrices of all fuzzy rules are θ = (θ[1], θ[2], …, θ[r])∈R r×(n+1)×k . Each fuzzy rule shares the antecedent parameters of the model, and the output of the current layer is:

[0123]

[0124] where, is the normalization of the weight of the lth rule, is the lth rule;

[0125] The fifth layer: the output layer, which is used to calculate the error of the whole model, and the output is:

[0126]

[0127] where x i is the i-th dimensional feature, is the consequent parameter vector of the model, and ∑ is the summation operation Finally, the output of the multi-output adaptive neural fuzzy inference system model is

[0128] Further, the Stochastic Fractal Search (SFS) algorithm is an intelligent optimization algorithm proposed by Salimi by imitating the physical diffusion phenomenon in nature. This algorithm is based on the concept of Diffusion Limited Aggregation (DLA) and further generates a random fractal structure. For simplicity, consider forming such a cluster in a plane, with the initial (seed) particle located at the origin. Then generate other particles randomly around the origin and cause diffusion. The diffusing particles stick to the seed particles made by them. Repeat this process until a cluster is formed, showing a branched structure, as shown in Figure 3 .

[0129] The SFS algorithm has two main phases: the fractal phase and the update phase. In the former, each particle diffuses around its current position, increasing the chance of finding the global minimum of the optimization function and avoiding getting trapped in local minima. In the latter, the remaining particles in the population update their positions based on the position of the best particle generated during the diffusion process and the position of a random particle.

[0130] The optimization object is defined as a global optimization function f(x1,x2,…,x) containing d-dimensional variables. d The optimization objective is to minimize the fitness value of the function f.

[0131] First, for a swarm of N particles P∈R d×N Perform random initialization, using the following formula:

[0132] P i =LB+ε×(UB-LB),i=1,2,…,N (6)

[0133] Where LB and UB are the lower and upper bounds of the particle position distribution, respectively, and ε is a uniformly distributed random number in the range [0,1]. After initializing all particles, the fitness value of each particle is calculated, and the particle with the minimum fitness value is recorded as the optimal particle BP.

[0134] Then, in the fractal stage, SFS uses a Gaussian distribution as a random walk method to generate new particles, as shown in the following formula:

[0135]

[0136] Where ε and ε′ are uniformly distributed random numbers between [0,1], and BP and P i These are the optimal particle and the i-th particle, respectively, μ BP and μ P The values ​​are respectively |μ BP |、 The standard deviation in a Gaussian distribution is calculated as follows:

[0137]

[0138] Where g represents the number of iterations. This indicates that as the number of iterations increases, the jump pace of the Gaussian function decreases; it allows particles to perform a more localized search, avoiding jumping out of the optimal solution.

[0139] Finally, in the update phase, the first step is to sort all particles according to their fitness values ​​and assign each particle a probability value that follows a uniform distribution, as shown in the following formula:

[0140]

[0141] The better the particle, the smaller its fitness value and its Pa. i The larger the value, the more likely it is to be retained in the next iteration. For particles that haven't obtained a good solution, the chance of being updated is greater. When Pa... i When ε < ε, the value of the j-th dimension in the particle is updated as follows:

[0142] P′ i (j)=P r (j)-ε×(P t (j)-P i (j)), j=1,2,…,d (10)

[0143] Among them, P i ′ is particle P i Updated position, P r and P t Two particles are randomly selected from the particle swarm;

[0144] Before starting the second step, again according to the formula The particles obtained from the first step of the statistical analysis are sorted, and when Pa i When ε < ε, the particle is updated according to the following formula; otherwise, it is not updated. The formula is as follows:

[0145]

[0146] Among them, P r and P t These are two particles randomly selected from the particle swarm obtained in the first step. These are random numbers generated from a Gaussian normal distribution. If P″ i The corresponding function fitness value is better than P′. i Then P″ i Replace P′ i .

[0147] Furthermore, in the second stage update process of the SFS algorithm, the entire particle swarm is updated based on BP particles and randomly selected particles. This makes it difficult to find the global optimum when identifying and optimizing the MOANFIS model parameters. As the dimensionality of the optimization problem increases, on the one hand, particles need to search in a larger space, leading to an increase in the algorithm's time complexity; on the other hand, the number of local optima increases, making particles prone to getting trapped in local optima and causing the algorithm to converge prematurely. To address this problem, this invention combines the hierarchical learning strategy in the Hierarchical Learning Particle Swarm Optimization (LLSO) algorithm with Lévy flight, improving the particle update method in the second stage of SFS, and proposing the Hierarchical Learning Stochastic Fractal Search (LLSFS) algorithm.

[0148] The hierarchical learning strategy comprises the following steps:

[0149] The particles are divided into different levels according to the fitness values of the corresponding functions, and the particle groups with a number of N are divided into NL levels, each level is represented as L i (1≤i≤NL), each level has the same number of particles. Good particles have a higher level, corresponding to a smaller level number. That is, L1 is the highest level, and L NL is the lowest level. The common level number is 4, and the hierarchical learning strategy is as shown in Figure 4 .

[0150] Firstly, the particles in the particle group are arranged in ascending order according to the corresponding fitness values of the functions, and are divided into four levels (L1-L4);

[0151] Then, the particles in L4 learn from the particles in L1-L3, the particles in L3 learn from the particles in L1-L2, the particles in L2 learn from the particles in L1, and the particles in L1 do not update and directly enter the next cycle, and the updating mode of the particles is as follows:

[0152]

[0153] Wherein, P i,j is the jth particle in the L i layer, V i,j is the speed thereof; and are two particles randomly selected from the two higher layers rl1 and rl2, and k1 and k2 represent the index values of the particles in the respective layers; The level is higher than the level, that is, the particle is better than the particle; r1, r2 and r3 are random numbers between 0 and 1, is a random number between 0 and 1.

[0154] In order to better improve the global search ability of the algorithm and avoid premature convergence, the present application integrates the Levy flight into the hierarchical learning strategy. The Levy flight is a random search method subject to the Levy distribution, and has good performance in the optimization of large-scale space. The function expression is as follows:

[0155]

[0156] Wherein, r a and r b are random numbers between 0 and 1, and β is a constant with a value of 1.5. The standard deviation of the Levy flight is calculated as follows:

[0157]

[0158] where Γ(x) = (x-1)!.

[0159] The particles in L3 and L4 are further optimized by using Levy flight, and the particle updating method is as follows:

[0160]

[0161] where P i,j is the jth particle in the L i th layer, V i,j is the velocity of the particle; and are two particles randomly selected from two higher layers rl1 and rl2, and k1 and k2 represent the index values of the particles in the respective layers; the layer level is higher than the layer level, i.e. the particle is better than the particle; r1, r2 and r3 are random numbers between 0 and 1, is a random number between 0 and 1, LB and UB are the lower and upper bounds of the particle position distribution, respectively, and levy() is the Levy flight function.

[0162] The hierarchical learning strategy with Levy flight is integrated into SFS, i.e., a hierarchical learning random fractal search algorithm (LLSFS) is obtained, and the algorithm steps are as follows:

[0163] step1: initialize the particle swarm P, and calculate the function fitness value F corresponding to each particle;

[0164] step2: enter the iteration loop;

[0165] step3: generate a new particle for each particle according to formula (7), and update the function fitness value corresponding to the new particle;

[0166] step4: sort the particle swarm according to formula (9), if Pa i <ε, update the particle P i according to formula (10), otherwise, the particle P i is retained;

[0167] step5: according to the function fitness value, divide the particle swarm P into four levels L1-L4, wherein the particles in L1 are not updated, the particles in L2 are updated according to formula (12), the particles in L3 and L4 are updated according to formula (15), and a new particle swarm P new is obtained;

[0168] step6: calculate the function fitness value of the particle swarm P newThe function fitness value F corresponding to each particle new ;

[0169] Step 7: If F new <F, the iteration continues, otherwise the iteration ends.

[0170] The hierarchical learning random fractal search algorithm is used for identification and optimization of model parameters of a multi-output adaptive neural fuzzy inference system model, so as to obtain an optimized multi-output adaptive neural fuzzy inference system model.

[0171] Further, in step S4:

[0172] The optimized multi-output adaptive neural fuzzy inference system model is used to realize prediction of the fermentation state of yellow rice wine.

[0173] The advantages of the method of the application will be described below through specific experiments.

[0174] I. Yellow rice wine fermentation data set

[0175] The data of the application come from real data collected in the pre-fermentation process of yellow rice wine production in the Daughter Red Wine Factory, and a total of 212 data samples of five different production batches. The total of 170 fermentation data of the first four fermentation tanks in the above data samples are used as a training set, and the 42 data of the last fermentation tank are used as a test set. Since there is a large difference between the data of each dimension, the data is first normalized, with a value range of [0, 1].

[0176] II. Experimental setup

[0177] The experimental platform used for the establishment and optimization of the yellow rice wine fermentation model is as follows: Window 10 operating system, Intel i5 processor with a main frequency of 2.9 GHz, memory of 8 GB, and programming environment of MATLAB2020b.

[0178] For MOANFIS, the related parameters that can be manually set are as follows:

[0179] (1) The grid method is used to obtain the antecedent parameters of the MOANFIS model, that is, each input membership function combination corresponds to a fuzzy rule;

[0180] (2) The fitness function is used to evaluate the training effect of the model, and the weighted sum of the error squares between the predicted value and the true value of the model is used, that is, Where MSE represents the mean square error.

[0181] In the MOANFIS model training process, the calculation of the fitness function F is based on the normalized data. Since the training data set contains different batches of yellow rice fermentation data, the de-normalization of the data will bring some error to the fitness function F, even if the tiny error will be amplified in the training iteration, which will affect the training of the model, so the data is not de-normalized in the training process of the MOANFIS. At the same time, the content of organic acid produced in the fermentation process of yellow rice is smaller than that of total sugar and alcohol, and the data distribution is concentrated in a small range. In order to improve the prediction ability of the model to organic acid, the weight of the organic acid in the fitness function formula is increased, and the values of k1 and k3 are set to 1, and the value of k2 is set to 5. That is, F=MSE(1)+5xMSE(2)+MSE(3).

[0182] In this paper, PSO (Particle Swarm Optimization), SHADE (Success-History based Adaptive DE), WOA (Whale Optimization Algorithm), SSA (Sparrow Search Algorithm), AVOA (African vultures optimization algorithm), SFS and LLSFS are used for experimental comparison, and the parameters of the MOANFIS model are optimized. For the algorithm LLSFS, in order to ensure that the number of particles in each layer is the same in hierarchical learning, the population size is set to 48, and the population size of the remaining algorithms is 50. The number of iterations is 200 times. The parameters in each algorithm are shown in Table 1:

[0183] Table 1

[0184]

[0185]

[0186] III. Experimental results and analysis

[0187] The present application uses each of the above algorithms to independently experiment on the fitness function F of the MOANFIS model for 10 times. The optimization ability and robustness of each algorithm are evaluated from the optimal value, average value and variance of the experimental results. The experimental results show that the LLSFS algorithm has better performance in the three evaluation indexes. Compared with the SFS algorithm, the LLSFS algorithm has an improvement of 4.5%, 4.4% and 10.5% in the three evaluation indexes respectively, has a faster convergence speed, and the global search ability and robustness are further improved. The experimental results of each algorithm are shown in Table 2, Figure 5The convergence curves of the algorithms are shown.

[0188] Table 2

[0189]

[0190] The performance of the model mainly depends on its performance on the test set. The present application respectively from the mean square error MSE, mean absolute error MAE, coefficient of determination R 2 Three evaluation indexes are used to evaluate the accuracy and generalization ability of the MOANFIS model, which are the prediction of the total sugar concentration at the next moment, the total acid concentration at the next moment, and the alcohol concentration at the next moment. First, the model prediction data is processed by inverse normalization to restore the real scale fermentation data, and then each evaluation index is analyzed. The calculation methods of MSE, MAE and R 2 are as follows:

[0191]

[0192]

[0193]

[0194] wherein the true output is Y = (y1, y2, …, y N ) T , and the prediction output of the MOANFIS model is

[0195] The experimental results show that, compared with other algorithms, the LLSFS-MOANFIS model has more accurate prediction results for the total sugar concentration, total acid concentration, and alcohol concentration at the next moment. Again, it is shown that LLSFS has strong global search and optimization ability when facing MOANFIS models with large parameter quantities and high data dimensions, effectively improving the accuracy of the model. The experimental results are shown in Table 3.

[0196] Table 3

[0197]

[0198] In order to further evaluate the performance of the model, the LLSFS-MOANFIS model is compared with the SSF model based on fermentation kinetics. The SSF model is solved and optimized for each of the three batches of yellow rice fermentation data, and the generalization ability of the model is not discussed, but it also has good reference value. The test set in this paper is the same batch of data as the 12 jars of fermentation data in the paper, so this batch of data is selected for comparative analysis of the experimental results. In addition to the total acid concentration R 2In addition to poor performance, the fitting effect of LLSFS-MOANFIS model in each index is better than that of SSF model, indicating that the LLSFS-MOANFIS model has better prediction ability. The results are shown in Table 4.

[0199] Table 4

[0200]

[0201]

[0202] As Figure 6 shown, the prediction effect of LLSFS-MOANFIS model on the test set is shown. As can be seen from the figure, the fitting effect of LLSFS-MOANFIS on the total sugar concentration, alcohol content at the next moment is good, but the prediction performance of the total acid concentration at the next moment still has room for improvement. On the one hand, the organic acid produced in the fermentation process of yellow rice wine is a trace substance, and the total acid concentration between different batches is quite different; on the other hand, the data distribution of total acid concentration is relatively dispersed, and there are many outliers, which affects the fitting effect.

[0203] In summary, the present application proposes a MOANFIS model with a multi-output structure in view of the characteristics of the monitoring target product in the fermentation mash of yellow rice wine, and establishes a model structure with the fermentation data at the current moment as input and the key fermentation data at the next moment as output. At the same time, the LLSFS algorithm proposed by the present application identifies and optimizes the parameters of the MOANFIS model, and the experimental results show that the LLSFS algorithm has better global search ability and robustness, which improves the precision of MOANFIS. The research work of yellow rice wine fermentation modeling based on fuzzy system theory opens up a new direction for yellow rice wine modeling.

[0204] As Figure 7 shown, the present application also provides a yellow rice wine fermentation prediction system based on ANFIS and random fractal search algorithm, which comprises:

[0205] A data acquisition module 100 is used to acquire data samples of the pre-fermentation process in the production of different batches of yellow rice wine;

[0206] A data preprocessing module 200 is used to divide the data samples into a training set and a test set, and to normalize the data samples;

[0207] A model optimization module 300 is used to input the processed data samples into the multi-output adaptive neural fuzzy inference system model constructed in advance, identify and optimize the model parameters of the multi-output adaptive neural fuzzy inference system model by using the hierarchical learning random fractal search algorithm, and obtain the optimized multi-output adaptive neural fuzzy inference system model;

[0208] The model prediction module 400 is configured to predict the fermentation state of the yellow rice wine.

[0209] The hierarchical learning random fractal search algorithm comprises the following steps:

[0210] The hierarchical learning strategy with Levy flight is obtained by integrating the Levy flight into the hierarchical learning strategy.

[0211] The hierarchical learning strategy with Levy flight is integrated into the random fractal search algorithm to obtain the hierarchical learning random fractal search algorithm.

[0212] The system is used to realize the yellow rice wine fermentation prediction method based on the ANFIS and the random fractal search algorithm as described above, and in order to avoid redundancy, it will not be repeated here.

[0213] Note that the above are only preferred embodiments of the present application and the technical principles applied. Those skilled in the art will understand that the present application is not limited to the specific embodiments described herein, and those skilled in the art can make various obvious changes, readjustments and substitutions without departing from the scope of the present application. Therefore, although the present application has been described in more detail through the above embodiments, the present application is not limited to the above embodiments, and can include more other equivalent embodiments without departing from the concept of the present application, and the scope of the present application is determined by the scope of the appended claims.

[0214] Those skilled in the art will understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0215] The present application is described with reference to flowcharts and / or block diagrams according to the method, device (system), and computer program product of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine, so that the instructions executed by the computer or other programmable data processing apparatus produce a device that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The device for performing the functions specified in one flow or multiple flows and / or blocks Figure 1 The device for performing the functions specified in one flow or multiple flows and / or blocks

[0216] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the Figure 1 function specified in the flow or flows and / or blocks of the block or blocks. These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the function specified in the flow or flows and / or blocks of the block or blocks. Figure 1 Figure 1 function specified in the flow or flows and / or blocks of the block or blocks. These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the function specified in the flow or flows and / or blocks of the block or blocks. Figure 1

[0217] It is apparent that the above-described embodiments are only examples for clarity and are not limiting on the implementation. For ordinary skilled in the art, other different forms of changes or variations can be made on the basis of the above description. Here, it is not necessary and also impossible to enumerate all the implementations. The obvious changes or variations derived therefrom are still within the protection scope of the present application.​​

Claims

1. A yellow rice wine fermentation prediction method based on ANFIS and random fractal search algorithm, characterized in that, The method comprises the following steps: S1: collecting data samples of the pre-fermentation process of yellow rice wine of different production batches; S2: dividing the data samples into a training set and a test set, and normalizing the data samples; S3: inputting the processed data samples into a multi-output adaptive neural fuzzy inference system model constructed in advance, identifying and optimizing the model parameters of the multi-output adaptive neural fuzzy inference system model by using a hierarchical learning random fractal search algorithm, and obtaining an optimized multi-output adaptive neural fuzzy inference system model; S4: predicting the fermentation state of yellow rice wine; The hierarchical learning random fractal search algorithm comprises the following steps: Levy flight is integrated into the hierarchical learning strategy to obtain a hierarchical learning strategy with Levy flight; The hierarchical learning strategy with Levy flight is integrated into a random fractal search algorithm to obtain a hierarchical learning random fractal search algorithm. 2.The yellow rice wine fermentation prediction method based on ANFIS and random fractal search algorithm according to claim 1, characterized in that, The input data of the multi-output adaptive neural fuzzy inference system model comprises sampling time, temperature, pH, total sugar concentration, total acid concentration and alcohol content, and the output data comprises total sugar concentration, total acid concentration and alcohol content at the next moment. 3.The yellow rice wine fermentation prediction method based on ANFIS and random fractal search algorithm according to claim 1, characterized in that, The training set of the multi-output adaptive neural fuzzy inference system model is set as D=(X, Y), Where X = (x1, x2, ..., x N )∈R n×N The input data is Y = (y1, y2, ..., y3). k )∈R k×N These are the actual output data, with a training set size of N. It is the n-dimensional feature vector of the i-th instance, where i = 1, ..., N, indicating that the input data is n-dimensional. The i-th instance is a k-dimensional label vector, i = 1, ..., N, indicating that the output data is k-dimensional. The output of the j-th layer node is denoted as O. j j = 1, 2, ..., 5. 4.The yellow rice wine fermentation prediction method based on ANFIS and random fractal search algorithm according to claim 1, characterized in that, The multi-output adaptive neural fuzzy inference system model comprises the following layers: The first layer is a fuzzy layer, which is used for fuzzy operation on the input data, calculates the membership degree of each dimensional feature, and outputs: where A i，j (i = 1, …, n; j = 1, 2) is the fuzzy set corresponding to each feature, x i is the ith dimension feature, is the membership function, exp[] is the Gaussian function, c ij , and σ ij are the mean and standard deviation of the Gaussian function, respectively; The second layer is a rule weight layer, which is used for implementing a fuzzy inference process, and the output of each node represents the credibility of a rule, and the output is: where w l is the degree of confidence of the ith rule, is the membership function, and Π is the product operation. The third layer is a rule weight normalization layer, which is used for concentrating the numerical analysis range of the rule weight in 0-1 and making the sum of all weights equal to 1, and the output is: wherein is a normalization of the weight of the first rule, and ∑ is a summation operation; The fourth layer is a fuzzy rule output layer, which is used for outputting the result of the operation of the rule and each weight, and the output is: wherein, is a normalization of the weight of the lth rule, is the lth rule; The fifth layer is an output layer, which is used for calculating the error of the whole model, and the output is: where x i is the ith dimension feature, is the consequent parameter vector of the model, and ∑ is the summation operation. Finally, the output of the multi-output adaptive neuro-fuzzy inference system model is 5.The yellow rice wine fermentation prediction method based on ANFIS and random fractal search algorithm according to claim 1, characterized in that, The random fractal search algorithm comprises the following steps: The optimization object is a global optimization function f(x1, x2, …, x d ) containing d-dimensional variables, and the optimization goal is to minimize the fitness value of the function f. First, a particle group P e R d×N is randomly initialized, and the initialization formula is as follows: P i = LB + ε x (UB - LB ), i = 1, 2, …, N (6) Wherein, LB and UB are the lower bound and upper bound of the particle position distribution respectively, and ε is a uniformly distributed random number in the range of [0, 1]; After initializing all particles, the function fitness value corresponding to each particle is calculated, wherein the particle with the minimum value is recorded as the optimal particle BP; Then, in the fractal stage, Gaussian distribution is used as the random walk method to generate new particles, and the formula is as follows: where ε and ε' are uniformly distributed random numbers between [0, 1], BPand P i are the best particle and the i-th particle, respectively, μ BP and μ P are the values of |μ BP | and |μ |, respectively. The standard deviation in the Gaussian distribution is calculated as follows: where g denotes the iteration number, This indicates that the step size of the Gaussian function is decreasing as the iteration number increases. Finally, in the update stage, the first step is to sort all particles according to the function fitness value, and give each particle a probability value, which is subject to uniform distribution, and the formula is as follows: When The j-th dimensional value in the particle is updated as follows: P' i (j) = P r (j) - ε x (P t (j) - P i (j)), j = 1, 2,..., d (10) where P i is the particle P i updated position, P r and P t are two particles randomly selected from the swarm of particles; Before starting the second step, the particles are sorted again according to the formula The particles obtained by the first step statistics are sorted, and when Pa i When < ε, the particles are updated according to the following formula, otherwise they are not updated: where P r and P t are two particles randomly selected from the swarm of particles obtained in the first step, is a random number generated by a Gaussian normal distribution. 6.The yellow rice wine fermentation prediction method based on ANFIS and random fractal search algorithm according to claim 5, characterized in that, The hierarchical learning strategy comprises the following steps: The particles are divided into different levels according to the fitness values of the functions corresponding to the particles, and the particle groups with a set number N are divided into NL levels, each level is represented as L i (1≤i≤NL), each level has the same number of particles; Firstly, the particles in the particle swarm are arranged in ascending order according to their corresponding function fitness values, and divided into four levels (L1-L4); Then, the particles in L4 learn from the particles in L1-L3, the particles in L3 learn from the particles in L1-L2, the particles in L2 learn from the particles in L1, and the particles in L1 do not update and directly enter the next cycle, and the updating mode of the particles is as follows: where P i，j is the velocity of the jth particle in the L i th layer, V i，j is its velocity; and are two particles randomly chosen from the two higher layers rl1 and rl2, and k1 and k2 represent the index values of the particles in the respective layers; higher in the hierarchy than higher in the hierarchy, i.e. better than better than; r1, r2 and r3 are random numbers between [0, 1], is a random number between [0, 1].

7. The ANFIS and random fractal search algorithm-based yellow rice wine fermentation prediction method according to claim 6, characterized in that, The Levy flight function expression is as follows: where r a and r b are random numbers between 0 and 1, β is a constant with value 1.5, and the standard deviation of the Levy flight is calculated as follows: Wherein, Γ(x)=(x-1)!. 8.The yellow rice wine fermentation prediction method based on ANFIS and random fractal search algorithm according to claim 7, characterized in that, The Lévy flight is integrated into the hierarchical learning strategy to obtain a hierarchical learning strategy with Lévy flight, which comprises the following steps: The particles in L3 and L4 are further optimized by using the Lévy flight, and the particle updating mode is as follows: where P i，j is the velocity of the jth particle in the L i th layer, V i，j is its velocity; and are two particles randomly selected from the two higher layers rl1 and rl2, and k1 and k2 represent the index values of the particles in the respective layers; the layer level is higher than the layer level, i.e. the particle is better than the particle; r1, r2 and r3 are random numbers between [0, 1], is a random number between [0, 1], LB and UB are the lower and upper bounds of the particle position distribution, and levy() is a Levy flight function. 9.The yellow rice wine fermentation prediction method based on ANFIS and random fractal search algorithm according to claim 8, characterized in that, The hierarchical learning strategy with Lévy flight is integrated into the random fractal search algorithm to obtain a hierarchical learning random fractal search algorithm, which comprises the following steps: step1: initializing a particle swarm P, and calculating the function fitness value F corresponding to each particle; step2: entering an iterative loop; step3: generating a new particle from each particle according to formula (7), and updating the function fitness value corresponding to the new particle; Step 4: Sort the particles according to equation (9) if The particle P i is updated according to equation (10), otherwise the particle P i is kept; Step 5: According to the function fitness value size, the particle swarm P is divided into four levels L1~L4, wherein the particles in L1 are not updated, the particles in L2 are updated according to formula (12), the particles in L3 and L4 are updated according to formula (15), and a new particle swarm P is obtained new ; Step 6: Calculate the particle swarm P new The fitness value F of the function corresponding to each particle new ; step7: if F new <F, iteration continues, otherwise iteration ends.

10. A yellow rice wine fermentation prediction system based on ANFIS and random fractal search algorithm, characterized in that, comprise: The data acquisition module is configured to collect data samples of the pre-fermentation process of different production batches of yellow rice wine; The data preprocessing module is configured to divide the data samples into a training set and a test set, and to normalize the data samples; The model optimization module is configured to input the processed data samples into a multi-output adaptive neural fuzzy inference system model constructed in advance, and to identify and optimize the model parameters of the multi-output adaptive neural fuzzy inference system model by using the hierarchical learning random fractal search algorithm, thereby obtaining an optimized multi-output adaptive neural fuzzy inference system model; The model prediction module is configured to predict the fermentation state of the yellow rice wine; The hierarchical learning random fractal search algorithm comprises: The Lévy flight is integrated into the hierarchical learning strategy to obtain a hierarchical learning strategy with Lévy flight; The hierarchical learning strategy with Lévy flight is integrated into the random fractal search algorithm to obtain a hierarchical learning random fractal search algorithm.

Images