An optimization method for the granulation process of a mixed material

Abstract

The present invention discloses an optimization method for a granulation process of a mixed material, which relates to the technical field of process optimization. The method constructs a plurality of training sets based on historical production data after correcting historical dynamic parameters, and trains a plurality of feeding prediction models for historical mixed materials. Measure the material ratio and moisture content of the current mixed material, construct a material combination including at least one set of historical mixed materials, and form a feeding combination model of the current mixed material according to the material combination. Dynamically adjust the granulation production line according to the feeding combination model, and at the same time determine the accuracy of the feeding combination model by comparing the current target hardness with the current finished product hardness, and construct a second loss function to continuously update the feeding combination model. Further, a granulation model of the granulation production line is introduced in the construction process of the feeding prediction model to reflect the influence of the equipment operation state on the production process.

Description

Technical Field

[0001] The present invention relates to the technical field of process optimization, and particularly to an optimization method for the granulation process of mixed materials. Background Art

[0002] By adjusting process parameters, the hardness of pellet feed can be changed. Different hardnesses will affect the digestion and absorption rate of the feed. Therefore, it is necessary to adapt different feed hardnesses according to animals at different stages. When determining process parameters, historical granulation data can be referred to. Chinese Patent Application Publication No. CN118244727A discloses a feed production control method, system and storage medium. This method analyzes the adjustment amount and adjustment direction of process parameters in the historical granulation process based on the particle hardness, particle size, nutrient composition ratio and material properties of the target finished feed. This method determines the adjustment direction and adjustment amount by comparing the granulation force coefficient with a preset range. Although it can ensure that the coefficient is maintained within an appropriate range, there are obvious limitations. First, the relationship function between the preset granulation force coefficient deviation and the adjustment amount is mostly based on fixed rules or empirical formulas, and it is difficult to adapt to the dynamic changes of the production scenario. Moreover, the composition and properties of agricultural product materials change greatly, and the prior art does not consider the influence of factors such as material properties on process parameters, so the optimization efficiency of the granulation process and the parameter adaptation ability are limited. Therefore, the prior art needs to be further improved. Summary of the Invention

[0003] In view of the above problems, the present invention provides an optimization method for the granulation process of mixed materials. First, a feeding prediction model for each historical mixed material is generated, then a feeding combination model under the current material ratio and moisture content is constructed, and then the optimized feeding variables are determined according to the feeding combination model, so as to improve the adaptability of the granulation process to different materials. Further, the model feeding variables in the ideal state are calculated through the granulation model of the granulation production line, and the loss function of the feeding prediction model is generated from the model feeding variables, so as to improve the training accuracy of the feeding prediction model.

[0004] The inventive object of the present application can be achieved by the following technical means:

[0005] An optimization method for the granulation process of mixed materials, comprising the following steps:

[0006] Step 1: Measure the material ratio and moisture content of the historical mixed material, collect the historical production data of the granulation production line for processing the historical mixed material, and generate the granulation model of the granulation production line;

[0007] Step 2: Construct multiple groups of training sets according to the historical production data and the historical target hardness, and label the historical feeding variables in the training sets based on the historical target hardness;

[0008] Step 3: Calculate the first loss function according to the granulation model, and then train multiple feeding prediction models with multiple groups of labeled training sets and the first loss function;

[0009] Step 4: Measure the material ratio and moisture content of the current mixed material, and construct a material combination including at least one group of historical mixed materials;

[0010] Step 5: Extract the feeding prediction model corresponding to each historical mixed material in the material combination to generate a feeding combination model for the current mixed material;

[0011] Step 6: Feed the current mixed material into the granulation production line, collect the current production data of the granulation production line, input the current production data and the current target hardness into the feeding combination model, output the predicted feeding variable, and control the granulation production line according to the predicted feeding variable;

[0012] Step 7: Collect the current finished product hardness of the granulation production line. If the hardness error of the current finished product hardness exceeds the error range, go to Step 8; otherwise, return to Step 6;

[0013] Step 8: Generate the second loss function of the feeding combination model based on the predicted feeding variable and the current target hardness, update the feeding combination model based on the second loss function, and return to Step 6.

[0014] In the present invention, in Step 1, the historical mixed material includes one or several material components of corn flour, expanded soybean meal, wheat bran powder, fermented soybean meal, meat and bone meal, and the material ratio is the mass percentage of each material component in the historical mixed material.

[0015] In the present invention, the historical production data includes historical dynamic parameters and historical static parameters. The historical dynamic parameters include historical feeding amount and historical finished product hardness. The input of the feeding prediction model is the historical production data and the historical target hardness, and the output is the predicted feeding variable.

[0016] In the present invention, in Step 1, a cavity material equation is generated according to the historical feeding amount, a cavity dynamics equation is obtained according to the historical static parameters, and then the granulation model is generated by combining the cavity kinematic equation.

[0017] In the present invention, in Step 2, the time series of the historical feeding amount and the historical finished product hardness are collected, the function value of the cross-correlation function of the historical feeding amount and the historical finished product hardness is calculated based on the time series, the lag time is generated based on the peak value of the function value, the time series of the historical feeding amount and the historical finished product hardness in the historical production data are aligned according to the lag time, and the historical feeding variable is labeled in the training set.

[0018] In the present invention, in step 3, the model feed variable is calculated according to the granulation model and the historical target hardness, and the physical constraint loss Y between the feed prediction model and the granulation model is calculated according to the model feed variable. physics The first loss function Y1 = αY data +(1 - α)Y physics where Y data is the data-driven loss between the feed prediction model and the training set, and α is the weight parameter.

[0019] In the present invention, in step 4, a material combination including at least one set of historical mixed materials is selected, a proportion weight is assigned to each historical mixed material in the material combination, and a difference function between the material combination and the current mixed material is constructed based on the proportion weight, and the material combination is determined by minimizing the difference function.

[0020] In the present invention, in step 5, the feed prediction model is fused according to the proportion weight of each historical mixed material in the material combination to generate a feed combination model.

[0021] In the present invention, in step 6, the current dynamic parameters and the current static parameters of the granulation production line are collected. The current dynamic parameters and the current static parameters form the current production data. Among them, the current dynamic parameters include the current feed rate and the current finished product hardness, and the current static parameters include the die hole opening rate, the number of dies, the die roller rotation speed, the die radius, the die roller radius, the die width, the rated power, and the material bulk density.

[0022] In the present invention, in step 8, first, the model feed variable ΔM4 is determined according to the granulation model and the current target hardness, and then the second loss function (ΔM5 - ΔM4) is calculated 2 , where ΔM5 is the predicted feed variable output by the feed combination model. Finally, the model parameter θ of the feed combination model is adjusted, θ = θ - η▽ θ (ΔM5 - ΔM4) 2 , η is the learning rate of the feed combination model, and ▽ θ represents the gradient of the second loss function (ΔM5 - ΔM4) 2 with respect to the model parameter θ.

[0023] The optimization method for the granulation process of the mixed materials of the present invention has the beneficial effects that: since the material ratio and moisture content of each batch of materials are different, which affects the prediction of the hardness of the finished product, the present invention measures the material ratio and moisture content of the current mixed materials and constructs a material combination including at least one set of historical mixed materials. The feeding prediction models of different historical mixed materials form the feeding combination model of the current mixed materials, and then the granulation production line is controlled according to the feeding combination model. Finally, the accuracy of the feeding combination model is determined by combining the current target hardness and the current hardness of the finished product, and a second loss function is constructed to update the feeding combination model. The feeding combination model of the present invention not only considers the differences in the material ratio and moisture content of different batches of materials, improves the accuracy of the model, but also improves the adaptability of the granulation process to different materials.

[0024] Furthermore, the present invention generates a granulation model between the feeding amount and the granulation pressure by analyzing the structural characteristics of the granulation production line. According to the granulation model and the target hardness, the model feeding variable at the target hardness can be obtained. The loss function of the feeding prediction model is generated from the model feeding variable, and then the feeding prediction model trained previously is updated by the loss function, so as to accelerate the iteration efficiency of the feeding prediction model and improve the training accuracy of the feeding prediction model. Brief Description of the Drawings

[0025] Figure 1 is the flowchart of the optimization method for the granulation process of the mixed materials of the present invention;

[0026] Figure 2 is the schematic diagram of the cross-correlation function of the present invention;

[0027] Figure 3 is the schematic diagram before the correction of the historical dynamic parameters of the present invention;

[0028] Figure 4 is the schematic diagram after the correction of the historical dynamic parameters of the present invention;

[0029] Figure 5 is the schematic diagram of the feeding prediction model of the present invention;

[0030] Figure 6 is the structural schematic diagram of the feeding prediction model of the present invention;

[0031] Figure 7 is the schematic diagram of generating the first loss function of the present invention;

[0032] Figure 8 is the schematic diagram of the die cavity of the granulation production line of the present invention;

[0033] Figure 9 is the schematic diagram of generating the material combination of the present invention. Detailed Embodiments

[0034] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0035] An appropriate feed hardness can not only reduce the breakage and pulverization of feed during transportation and storage, maintain integrity, but also help animals digest and absorb nutrients. In a ring die granulator, the feeding rate is a key factor in controlling the hardness of the finished product. The feeding rate affects the compression time and pressure of the material in the ring die, thereby affecting the particle hardness. Unstable feeding rate will lead to uneven particle hardness and affect the feed quality. At present, the precise control of the feeding rate needs to rely on manual experience, and manual operation is likely to introduce fluctuations, affecting the consistency of the finished product. The present invention trains multiple feeding prediction models through the historical production data of the granulation production line. According to the material ratio and moisture content of the current mixed material and the material ratio and moisture content of the historical mixed material, a material combination including at least one group of historical mixed materials is constructed. Based on the feeding prediction models corresponding to the historical mixed materials in the material combination, a feeding combination model for the current mixed material is constructed. The feeding combination model can predict the predicted feeding variable for adjusting the feeding rate according to the target hardness and production data, preventing the system from being unstable due to manual adjustment according to experience. Embodiment 1

[0036] Refer to Figure 1 , the optimization method of the mixed material granulation process of the present invention described in detail in this embodiment includes the following steps:

[0037] Step 1: Measure the material ratio and moisture content of the historical mixed material, collect the historical production data of the granulation production line for processing the historical mixed material, and generate a granulation model of the granulation production line. The historical mixed material includes one or several material components such as corn flour, extruded soybean meal, wheat bran flour, fermented soybean meal, and meat and bone meal. The material ratio is the mass percentage of each material component in the historical mixed material. The historical production data includes historical dynamic parameters and historical static parameters. The historical dynamic parameters include historical feeding rate and historical finished product hardness. According to the historical feeding rate, a cavity material equation is generated, and according to the historical static parameters, a cavity dynamics equation of the granulation production line is obtained. Then, combined with the cavity kinematics equation, the granulation model is generated. The granulation model of this embodiment specifically refers to Embodiment 3.

[0038] Step 2: Construct multiple groups of training sets according to the historical production data and historical target hardness, and label the historical feeding variables in the training sets based on the historical target hardness. The data response in agricultural product processing has obvious hysteresis, that is, the hardness change lags behind the change of the feeding rate. Refer to Figures 2 to 4 , the present invention can correct the historical dynamic parameters in the historical production data before constructing the training set: collect the time series of the historical feeding rate and the historical finished product hardness. The time series of the historical finished product hardness is D1(t), and the time series of the historical feeding rate is M(t). Then, the expression of the cross-correlation function is , where \(t\) is the time of the sampling point, \(T\) is the length of the time series, and \(\tau\) is the time shift. Calculate the function value \(E\) of the cross-correlation function. The time shift corresponding to the peak value of the function value \(E\) is the lag time between the historical feed rate and the historical finished product hardness. The lag time in this embodiment is, for example, 5 minutes. Align the historical feed rate and the finished product hardness time series according to the lag time.

[0039] Label the historical feed variables in the training set. In the actual production process, due to process fluctuations and human operation differences, the data of high-quality historical feed variables that can accurately match the historical target hardness are relatively scarce. To construct an effective training set, it is necessary to label the historical feed variables based on the historical target data. The following labeling method is adopted: when the historical finished product hardness produced by the granulation production line at the historical feed rate \(M1\) is \(D1\), the required historical target hardness at this time is \(D1'\); and when the historical finished product hardness produced at another historical feed rate \(M2\) is \(D2\), if the relative error between \(D1'\) and \(D2\) is less than \(1\% D1'\) (i.e., \(|D1' - D2| < 1\% D1'\)), then label the historical feed variable when the historical feed rate is \(M1\) as \(|M2 - M1|\). Label the historical feed variables in the training set based on the above labeling method.

[0040] Step 3: Calculate the first loss function according to the granulation model, and then train multiple groups of feed prediction models with multiple groups of labeled training sets and the first loss function. Specifically, generate a feed prediction model according to multiple groups of labeled training sets, and then update and iterate the feed prediction model with the first loss function, so as to complete the training of the feed prediction model. The granulation model can calculate the model feed variable based on the historical target hardness, the feed prediction model can obtain the predicted feed variable, and the first loss function includes the physical constraint loss between the feed variables of the feed prediction model and the granulation model, as described in Embodiment 2. The present invention does not limit the specific structure of the feed prediction model, which can be the one-dimensional convolutional neural network described in Embodiment 2 or the Transformer architecture.

[0041] Step 4: Measure the material ratio and moisture content of the current mixed material, and construct a material combination including at least one group of historical mixed materials. Specifically, select a material combination including at least one group of historical mixed materials, assign a proportion weight to each historical mixed material in the material combination, construct a difference function between the material combination and the current mixed material based on the proportion weight, and minimize the difference function to determine the material combination.

[0042] In this embodiment, first select \(I\) groups of historical mixed materials according to the material components in the current mixed material. The material combination includes \(I\) groups of historical mixed materials, and construct \(I\) groups of historical mixed material vectors and a group of proportion weight vectors according to the \(I\) groups of historical mixed materials. Among them, the historical mixed material vector of the historical mixed material \(i\) is \((q 1i ,q 2i,…,q ni ,…,q Ni ), q ni is the proportion of the nth material component in the historical mixed material i, and N is the number of material components. The proportion weight vector is (w1, w2, …, w i ,…, w I ), w i is the proportion weight of the historical mixed material i, where i = 1, 2, …, I. , .

[0043] Reconstruct the difference function and determine the material combination. Reconstruct the difference function . Among them, (q'1, q'2, …, q' n ,…, q' N ) is the current mixed material vector, H i is the moisture content of the historical mixed material i, H' is the moisture content of the current mixed material, and γ is the adjustment coefficient of the material ratio and moisture content, where γ ∈ [0, 1]. Minimize the difference function to determine the material combination. For example, use the least squares method to minimize the difference function to obtain the optimal proportion weight vector (when the difference function takes the minimum value), and determine the material combination according to the proportion weights of this proportion weight vector. Specifically, refer to Embodiment 4.

[0044] Step 5: Extract the feeding prediction model corresponding to each historical mixed material in the material combination, and generate the feeding combination model of the current mixed material. According to the proportion weights of each historical mixed material in the material combination, fuse the feeding prediction models to generate the feeding combination model. Specifically, the present invention uses a linear weighted fusion method to calculate the feeding combination model G. The feeding combination model , G i is the feeding prediction model of the historical material combination i. It can be understood that if there is only one set of historical mixed materials in the material combination, directly use the corresponding feeding prediction model as the feeding combination model of the current mixed material. The feeding combination model has the same structure as the feeding prediction model.

[0045] Step 6: Feed the current mixed material into the granulation production line, collect the current production data of the granulation production line, input the current production data and the current target hardness into the feeding combination model, output the predicted feeding variables, and control the granulation production line according to the predicted feeding variables. Collect the current dynamic parameters and current static parameters of the granulation production line. The current dynamic parameters and current static parameters constitute the current production data. The current dynamic parameters include the current feeding amount and the current finished product hardness, and the current static parameters include the die hole opening rate, the number of die holes, the die roller rotation speed, the die radius, the roller radius, the die width, the rated power, and the material bulk density.

[0046] Step 7: Collect the current hardness of the finished product on the granulation production line. If the hardness error of the current hardness of the finished product exceeds the error range, go to Step 8; otherwise, return to Step 6. To improve the model adaptability, when the hardness error of the current hardness of the finished product exceeds the error range, enter the update step of the feeding combination model. The hardness error is the difference between the current hardness of the finished product and the current target hardness. The specific value of the error range is not limited in this embodiment. Generally, if the error range is too large, the production accuracy will be reduced; if the error range is too small, the update of the feeding combination model will be too frequent. The error range can be set according to the actual production standard. In this embodiment, the current target hardness is D2', and the production standard is that a deviation of ±0.04D2' in the hardness of the finished product is allowed, so the error range can be set to ±0.04D2'.

[0047] Step 8: Generate the second loss function of the feeding combination model based on the predicted feeding variable and the current target hardness, update the feeding combination model based on the second loss function, and return to Step 6. Specifically, first determine the model feeding variable ΔM4 according to the granulation model and the current target hardness, and then calculate the second loss function (ΔM5 - ΔM4) 2 . Wherein, ΔM5 is the predicted feeding variable output by the feeding combination model. Finally, adjust the model parameter θ of the feeding combination model, θ = θ - η▽ θ (ΔM5 - ΔM4) 2 , η is the learning rate of the feeding combination model, and ▽ θ represents the gradient of the second loss function (ΔM5 - ΔM4) 2 with respect to the model parameter θ. The model parameter θ is a multi-dimensional vector containing all trainable parameters in the feeding combination model. The calculation method of the model feeding variable ΔM4 can be similarly referred to the calculation method of the model feeding variable ΔM3 in Embodiment 3. Embodiment 2

[0048] Refer to Figures 5 to 7 , this embodiment further discloses a preferred method for training the feeding prediction model in Step 3. For example, the present invention uses a 1D-CNN (one-dimensional convolutional neural network) architecture to train the feeding prediction model. The 1D-CNN has a simple structure and fewer model parameters, which can effectively reduce the risk of overfitting, especially suitable for the situation where high-quality data is scarce in the present invention. The feeding prediction model includes an input layer, a convolutional layer, a fully connected layer, and an output layer, and its output is the predicted feeding variable. The role of the pooling layer is to reduce the data dimension. In this embodiment, the data dimension of the training set is low, so the pooling layer is not used in this embodiment.

[0049] Step 201: Input the data of the training set into the input layer one by one. Through a fully connected layer, map the corresponding historical static parameters, historical target hardness, and historical feeding variables to the same feature space as the historical dynamic parameters. The input layer concatenates the historical dynamic parameters, the mapped historical static parameters, the historical target hardness, and the historical feeding variables according to the dimension of the feature space to generate an input matrix. The convolution kernel size in the convolutional layer is 2×2, and the convolutional kernel performs a convolution operation on the input matrix to extract the local features of the time series in the input matrix. The fully connected layer performs a non-linear transformation on the output of the convolutional layer to integrate the global features, and the output layer maps the global features to the predicted feeding variable through a fully connected layer.

[0050] Step 202: The predicted feeding variable output by the feeding prediction model is ΔM2. Use the mean square error to calculate the data-driven loss Y between the feeding prediction model and the training set data , that is, Y data =(ΔM1 - ΔM1) 2 , where ΔM1 is the historical feeding variable in the training set.

[0051] Step 203: Use the mean square error to calculate the physical constraint loss Y between the feeding prediction model and the granulation model physics , that is, Y physics =(ΔM2 - ΔM3) 2 , and the method for calculating the model feeding variable ΔM3 based on the granulation model and the historical target hardness is as described in Embodiment III.

[0052] Step 204: The first loss function Y1 = αY data +(1 - α)Y physics , where α is a weight parameter used to balance the data-driven loss and the physical constraint loss. Update the feeding prediction model according to the first loss function. If the input of the training set data is completed, end the model training; otherwise, return to Step 201 and continue to train the feeding prediction model. Embodiment III

[0053] In Embodiment II, the model feeding variable ΔM3 is used to calculate the first loss function. This embodiment further gives a preferred method for calculating the model feeding variable ΔM3 based on the granulation model and the historical target hardness.

[0054] Collect historical dynamic parameters and historical static parameters. The historical dynamic parameters include the historical feeding amount M, and the historical static parameters include: the ring die opening ratio λ, the number of ring dies A1, the die-roller rotation speed A2, the ring die radius r1, the die-roller radius r2, the ring die width L2, the rated power P2, and the material bulk density ρ.

[0055] Generate a die cavity material equation according to the historical feeding amount M. The die cavity material equation is M / ρ = 6×10 -11 πλA1A2L2[r12 -(r1 - L1) 2 , that is, the material thickness , where A1 is usually 2, and ρ is usually 5 - 12 kN / m³.

[0056] Obtain the cavity dynamics equation based on historical static parameters. Refer to Figure 8 , the center of the cavity is O1, the center of the ring die is O2, the central angle of the extrusion zone is β, and the total pressure F is applied to the extrusion zone. The area of the extrusion zone S = r1L2β, and the central angle β = arccos[(r1 - r2) 2 +(r1 - L1) 2 - r2 2 / [2(r1 - r2)(r1 - L1)], then the cavity kinematic equation is S = r1L2arccos[(r1 - r2) 2 +(r1 - L1) 2 - r2 2 / [2(r1 - r2)(r1 - L1)].

[0057] Generate the granulation model by combining the cavity kinematic equation. The total pressure F of the extrusion zone = 9550P2 / (μr1A2), where μ is the friction coefficient between the historical mixed material and the cavity, and S is the area of the extrusion zone. The granulation pressure P1 = F / S, so the cavity dynamics equation is P1 = 9550P2 / (μr1A2S). The granulation model consists of the cavity material equation, the cavity kinematic equation, and the cavity dynamics equation. Therefore, the granulation model can determine the functional relationship between the historical feed rate M and the granulation pressure P1.

[0058] Calculate the model feed variable according to the granulation model and the historical target hardness. During the ring die granulation process, the granulation pressure in the granulation model is related to the historical target hardness. In one embodiment, construct a linear relationship between the historical target hardness D1' and the granulation pressure P1: D1' = kP1 + C, where k is the sensitivity coefficient between the historical target hardness and the granulation pressure, and C is the reference offset. Substitute the historical target hardness D1' into the linear relationship to obtain the granulation pressure P1 corresponding to the historical target hardness D1'. In another embodiment, draw a relationship table of the historical target hardness and the granulation pressure, and obtain the granulation pressure P1 corresponding to the historical target hardness D1' by looking up the table. Then, combine the granulation model and the granulation pressure to obtain the model feed rate M'. The model feed variable ΔM3 is the difference between the model feed rate M' and the historical feed rate M, that is, ΔM3 = M' - M. Example 4

[0059] Refer to Figure 9 , this embodiment further discloses a preferred method for generating the feed combination model of the current mixed material in step 5.

[0060] The material ratio of the current mixed material is 48% corn flour, 23% expanded soybean meal, 5% wheat bran powder, 8% fermented soybean meal, 7% meat and bone meal, and the remaining auxiliary materials are 9%. The moisture content of the current mixed material is 15%.

[0061] The material ratio and moisture content of the historical mixed materials are as shown in the following table. Among them, historical mixed material 1 and historical mixed material 2 are fattening pig feeds. Historical mixed material 3 is a sow feed. Historical mixed material 4 is a weaned piglet feed. Historical mixed material 5 is a nursery pig feed.

[0062]

[0063] Construct a difference function: , for example, γ takes 0.5, I = 5, N = 6. The least squares method is used to solve the difference function, and the optimal proportion weight vector (0, 0.5, 0.25, 0, 0.25) is obtained. Therefore, the determined material combination includes 50% of historical mixed material 2, 25% of historical mixed material 3, and 25% of historical mixed material 5.

[0064] Construct a feeding combination model. Combine the feeding combination model of Example 1 , so in this embodiment, the feeding combination model G = 50%G2 + 25%G3 + 25%G5 can be determined. It can be understood that if there is only one set of historical mixed materials in the material combination, directly use the corresponding feeding prediction model as the feeding combination model of the current mixed material. The structure of the feeding combination model is the same as that of the feeding prediction model.

[0065] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for optimizing a mixed material granulation process, characterized in that: The following steps are involved: Step 1: Determine the material ratio and moisture content of the historical mixed material, collect the historical production data of the granulation production line processing the historical mixed material, and generate a granulation model of the granulation production line; Step 2: Construct multiple training sets based on historical production data and historical target hardness, and label historical feeding variables in the training sets based on historical target hardness; Step 3: Calculate the first loss function according to the granulation model, and then train multiple groups of feeding prediction models using multiple groups of labeled training sets and the first loss function; Step 4: Determine the material ratio and moisture content of the current mixed material, and construct a material combination containing at least one set of historical mixed materials; Step 5: Extract the feeding prediction model corresponding to each historical mixed material in the material combination to generate a feeding combination model of the current mixed material; Step 6: feeding the current mixed material into the granulation production line, collecting the current production data of the granulation production line, inputting the current production data and the current target hardness into the feeding combination model, outputting the predicted feeding variables, and controlling the granulation production line according to the predicted feeding variables; Step 7: Collect the hardness of the current finished product of the granulation production line. If the hardness error of the current finished product exceeds the error range, proceed to step 8, otherwise return to step 6; Step 8: Generate a second loss function of the feeding combination model based on the predicted feeding variables and the current target hardness, update the feeding combination model based on the second loss function, and return to step 6.

2. The method for optimizing the mixed material granulation process according to claim 1, characterized in that: In step 1, the historical mixed material includes one or more material components of corn flour, puffed soybean meal, wheat bran powder, fermented soybean meal, and meat and bone meal, and the material ratio is the mass percentage of each material component in the historical mixed material.

3. The optimization method of the mixed material granulation process according to claim 1, characterized in that: The historical production data includes historical dynamic parameters and historical static parameters. The historical dynamic parameters include historical feeding amount and historical finished product hardness. The input of the feeding prediction model is the historical production data and the historical target hardness, and the output is the predicted feeding variable.

4. The method for optimizing the mixed material granulation process according to claim 3, characterized in that: In step 1, a cavity material equation is generated according to the historical feed rate, a cavity dynamic equation is obtained according to the historical static parameters, and then the granulation model is generated in combination with the cavity kinematic equation.

5. The method for optimizing the mixed material granulation process according to claim 3, characterized in that: In step 2, the time series of historical feed amount and historical finished product hardness are collected, and the function value of the cross-correlation function of the historical feed amount and the historical finished product hardness is calculated based on the time series. The lag time is generated based on the peak value of the function value, and the time series of the historical feed amount and the historical finished product hardness in the historical production data are aligned according to the lag time, and then the historical feeding variables are labeled in the training set.

6. The method for optimizing the mixed material granulation process according to claim 1, characterized in that: In step 3, the model feeding variables are calculated based on the pelletizing model and the historical target hardness, and the physical constraint loss Y between the feeding prediction model and the pelletizing model is calculated based on the model feeding variables. physics , the first loss function Y1=αY data +(1-α)Y physics , Y data is the data-driven loss between the feeding prediction model and the training set, and α is the weight parameter.

7. The method for optimizing the mixed material granulation process according to claim 1, characterized in that: In step 4, a material combination including at least one set of historical mixed materials is selected, a proportion weight is assigned to each historical mixed material in the material combination, a difference function between the material combination and the current mixed material is constructed based on the proportion weight, and the material combination is determined by minimizing the difference function.

8. The method for optimizing the mixed material granulation process according to claim 7, characterized in that: In step 5, the feed prediction model is integrated according to the weight of each historical mixed material in the material combination to generate a feed combination model.

9. The method for optimizing the mixed material granulation process according to claim 1, characterized in that: In step 6, the current dynamic parameters and the current static parameters of the granulation production line are collected, and the current dynamic parameters and the current static parameters constitute the current production data, wherein the current dynamic parameters include the current feed amount and the current finished product hardness, and the current static parameters include the ring die opening rate, the number of ring dies, the die roller speed, the ring die radius, the die roller radius, the ring die width, the rated power, and the material bulk density.

10. The method for optimizing the mixed material granulation process according to claim 1, characterized in that: In step 8, the model feed variable ΔM4 is first determined according to the granulation model and the current target hardness, and then the second loss function (ΔM5-ΔM4) is calculated. 2 , where ΔM5 is the predicted feeding variable output by the feeding combination model. Finally, the model parameter θ of the feeding combination model is adjusted, θ=θ-η▽ θ (ΔM5-ΔM4) 2 , η is the learning rate of the feeding combination model, ▽ θ Represents the second loss function (ΔM5-ΔM4) 2 The gradient of the model parameters θ.

Images