Electric induction furnace electrical parameter identification method and device based on hierarchical physical information learning model, medium and product

Abstract

The invention discloses an electric induction furnace electrical parameter identification method based on a layered physical information learning model. The method comprises the following steps: firstly, establishing a series of linear physical models at a plurality of working points of an electric induction furnace for guiding data driving models of the corresponding working points to jointly form a bottom layer model so as to comprehensively capture data characteristics and electrical dynamics of equipment; secondly, in upper-layer modeling, global nonlinear constraint based on physical priori knowledge is introduced, and the accuracy and physical consistency of the model are guaranteed; and finally, converting a layered modeling problem into a multi-task learning optimization problem with global nonlinear constraints, and realizing electric parameter identification of the electric induction furnace by utilizing a shared layer-task layer cascade model architecture. According to the electric parameter identification method for the electric induction furnace based on the hierarchical physical information learning model, the electric parameters capable of effectively reflecting the running state of equipment can be accurately identified according to input and output real-time sampling data in the electric furnace smelting process, and reference and guidance are provided for worker operation and automatic optimization.

Description

Technical Field

[0001] The present invention relates to a method for identifying electrical parameters of an induction furnace, and more particularly to a method, device, medium and product for identifying electrical parameters of an induction furnace based on a hierarchical physics-informed learning model. Background Art

[0002] The development and progress of the automation control technology of induction furnaces have solved the problems of a large number of manual operations and large operation fluctuations in the traditional smelting process, making the material intake and energy consumption in the smelting process more accurate, and the quality of the molten steel has been greatly improved, achieving good economic and social benefits. With the continuous upgrading of industrial intelligence, the manufacturing system is developing in the direction of low cost, high efficiency and digitization. However, there are some challenges in the intelligent upgrading of induction melting: the process mechanism of induction furnaces is extremely complex, with severe coupling of multiple fields and multiple phases, making it difficult to achieve accurate modeling; induction furnaces have strong non-linear time-varying characteristics, and the operating conditions are dynamically variable, without a full-process direct on-line detection method. This leads to difficulties in perceiving the state of molten steel and equipment, resulting in obvious "black box" characteristics. At the same time, a large amount of valuable data generated during the smelting process has not been effectively utilized, which contains the operating characteristics, health status, etc. of induction furnaces. The working state of an induction furnace depends to a large extent on its electrical dynamics at each working point. Therefore, it is necessary to design a method that can automatically identify electrical parameters.

[0003] Patent CN107340040B discloses an on-line weighing method for hot metal in an induction furnace based on a distributed parameter model, which uses the input and output sampling data of the furnace to identify electrical parameters by the least square method. However, this method is highly dependent on data quality and recursion step size, and has poor robustness and accuracy. In addition, this method does not fully consider the physical consistency problem, and abnormal prediction results may occur in practical applications, weakening its reliability. Summary of the Invention

[0004] To solve the problems and requirements in the background art, the present invention provides a method for identifying electrical parameters of an induction furnace based on a hierarchical physics-informed learning model. A linear physical model is established for each working point at the bottom layer to guide the data-driven model, and a global non-linear constraint based on physical prior knowledge is designed at the upper layer to ensure the accuracy and physical consistency of the model. The input and output sampling data of the induction furnace are used to automatically identify the electrical parameters of the system at each moment.

[0005] The technical solution of the present invention is as follows:

[0006] A method for identifying electrical parameters of an induction furnace based on a hierarchical physics-informed learning model, comprising the following steps:

[0007] S1: Establish a series of linear physical models at multiple working points of the induction furnace to guide the data-driven models at the corresponding working points, and jointly form the underlying models;

[0008] S2: In the upper-layer modeling, introduce global non-linear constraints based on physical prior knowledge to ensure the accuracy and physical consistency of the model;

[0009] S3: Transform the hierarchical modeling problem based on a series of underlying models and upper-layer models into a multi-task learning optimization problem with global non-linear constraints, and use the shared layer-task layer model architecture to realize the identification of the electrical parameters of the induction furnace.

[0010] In the above S1, the physical equation of the system established at the working point i of the induction furnace is as follows:

[0011]

[0012] The above differential equation is simplified to the following form:

[0013] f i = Y d,i P i

[0014]

[0015] P i = [1 P 1,i P 2,i P 3,i P 4,i P 5,i P 6,i P 7,i T

[0016]

[0017] Among them, C 1 and C 2 are the parallel and series capacitors respectively, L 0 is the leakage inductance, M i is the mutual inductance, R M is the iron loss resistance, R 0 is the copper loss resistance, R i is the melting resistance, and I i are the voltage and current sampled at the working point i respectively, t is the melting time, f i is the linear physical control equation, Y d,i is the vector composed of variables and their derivative terms, and P i is the electrical parameter vector.

[0018] ​The formula for the underlying data-driven model of the system established at the working point i of the induction furnace is as follows:

[0019]

[0020] Among them, is the model input, is the solution of the model, and the data operator is used to track the observed data at the current working point Ω is the solution domain. The calculation formula for the k-th hidden layer of the data-driven model is as follows:

[0021] z k = σ(W k z k-1 + b k ), k ∈ {1, …, N L}

[0022] Among them, z k represents the output of the hidden layer, z k-1 is the input of the hidden layer, W k is the weight matrix, b k is the bias vector, σ is the non-linear activation function, and N L is the total number of hidden layers.

[0023] The formula for the underlying linear physical model of the system established at the working point i of the induction furnace is as follows:

[0024]

[0025] Among them, the linear physical operator represents the electrical dynamics of the induction furnace at a specific working point i, and it follows the linear physical control equation f i , λ i is the true electrical parameter (unknown), is the estimated value of λ i , is the vector composed of the solution of the data model and its differential, and its form is the same as Y d,i , is the estimated value of the electrical parameter vector P i .

[0026] In S2, according to engineering experience, there is a certain proportional relationship between the mutual inductance M, the melting resistance R, the liquid level height, the furnace lining radius, and the resonance frequency f 0 . In the same melting batch, the furnace lining loss can be ignored, that is, the furnace lining radius is a constant. Therefore, the following proportional relationship can be obtained:

[0027]

[0028] Among them, K is a constant.

[0029] Using this proportional relationship, the formula for the global non-linear constraint based on physical prior knowledge in the upper layer modeling of the induction furnace is as follows:

[0030]

[0031] Among them, is the lining conservation constraint operator for a single furnace campaign, S is the label set of all working points, j represents an element different from i, and both i and j belong to S.

[0032] The specific S3 is as follows:

[0033] First of all, the modeling of each working point can be regarded as multiple related but different tasks. The tasks are encoded according to the resonance frequency of the system at each working point. For multiple tasks The formula for the multi-task learning model is as follows:

[0034]

[0035] Among them, i = 1, 2,..., N t is the task label, θ s is the parameter of the shared layer, θ t,i is the parameter of the task layer, is the electrical parameter to be identified for model parameterization.

[0036] Then, the formula for the loss function of this model is as follows:

[0037]

[0038] Among them, is the average underlying data modeling loss of all tasks, is the average underlying linear physical modeling loss of all tasks, is the upper layer non-linear physical prior knowledge modeling loss.

[0039] Finally, the formulas for the total loss function and the total optimization objective are as follows:

[0040]

[0041] Among them, λ d is the weight of the underlying data modeling loss, λ p is the weight of the underlying linear physical modeling loss, λ k is the weight of the upper layer non-linear physical prior knowledge modeling loss.

[0042] Another technical solution of the present invention is:

[0043] A computer device, the computer device includes a memory and a processor, the memory stores a computer program, and when the processor executes the computer program, the steps of the method are implemented.

[0044] Another technical solution of the present invention is:

[0045] A computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the method are implemented.

[0046] Another technical solution of the present invention is:

[0047] A computer program product, the computer program product includes a computer program / instructions, and when the computer program / instructions are executed by a processor, the steps of the method are implemented.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] (1) The hierarchical physical information learning model designed by the present invention combines the physical model with the neural network at the bottom layer, improving the capture accuracy of the model for data features and electrical dynamics.

[0050] (2) The hierarchical physical information learning model designed by the present invention introduces global non-linear constraints based on physical prior knowledge at the upper layer, ensuring the accuracy and physical consistency of the model.

[0051] (3) The hierarchical physical information learning model designed by the present invention can automatically identify the electrical parameters of the system according to the input information, which is simple and effective, requires fewer parameters to be adjusted, and is suitable for practical engineering applications. Description of the Drawings

[0052] Figure 1 It is the induction furnace circuit structure targeted by the present invention;

[0053] Figure 2 It is the structure diagram of the hierarchical physical information learning model proposed by the method of the present invention;

[0054] Figure 3 It is the loss change curve of the training and testing of the hierarchical physical information learning model proposed by the method of the present invention in the embodiment;

[0055] Figure 4 It is the prediction accuracy diagram of the hierarchical physical information learning model proposed by the method of the present invention in the embodiment;

[0056] Figure 5 It is the voltage and current prediction result diagram corresponding to some tasks of the hierarchical physical information learning model proposed by the method of the present invention in the embodiment;

[0057] Figure 6The electrical parameter identification result diagram of the hierarchical physical information learning model proposed by the method of the present invention in the embodiment. Detailed implementation manners

[0058] To enable those of ordinary skill in the art to more clearly understand the purpose, technical solution and advantages of the present invention, the present invention will be further described below in conjunction with the accompanying drawings and embodiments, but the present invention is not limited to the following embodiments.

[0059] In this embodiment, an intermediate frequency induction furnace with a melting speed of 35 t / h is selected for research.

[0060] An induction furnace electrical parameter identification method based on a hierarchical physical information learning model of the present invention includes the following steps:

[0061] S1: Establish a series of linear physical models at multiple working points of the induction furnace to guide the data-driven models at the corresponding working points, and jointly form the underlying model;

[0062] Specifically:

[0063] The circuit structure of the induction furnace is as Figure 1 shown. By analyzing the loss characteristics in the furnace loop, according to Kirchhoff's law, the differential equation at the working point i of the induction furnace can be obtained as follows:

[0064]

[0065]

[0066] Among them, C 1 and C 2 are the parallel and series capacitors respectively, L 0 is the leakage inductance, M i is the mutual inductance, R M is the iron loss resistance, R 0 is the copper loss resistance, R i is the smelting resistance, and I i are the voltage and current sampled at the working point i respectively, t is the smelting time, and d represents the differential.

[0067] Furthermore, the differential equation can be simplified to the following form:

[0068] f i = Y d,i P i

[0069]

[0070] P i = [1P 1,i P 2,i P3,i P 4,i P 5,i P 6,i P 7,i T

[0071]

[0072] Among them, f i is a linear physical control equation, Y d,i is a vector composed of variables ( and I i ) and their derivative terms, P i is a vector of electrical parameters, and T is the transpose symbol of the matrix.

[0073] Next, a bottom-layer data model is established to capture the data characteristics in the smelting process of the induction furnace. The formula for the bottom-layer data-driven model of the system established at the working point i of the induction furnace is as follows:

[0074]

[0075] Among them, is the model input, is the solution of the model, and the data operator is used to track the observed data at the current working point Ω is the solution domain. The calculation formula for the k-th hidden layer of the data-driven model is as follows:

[0076] z k = σ(W k z k-1 + b k ), k ∈ {1, …, N L}

[0077] Among them, z k represents the output of the hidden layer, z k-1 is the input of the hidden layer, W k is the weight matrix, b k is the bias vector, σ is the non-linear activation function, and N L is the total number of hidden layers.

[0078] Then, a bottom-layer linear physical model is established to capture the electrical dynamics in the smelting process of the induction furnace. The formula for the bottom-layer linear physical model of the system established at the working point i of the induction furnace is as follows:

[0079]

[0080] Among them, the linear physical operator represents the electrical dynamics of the induction furnace at a specific working point i, and it follows the linear physical control equation f i , λ​i is the true electrical parameter (unknown), is λ i the estimated value of, is the vector formed by the solution of the data model and its differential, with the same form as Y d,i the same, is the electrical parameter vector P i the estimated value of.

[0081] S2: In the upper-layer modeling, introduce global non-linear constraints based on physical prior knowledge to ensure the accuracy and physical consistency of the model;

[0082] Specifically:

[0083] According to engineering experience, there is a certain proportional relationship between the mutual inductance M, the melting resistance R, the liquid level height, the furnace lining radius, and the resonance frequency f 0 There is a certain proportional relationship. In the same melting batch, the furnace lining loss can be ignored, that is, the furnace lining radius is a constant. Therefore, the following proportional relationship can be obtained:

[0084]

[0085] where K is a constant.

[0086] Using this proportional relationship, establish non-linear constraints based on physical prior knowledge in the upper layer, and its description is as follows:

[0087]

[0088] where, is the single-furnace lining conservation constraint operator, S is the label set of all working points. Embed this physical prior knowledge into the neural solver to make the identification result more in line with physical laws.

[0089] S3: Transform the hierarchical modeling problem based on a series of underlying models and upper-layer models into a multi-task learning optimization problem with global non-linear constraints, and use the shared layer-task layer model architecture to realize the identification of the electrical parameters of the induction furnace.

[0090] S3 is specifically as follows:

[0091] First, the modeling of each working point can be regarded as multiple related but different tasks. Encode the tasks according to the resonance frequency of the system at each working point. For multiple tasks The hierarchical physical information learning model architecture proposed by the present invention is as Figure 2 shown, and the formula of the multi-task learning model is as follows:

[0092]

[0093] where i = 1, 2,..., Nt is the task label, θ s is the parameter of the shared layer, θ t,i is the parameter of the task layer, is the electrical parameter to be estimated for model parameterization, is the data-driven model operator, which represents the data model in the form of a function.

[0094] Next, the formula for the loss function of the model is as follows:

[0095]

[0096] Among them, is the average underlying data modeling loss of all tasks, is the average underlying linear physical modeling loss of all tasks, is the upper-layer non-linear physical prior knowledge modeling loss, and n ranges from 1 to N i , which means that each task contains a total of N i sampling points, and i ranges from 1 to N t , which means a total of N t tasks.

[0097] Finally, the formulas for the total loss function and the total optimization objective are as follows:

[0098]

[0099] Among them, λ d is the weight of the underlying data modeling loss, λ p is the weight of the underlying linear physical modeling loss, λ k is the weight of the upper-layer non-linear physical prior knowledge modeling loss.

[0100] Next, the technical solution of the present invention will be elaborated in detail with specific embodiments.

[0101] The loss function proposed by the present invention presents a highly non-convex multi-parameter optimization problem, and the Adam optimizer is used to converge the loss during training. The sampling frequency of the sampling data of the induction furnace operation used in the present invention is 20000Hz, the sampling interval of the working point is 1min, and the number of data samplings at each working point is 1024. The data used for training the hierarchical physical information learning model proposed by the present invention is the data of the first two cycles before each working point, and the data used for testing is the data of the third and fourth cycles of each working point. The architecture of the hierarchical physical information learning model proposed by the present invention is a shared-layer neural network with 4 hidden layers, each hidden layer contains 128 neurons, and N t cascaded task-layer neural networks with 2 hidden layers, each hidden layer contains 128 neurons. Other parameters are shown in Table 1. For this embodiment, C1 = 14148 μF, C 2 = 12576 μF, total inductance L = L 0 + M = 458.5 μH, R 0 = 1 mΩ, R M = 40 Ω, K = 2.398×10 -8 . The loss change curves of the training and testing of the hierarchical physics-informed learning model proposed by the present invention are as shown in Figure 3 . The training and testing results show that as the number of training cycles increases, the loss gradually decreases, indicating that the model has good convergence. The present invention defines the following mean absolute error (MAE) and root mean square error (RMSE) to evaluate the prediction performance of the model, and the formulas are as follows:

[0102]

[0103] where y is the observed value, is the predicted value. The prediction accuracy results of the hierarchical physics-informed learning model proposed by the present invention are as shown in Figure 4 . The evaluation results show that the method proposed by the present invention can effectively capture the data characteristics and electrical dynamics of the induction furnace during the entire melting process. The predicted results of the voltage and current corresponding to some tasks of the hierarchical physics-informed learning model proposed by the present invention are as shown in Figure 5 , Figure 5 (a) of Figure 5 (b) of Figure 5 (c) of Figure 5 (d) of Figure 5 (e) of Figure 5 (f) of Figure 6 are 98 Hz, 100 Hz, 110 Hz, 120 Hz, 130 Hz, and 139 Hz respectively. The prediction results further verify the accuracy and effectiveness of the method proposed by the present invention. The electrical parameter identification results of the hierarchical physics-informed learning model proposed by the present invention are as shown in Figure 6 . The identification results show that the method proposed by the present invention can accurately identify the R-M characteristics of the induction furnace at each working point, and once again verify the effectiveness of the method proposed by the present invention.

[0104] Table 1 is the specific configuration table for the training of the hierarchical physics-informed learning model

[0105]

[0106] Finally, it should be noted that the above embodiments and descriptions are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the disclosure of the technical solutions of the present invention, and they should all be covered by the protection scope of the claims of the present invention.

Claims

1. A method for identifying electrical parameters of an induction furnace based on a hierarchical physical information learning model, characterized in that: The following steps are involved: S1: Establish a series of linear physical models at multiple working points of the induction furnace to guide the data-driven models of the corresponding working points, which together constitute the underlying model; S2: In the upper-level modeling, global nonlinear constraints based on physical prior knowledge are introduced to ensure the accuracy and physical consistency of the model; S3: The hierarchical modeling problem based on a series of underlying models and upper-level models is transformed into a multi-task learning optimization problem with global nonlinear constraints, and the electrical parameter identification of the induction furnace is realized by using the shared layer-task layer model architecture.

2. The method for identifying electrical parameters of an induction furnace based on a hierarchical physical information learning model according to claim 1, characterized in that: In S1, the physical equation of the system established at the working point i of the induction furnace is as follows: The above differential equation is simplified to the following form: f i =Y d,i P i P i =[1 P 1,i P 2,i P 3,i P 4,i P 5,i P 6,i P 7,i ] T Among them, C1 and C2 are parallel and series capacitors respectively, L0 is the leakage inductance, M i is the mutual inductance, R M is the iron loss resistance, R0 is the copper loss resistance, R i is the melting resistance, and I i are the voltage and current sampled at working point i, t is the melting time, f i is the linear physical governing equation, Y d,i is a vector consisting of variables and their derivatives, P i is the electrical parameter vector.

3. The method for identifying electrical parameters of an induction furnace based on a hierarchical physical information learning model according to claim 1, characterized in that: In S1, the formula of the underlying data-driven model of the system established at the induction furnace working point i is as follows: in, is the model input, is the solution of the model, the data operator Used to track the observation data of the current working point, Ω is the solution domain; the calculation formula for the kth hidden layer of the data-driven model is as follows: With k =σ(W k With k-1 +b k ),k∈{1,...,N L } Among them, z k represents the hidden layer output, z k-1 is the hidden layer input, W k is the weight matrix, b k is the bias vector, σ is the nonlinear activation function, N L is the total number of hidden layers.

4. The method for identifying electrical parameters of an induction furnace based on a hierarchical physical information learning model according to claim 1, characterized in that: In S1, the formula of the underlying linear physical model of the system established at the induction furnace working point i is as follows: Among them, the linear physical operator represents the electrical dynamics of the induction furnace at a specific operating point i, which follows the linear physical control equation f i ,λ i is the unknown true electrical parameter, is λ i The estimated value of is the vector consisting of the solution of the data model and its differential, in the same form as Y d,i same, is the electrical parameter vector P i The estimated value of .

5. The method for identifying electrical parameters of an induction furnace based on a hierarchical physical information learning model according to claim 1, characterized in that: In S2, according to engineering experience, there is a certain proportional relationship between the mutual inductance M and the smelting resistance R and the liquid level, the furnace radius and the resonant frequency f0. In the same smelting batch, the lining loss can be ignored, that is, the furnace radius K is a constant. Therefore, the following proportional relationship can be obtained: Where K is a constant. Using this proportional relationship, the formula for the global nonlinear constraint based on physical prior knowledge in the upper layer modeling of the induction furnace is as follows: in, is the single-furnace lining conservation constraint operator, S is the label set of all working points, j represents an element different from i, and i and j both belong to S.

6. The method for identifying electrical parameters of an induction furnace based on a hierarchical physical information learning model according to claim 1, characterized in that: The S3 is specifically: First, the modeling of each operating point can be regarded as multiple related but different tasks. The tasks are encoded according to the resonant frequency of the system at each operating point. The formula of the multi-task learning model is as follows: Where i = 1, 2, ..., N t is the task number, θ s is the shared layer parameter, θ t,i is the task layer parameter, is the electrical parameter to be identified by model parameterization; Next, the formula for the loss function of the model is as follows: in, is the average underlying data modeling loss across all tasks, is the average underlying linear physical modeling loss across all tasks, It is the upper nonlinear physical prior knowledge modeling loss; Finally, the formula for the total loss function and the overall optimization objective is as follows: Among them, λ d is the weight of the underlying data modeling loss, λ p is the weight of the underlying linear physical modeling loss, λ k is the weight of the upper-level nonlinear physical prior knowledge modeling loss.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images