Rare earth extraction simulation model construction method and device, rare earth extraction simulation model application method and device and medium

Abstract

The invention discloses a construction and application method and device of a rare earth extraction simulation model and a medium, and relates to the technical field of chemical process simulation, and the method comprises the steps: obtaining rare earth extraction historical data sets under different working conditions; each sample in the rare earth extraction historical data set comprises historical production process characteristic parameters of rare earth extraction and a corresponding component content actual value of each separation stage; training a multi-branch generative adversarial network by using the rare earth extraction historical data set to obtain a rare earth extraction simulation model; the multi-branch generative adversarial network comprises a main network and a plurality of branch networks; each separation stage in the rare earth extraction process is correspondingly provided with a branch network; according to the method, the multi-branch generative adversarial network for different stages of extraction tanks is constructed, the prediction precision and accuracy of rare earth extraction simulation are improved, and the step-by-step hierarchical output characteristic of the component content in actual rare earth production is more appropriately reflected.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of chemical process simulation, in particular to a rare earth extraction simulation model construction method, application method, device and medium. BACKGROUND

[0002] Rare earths include 17 elements in the periodic table, including lanthanides, scandium (Sc) and yttrium (Y). In the actual production process of rare earth extraction, the traditional component content detection method is to manually take samples from a certain stage extraction tank and then conduct offline testing. The working personnel adjusts the on-site production process characteristic parameters according to the test results. This method has poor timeliness. Therefore, it is necessary to accurately predict the rare earth element component content of the extraction tank in the extraction production process and timely adjust the operation variables to improve product quality.

[0003] There are mainly two types of research on chemical process simulation methods: one is static modeling based on in-depth mechanism understanding, which constructs a mathematical model by analyzing the physical and chemical properties of the extraction process in detail. This process highly depends on the participation of experts with rich field knowledge to ensure that the model can reflect the actual process. The other is dynamic modeling driven by data, which uses a large amount of process data to train the model to realize real-time simulation and prediction of the process. However, due to the complex physical and chemical changes in the rare earth extraction process, there are characteristics such as complex mechanism, multiple stages and strong coupling between stages. Mechanism modeling ignores many factors, so the mechanism model cannot well predict the changes of rare earth element component content in the extraction process. So far, there is no mechanism model that can fully and accurately describe all the details of the rare earth extraction and separation process. They all contain certain idealized assumptions. At the same time, with the continuous progress of data processing technology and data acquisition system, data-driven modeling methods based on historical data are increasingly favored. This method does not need to rely on deep field expert experience, but only needs sufficient production historical data to effectively train the model.

[0004] In the data-driven modeling method, due to the superior performance of the neural network algorithm, there are currently a large number of research cases of applying neural networks to rare earth extraction modeling. However, the rare earth extraction process is accompanied by complex physical and chemical changes, has the characteristics of complex mechanism, multiple stages and strong coupling between stages, and part of the mechanism process has not been clear so far. In the existing data-driven method, a shallow network, which is a black box model, is used to model the extraction separation process of dozens or even hundreds of stages. The prediction result is output in a single dimension at the final output layer of the network. For the complex industrial scene with variable working conditions, the rare earth extraction industry has a phenomenon of data scarcity, which leads to the inability of the neural network to handle the actual complex nonlinear relationship due to the scarcity of samples, insufficient learning ability, and inability to accurately describe the multi-level output of the component content. And it does not consider the different mapping relationships between the component content of each stage of the extraction tank and the input features, as well as the actual characteristics that the transfer exchange and cascade effect of the components will cause coupling between adjacent extraction tanks, which will lead to the component content value obtained cannot simulate the actual change of the component content in different extraction tanks. Therefore, for the process simulation of the multi-stage complex rare earth extraction process, the existing method of using a shallow neural network for modeling has limitations and insufficient prediction accuracy. SUMMARY

[0005] The purpose of the present application is to provide a rare earth extraction simulation model construction, application method, device and medium, which can construct a rare earth extraction simulation model capable of accurately simulating the actual change of the component content in different extraction tanks, thereby improving the accuracy of rare earth extraction simulation.

[0006] To achieve the above purpose, the present application provides the following scheme: In a first aspect, the present application provides a rare earth extraction simulation model construction method, comprising: obtaining a rare earth extraction historical data set under different working conditions; each sample in the rare earth extraction historical data set includes historical production process feature parameters of rare earth extraction and actual values of component content of each separation stage; training a multi-branch generative adversarial network using the rare earth extraction historical data set to obtain a rare earth extraction simulation model; the multi-branch generative adversarial network includes a main network and a plurality of first branch networks; each separation stage in the extraction process corresponds to a first branch network; the main network includes an input layer and a plurality of hidden layers connected in series; the output of each hidden layer is connected to the input of each first branch network one by one; each first branch network includes a generator, a discriminator and a regressor; the input of the generator is connected to the output of the corresponding hidden layer; the input of the regressor is connected to the output of the generator, and the output of the regressor is the simulation result of the component content of the corresponding separation stage in the extraction process; during training, the input of the regressor further includes the actual value of the component content of the corresponding separation stage.

[0007] In a second aspect, the application provides an application method of the rare earth extraction simulation model, comprising: obtaining the characteristic parameters of the rare earth extraction production process to be simulated; inputting the characteristic parameters of the rare earth extraction production process to be simulated into the rare earth extraction simulation model to obtain the simulation results of the component content of each stage in the rare earth extraction process; the rare earth extraction simulation model is constructed based on the construction method of the rare earth extraction simulation model described above; the rare earth extraction simulation model comprises a main network and a second branch network in a multi-branch generative adversarial network; the second branch network comprises a generator and a regressor of the first branch network in the multi-branch generative adversarial network.

[0008] In a third aspect, the application provides a construction device of a rare earth extraction simulation model, comprising: a sample acquisition module, configured to acquire a rare earth extraction historical data set under different working conditions; each sample in the rare earth extraction historical data set comprises historical production process characteristic parameters of rare earth extraction and actual values of component content of each separation stage; a model construction module, configured to train a multi-branch generative adversarial network by using the rare earth extraction historical data set to obtain a rare earth extraction simulation model; the multi-branch generative adversarial network comprises a main network and a plurality of first branch networks; each separation stage in the extraction process corresponds to a first branch network; the main network comprises an input layer and a plurality of hidden layers connected in series; the output of each hidden layer is connected to the input of each first branch network one by one; each first branch network comprises a generator, a discriminator and a regressor; the input of the generator is connected to the output of the corresponding hidden layer; the input of the regressor is connected to the output of the generator, and the output of the regressor is the simulation result of the component content of the corresponding separation stage in the extraction process; during training, the input of the regressor further comprises the actual value of the component content of the corresponding separation stage.

[0009] In a fourth aspect, the application provides an application device of a rare earth extraction simulation model, comprising: a parameter acquisition module, configured to acquire the characteristic parameters of the rare earth extraction production process to be simulated; a rare earth extraction simulation module, configured to input the characteristic parameters of the rare earth extraction production process to be simulated into the rare earth extraction simulation model to obtain the simulation results of the component content of each stage in the rare earth extraction process; the rare earth extraction simulation model is constructed based on the construction method of the rare earth extraction simulation model described above; the rare earth extraction simulation model comprises a main network and a second branch network in a multi-branch generative adversarial network; the second branch network comprises a generator and a regressor of the first branch network in the multi-branch generative adversarial network.

[0010] In a fifth aspect, the present application provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method for constructing a rare earth extraction simulation model or the method for applying the rare earth extraction simulation model.

[0011] In a sixth aspect, the present application provides a computer readable storage medium, which stores a computer program, wherein the computer program is executed by a processor to implement the method for constructing a rare earth extraction simulation model or the method for applying the rare earth extraction simulation model.

[0012] According to the specific embodiments provided in the present application, the following technical effects are disclosed: The present application provides a method and device for constructing and applying a rare earth extraction simulation model, and designs a multi-branch generative adversarial network, each branch network corresponds to a first extraction tank, and each branch is connected together through each hidden layer in the main network to jointly model different stages of extraction tanks. The main network uses multiple hidden layers to learn the original features step by step, and the main network has the same number of hidden layers as the stages of extraction. The process is consistent with the hierarchical output of component content in rare earth production. A plurality of independent branch networks are introduced under the main network. Each branch network first introduces a generator to map to the latent space step by step to generate samples, so that the generator of each branch network can extract the features learned by the hidden layer of the previous branch network. The multi-branch generative adversarial network designed in the present application considers the different mapping relationships between the component content of each stage of the extraction tank and the input features, as well as the actual characteristics that the transmission and exchange of components and the cascade effect will cause the coupling of adjacent extraction tanks. Therefore, the application of the multi-branch generative adversarial network designed in the present application can improve the accuracy of rare earth extraction simulation. BRIEF DESCRIPTION OF DRAWINGS

[0013] 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 introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0014] Figure 1 The application environment diagram of the method for constructing a rare earth extraction simulation model in an embodiment of the present application; Figure 2 The flowchart of the method for constructing a rare earth extraction simulation model provided in an embodiment of the present application; Figure 3 The technical concept diagram of the method for constructing a rare earth extraction simulation model provided in an embodiment of the present application; Figure 4A structural schematic diagram of a multi-branch generative adversarial network provided by an embodiment of the present application is shown in FIG. 1. Figure 5 A structural schematic diagram of a main network provided by an embodiment of the present application is shown in FIG. 2. Figure 6 A structural schematic diagram of a discriminator and a regressor network provided by an embodiment of the present application is shown in FIG. 3. Figure 7 A rare earth extraction process production flowchart provided by an embodiment of the present application is shown in FIG. 4. Figure 8 A simulated prediction curve diagram of a multi-branch generative adversarial network for the content of the 25th stage organic phase in a test set provided by an embodiment of the present application is shown in FIG. 5. Figure 9 A simulated prediction curve diagram of a multi-branch generative adversarial network for the content of the 25th stage water phase in a test set provided by an embodiment of the present application is shown in FIG. 6. Figure 10 A Pr component content distribution comparison diagram of the 50th example and model prediction provided by an embodiment of the present application is shown in FIG. 7. Figure 11 A rare earth extraction process simulation result diagram of a test set sample of the 36th example provided by an embodiment of the present application is shown in FIG. 8. Figure 12 A rare earth extraction process simulation result diagram of a test set sample of the 50th example provided by an embodiment of the present application is shown in FIG. 9. Figure 13 A rare earth extraction process simulation result diagram of a test set sample of the 142th example provided by an embodiment of the present application is shown in FIG. 10. Figure 14 A functional module schematic diagram of a construction device of a rare earth extraction simulation model provided by an embodiment of the present application is shown in FIG. 11. Figure 15 A flowchart of an application method of a rare earth extraction simulation model provided by an embodiment of the present application is shown in FIG. 12. Figure 16 A functional module schematic diagram of an application device of a rare earth extraction simulation model provided by an embodiment of the present application is shown in FIG. 13. Figure 17 A structural schematic diagram of a computer device provided by an embodiment of the present application is shown in FIG. 14. DETAILED DESCRIPTION

[0015] 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 only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0016] The above objects, characteristics and advantages of the present application will be more apparent and understandable from the following detailed description of the present application in conjunction with the accompanying drawings and specific embodiments.

[0017] The method for constructing the rare earth extraction simulation model provided in the embodiments of the present application can be applied to the application environment as shown in Figure 1 The terminal communicates with the server through the network. The data storage system can store the data required to be processed by the server. The data storage system can be separately arranged, integrated on the server, or placed on the cloud or other servers. The terminal can send the rare earth extraction historical data set under different working conditions (each sample in the rare earth extraction historical data set includes the historical production process characteristic parameters of rare earth extraction and the actual values of the component content of each separation stage) to the server. After receiving the rare earth extraction historical data set under different working conditions, the server trains a multi-branch generative adversarial network using the rare earth extraction historical data set to obtain a rare earth extraction simulation model. The multi-branch generative adversarial network includes a main network and a plurality of first branch networks. Each separation stage in the extraction process corresponds to a first branch network. The main network includes an input layer and a plurality of hidden layers connected in series. The output of each hidden layer is connected to the input of each first branch network one by one. Each first branch network includes a generator, a discriminator, and a regressor. The input of the generator is connected to the output of the corresponding hidden layer. The input of the regressor is connected to the output of the generator, and the output of the regressor is the simulation result of the component content of the corresponding separation stage in the extraction process. During training, the input of the regressor also includes the actual value of the component content of the corresponding separation stage. In addition, in some embodiments, the method for constructing the rare earth extraction simulation model can also be implemented by the server or the terminal alone, such as the terminal directly constructing the model for the rare earth extraction historical data set under different working conditions, or the server obtaining the rare earth extraction historical data set under different working conditions from the data storage system and constructing the model.

[0018] The terminal can be, but is not limited to, various desktop computers, notebook computers, smart phones, tablet computers, Internet of Things devices, and portable wearable devices. The server can be implemented by an independent server or a server cluster composed of multiple servers, and can also be a cloud server.

[0019] In an exemplary embodiment, as shown in Figure 2 and Figure 3 A method for constructing a rare earth extraction simulation model is provided, which is executed by a computer device, specifically by a terminal or a server, or by both the terminal and the server. In the embodiments of the present application, the method is applied to the server in Figure 1 The method includes the following steps 101 to 102.

[0020] Step 101, obtaining a rare earth extraction historical data set (i.e. a rare earth original data set in Figure 3 ) under different working conditions; each sample in the rare earth extraction historical data set includes a historical production process characteristic parameter of rare earth extraction and a corresponding component content actual value of each separation stage.

[0021] For a rare earth extraction actual production process, according to production indexes and process requirements, a rare earth extraction historical data set is obtained, which includes a component content of a feed liquid, a relative separation coefficient between each component, a scrubbing agent flow, an extractant flow, a feeding mode, an outlet index and an outlet separation coefficient under each production working condition, and then a component content value of each stage when extraction separation reaches dynamic equilibrium is obtained through sampling and testing, and then a rare earth extraction historical data set is constructed, wherein a sample includes a group of production process characteristic parameters and a corresponding component content actual value of each separation stage. The rare earth extraction historical data set includes sufficient extraction data under different production working conditions.

[0022] Step 102, training a multi-branch generative adversarial network using the rare earth extraction historical data set to obtain a rare earth extraction simulation model; as shown in Figure 4 and Figure 5 , the multi-branch generative adversarial network includes a main network and a plurality of first branch networks; each separation stage in the extraction process corresponds to a first branch network; the main network includes an input layer and a plurality of hidden layers connected in series; the output of each hidden layer is connected to the input of each first branch network one by one; each first branch network includes a generator, a discriminator and a regressor; the input of the generator is connected to the output of the corresponding hidden layer, the output of the generator is respectively connected to the input of the discriminator and the input of the regressor, the input of the discriminator further includes a component content actual value of the corresponding separation stage, the output of the discriminator is connected to the input of the generator, and the output of the regressor is a simulation result of the component content of the corresponding separation stage in the extraction process; during training, the input of the regressor further includes the component content actual value of the corresponding separation stage.

[0023] As shown in Figure 6 , the input of the discriminator D is gradually extracted through three layers of convolution (the number of channels is 32, 64 and 128 in turn) and LeakyReLU activation function, and then input into a full connection layer (220 dimensions and 1 dimension) after flattening (corresponding to FLATTEN in Figure 6 ), and then output a probability through a Sigmoid activation function (corresponding to Figure 6 in ). The input of the regressor R is flattened after being processed by a layer of convolution (output channel 32) and LeakyReLU activation, and then output a simulation result of the component content of the corresponding separation stage through two full connection layers (1024 dimensions and 2 dimensions) (corresponding to Figure 6 in ). Figure 6In this context, "Bs" represents the batch size, and "Dense" represents the fully connected layer. The parameters of the first convolutional layer of the regressor R are the same as those of the first convolutional layer of the discriminator D, i.e., they share convolutional layers. Therefore, the discriminator D consists of a first convolutional layer, a second convolutional layer, a third convolutional layer, a first fully connected layer, and a second fully connected layer connected in sequence. The regressor R consists of a fourth convolutional layer, a third fully connected layer, and a fourth fully connected layer connected in sequence.

[0024] like Figure 4 , Figure 5 , Figure 6 and Figure 7 As shown, for the rare earth extraction process, based on the process design, the number of stages for the extraction section and the washing section are determined to be as follows: and Then the total number of separation levels is , No. Each stage is a feed stage. Each stage corresponds to one branch of a multi-branch generative adversarial network (GAN). Each stage contains an organic phase and an aqueous phase. Since the sum of the contents of the difficult-to-extract and easy-to-extract components in the same phase is 1, this application simplifies the structure of the GAN by considering only the easy-to-extract component A in the organic phase and the difficult-to-extract component B in the aqueous phase as the outputs of the GAN. Each first branch network has two outputs: the easy-to-extract component value in the organic phase and the difficult-to-extract component value in the aqueous phase. Considering the actual characteristics of rare earth extraction sites, component transfer and cascade effects can cause coupling between adjacent extraction cells. Therefore, a multi-branch GAN is designed, where each branch network corresponds to one extraction cell. The branch networks are connected in series through the hidden layers in the main network to jointly model extraction cells of different stages. The main network uses multiple hidden layers to learn the original features step by step. A hidden layer, the process closely mirroring the hierarchical output of component content in rare earth production. An additional layer is introduced under the main network. Each first branch network (first branch network) first introduces a generator G structure (e.g., a fully connected layer) to progressively map to the latent space to generate samples. This allows the generator of each first branch network to extract features learned from the hidden layers of the previous first branch network. An independent discriminator D is then introduced under each first branch network. Each discriminator, constrained by its own loss function, forces the generator to learn different patterns in the dataset. A regressor R is added after the discriminator in each first branch network. The discriminator and regressor share a shallow network; that is, the parameters of the first convolutional layer of the regressor and the discriminator are shared. Low-level features of each first branch network can be captured and shared by the shallow network. The regressor of each first branch network has two outputs: easily extractable component A in the organic phase and difficult-to-extract component B in the aqueous phase. Assume that a total of [number missing] samples are extracted from the historical rare earth extraction dataset. The input features, i.e., the input dimensions of the rare earth extraction historical dataset, are: The specific formula of the analog multi-input multi-output model and the multi-branch generative adversarial network is as follows.

[0025]

[0026] wherein the production process characteristic parameters in the extraction process are taken as the conditional variables, and the conditional variables are represented as The conditional variables are spliced with noise to form as the input of the multi-branch generative adversarial network, ; the dependent variables are the real data of , wherein represent the actual values of the component contents of the first-stage to the stage extraction; denotes the easy-to-extract component A in the organic phase output by the branch network j; denotes the difficult-to-extract component B in the aqueous phase output by the branch network j. denotes the jth-stage rare earth extraction simulation process, .

[0027] wherein the specific formula of the main network and the generator G is as follows:

[0028] wherein represents the dimensional input of the main network, wherein there are ; represents the feature transmission process from the input layer to the first hidden layer in the main network; represents the feature transmission process from the first hidden layer to the second hidden layer in the main network; wherein represents the nonlinear activation function of the first hidden layer; is the bias vector of the first hidden layer; represents the weight parameter matrix of the first hidden layer; represents the nonlinear activation function of the second hidden layer; is the bias vector of the second hidden layer; represents the weight parameter matrix of the second hidden layer; is the output of the second hidden layer, represents the number of hidden layer neurons; represents the activation function of the generator in the first branch network ; ​​​​​​and represents the first branch network generator in the jth branch network. In order to prevent the drift of features, the activation function of the hidden layer and the generator is selected as , assuming that the weighted input of feature transmission is , the activation function is , wherein represents .

[0029] wherein, the specific formula of the discriminator and the regressor in the first branch network j is as follows:

[0030]

[0031] wherein, represents the one-dimensional sequence of the generated data of the generator in the first branch network j and the corresponding real data , represents the actual value of the component content of the jth stage in the extraction process; , , , , respectively represent the weight matrix of the first layer to the fifth layer of the discriminator ; represents the mutual information lower bound value between the generated data of the generator in the branch j network and the input feature variable ; , , , , respectively represent the bias of the first layer to the fifth layer of the discriminator ; RELU() and LeakyRELU() represent the activation function; BatchNorm() represents the batch normalization operation; TANH() represents the tanh activation function; , , respectively represent the weight matrix of the first layer and the third layer of the regressor ; , , respectively represent the bias of the first layer to the third layer of the regressor .

[0032] In another example embodiment of the present application, in step 102, considering that the original feature parameters in the original rare earth extraction history data set have different dimensions and different numerical values, which is not conducive to the application of the stochastic gradient descent algorithm in the multi-branch generative adversarial network training. In order to improve the accuracy of the multi-branch generative adversarial network and the convergence speed of the target loss optimization function during training, it is necessary to eliminate the limitation of different dimensions and also to remove the sample data that does not conform to the actual situation after data cleaning and other methods on the original data set. Among them, the production process feature parameters in the original rare earth extraction history data set are standardized to make the original data set conform to the standard normal distribution, that is, the mean is 0. According to The standardization of the features in the original data set is as follows: is the production process feature parameter in the original rare earth extraction history data set, and are the mean and variance of the production process feature parameter , respectively, is the standardized production process feature parameter. Therefore, in step 102, the multi-branch generative adversarial network is trained using the rare earth extraction history data set to obtain a rare earth extraction simulation model, including: (a1) Data cleaning and standardization preprocessing are performed on the rare earth extraction history data set to obtain a preprocessed rare earth extraction history data set.

[0033] (a2) The multi-branch generative adversarial network is trained using the preprocessed rare earth extraction history data set to obtain a rare earth extraction simulation model.

[0034] In another example embodiment of the present application, in step 102, during the training of the multi-branch generative adversarial network, the back propagation technique and the stochastic gradient descent algorithm are used to iteratively update based on the target loss function defined by each first branch network. The training continues until the target loss function of each branch meets the preset convergence standard, at which time it is considered that the training is completed, and the optimized multi-branch generative adversarial network is obtained. In order to avoid the potential overfitting risk during training and the low training efficiency problem caused by the fluctuation of the target loss function learning curve, the present application reduces the initial learning rate by a predetermined proportion after a certain number of training iterations. The termination condition of training is set as reaching the preset maximum number of iterations or the target loss function of each branch converging to a specified threshold. Therefore, in step 102, the multi-branch generative adversarial network is trained using the rare earth extraction history data set to obtain a rare earth extraction simulation model, including: (b1) Construct a training set and a test set according to the rare earth extraction history data set, and set an initial learning rate.

[0035] (b2) inputting the training set into the multi-branch generative adversarial network to output simulation results of component contents of corresponding separation stages in the extraction process.

[0036] (b3) calculating loss function gradients of each hidden layer and the corresponding first branch network according to simulation results of component contents of each separation stage and actual values of component contents of the corresponding separation stage.

[0037] (b4) adjusting model parameters of each hidden layer and the corresponding first branch network according to the loss function gradients.

[0038] (b5) determining whether the current iteration number reaches a preset iteration number, if yes, reducing the initial learning rate by a preset proportion to obtain a reduced learning rate.

[0039] (b6) replacing the initial learning rate with the reduced learning rate, and returning to the step of inputting the training set into the multi-branch generative adversarial network until each loss function gradient converges, to obtain the trained multi-branch generative adversarial network, i.e. the rare earth extraction simulation model.

[0040] Before training the multi-branch generative adversarial network, a target optimization function of the multi-branch generative adversarial network is determined, which is specifically as follows:

[0041] wherein, represents expectation; represents a weight parameter of a discriminator ; represents a weight parameter of a generator and the jth hidden layer; is input data of the generative adversarial network; is a real label value, i.e. an actual value of component content of each stage; is a generated sample generated by the generator ; and respectively represent output scores of the discriminator on the generated sample and the real sample. wherein, represents a new value point generated by interpolating sampling between a real data distribution and a generated data distribution in the following manner, . . is a norm of a gradient of the discriminator at ; represents a data distribution composed of noise z and a condition variable C, z is random Gaussian noise; the hyperparameter This represents the weights of the constraints imposed on the generator G by the balanced discriminator D and the regressor R. This indicates that when its input is a real sample At that time, the predicted value made by the regression model R; Regressor Network weight parameters; This indicates that when the input to the regressor R is the generated data... At that time, the variational method is used to estimate With input feature variables Mutual information between The lower bound of mutual information is determined by minimizing... The negative log-likelihood is obtained, and the specific process is as follows: formula.

[0042]

[0043] in, Indicates the total number of samples; This represents the input feature variable (production process feature parameter) corresponding to the h-th sample.

[0044] The specific formulas for the loss optimization functions of each first branch network are as follows.

[0045]

[0046]

[0047] in, This represents the loss optimization function of the discriminator D. This represents the loss optimization function of the discriminator G; This represents the actual value of the component content in stage j of the extraction process. ; Indicates the first The predicted component content of the h-th sample output by the first branch network; and These represent the discriminator generating the corresponding samples. and real samples The output score.

[0048] Training is performed using backpropagation and stochastic gradient descent algorithms. During backpropagation, the neural network models of the discriminator D and generator G are trained and parameterized. According to the chain rule The gradient can be propagated from the output layer to the input layer layer by layer through a chain-like computation, and correspondingly there are... Indicates the first The cumulative weight matrix and bias vector of the generator and its corresponding hidden layer in the first branch network. Indicates the first The weight matrix and bias vector accumulated from the convolutional and fully connected layers of the discriminator in the first branch network. Therefore, the loss functions of the discriminator D and the generator G can be calculated. and Relative to parameters and parameters gradient value and Finally, gradient descent is used to update the value, and the formula for calculating the gradient is shown below.

[0049]

[0050] in, and Let represent the gradient of the loss function of the discriminator D and the gradient of the loss function of the generator G, respectively. and The learning rate; These represent the weight parameters of the discriminator; The weight parameters represent the generators (including the hidden layers of the main network and the generators of the first branch network); and This represents the gradient.

[0051] Maximize the first branch network through supervised learning. The loss function of the discriminator Minimize the first branch network The loss function of the generator The gradient descent algorithm is used to update the weight parameters of the discriminator. and generator weight parameters When the first The weight matrix of the hidden layer corresponding to the first branch network and the weight matrix of the first branch network. When the bias vectors of the hidden layers corresponding to the first branch networks all converge to the set threshold, the th... The first branch of the network stops training. (Judgment) Is it greater than or equal to? ,if Greater than or equal to This indicates that the target loss of each first branch network has converged, and the trained multi-branch generative adversarial network is output; if Less than Then let Continue training.

[0052] In another exemplary embodiment of this application, after performing the step "training a multi-branch generative adversarial network using a rare earth extraction historical dataset to obtain a rare earth extraction simulation model", the method for constructing the rare earth extraction simulation model further includes: The rare earth extraction simulation model is simulated by using the test set to obtain test evaluation indexes. When the simulation accuracy reaches the set accuracy, the simulation is stopped, and the final multi-branch generative adversarial network is obtained.

[0053] In the rare earth extraction process, there are complex physical and chemical changes, complex mechanisms, multiple stages, and strong coupling between stages. Some mechanism processes have not been clearly defined so far. For data-driven modeling, the existing shallow neural network modeling method is a black box model, and the prediction result is output in a single dimension at the final output layer of the network. For complex industrial scenes with varying working conditions, the rare earth extraction industry has a phenomenon of data scarcity, which leads to the inability of neural networks to handle actual complex nonlinear relationships due to sample scarcity, insufficient learning ability, and inability to accurately describe the multilevel output of component content. The present application proposes a multi-branch generative adversarial network to solve the problem of insufficient prediction accuracy of existing neural network modeling methods. It considers the different mapping relationships between the component content of each stage of the extraction tank and the input features, as well as the actual characteristics that the transfer exchange and cascade effect of components will cause coupling between adjacent extraction tanks. It also considers the problem of insufficient learning ability of shallow neural networks due to the scarcity of on-site data, which leads to complex nonlinear relationships. Through the first branch network connected in series in each hidden layer, the different stages of the extraction tank are jointly modeled to construct a multi-branch generative adversarial network, which includes a main network (including multiple series-connected hidden layers) and multiple first branch networks (each branch network includes a generator, a discriminator, and a regressor). Then, a target loss optimization function is established for each first branch network and the corresponding hidden layer. Finally, the backpropagation technique and the stochastic gradient descent algorithm are used to iteratively update the target loss function defined for each first branch network in the model to obtain the trained multi-branch generative adversarial network, which ultimately makes the simulated component content of each stage closer to the actual value in the extraction production site, and the model process accurately reflects the hierarchical output of component content in actual rare earth production.

[0054] In this application, the extraction process has a total of stages, and the multi-branch generative adversarial network has A first branch network is considered. The actual characteristics of the adjacent extraction tanks caused by the different mapping relationships of the component content and the input characteristics of each stage of the extraction tank, the transfer exchange and cascade effect of the components, and the insufficient learning ability caused by the lack of field data. According to the order from the first stage to the last stage, a first branch network is introduced from every other hidden layer in the main network, and a shallow feature sharing mechanism (i.e., sharing the first layer of convolutional network parameters) between the discriminator and the regressor is proposed, which improves the performance of the regressor and reduces the redundancy of the model. The hidden layer corresponding to each first branch network constitutes a sub-adversarial model, which prevents the generator from receiving fixed feedback from the discriminator, so that the generator uses its deep feature learning ability to generate training samples containing feature information of each branch step by step. The independent regressors of each branch can predict the component content, which conforms to the complexity distribution relationship of the extraction separation reaction. The target loss optimization function is established for each branch to train the multi-branch generative adversarial network. The output of the generative adversarial network is the organic phase easy-extraction component content and the aqueous phase difficult-extraction component content corresponding to each branch. In the training process of the generative adversarial network, the generator can generate new generated samples through adversarial training based on the originally scarce and insufficient original data set to achieve data expansion, which is more reasonable than using only the scarce and insufficient original data set in the training process of other networks, and solves the problem of insufficient learning ability caused by the lack of field data.

[0055] An example is given below to verify the effectiveness of the method of the present application. The example is the separation of four-component elements of lanthanum, cerium, praseodymium and neodymium (La, Ce, Pr, Nd), and the feed amount MF is 1.00 mol. After testing, the molar fractions fa, fb, fc, and fd of each element in the feed liquid are obtained, and the condition is fa+fb+fc+fd=MF. The first stage needs to be considered in the separation process, that is, the cutting component is cerium, and the lanthanum-cerium and praseodymium-neodymium products are obtained at the two ends. As shown in Figure 10 , the separation process has a total of 58 stages, and the 39th stage is the feed stage, and the component distribution of the feed liquid is [0.2543, 0.4912, 0.0486, 0.2012]. In addition, it can be seen that the predicted value of the component distribution basically fits the true value curve, which proves the effectiveness of the method of the present application.

[0056] As shown in Figure 8 and Figure 9 , in the extraction process, the fluctuation and change of the component content of the middle-stage extraction tank are more complex and frequent than those of the extraction tanks at the two ends. The predicted values of the organic phase easy-extraction component A and the aqueous phase difficult-extraction component B of the 25th stage in 200 test samples are compared using the multi-branch generative adversarial network. The multi-branch generative adversarial network has the best fitting degree on the predicted values of the organic phase and the aqueous phase in the test set, which reflects high accuracy. It is shown that the multi-branch generative adversarial network proposed in the present application can learn the complex nonlinear relationship between the working condition parameters of the rare earth extraction process and the component content at each stage with high accuracy.Figure 8 and Figure 9 In this context, TRUE refers to the actual true value of the component content data; BPNN, CGAN, and MBRDN represent BP neural network, conditional generative adversarial network, and multi-branch residual deep network, respectively; MB-RGAN represents the multi-branch generative adversarial network proposed in this application.

[0057] Figure 11 , Figure 12 , Figure 13 The output simulated curves of the component content of each branch of the multi-branch generative adversarial network are shown, corresponding to the feed stage in the example sample. The values ​​are 36, 105, and 142. From Figure 8 , Figure 9 , Figure 10 It can be seen from this that the easily extracted component A is in The content of the organic phase outlet component is close to 1, and the content of the difficult-to-extract component B at the first-stage aqueous phase outlet component is close to 1. This simulates the actual changes in component content in different extraction tanks, and the numerical error is within an acceptable range. The simulation reflects the changes in component content at each stage during the actual extraction process.

[0058] Based on the above, the content of each component obtained by simulating the rare earth extraction process using the method proposed in this application is close to the actual value on the extraction production site. It can truly reflect the rare earth extraction process and enable production technicians to judge the production situation and adjust the production process characteristic parameters in a timely manner.

[0059] In one exemplary embodiment, such as Figure 14 As shown, a device for constructing a rare earth extraction simulation model is proposed, comprising: The sample acquisition module M1 is used to acquire historical datasets of rare earth extraction under different operating conditions; each sample in the historical dataset of rare earth extraction includes historical production process characteristic parameters of rare earth extraction and the actual values ​​of component content of each separation stage.

[0060] The model construction module M2 is configured to train a multi-branch generative adversarial network by using the rare earth extraction historical data set to obtain a rare earth extraction simulation model; the multi-branch generative adversarial network comprises a main network and a plurality of first branch networks; each separation stage in the extraction process corresponds to a first branch network; the main network comprises an input layer and a plurality of hidden layers connected in series; the output of each hidden layer is connected to the input of each first branch network one by one; each first branch network comprises a generator, a discriminator and a regressor; the input of the generator is connected to the output of the corresponding hidden layer, the output of the generator is connected to the input of the discriminator and the input of the regressor respectively, the input of the discriminator further comprises the actual value of the component content of the corresponding separation stage, the output of the discriminator is connected to the input of the generator, and the output of the regressor is the simulation result of the component content of the corresponding separation stage in the extraction process; during the training, the input of the regressor further comprises the actual value of the component content of the corresponding separation stage.

[0061] In another exemplary embodiment of the present application, an application method of a rare earth extraction simulation model is provided, as shown in Figure 15 The application method comprises the following steps: Step 201: obtaining the characteristic parameters of a rare earth extraction production process to be simulated.

[0062] Step 202: inputting the characteristic parameters of the rare earth extraction production process to be simulated into the rare earth extraction simulation model to obtain the simulation result of the component content of each stage in the rare earth extraction process; the rare earth extraction simulation model is constructed based on the construction method of the rare earth extraction simulation model described above; the rare earth extraction simulation model comprises the main network and the second branch network in the multi-branch generative adversarial network; the second branch network comprises the generator and the regressor of the first branch network in the multi-branch generative adversarial network.

[0063] Based on the same inventive concept, the present application also provides an application device of a rare earth extraction simulation model for implementing the application method of the rare earth extraction simulation model. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme described in the above method, and therefore the specific limitations in one or more rare earth extraction simulation model application device embodiments provided below can be referred to the limitations of the rare earth extraction simulation model application method described above, which will not be described herein again.

[0064] In an exemplary embodiment, as shown in Figure 16 The application device of the rare earth extraction simulation model comprises the following modules: The parameter acquisition module T1 is configured to obtain the characteristic parameters of a rare earth extraction production process to be simulated.

[0065] The rare earth extraction simulation module T2 is used for inputting the characteristic parameters of the rare earth extraction production process to be simulated into a rare earth extraction simulation model to obtain simulation results of the component content at each stage in the rare earth extraction process. The rare earth extraction simulation model is constructed based on the construction method of the rare earth extraction simulation model described above. The rare earth extraction simulation model includes a main network and a second branch network in the multi-branch generative adversarial network. The second branch network includes a generator and a regressor of the first branch network in the multi-branch generative adversarial network.

[0066] The application further provides an application scenario of the application method of the rare earth extraction simulation model. Specifically, the application method of the rare earth extraction simulation model provided in this embodiment can be applied in a rare earth extraction simulation scenario. The scenario includes a data acquisition link, a rare earth extraction simulation link and a rare earth extraction process control link. The data acquisition link is used for acquiring the production process characteristic parameters of the rare earth extraction production. The rare earth extraction simulation link is used for rare earth extraction simulation according to the acquired production process characteristic parameters. The rare earth extraction process control link is used for adjusting the production process characteristic parameters in the rare earth extraction production process according to the rare earth extraction simulation results. The application method of the rare earth extraction simulation model provided in this embodiment belongs to the rare earth extraction simulation link.

[0067] In an exemplary embodiment, a computer device, which can be a server or a terminal, has an internal structure as shown in Figure 17 The computer device (i.e. computer equipment) includes a processor, a memory, an input / output interface (I / O) and a communication interface. The processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store video tag processing data. The input / output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with external terminals through network connection. The computer program is executed by the processor to implement a construction method of a rare earth extraction simulation model or an application method of a rare earth extraction simulation model.

[0068] Those skilled in the art can understand, Figure 17The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components. In an exemplary embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, and the processor implementing the steps in the above method embodiments when executing the computer program.

[0069] In an exemplary embodiment, a computer readable storage medium is provided, storing a computer program, and the computer program is executed by a processor to implement the steps in the above method embodiments.

[0070] It should be noted that the data involved in the present application (including but not limited to data for analysis, stored data, displayed data, etc.) are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data need to comply with relevant regulations.

[0071] The technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above embodiments are described, but as long as the combinations of the technical features do not exist contradictory, they should be considered as the scope of the present application.

[0072] The principles and implementation modes of the present application are described by applying specific examples herein, and the above embodiment descriptions are only used to help understand the method of the present application and its core idea; at the same time, for those skilled in the art, according to the idea of the present application, the specific implementation mode and application range will also be changed. In conclusion, the content of the present application should not be understood as a limitation.

Claims

1. A method for constructing a rare earth extraction simulation model, characterized in that, The method for constructing the rare earth extraction simulation model includes: Obtain historical datasets of rare earth extraction under different operating conditions; each sample in the historical dataset of rare earth extraction includes historical production process characteristic parameters of rare earth extraction and the actual values ​​of component content of each separation stage. A multi-branch generative adversarial network (GAN) is trained using a historical dataset of rare earth extraction to obtain a rare earth extraction simulation model. The GAN includes a main network and multiple first branch networks. Each separation stage in the extraction process corresponds to a first branch network. The main network includes a series of input layers and multiple hidden layers. The output of each hidden layer is connected to the input of each first branch network. Each first branch network includes a generator, a discriminator, and a regressor. The input of the generator is connected to the output of the corresponding hidden layer. The input of the regressor is connected to the output of the generator, and the output of the regressor is the simulated result of the component content of the corresponding separation stage in the extraction process. During training, the input of the regressor also includes the actual value of the component content of the corresponding separation stage.

2. The method for constructing a rare earth extraction simulation model according to claim 1, characterized in that, The output of the generator is also connected to the input of the discriminator, whose input includes the actual values ​​of the component content corresponding to the separation stage. The output of the discriminator is connected to the input of the generator.

3. The method for constructing a rare earth extraction simulation model according to claim 1, characterized in that, A multi-branch generative adversarial network was trained using a historical dataset of rare earth extraction to obtain a rare earth extraction simulation model, including: Training and test sets were constructed based on historical datasets of rare earth extraction, and an initial learning rate was set. The training set is input into a multi-branch generative adversarial network, which outputs the simulated results of the component content at the corresponding separation stage during the extraction process; The gradient of the loss function of each hidden layer and the corresponding first branch network is calculated based on the simulation results of the component content of each separation stage and the actual values ​​of the component content of the corresponding separation stage. The model parameters of each hidden layer and the corresponding first branch network are adjusted by backpropagation based on the gradient of the loss function. Determine whether the current iteration count has reached the preset iteration count. If so, reduce the initial learning rate according to the preset ratio to obtain the reduced learning rate. The initial learning rate is replaced with the reduced learning rate, and the process returns to the step "input the training set into the multi-branch generative adversarial network" until the gradient of each loss function converges or the maximum number of iterations is reached, thus obtaining the trained multi-branch generative adversarial network, i.e., the rare earth extraction simulation model.

4. The method for constructing a rare earth extraction simulation model according to claim 3, characterized in that, After performing the step "training a multi-branch generative adversarial network using a historical dataset of rare earth extraction to obtain a rare earth extraction simulation model", the method for constructing the rare earth extraction simulation model further includes: The rare earth extraction simulation model was tested using a test set, and test evaluation indicators were obtained.

5. The method for constructing a rare earth extraction simulation model according to claim 1, characterized in that, A multi-branch generative adversarial network was trained using a historical dataset of rare earth extraction to obtain a rare earth extraction simulation model, including: Data cleaning and standardization preprocessing were performed on the historical dataset of rare earth extraction to obtain the preprocessed historical dataset of rare earth extraction. A multi-branch generative adversarial network was trained using a preprocessed historical dataset of rare earth extraction to obtain a rare earth extraction simulation model.

6. A method for applying a rare earth extraction simulation model, characterized in that, The application methods of the rare earth extraction simulation model include: Obtain the characteristic parameters of the rare earth extraction production process to be simulated; The characteristic parameters of the rare earth extraction production process to be simulated are input into the rare earth extraction simulation model to obtain the simulation results of the content of each component in the rare earth extraction process; the rare earth extraction simulation model is constructed based on the construction method of the rare earth extraction simulation model according to any one of claims 1 to 5; the rare earth extraction simulation model includes a main network and a second branch network in a multi-branch generative adversarial network; the second branch network includes the generator and regressor of the first branch network in the multi-branch generative adversarial network.

7. A device for constructing a rare earth extraction simulation model, characterized in that, The apparatus for constructing the rare earth extraction simulation model includes: The sample acquisition module is used to acquire historical datasets of rare earth extraction under different operating conditions; each sample in the historical dataset of rare earth extraction includes historical production process characteristic parameters of rare earth extraction and the actual values ​​of component content of each separation stage. The model building module is used to train a multi-branch generative adversarial network (GAN) using a historical dataset of rare earth extraction to obtain a rare earth extraction simulation model. The multi-branch GAN includes a main network and multiple first branch networks. Each separation stage in the extraction process corresponds to a first branch network. The main network includes a series of input layers and multiple hidden layers. The output of each hidden layer is connected to the input of each first branch network. Each first branch network includes a generator, a discriminator, and a regressor. The input of the generator is connected to the output of the corresponding hidden layer. The input of the regressor is connected to the output of the generator. The output of the regressor is the simulated result of the component content of the corresponding separation stage in the extraction process. During training, the input of the regressor also includes the actual value of the component content of the corresponding separation stage.

8. An application device for a rare earth extraction simulation model, characterized in that, The application device for the rare earth extraction simulation model includes: The parameter acquisition module is used to acquire the characteristic parameters of the rare earth extraction production process to be simulated. A rare earth extraction simulation module is used to input the characteristic parameters of the rare earth extraction production process to be simulated into a rare earth extraction simulation model to obtain the simulation results of the content of each component in the rare earth extraction process; the rare earth extraction simulation model is constructed based on the construction method of the rare earth extraction simulation model according to any one of claims 1 to 5; the rare earth extraction simulation model includes a main network and a second branch network in a multi-branch generative adversarial network; the second branch network includes the generator and regressor of the first branch network in the multi-branch generative adversarial network.

9. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the method for constructing a rare earth extraction simulation model according to any one of claims 1-5, or the method for applying the rare earth extraction simulation model according to claim 6.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the method for constructing the rare earth extraction simulation model according to any one of claims 1-5, or the method for applying the rare earth extraction simulation model according to claim 6.

Images