Method for soft measurement of internal magnetic field of aluminum electrolysis cell driven by combination of moduli and moduli

Abstract

The invention discloses an aluminum electrolysis cell internal magnetic field soft measurement method driven by modulus combination. The method comprises the following steps: S1, constructing a model driving simulation model of an internal magnetic field of a target aluminum electrolysis cell; s2, constructing an aluminum electrolysis cell internal magnetic field data driving model: based on the model driving simulation model of the target aluminum electrolysis cell internal magnetic field, obtaining a simulation data set, training an MLP network model by using the simulation data set, and predicting the magnetic field intensity of each point in the space through the trained MLP network model; s3, after model training is completed, performing model accuracy proofreading and evaluation on the test set according to prediction data and actual measurement data of the soft measurement model; and S4, inputting the actual spatial characteristics and the anode current I of the aluminum electrolysis cell into the trained MLP network model to obtain the corresponding predicted magnetic field intensity.

Description

Technical Field

[0001] The invention relates to the technical field of aluminum electrolysis process, and in particular to a soft measurement method for the internal magnetic field of an aluminum electrolysis cell driven by analog and digital combination. Background Art

[0002] In the aluminum electrolysis process, the measurement of the magnetic field distribution in the cell is extremely critical to the entire aluminum production process. With the rapid development of the aluminum electrolysis industry and the expansion of the scale of aluminum electrolysis production, the physical field of the aluminum electrolysis cell has an increasingly significant impact on the electrolysis process. In the process of aluminum electrolysis, the magnetic field intensity caused by high current is huge, and the electromagnetic force generated by the uneven distribution of the magnetic field has a profound impact on the flow of electrolyte and aluminum liquid, which directly affects the current efficiency of the electrolysis process, the service life of the electrolysis cell and other key parameters. Accurate magnetic field measurement can not only greatly improve production efficiency and optimize energy consumption, but also significantly improve product quality.

[0003] By accurately measuring and adjusting the magnetic field distribution in the electrolytic cell, the uniformity of current and heat distribution can be improved, thereby enhancing the consistency and efficiency of the electrolytic reaction, reducing energy consumption, and increasing aluminum production. In addition, proper adjustment of the magnetic field helps reduce the uneven heating of aluminum liquid, improves the physical and chemical properties of aluminum, and thus improves the consistency and quality of the product.

[0004] In summary, the measurement and control of the magnetic field inside the aluminum reduction cell is one of the key technologies to improve aluminum production efficiency and product quality.

[0005] In recent years, soft sensing technology has gradually become an important tool for industrial physical quantity measurement. Soft sensing refers to methods and algorithms that use existing measurement data and knowledge to estimate or predict certain physical quantities or product quality in industrial processes. Soft sensing models can generally be divided into model-driven and data-driven.

[0006] (1) Model-driven model: Also known as white box model or first principle model, it is a modeling method based on the physical and chemical mechanisms that describe industrial processes. It can reflect the process mechanism, does not require a large amount of process data, and has the advantages of high reliability and strong interpretability. Generally, for objects or processes with clear basic laws, the mechanism soft measurement model can accurately describe the relationship between the two. With the development of computer simulation software, many scholars have used finite element simulation technology to simulate the magnetic field distribution inside the aluminum electrolytic cell and establish an aluminum electrolytic cell model-driven model. Finite element simulation is an effective method for establishing an aluminum electrolytic cell model-driven model, but the solution process is complex, the development is difficult, and the calculation cost is high. The soft measurement method established by finite element simulation cannot complete real-time prediction.

[0007] (2) Data-driven model: also known as black box model. It is based on the empirical observation of the process through statistical or regression methods and does not reflect any information about the internal mechanism of the process. Due to its simple mechanism and strong universality, the data-driven model has shown certain advantages in the prediction of the magnetic field of aluminum electrolytic cells in recent years. A large amount of magnetic field data is collected from the actual operating aluminum electrolytic cells, including the current density distribution and electrolyte flow conditions under different operating conditions. Then, machine learning or deep learning methods are used to process the information of spatial and temporal series, and magnetic field prediction is performed after the network training is completed. However, data-driven models usually require a large amount of labeled data for training in order to learn and generalize patterns. It is difficult to collect data inside aluminum electrolytic cells, and there are generally small amounts of data and low data quality, which will affect the network training effect and prediction ability. In addition, the prediction results are poorly interpretable, and it is impossible to understand the physical rationality and causal relationship of the results, which limits the application of this method in engineering practice.

[0008] At present, in the field of soft measurement of magnetic field inside aluminum electrolysis cells, existing methods are all single model / data driven models, which usually have problems such as poor real-time performance, poor adaptability, high computational cost, data scarcity, poor interpretability of physical laws, etc., which affect the actual application of engineering sites. The present invention provides a soft measurement method for magnetic field inside aluminum electrolysis cells driven by analog and digital combination, which is used to at least solve the above problems in the prior art and meet the real-time and accuracy application requirements of engineering. Summary of the invention

[0009] In view of this, the present invention provides a soft measurement method for the internal magnetic field of an aluminum electrolysis cell driven by a combination of analog and digital, which is used to at least solve the problems in the prior art that the model-driven method has a complex solution process, great development difficulty, high calculation cost, and poor real-time performance; and the data-driven method has difficulties in obtaining training data, low prediction accuracy, poor generalization ability, and poor physical interpretability.

[0010] In order to achieve the above object, the present invention adopts the following technical solution:

[0011] A soft measurement method for magnetic field inside an aluminum electrolysis cell driven by analog and digital combination comprises the following steps:

[0012] S1. Construct a model-driven simulation model of the internal magnetic field of the target aluminum electrolysis cell;

[0013] S2. Construct a data-driven model of the internal magnetic field of an aluminum reduction cell: obtain a simulation data set based on a model-driven simulation model of the internal magnetic field of a target aluminum reduction cell, use the simulation data set to train an MLP network model, and use the trained MLP network model to predict the magnetic field intensity at each point in space;

[0014] S3. After the model training is completed, the model accuracy is checked and evaluated on the test set based on the predicted data and measured data of the soft measurement model;

[0015] S4. Input the actual spatial features and the anode current I of the aluminum electrolysis cell into the trained MLP network model to obtain the corresponding predicted magnetic field strength.

[0016] Preferably, the specific content of S1 includes:

[0017] S11. Construct a physical model of aluminum electrolysis cell;

[0018] S12. defining the material properties of each part of the aluminum electrolysis cell physical model, and meshing the aluminum electrolysis cell physical model to obtain the spatial coordinates of each point in the aluminum electrolysis cell physical model;

[0019] S13. Based on the material properties of each part of the aluminum electrolysis cell physical model defined in S12, namely, magnetic permeability, resistivity and thermal conductivity, the electric-magnetic field equation and thermal field equation of the aluminum electrolysis cell physical model are constructed;

[0020] S14. Boundary condition definition: Couple the electric-magnetic field equation and the thermal field equation to obtain the electric-magnetic-thermal field, and set the initial boundary conditions for the heat transfer, current, and voltage interfaces;

[0021] S15. Solve and calculate the electromagnetic-magnetic-thermal coupled field until the result converges.

[0022] Preferably, the electric field equation inside the electrolytic cell satisfies the three-dimensional conductive Laplace equation and Joule's law:

[0023]

[0024] V=IR

[0025]

[0026] Where: x , y and ρ z is the resistivity of each point in the aluminum reduction cell along the x, y and z axes; V is the potential; I is the current; R is the resistance; J is the current density vector, is the Hamiltonian operator; σ is the conductivity, is the gradient of the electric field strength, J e represents the external current density;

[0027] The thermal field equation is:

[0028]

[0029] Where: x ,y and λ z is the thermal conductivity of each part of the aluminum electrolytic cell along the x, y, and z axes; T is the temperature; Q is the Joule heat generated by the conductive part.

[0030] Preferably, the specific content of S2 includes:

[0031] S21. Construct an MLP network model and generate a sampling data set by simulating the model-driven simulation model;

[0032] S22. Increase the sampling density in the anode-aluminum liquid-electrolyte-cathode region where the electromagnetic field changes dramatically. The sampling data includes the spatial coordinates u(x, y, z) of the sampling point and the magnetic induction intensity B. Generate a simulated data set of the magnetic field inside the aluminum electrolysis cell, and randomly extract data sets from the simulated data set of the magnetic field inside the aluminum electrolysis cell as training sets and test sets.

[0033] S23. Train the MLP network model using the training set; the MLP network model includes: an input layer, a fully connected layer, and an output layer;

[0034] The spatial features u(x, y, z) and the anode current I of the aluminum electrolytic cell are input into the MLP network model through the input layer. In order to eliminate the dimensional differences between different variables and accelerate the convergence of training losses, the input data is normalized to [0, 1] through minimum-maximum normalization:

[0035]

[0036] Among them, x i is the original input data, x norm is the normalized input data, x min and x max are the minimum and maximum values ​​of the original input data;

[0037] The MLP network training uses random numbers in the interval [-1, 1] as the initial weights and sends the normalized data to the input layer;

[0038] The fully connected layer uses ReLu as the activation function;

[0039] The output layer outputs the magnetic induction intensity B (x, y, z) of any coordinate point inside the aluminum electrolysis cell;

[0040] The output value under the current weight and activation function is calculated through forward propagation; the Adam optimizer is used to adjust the attenuation gradient and learning rate, update the weight parameters, and repeat the above training process through back propagation until the loss function is reduced to less than the set threshold ε, and the iteration is stopped.

[0041] Preferably, the specific content of S3 includes:

[0042] Model evaluation and verification is performed using field data measured by sensors or magnetic field data simulated and calculated using the finite element method and magnetic field data predicted by the MLP network: verification is performed using the test set and the results predicted by the MLP network, using the mean absolute error (MAE) and the coefficient of determination (R) 2 To evaluate the performance of the model:

[0043] MAE measures the average absolute difference between predicted and actual values:

[0044]

[0045] Where n is the number of samples, y i is the actual value, is the predicted value;

[0046] R 2 Measures the model's ability to explain variation in the data:

[0047]

[0048] in, is the mean of the actual values, the numerator is the residual sum of squares of the model, and the denominator is the total sum of squares of the actual values.

[0049] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a soft measurement method for the magnetic field inside an aluminum electrolysis cell driven by analog and digital combination, which has the following beneficial effects:

[0050] Aiming at the measurement of the magnetic field inside an aluminum reduction cell, the present invention proposes a soft measurement model of the magnetic field inside an aluminum reduction cell driven by a combination of analog and digital. Compared with the existing method, the soft measurement model combining model drive and data drive is adopted to measure the magnetic field inside the aluminum reduction cell. After the MLP network training is completed, the time taken for predicting the magnetic induction intensity of a certain spatial position is in the millisecond level, which fully meets the real-time prediction requirements in the project and occupies less computing resources. The problems of poor real-time performance, large amount of computing resources and time consumption, poor generalization ability, and the like of traditional model-driven methods and experimental measurement methods are solved. A self-supervised learning network is constructed. For the measurement of the magnetic field inside an aluminum reduction cell, pseudo labels are generated through model-driven simulation data to enhance the performance of the data-driven model. The problems of small amount of magnetic field data inside the reduction cell, poor quality, and difficulty in training resulting in network inability to be trained and poor prediction accuracy are overcome when training the data-driven model.

[0051] It can be seen that this scheme can effectively reduce the amount of calculation for measuring the magnetic field inside the electrolytic cell under complex working conditions and improve the speed and accuracy of measuring the magnetic field inside the aluminum electrolytic cell. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0053] Figure 1 A flow chart of the soft sensing method combining model drive and data drive provided by the present invention;

[0054] Figure 2 The present invention provides a model-driven model simulation modeling flow chart. DETAILED DESCRIPTION

[0055] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0056] The present invention provides Figure 1 The present invention proposes a soft-sensing model of the magnetic field inside an aluminum electrolysis cell driven by analog and digital combination. The model framework specifically includes the following parts:

[0057] 1. Aluminum electrolysis cell internal magnetic field model driving model

[0058] The method of the present invention considers the magnetic field generation in the electrolytic cell under general conditions and makes the following simplifications: 1. Ignore the influence of the busbar on the magnetic field inside the electrolytic cell, and only focus on the magnetic field change in the reaction area inside the electrolytic cell. 2. Ignore the effect of electrolyte flow on the magnetic field.

[0059] like Figure 2 As shown in the flow chart, the present invention comprises the following steps:

[0060] S1. Using the finite element simulation method, COMSOL software is used to construct a model-driven simulation model of the internal magnetic field of the target aluminum electrolytic cell according to the production process parameters of the aluminum electrolytic cell and historical prior knowledge;

[0061] S2. Build a data-driven model of the magnetic field inside the aluminum electrolysis cell: Build a self-supervised learning network that can generate pseudo labels from unlabeled data for model training. For the magnetic field measurement inside the aluminum electrolysis cell, pseudo labels can be generated from finite element simulation data to enhance the performance of the data-driven model.

[0062] In order to improve the prediction accuracy of the model for key areas, a non-uniform sampling strategy can be introduced in the simulation data generation and sampling stage. The MLP (multi-layer perceptron) network model is trained using the simulation data set, and the trained MLP network model is used to predict the magnetic field intensity at each point in the space;

[0063] S3. After the model training is completed, the model accuracy is checked and evaluated on the test set based on the predicted data and simulation data of the soft measurement model.

[0064] S4. Input the actual spatial features and the anode current I of the aluminum electrolysis cell into the trained MLP network model to obtain the corresponding predicted magnetic field strength.

[0065] In order to further implement the above technical solution, the specific contents of S1 include:

[0066] S11. Obtain the production process parameters of the aluminum electrolytic cell, use CAD software to draw, and establish the physical model of the aluminum electrolytic cell. The structure of this model mainly includes: aluminum guide rods, phosphorus pig iron, steel claws, steel claw welding sheets, anode carbon blocks, aluminum liquid, electrolyte, anode feed, cathode carbon blocks, cathode steel rods, carbon inter-paste, tying paste, cathode steel rods, dry anti-seepage materials, calcium carbonate classes, insulation bricks, refractory bricks, tank shells, cradle frames and other geometric bodies. Use the CAD interface provided by COMSOL software to import, optimize the conflict structure of the imported model, and ensure the subsequent meshing and solution accuracy;

[0067] S12. The material library provided by COMSOL contains more than 500 established material properties, including various metals, alloys, minerals, piezoelectric materials, thermal insulation materials, etc. It also supports full control of defining and adding material properties. A variety of materials are used in the modeling of aluminum electrolytic cells. The magnetic permeability of the materials used in the model established by the present invention. Thermal conductivity and resistivity change significantly with temperature. Therefore, it is necessary to define the material properties of each part of the physical model of the aluminum electrolytic cell separately;

[0068] S13. After the material properties are defined, the model needs to be meshed. In COMSOL finite element simulation analysis, meshing is a very critical step. Meshing is a discretization method in simulation calculations. It is used to discretize continuous physical areas into a series of points, lines, surfaces or body elements so that calculations can be performed at these discrete points. The quality of meshing directly affects the simulation results.

[0069] The tetrahedral meshing used in this invention is applicable to almost all geometric bodies, and it is easy to use curvature and close-range meshes in key areas. Set the smoothness to medium, the unit size to 0.003m, the expansion selects smooth transition, uses the pre-processing expansion algorithm, the transition ratio is 0.272, the maximum number of layers is 5, the growth rate is 1.2, and the rest of the dimensions are meshed by default for subsequent finite element simulation calculations;

[0070] S14. All materials in the aluminum electrolytic cell can conduct heat, but the main ones that conduct electricity are the aluminum guide rod, anode carbon block, steel claw, cathode steel rod, electrolyte, aluminum liquid, and cathode carbon block. Based on the magnetic permeability, resistivity, and thermal conductivity of each part of the material in the aluminum electrolytic cell physical model defined by the material properties in S13, the electric-magnetic field equation and thermal field equation of this model are constructed;

[0071] S15. Boundary condition definition: The electric-magnetic field equation and the thermal field equation are coupled to obtain the electric-magnetic-thermal field, and the initial boundary conditions of the corresponding interfaces of heat transfer, current, and voltage are set using the multi-field coupling interface of the COMSOL software;

[0072] S16. Finally, the solver settings can be set by setting different solvers, error factors, number of iterations, etc., and finally performing a steady-state solution to calculate the electromagnetic-magnetic-thermal coupling field until the results converge.

[0073] In order to further implement the above technical solution, the electric-magnetic field equation in S13 is:

[0074] The electric field equation inside the electrolytic cell satisfies the three-dimensional conductive Laplace equation and Joule's law:

[0075]

[0076] V=IR

[0077]

[0078] Where: x , y and ρ z is the resistivity of each part of the aluminum reduction cell along the x, y and z axes; V is the potential; I is the current; R is the resistance; J is the current density vector, σ is the conductivity, is the gradient of the electric field strength, J e represents the external current density;

[0079] The thermal field equation is:

[0080]

[0081] Where: x , y and λ z is the thermal conductivity of each part of the aluminum electrolytic cell along the x, y, and z axes; T is the temperature; Q is the Joule heat generated by the conductive part.

[0082] In order to further implement the above technical solution, the specific contents of S2 include:

[0083] S21. Build a self-supervised learning network that can generate pseudo labels from unlabeled data for model training. For the measurement of the magnetic field inside the aluminum electrolysis cell, since it is impossible to directly measure the magnetic field inside the closed aluminum electrolysis cell, the finite element simulation model built by S1 can be used to sample data and generate pseudo labels using the finite element simulation data to enhance the performance of the data-driven model.

[0084] S22. In order to improve the prediction accuracy of the model for key areas, a non-uniform sampling strategy can be introduced in the simulation data generation and sampling stage. Increase the sampling density in the anode-aluminum liquid-electrolyte-cathode area where the electromagnetic field changes violently, so that the model can obtain more relevant data for learning, and reduce the sampling points in other areas with relatively stable changes to save computing resources and data processing time. The sampling data includes the x, y, z spatial coordinates of the sampling point and the magnetic induction intensity B, and generate a simulation data set of the magnetic field inside the aluminum electrolytic cell, and randomly extract the data set as the training set (80% of the data set) and the test set (20% of the data set). ;

[0085] S23. The MLP network model is trained using the simulation data set. The MLP network is divided into an input layer, a fully connected layer, and an output layer. The model input in the input layer is the spatial feature u(x, y, z) and the anode current I of the aluminum electrolytic cell. In order to eliminate the dimensional differences between different variables and accelerate the convergence of the training loss, the input data is normalized to [0, 1] through minimum-maximum normalization:

[0086]

[0087] Among them, x i is the original input data, x norm is the normalized input data, x min and x max are the minimum and maximum values ​​of the input data.

[0088] The MLP network training uses random numbers in the interval [-1, 1] as the initial weights and sends the normalized data to the input layer.

[0089] The fully connected layer consists of 5 hidden layers, each containing 64 neurons, and ReLu is used as the activation function; the output layer outputs the magnetic induction intensity B(x, y, z) of a certain coordinate point in the internal space of the aluminum electrolytic cell. Then, the output value under the current weight and activation function is calculated through forward propagation; the Adam optimizer is used to adjust the attenuation gradient and learning rate, update the weight parameters, and repeat the above training process through back propagation until the loss function is reduced to less than the set threshold ε, and the iteration is stopped.

[0090] In order to further implement the above technical solution, the specific contents of S3 include:

[0091] After the model training is completed, we need to evaluate and verify the model through field data measured by sensors or magnetic field data simulated by finite element method. The test set (20% of the data set) divided in S22 is used for verification with the results predicted by the MLP network. The mean absolute error MAE and the coefficient of determination R 2 To evaluate the performance of the model.

[0092] MAE is used to measure the average absolute difference between the predicted value and the actual value. The formula is:

[0093]

[0094] Where n is the number of samples, y i is the actual value, is the predicted value.

[0095] R 2 Measures the model's ability to explain data variation. The formula is:

[0096]

[0097] in, is the mean of the actual values, the numerator is the residual sum of squares of the model, and the denominator is the total sum of squares of the actual values.

[0098] In order to further implement the above technical solution, the specific contents of S4 include:

[0099] An online prediction software is built, and the online data and coordinate data are sent to the soft measurement device to generate a real-time predicted magnetic field.

[0100] The input variable I is collected by physical sensors, and x, y, and z are coordinates specified by the user. After the network training is completed, the prediction speed can meet the online real-time prediction requirements. Online prediction software can be built, and the online data and coordinate data are sent to the soft meter to generate a real-time predicted magnetic field.

[0101] The present invention aims at measuring the magnetic field inside the aluminum electrolytic cell, and proposes a soft measurement model of the magnetic field inside the aluminum electrolytic cell driven by a combination of analog and digital. Compared with the existing method, the soft measurement model combining model drive and data drive is used to measure the magnetic field inside the aluminum electrolytic cell, and its prediction speed is very fast, which meets the real-time prediction requirements. It solves the problems of poor real-time performance of traditional model-driven methods, experimental measurement methods, etc., requiring a lot of computing resources and time, difficult installation of internal magnetic field measurement instruments, high wear and tear costs, and inaccurate measurements; it also overcomes the defects of poor measurement accuracy caused by small amount of data, poor quality, and difficulty in training when training the data-driven model.

[0102] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A soft measurement method for the magnetic field inside an aluminum electrolysis cell driven by analog and digital combination, characterized in that: The following steps are involved: S1. Construct a model-driven simulation model of the internal magnetic field of the target aluminum electrolysis cell; S2. Construct a data-driven model of the internal magnetic field of an aluminum reduction cell: obtain a simulation data set based on a model-driven simulation model of the internal magnetic field of a target aluminum reduction cell, use the simulation data set to train an MLP network model, and use the trained MLP network model to predict the magnetic field intensity at each point in space; S3. After the model training is completed, the model accuracy is checked and evaluated on the test set based on the predicted data and measured data of the soft measurement model; S4. Input the actual spatial features and the anode current I of the aluminum electrolysis cell into the trained MLP network model to obtain the corresponding predicted magnetic field strength.

2. The method for soft-sensing the magnetic field inside an aluminum electrolysis cell driven by analog-digital combination according to claim 1 is characterized in that: The specific contents of S1 include: S11. Construct a physical model of aluminum electrolysis cell; S12. defining the material properties of each part of the aluminum electrolysis cell physical model, and meshing the aluminum electrolysis cell physical model to obtain the spatial coordinates of each point in the aluminum electrolysis cell physical model; S13. Based on the material properties of each part of the aluminum electrolysis cell physical model defined in S12, namely, magnetic permeability, resistivity and thermal conductivity, the electric-magnetic field equation and thermal field equation of the aluminum electrolysis cell physical model are constructed; S14. Boundary condition definition: Couple the electric-magnetic field equation and the thermal field equation to obtain the electric-magnetic-thermal field, and set the initial boundary conditions for the heat transfer, current, and voltage interfaces; S15. Solve and calculate the electromagnetic-magnetic-thermal coupled field until the result converges.

3. The method for soft-sensing the magnetic field inside an aluminum electrolysis cell driven by analog-digital combination according to claim 2 is characterized in that: The electric-magnetic field equation in S13 is: The electric field equation inside the electrolytic cell satisfies the three-dimensional conductive Laplace equation and Joule's law: V=IR J=-σ▽E+J e Where: x , y and ρ z is the resistivity of each point in the aluminum reduction cell along the x, y and z axes; V is the potential; I is the current; R is the resistance; J is the current density vector, ▽ is the Hamiltonian operator; σ is the conductivity, ▽E is the gradient of the electric field intensity, J e represents the external current density; The thermal field equation is: Where: x , y and λ z is the thermal conductivity of each part of the aluminum electrolytic cell along the x, y, and z axes; T is the temperature; Q is the Joule heat generated by the conductive part.

4. The method for soft-sensing the magnetic field inside an aluminum electrolysis cell driven by analog-digital combination according to claim 1 is characterized in that: The specific contents of S2 include: S21. Construct an MLP network model and generate a sampling data set by simulating the model-driven simulation model; S22. Increase the sampling density in the anode-aluminum liquid-electrolyte-cathode region where the electromagnetic field changes dramatically. The sampling data includes the spatial coordinates u(x, y, z) of the sampling point and the magnetic induction intensity B. Generate a simulated data set of the magnetic field inside the aluminum electrolysis cell, and randomly extract data sets from the simulated data set of the magnetic field inside the aluminum electrolysis cell as training sets and test sets. S23. Train the MLP network model using the training set; the MLP network model includes: an input layer, a fully connected layer, and an output layer; The spatial features u(x, y, z) and the anode current I of the aluminum electrolytic cell are input into the MLP network model through the input layer. In order to eliminate the dimensional differences between different variables and accelerate the convergence of training losses, the input data is normalized to [0, 1] through minimum-maximum normalization: Among them, x i is the original input data, x norm is the normalized input data, x min and x max are the minimum and maximum values ​​of the original input data; The MLP network training uses random numbers in the interval [-1, 1] as the initial weights and sends the normalized data to the input layer; The fully connected layer uses ReLu as the activation function; The output layer outputs the magnetic induction intensity B (x, y, z) of any coordinate point inside the aluminum electrolysis cell; The output value under the current weight and activation function is calculated through forward propagation; the Adam optimizer is used to adjust the attenuation gradient and learning rate, update the weight parameters, and repeat the above training process through back propagation until the loss function is reduced to less than the set threshold ε, and the iteration is stopped.

5. The method for soft-sensing the magnetic field inside an aluminum electrolysis cell driven by analog-digital combination according to claim 4 is characterized in that: The specific contents of S3 include: Model evaluation and verification is performed using field data measured by sensors or magnetic field data simulated and calculated using the finite element method and magnetic field data predicted by the MLP network: verification is performed using the test set and the results predicted by the MLP network, using the mean absolute error (MAE) and the coefficient of determination (R) 2 To evaluate the performance of the model: MAE measures the average absolute difference between predicted and actual values: Where n is the number of samples, y i is the actual value, is the predicted value; R 2 Measures the model's ability to explain variation in the data: in, is the mean of the actual values, the numerator is the residual sum of squares of the model, and the denominator is the total sum of squares of the actual values.

Images