CN115984209B - Rare earth element component content prediction method for concentration and component content synergistic optimization - Google Patents

Rare earth element component content prediction method for concentration and component content synergistic optimization Download PDF

Info

Publication number
CN115984209B
CN115984209B CN202211687100.7A CN202211687100A CN115984209B CN 115984209 B CN115984209 B CN 115984209B CN 202211687100 A CN202211687100 A CN 202211687100A CN 115984209 B CN115984209 B CN 115984209B
Authority
CN
China
Prior art keywords
rare earth
content
optimization
concentration
prediction
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202211687100.7A
Other languages
Chinese (zh)
Other versions
CN115984209A (en
Inventor
张水平
张奇涵
王碧
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Jiangxi University of Science and Technology
Original Assignee
Jiangxi University of Science and Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Jiangxi University of Science and Technology filed Critical Jiangxi University of Science and Technology
Priority to CN202211687100.7A priority Critical patent/CN115984209B/en
Publication of CN115984209A publication Critical patent/CN115984209A/en
Application granted granted Critical
Publication of CN115984209B publication Critical patent/CN115984209B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
    • Y02P10/00—Technologies related to metal processing
    • Y02P10/20—Recycling
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
    • Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
    • Y02P90/30—Computing systems specially adapted for manufacturing

Landscapes

  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The application discloses a method for predicting the content of a rare earth element component by synergistically optimizing the concentration and the component content, which comprises the following steps: acquiring detection points based on a rare earth extraction process; acquiring image data of the rare earth mixed extraction solution based on the detection points; constructing a prediction model, inputting image data of the rare earth mixed extraction solution into the prediction model for training, and generating an optimization model; and predicting the content of the rare earth element component based on the optimization model to generate a prediction result. The application firstly builds the multi-task deep neural network, and improves the generalization capability and robustness of the model. Secondly, a method for predicting the component content and the concentration of each rare earth element in the mixed extraction solution based on a multi-objective optimization algorithm is provided, and the prediction precision of each task is improved by searching pareto optimal. A plurality of groups of comparison experiments show that the method has the best performance when the content of the multi-element components or the content and the concentration of the multi-element components are trained simultaneously, and can meet the accuracy and the instantaneity of the online detection of the content of the rare earth elements.

Description

Rare earth element component content prediction method for concentration and component content synergistic optimization
Technical Field
The application belongs to the field of rare earth content detection, and particularly relates to a method for predicting the content of a rare earth element component by synergistically optimizing the concentration and the component content.
Background
Rare earth becomes a necessary key element of materials such as permanent magnetic materials, catalytic materials, hydrogen storage materials and the like due to the special optical, electric, magnetic and other properties, and the high-performance rare earth materials play a core role in the high and new technical fields such as metallurgy, chemical industry, aerospace, national defense technology and the like, and are widely focused by the national and international scientific circles. China is a large-scale country of rare earth reserves, the industrial scale of rare earth separation and the yield are the first world, however, the mode of obtaining the component content value in the automatic production process of the rare earth separation industry is basically remained in the low-level stage of offline analysis. The rare earth linkage extraction separation process mainly relies on online detection of the content of each element component of rare earth at a monitoring point in a continuous production process, and the material flow of feed liquid, organic phase, washing liquid and the like is optimally controlled by detection data, so that the high purification of separated products is ensured.
Domestic students study the multi-objective optimization problem, zhang Bo and the like decompose the multi-objective optimization problem into a plurality of two-objective optimization problems on the basis of two algorithms of multiple gradient descent and Frank-Wolfe, and perform gradient update and linear search on the gradient descent direction. Zhou Xiaojun and the like, a multi-gradient descent algorithm is used for multi-license plate recognition, and an end-to-end license plate recognition method based on multi-objective optimization multi-task learning is provided; zhao Jiaming and the like are used for identifying working conditions of the froth flotation process, so that a good prediction effect is obtained, but iteration time of the training process is not counted and compared in Zhao and Zhao, and time performance of different algorithms in different task training iteration processes cannot be known.
Disclosure of Invention
The application aims to provide a method for predicting the content of a rare earth element component by synergistically optimizing the concentration and the component content, so as to solve the problems in the prior art.
In order to achieve the above object, the present application provides a method for predicting the content of a rare earth element component by synergistically optimizing the concentration and the component content, comprising:
rare earth is obtained, and extraction is carried out on the rare earth to obtain extraction pure solution;
setting detection points for the extraction pure solution, and obtaining rare earth mixed extraction solution image data;
constructing a prediction model, inputting the image data of the rare earth mixed extraction solution into the prediction model for training, and generating an optimization model;
and predicting the content of the rare earth element component based on the optimization model to generate a prediction result.
Preferably, the process of obtaining an extraction pure solution comprises:
obtaining rare earth, and fusing the rare earth with an organic liquid to generate a fused solution;
setting a separation outlet to generate a related separation unit;
and extracting the fusion solution through the related separation unit to generate an extraction pure solution.
Preferably, the process of acquiring the image data of the rare earth mixed extraction solution comprises the following steps:
setting a solution detection point, detecting the extraction pure solution based on the solution detection point and optical imaging conditions, and obtaining a rare earth mixed extraction solution image;
and classifying the characteristics of the rare earth mixed extraction solution image to obtain the rare earth mixed extraction solution image data.
Preferably, the process of constructing the prediction model includes:
constructing a ResNet18 network, wherein a residual error module in the ResNet18 network consists of a convolution layer, a nonlinear activation function layer and a batch normalization layer;
and connecting a specific task layer in the ResNet18 network with a corresponding full connection layer to generate a prediction model.
Preferably, the process of generating the optimization model includes:
inputting the image data of the rare earth mixed extraction solution to the prediction model for loading to obtain a solution parameter data set;
acquiring a centralized data point based on the solution parameter dataset;
and performing multi-objective optimization on the prediction model based on the concentrated data points to generate an optimization model.
Preferably, the process of generating the optimization model further includes:
initializing network parameters in the prediction model and setting model parameters;
performing multi-objective optimization on the upper boundary of the objective, and obtaining a specific task gradient;
and solving a multi-objective optimization solution based on the specific task gradient, and performing multi-objective optimization on the prediction model based on the multi-objective optimization solution to generate an optimization model.
Preferably, the network parameters comprise real label values corresponding to the image samples, shared layer network parameters and specific task layer network parameters;
the model parameters include a loss function, a maximum number of iterations, a batch size, an optimizer parameter, and a learning rate.
Preferably, the process of performing multi-objective optimization on the prediction model based on the multi-objective optimization solution includes:
acquiring stationary points for predicting a plurality of tasks based on the specific task gradient;
updating the shared layer network parameters and the task layer network parameters based on the stable points to generate optimized parameters;
and optimizing the descending direction of the specific task based on the optimization parameters to finish multi-objective optimization.
Preferably, the process of generating the prediction result includes:
acquiring shared layer network parameters and specific layer network parameters in an optimization model;
and predicting the content and the concentration of components in the rare earth mixed extraction solution image to be detected based on the shared layer network parameters and the specific layer network parameters to generate a prediction result.
The application has the technical effects that:
1) The application proves that the multi-task learning can find the commonality among a plurality of tasks, and improves the overall generalization force and robustness of the model;
2) The iterative time length is close to the single task learning and the homodyne uncertainty with the lowest time consumption in the training process, is far lower than that of a multiple gradient descent method, does not increase along with the increase of the number of tasks, and has the advantage of short training time consumption in a multiple task optimization method;
3) The overall performance of the method is inferior to the multiple gradient reduction in the double-task learning; the overall performance in three-task learning is inferior to that of single-task learning; the overall performance in four-task learning is superior to that of other optimization methods, and the method is more suitable for simultaneous training and prediction of component content and concentration compared with other multi-task optimization methods. Therefore, compared with other multitask optimization methods, the method is more suitable for predicting the content and concentration of the rare earth multielement component;
4) The application proves that the component content prediction precision of the double-task learning (Pr and Nd component content prediction) is slightly higher than that of the four-task learning (Pr and Nd component content, pr and Nd concentration prediction), and the component content and concentration simultaneous training can not obviously improve the component content prediction precision;
5) The maximum relative error of the component content prediction in each task is within +/-5% of the maximum relative error of the component content detection in the rare earth extraction production, the prediction time cost is within 3s, the requirement of on-line soft measurement of the multi-element component content in the rare earth linkage extraction process is met, and a new thought is provided for the rare earth element component content soft measurement method.
Drawings
The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and together with the description serve to explain the application. In the drawings:
FIG. 1 is a schematic diagram of a soft measurement method of the rare earth component content in an embodiment of the present application;
FIG. 2 is a flow chart of La/Ce/Pr/Nd four-element linkage extraction separation in an embodiment of the application;
FIG. 3 is a diagram showing a rare earth multi-element component content and concentration prediction model based on multi-task learning in an embodiment of the application;
FIG. 4 is a flow chart of the Frank-Wolfe algorithm in an embodiment of the present application.
Detailed Description
It should be noted that, without conflict, the embodiments of the present application and features of the embodiments may be combined with each other. The application will be described in detail below with reference to the drawings in connection with embodiments.
It should be noted that the steps illustrated in the flowcharts of the figures may be performed in a computer system such as a set of computer executable instructions, and that although a logical order is illustrated in the flowcharts, in some cases the steps illustrated or described may be performed in an order other than that illustrated herein.
Example 1
As shown in fig. 1, in this embodiment, a method for predicting the content of a rare earth element component by synergistically optimizing the concentration and the content of the component is provided, which includes:
in the rare earth linkage extraction and separation process, in order to reduce consumption of materials (organic phase, washing liquid and the like) required by the extraction process, related separation units in the extraction and separation process are associated in a transverse and longitudinal connection mode, and materials are provided among the units. The correlation mode ensures that each material in the whole rare earth separation process can be provided with each outlet for simultaneously producing the pure product of the required extraction element only by one input port. The linkage extraction and separation process of the La/Ce/Pr/Nd four rare earth elements is shown in figure 2, and six separation units are provided: six units of (LaCePr)/(CePrNd), (LaCe)/(CePr), (CePr)/(PrNd), la/Ce, ce/Pr and Pr/Nd are mutually connected horizontally and longitudinally, and mixed rare earth feed liquid, organic phase and washing liquid which need extraction and separation are respectively input through the three units of (La Ce Pr)/(Ce Pr Nd), la/Ce and Pr/Nd, so that pure solution after extraction of four elements can be obtained from the outlets of all elements.
In order to optimize the feeding process of the whole extraction flow, the flow rate of each feed liquid input by the three feeding units is calculated through detection values of the content of each element component in the extraction liquid by each monitoring point. And because part of rare earth elements (such as Pr and Nd) have unique ion color characteristics, the extraction solution has the characteristics of certain reflectivity and refraction to light, has good light transmittance, can present color characteristics in a visible light region, and meets the optical imaging condition of soft measurement of the content of the rare earth elements based on a machine vision technology. Therefore, detection points can be arranged at the positions of the contents of the rare earth elements to be detected in the rare earth linkage extraction separation process, and images of the rare earth mixed extraction solution are collected and used for model training and prediction. The soft measurement data of the rare earth element component content is calculated by a theoretical formula, so that the theoretical minimum extraction amount and minimum washing amount required by each separation unit and the optimized flow required by the linkage connection between the transverse and longitudinal sides of the separation units can be obtained, thereby optimizing and controlling the extraction flow, and further ensuring the high purification of the separation products.
In order to meet the requirement of rare earth linkage extraction on multi-element component content detection and explore commonalities between multi-element component contents and between component contents and concentrations, the embodiment provides a multi-task learning-based rare earth multi-element component content and concentration prediction model, which is used for simultaneously training the rare earth multi-element component contents and concentrations in an acquired rare earth mixed extraction solution image in a shared layer network and setting a plurality of task-specific layer networks to simultaneously output the prediction results of a plurality of tasks of component contents and concentrations. The structure of the multi-task learning model is shown in fig. 3, and the shared layer network in the model is composed of a ResNet18 network with an end full connection layer removed; and the network head of each specific task layer is connected with the shared layer network to realize parameter sharing, and a full-connection layer is respectively arranged for outputting a specific task predicted value.
The model is built on a hard parameter sharing structure based on multi-image classification tasks, and ResNet18 with good fitting effect on the data set prediction result is selected as a sharing backbone network in a comparison experiment through simultaneous training of single rare earth element and concentration. The residual error module in the ResNet18 network consists of a convolution layer, a nonlinear activation function layer and a batch normalization layer, the batch normalization (Batch Normalization) layer replaces the dropout layer, the output of the hidden layer of the network is subjected to standardization processing, the original data is maintained, the numerical value is more stable, and the overfitting is prevented. And the characteristic shorting operation of the residual error module enables the model to learn the characteristics of the deeper network. In particular, in order to obtain single accurate predicted values of the content and concentration of the rare earth element component, in the embodiment, the output of the full-connection layer in the forward propagation process in the specific task layer network is connected with a Softmax function, a regression model is built for the output of each class probability of the specific task and each class true value through a linear regression loss function by utilizing the Softmax function, then the loss function value is propagated in the opposite direction, and network parameters are optimized. The Loss functions commonly used in the linear regression model are L1Loss and L2Loss, and an average absolute error MAE (formula 1) and a mean square error MSE (formula 2) are respectively calculated;
wherein y is i As a result of the model predictive value,and (3) the real label value of the task corresponding to the predicted data point, wherein n is the number of the predicted data points. The L2Loss function squares the error compared to L1Loss, which can exacerbate ignoring minor errors. For example, the real element component content of a sample of a mixed extraction solution image is 0.001, and the predicted value is 0.011. Although the error is 10 times the true value, the value is smaller. If L2Loss is used, the difficulty of modifying the error is further increased. Therefore, the embodiment adopts L1los as the Loss function of each task of the model.
In the training process of the multi-task learning model for predicting the content and the concentration of the rare earth multi-element components, a conflict can occur among a plurality of specific tasks, so that the network parameters of the sharing layer are biased to the condition of one specific task. In the embodiment, after being inspired by the tuner, the situation is attempted to be solved through a multi-objective optimization algorithm, a common training process of a plurality of specific tasks is regarded as a multi-objective optimization process, competition among the tasks is balanced by finding Pareto solutions among the plurality of specific tasks, and the network parameters of a sharing layer cannot deviate to a specific task, so that the accuracy of predicting the content and concentration of the rare earth multi-element components is improved.
The input space in the multitask learning problem of the predication of the content and concentration of the rare earth multielement components is the acquired rare earth mixed extraction solution image,for the corresponding component content and concentration real value space in a group of rare earth mixed extraction solution images, the data point in the whole rare earth mixed extraction solution image data set is +.>Wherein T is the number of tasks of the content and concentration of the rare earth element components to be predicted, N is the number of image data points of the rare earth mixed extraction solution, and +.>And labeling the actual value of the ith data point corresponding to the content or concentration prediction task of the ith component. Still further consider the parameterized assumption for each specific task as +.>Wherein θ is sh To share layer network parameters, θ t Network parameters are task-specific. The empirical minimization formula for all the predictive tasks of the rare earth element component content or concentration can be expressed as formula (3):
wherein c t The calculated weights are measured for the specific tasks;the empirical loss function for the t-th prediction task can be specifically defined as:
assume that there are two sets of parametric solutions θ andat t for two specific tasks 1 And t 2 The following problems may occur under the conditions of (a): />And-> I.e. the parameter solution θ is more biased towards task t 1 But->More prefers task t 2 . To solve this problem, the present embodiment regards the multiple tasks together learning as a multi-objective optimization problem, finding Pareto solutions that balance between the multiple tasks so that the network parameters are not biased towards any particular task. The above problem can be translated into a solution to the following parameters:
equation (5) consists of a vector L containing rare earth element component content and concentration to predict loss functions corresponding to a plurality of tasks, with the objective of seeking to balance Pareto solutions between the plurality of tasks. The definition of Pareto solutions between tasks in a multi-objective optimization is given below:
definition 1 has all tasks tAnd is also provided withThen the parameter solution θ dominates +.>
Definition 2 if there is no parameter solution θ dominates θ * Then we call the parameter solution θ * Is the Pareto optimal solution.
The embodiment provides a multi-gradient descent algorithm based on an optimization upper bound, which is used for searching a Pareto optimal solution in the multi-task learning of the rare earth element component content and concentration prediction, and the multi-objective optimization is completed through a gradient descent mode in the multi-task learning model training. The following multi-objective optimization will be achieved using the Karush-Kuhn-Tucker (KKT) condition, which is also a necessary condition for multi-objective optimization. The KKT conditions required for sharing layer network parameters and specific task layer network parameters in the rare earth multi-element component content and concentration prediction model based on multitask learning are now expressed as follows:
1) Presence of alpha 1 ,…,α T Not less than 0, and hasAnd->
2) For the prediction task t of the content and concentration of any rare earth element component,
any solution that satisfies the above condition is called Pareto plateau, where the Loss function used by the model of this embodiment in completing the above optimization condition is L1Loss, which is mentioned in section 2.2. Consider the following optimization problem:
desidri et al show that: when the solution of the optimization problem is 0, the solution of the optimization problem meets the KKT condition; otherwise, the solution gives a descending direction, and all prediction tasks of the content and the concentration of the rare earth element component are optimized. However, the algorithm described by this problem requires computationThe back propagation is needed for the shared layer network parameter associated with each specific task, so that the gradient calculation of each specific task needs to be obtained in T times of back propagation and then forward propagation, and the consumption of memory and time is greatly increased.
In this embodiment, a multiple gradient descent algorithm based on an optimization upper bound is provided, which is different from the multiple gradient descent algorithm, optimizes the target upper bound, and can obtain gradients of all specific tasks in forward propagation by only one-pass back propagation, thereby saving memory and time consumption. Optimizing the upper bound requires combining the shared representation function with the decision function for the particular task, and the hypothesis class constraint can be defined as:
where g is a representation function shared by all tasks, f t To take the representation as a specific task function of the input, if the representation function is represented as z= (Z) 1 ,…,Z N ) Wherein Z is i ＝g(x i ；θ sh ) The upper bound can be expressed as the following formula, which is a direct result of the chain law:
wherein the method comprises the steps ofZ is equivalent to theta sh Jacobian matrix norms of (a). Two desirable properties of this upper bound are:
1)gradients for all specific tasks can be calculated in one back propagation;
2)not one about alpha 1 ,…,α T And therefore can be removed when it is the objective of optimization.
To obtain an approximate optimal solution, the two desirable properties are considered, one of the two desirable properties in equation (6)Replaced by the upper bound in formula (8) and delete +.>The term, then, converts the multitasking optimization into solving the following equation:
the algorithm for solving the above-mentioned algorithm is called a multiple gradient descent algorithm based on the optimization upper bound, and Sener proposes that whenIs full rank and alpha 1 ,…,α T In the case of the solution of formula (9), one of the following two conditions can be satisfied:
1)the current multitask learning model parameter is Pareto stable;
2)is the direction of descent of all target tasks
The algorithm finds a Pareto plateau and the computational overhead is negligible.
Equation (9) belongs to a convex quadratic problem with linear constraints, solving the optimization problem is equivalent to finding the minimum norm point in the convex hull of the input point set. We consider first the case of solving based on two tasks, the optimization problem can be defined as:
this is a one-dimensional quadratic function about α with analytical solution, as we will followSimplified as θ, will->Simplified to->Then the equation (10) is derived and the following equation (11) is solved.
Although the formula (11) is only suitable for solving two tasks, the present embodiment uses the Frank-Wolfe algorithm proposed by Jaggi, etc. to take the formula (10) as a sub-process of linear search, so as to solve the constraint optimization problem of two or more tasks. The Frank-Wolfe algorithm is shown in fig. 4. Alpha is obtained 1 ,…,α T The method is a solution of a multi-gradient descent algorithm based on an optimization upper bound, and the descent direction of all specific tasks can be optimized through the solution to obtain the Pareto optimal of the rare earth multi-element component content and concentration prediction multi-objective.
In the process of optimizing a plurality of targets for predicting the content and the concentration of the rare earth multielement components, the embodiment adopts a multiple gradient descent algorithm based on an optimization upper bound to solve the gradient of each task in model training, and the algorithm optimizes the upper bound of the targets, so that the gradient of all specific tasks can be obtained in the forward propagation process only through one-time back propagation. And substituting the gradient of each specific task into the Frank-Wolfe algorithm to obtain a multi-objective optimization solution, so as to obtain Pareto stable points for predicting a plurality of tasks with respect to the content and the concentration of the rare earth multi-element components, and realize multi-objective optimization. The detailed steps are described as follows:
input: and (3) inputting the standardized pre-processed rare earth mixed extraction solution image into a rare earth multi-element component content and concentration prediction model based on multi-task learning, and loading an image sample to correspond to a real label value. Initializing shared layer network parameters θ sh Specific task layer network parameter θ t The method comprises the steps of carrying out a first treatment on the surface of the And setting a loss function, the maximum iteration number, the batch size, an optimizer and the learning rate eta, and solving a multi-objective optimization problem in the shared layer network.
And (3) outputting: and outputting the predicted value of the content or concentration of the rare earth element component in each specific task layer.
Step 1. Initializing a multi-objective optimization upper bound according to equation (9)
Step 2, obtaining the gradient of each specific task under the optimization upper bound
Step 3, substituting gradients of all specific tasks into the Frank-Wolfe algorithm to obtain a multi-objective optimization solution, and obtaining a Pareto stable point alpha of predicting a plurality of tasks with respect to the content and concentration of the rare earth multi-element components 1 ,…,α T ；
Step 4. Utilize alpha 1 ,…,α T Updating shared layer network parameters
Step 5, updating the network parameters of the specific task layerOptimizing the descending directions of all specific tasks;
step 6, judging whether the maximum iteration times are reached, if yes, entering step 7, otherwise, returning to step 1;
step 7, according to the optimized shared layer network parameter theta sh And layer-specific network parameters θ t And (T epsilon T) predicting the content and concentration of the components in the rare earth mixed extraction solution image to be detected.
In order to embody the effectiveness of the multi-objective optimization algorithm provided by the embodiment on the predication of the content and the concentration of the rare earth multi-element components, the multi-objective optimization algorithm is prepared from the following steps of: 1) Verifying the loss function value change of the set and the average time length of single iteration of the training process; 2) And comparing the single task learning with different multi-task optimization methods at the two angles of error evaluation index values of the 10 test set sample predicted values and the true values.
In the embodiment, for the rare earth multi-element component content and concentration prediction based on a multi-objective optimization algorithm, two rare earth extraction solutions with ionic color characteristics, namely Pr (apple green) and Nd (mauve), are taken as examples, simulation experiments are carried out, and the Pr/Nd mixed extraction solution is obtained through the following ways:
measuring the component content of the raw solution: 1L1.8275mol/LPrCL with 99.9% purity was purchased from Gannan rare earth company 3 And 1L2.063mol/LNdCl 3 The rare earth concentration and the distribution of the extracted pure solution are provided by the national tungsten and rare earth product quality supervision and inspection center.
Solution dilution: the concentrations of the two element raw solutions are respectively diluted to 11 rare earth extraction pure solutions with different concentrations of 0.01mol/L to 0.50mol/L and good light transmittance.
Mixing the solution: 50ml of each solution of two rare earth elements with different concentrations is mixed with each other to obtain 121 groups of Pr/Nd mixed extraction solutions with different component contents and concentrations.
The 121-group rare earth mixed extraction solution with the Pr/Nd element component content changed from 1.96% -98.04% and the concentration changed from 0.005mol/L to 0.25mol/L is obtained by the steps, and the rare earth mixed extraction solution with the component content and the concentration different from each other is poured into a collecting dish for sealing and preservation. As the rare earth mixed extraction solution prepared in the laboratory has the practical characteristics of certain reflectivity and refraction to light, the light transmittance is better, and the optical imaging condition based on soft measurement of the rare earth element component content in the machine vision experiment is satisfied. Meanwhile, the ion color characteristic band formed in the container filled with the rare earth mixed solution provides a feasible path for adopting a rapid, accurate and continuously-detectable image recognition technology.
And pouring the rare earth mixed extraction solution into a quartz container with the length and width of 150 multiplied by 5 multiplied by 170mm until the container is filled with the solution during experimental image acquisition. And then the container was placed in a 60CM studio. Two LED light sources are built in a studio, the output voltage of the light sources is 24V, the power is 48W, and the maximum lumen of lamp beads is 15000LM; the background is pure white; the image acquisition equipment is a NIKON D700 camera; the final acquired image is a JPG format picture with 4256 multiplied by 2832 resolution.
Because the Pr/Nd mixed extraction solution image obtained by shooting contains a non-solution part outside the edge of the quartz container and a non-uniform color part inside the edge, the part filled with the color features in the shot image is cut, 10 pictures are cut according to the sequence from top to bottom and from left to right in each group of solution images, and 1210 rare earth mixed extraction solution images with uniform colors are obtained. Classifying the rare earth mixed extraction solution image into 121 categories according to the difference of the component content and the concentration of each element of Pr and Nd, respectively marking the real label values of the component content and the concentration of Pr and Nd on each category, dividing 70% of the rare earth mixed extraction solution image into a training set, 20% into a verification set and 10% into a prediction set, and preparing a complete data set for constructing a rare earth multi-element component content and concentration prediction model based on multi-task learning.
All model training and prediction experiments were carried out in the following experimental environment: the hardware environment is Windows10 operating system, CPU Intel Core i7-12700F (12 cores), GPU RTX30 (memory is 10G); the software environment is the PyTorch deep learning framework used on PyChram.
In order to verify the effectiveness of the rare earth multi-element component content and concentration prediction method based on the multi-objective optimization algorithm provided by the embodiment, the multi-objective optimization algorithm (multi-gradient descent based on the optimization upper bound) used by the embodiment is combined with single-task learning, multi-gradient descent and other multi-task optimization main stream methods: 1) Gradient normalization; 2) From the variance uncertainty, the angle of the mean duration of a single iteration of the training process from the loss function value change of the validation set is compared across a plurality of predictive tasks:
1. double-task learning: predicting Pr and Nd component content;
2. three-task learning: predicting Pr and Nd component content and Pr concentration;
3. four tasks are learned: predicting Pr and Nd component content and Pr and Nd concentration;
in order to ensure fairness of comparison implementation, all optimization methods of the embodiment are implemented in a rare earth multielement component content and concentration prediction model based on multitask learning, and unified main parameters of the model are set as shown in table 1, and the main parameters are general parameters for improving overall performance of the model. Thus, this is fair to all optimization methods in the comparative experiments.
TABLE 1
The verification set loss function value of Nd component content prediction tasks in double, triple and quadruple task learning can be seen to be the lowest; in the three-task learning, the Pr component content prediction task and the four-task learning, the Pr and Nd concentration prediction task are only inferior to the single-task learning with the lowest loss function value; in other multitask learning, the loss function values of Pr component, pr and Nd concentration prediction tasks are all close to the minimum value and are superior to most of optimization methods. The multiple gradient descent also presents better performance in all prediction tasks, and the loss function value is generally lower than that of other mainstream multi-objective optimization methods. In particular, gradient normalization has higher validation set loss values for Pr, nd concentration predictions in each multitask study, possibly due to the different magnitudes of the Pr, nd component content and concentration prediction tasks.
TABLE 2
The average time length of single iteration in the multi-task training of each optimization method is compared, wherein the average time length of single iteration in the single-task learning training is shown in table 2, and the average time length of single iteration in the multi-task learning training of different multi-task optimization methods is shown in table 3.
The average time length of single iteration of each task in table 2 is different from that of each method in table 3 in training all tasks at the same time, and the training is obtained by independently training each task through single task learning.
TABLE 3 Table 3
From tables 2-3, it can be known that the same-variance uncertainty training time is shortest in the multitask optimization method, and is close to the single-task learning training time; the consumption reduction under multiple gradients is longest, and the training time can be increased along with the increase of the number of tasks; the training time length of the method is only longer than that of the homodyne uncertainty and the single-task learning, and is better than that of other multi-objective optimization methods.
In summary, the loss function value of the method in the embodiment is generally lower in multi-task learning, and the loss value in the individual prediction task is only higher than the lowest value of the current task and is basically superior to other multi-task optimization methods; the iteration time length of the method in the training process is close to the single task learning with the lowest time consumption and the uncertainty of the same variance, and is far lower than that of a multiple gradient descent method, and the training time length cannot be increased along with the increase of the number of tasks; therefore, the multi-gradient descent algorithm based on the optimization upper bound has the great advantage of more heavy gradient descent algorithms in training time, and reduces the time consumption of training and the loss function value of each task.
In order to avoid the situation that the rare earth multielement component content and concentration prediction model based on multitask learning is trained and fitted, the section carries out simulation prediction on a test set sample, compares and analyzes errors of a predicted value and a true value, and judges whether the model is within the maximum fault tolerance range of the component content prediction error in the rare earth mixed extraction image in the rare earth linkage extraction process. The multi-objective optimization algorithm (multi-gradient descent based on the optimization upper bound) used in this embodiment is combined with one-task learning, multi-gradient descent and other multi-task optimization mainstream methods as follows: 1) Gradient normalization; 2) The same variance uncertainty, the error evaluation index value angles from the 10 test set sample predictions and the true value are compared on the following prediction tasks:
the relative error of each optimization method in the task of predicting the Pr and Nd component content is not great, but the relative error in the task of predicting the Pr and Nd concentration is great, the error fluctuation range of gradient normalization is great, and the phenomenon is consistent with the higher performance of gradient normalization in the loss function value of the verification set.
TABLE 4 Table 4
TABLE 5
TABLE 6
TABLE 7
From tables 4-7, the following 6 points can be seen:
1) In the two-task learning, the method of the present embodiment is compared with the one-task learning: in the method, the average relative error, the root mean square error, the maximum relative error absolute value and the Nd component maximum relative error absolute value of the Pr component prediction task are lower than those of single task learning, and the average relative error, the root mean square error and the single task learning of the Nd component prediction task are close to each other, so that the whole method is superior to the single task learning; compared with other multitasking optimization methods: in the method, the average relative error and the root mean square error of the Pr component content prediction task are only inferior to the multiple gradient decrease, the maximum relative error of the Pr component prediction task is the lowest in all optimization methods, and the maximum relative error of the Nd component prediction task is only lower than gradient normalization. Therefore, the overall performance of the method of the present embodiment is inferior to the multiple gradient drop in the double-task learning.
2) In three-task learning, the average relative error, root mean square error and maximum relative error absolute value of the Pr concentration prediction task are the lowest in all optimization methods, and the method is compared with the single-task learning: the error evaluation indexes except the lowest error evaluation indexes are higher than the single-task learning, and the whole error evaluation indexes are inferior to the single-task learning; comparing the method of the embodiment with other multitasking optimization methods: in the method of the embodiment, except the lowest, other error evaluation indexes are higher than gradient normalization, but the gradient normalization has higher error in Pr concentration prediction, and compared with other methods except the gradient normalization, the method of the embodiment has better overall performance. Therefore, the gradient normalization method is not suitable for simultaneous training and prediction of components and concentrations, and considering the situation, the overall performance of the method in the embodiment is inferior to that of single-task learning in three-task learning.
3) In four-task learning, the average relative error, root mean square error, maximum absolute value of relative error and average relative error in Pr component prediction task of Nd component and Nd concentration prediction task of the method of the embodiment are lower than those of other optimization methods, and the overall performance is better; comparing the text method with the single task learning: except the lowest error evaluation index values, the rest error evaluation index values are higher than those of single-task learning, but the overall performance is better than that of the single-task learning; compared with other multi-task optimization methods, the average relative error and root mean square error of the Pr concentration prediction task are lower than those of other optimization methods except the lowest. Therefore, in the four-task learning, the overall performance of the method of the embodiment is superior to that of other optimization methods, and the method is more suitable for simultaneous training and prediction of components and concentrations compared with other multi-task optimization methods.
4) Comparing the single-task learning with the multi-task learning: the single task learning only predicts that each index of the task is lower in comparison with the four task learning, and the other indexes are higher in comparison with all the multi-task learning, so that the multi-task learning can find the correlation between the components and the concentration tasks and jointly improve the overall performance of the model compared with the single task learning.
5) Comparing the multitask learning, firstly, comparing the Pr and Nd component prediction task error evaluation index values in the two, three and four task learning, and finding that the prediction task errors of the Pr and Nd components in the three task learning are higher than those of the two and four task learning, which indicates that only the simultaneous training and prediction of the element components or the simultaneous training and prediction of the element components and the concentration are effective in improving the prediction precision of the component content; secondly, comparing Pr concentration prediction tasks in three-task and four-task learning, the error evaluation index value of the Pr concentration prediction task in three-task learning can be found to be generally lower than that of four-task learning, so that simultaneous training and prediction of element components and concentration two tasks only can improve the precision of component prediction, but cannot improve concentration pre-prediction
The present application is not limited to the above-mentioned embodiments, and any changes or substitutions that can be easily understood by those skilled in the art within the technical scope of the present application are intended to be included in the scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims (3)

1. A method for predicting the content of a rare earth element component by synergistically optimizing the concentration and the component content is characterized by comprising the following steps:
acquiring detection points based on a rare earth extraction process;
acquiring image data of the rare earth mixed extraction solution based on the detection point;
constructing a prediction model, inputting the image data of the rare earth mixed extraction solution into the prediction model for training, and generating an optimization model;
predicting the content of the rare earth element component based on the optimization model to generate a prediction result;
the process for acquiring the image data of the rare earth mixed extraction solution comprises the following steps:
detecting the extraction pure solution based on the detection point and the optical imaging condition to obtain a rare earth mixed extraction solution image;
performing feature classification on the rare earth mixed extraction solution image to obtain rare earth mixed extraction solution image data;
the process for constructing the prediction model comprises the following steps:
constructing a ResNet18 network, wherein a residual error module in the ResNet18 network consists of a convolution layer, a nonlinear activation function layer and a batch normalization layer;
connecting a specific task layer in the ResNet18 network with a corresponding full-connection layer to generate a prediction model;
the process for generating the optimization model comprises the following steps:
inputting the image data of the rare earth mixed extraction solution into the prediction model for loading to obtain a solution parameter data set;
acquiring a centralized data point based on the solution parameter dataset;
performing multi-objective optimization on the prediction model based on the concentrated data points to generate an optimization model;
initializing network parameters in the prediction model and setting model parameters;
performing multi-objective optimization on the upper boundary of the objective, and obtaining the gradient of a specific task layer;
obtaining a multi-objective optimization solution based on the gradient brought into a Frank-Wolfe algorithm of the specific task layer, and carrying out multi-objective optimization on the prediction model based on the multi-objective optimization solution to generate an optimization model;
the process of performing multi-objective optimization on the prediction model based on the multi-objective optimization solution comprises the following steps:
acquiring the content and concentration of the rare earth multi-element components based on the gradient of the specific task layer, and predicting Pareto stable points of a plurality of tasks;
updating the network parameters of the shared layer and the network parameters of the specific task layer based on the stable points to generate optimized parameters;
optimizing the descending direction of a specific task layer based on the optimization parameters to finish multi-objective optimization;
the process for generating the prediction result comprises the following steps:
acquiring network parameters of a sharing layer and network parameters of a specific task layer in an optimization model;
and predicting the content and concentration of components in the rare earth mixed extraction solution image to be detected based on the network parameters of the shared layer and the network parameters of the specific task layer, and generating a prediction result.
2. The method for predicting the content of a rare earth element component by synergistically optimizing the concentration and the component content according to claim 1, wherein the obtaining a detection point in the rare earth extraction flow comprises:
setting a separation outlet to generate a related separation unit;
and detecting the rare earth extraction flow based on the related separation unit to generate a detection point.
3. The method for predicting the content of a rare earth element component according to claim 1, wherein the concentration and the component content are synergistically optimized,
the network parameters comprise real label values corresponding to the image samples, shared layer network parameters and network parameters of a specific task layer;
the model parameters include a loss function, a maximum number of iterations, a batch size, an optimizer parameter, and a learning rate.
CN202211687100.7A 2022-12-27 2022-12-27 Rare earth element component content prediction method for concentration and component content synergistic optimization Active CN115984209B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202211687100.7A CN115984209B (en) 2022-12-27 2022-12-27 Rare earth element component content prediction method for concentration and component content synergistic optimization

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202211687100.7A CN115984209B (en) 2022-12-27 2022-12-27 Rare earth element component content prediction method for concentration and component content synergistic optimization

Publications (2)

Publication Number Publication Date
CN115984209A CN115984209A (en) 2023-04-18
CN115984209B true CN115984209B (en) 2023-10-24

Family

ID=85969564

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202211687100.7A Active CN115984209B (en) 2022-12-27 2022-12-27 Rare earth element component content prediction method for concentration and component content synergistic optimization

Country Status (1)

Country Link
CN (1) CN115984209B (en)

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110223288A (en) * 2019-06-17 2019-09-10 华东交通大学 A kind of Rare-Earth Extraction Process multicomponent content prediction method and system
CN113538475A (en) * 2021-07-19 2021-10-22 中国科学院自动化研究所 Real-time multi-instrument segmentation method and system based on multi-task algorithm
CN115206455A (en) * 2022-07-28 2022-10-18 江西理工大学 Deep neural network-based rare earth element component content prediction method and system

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110223288A (en) * 2019-06-17 2019-09-10 华东交通大学 A kind of Rare-Earth Extraction Process multicomponent content prediction method and system
CN113538475A (en) * 2021-07-19 2021-10-22 中国科学院自动化研究所 Real-time multi-instrument segmentation method and system based on multi-task algorithm
CN115206455A (en) * 2022-07-28 2022-10-18 江西理工大学 Deep neural network-based rare earth element component content prediction method and system

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
Multi-Task Learning as Multi-Objective Optimization;Ozan Sener 等;《https://arxiv.org/pdf/1810.04650v1.pdf》;1-15 *
基于多任务学习多目标优化的稀土元素组分含量与浓度多维度软测量;张水平 等;《控制理论与应用》;1-13 *

Also Published As

Publication number Publication date
CN115984209A (en) 2023-04-18

Similar Documents

Publication Publication Date Title
CN110717553B (en) Traffic contraband identification method based on self-attenuation weight and multiple local constraints
Sommer et al. A deep learning and novelty detection framework for rapid phenotyping in high-content screening
CN112508104A (en) Cross-task image classification method based on rapid network architecture search
CN114299324B (en) Pathological image classification method and system based on multiscale domain countermeasure network
CN114660040A (en) Method, device, medium and equipment for identifying microbial unicell species
CN114121163A (en) Culture medium prediction system based on ensemble learning, training and culture medium prediction method
US20150242676A1 (en) Method for the Supervised Classification of Cells Included in Microscopy Images
Xiao et al. DEEPACC: automate chromosome classification based on metaphase images using deep learning framework fused with priori knowledge
CN115984209B (en) Rare earth element component content prediction method for concentration and component content synergistic optimization
CN113724195B (en) Quantitative analysis model and establishment method of protein based on immunofluorescence image
Yildiz et al. Nuclei segmentation in colon histology images by using the deep CNNs: a U-net based multi-class segmentation analysis
CN105740884A (en) Hyper-spectral image classification method based on singular value decomposition and neighborhood space information
Liu et al. Focusformer: Focusing on what we need via architecture sampler
Çakmak Grapevine Leaves Classification Using Transfer Learning and Fine Tuning
Zhang et al. Learning to search efficient densenet with layer-wise pruning
US11543357B2 (en) Metal sorting system using laser induced breakdown spectroscopy and operating method thereof
Xie et al. Sensitive Feature Selection for Industrial Flotation Process Soft Sensor based on Multi swarm PSO with Collaborative Search
CN112257787B (en) Image semi-supervised classification method based on generation type dual-condition confrontation network structure
CN115457337A (en) Image classification method containing fine-grained noise based on label distribution learning
CN113723536A (en) Power inspection target identification method and system
Pommé et al. NetPrune: a sparklines visualization for network pruning
Sun et al. Impact of various DIII-D diagnostics on the accuracy of neural network surrogates for kinetic EFIT reconstructions
Emami et al. Condensed-gradient boosting
Yaary et al. Comparing machine learning models for tau triggers
Jin et al. Semi-discrete Optimal Transport for Long-tailed Classification

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant