Intelligent recommendation method for non-oriented silicon steel production process

Abstract

The invention relates to an intelligent recommendation method for a non-oriented silicon steel production process, and the method comprises the steps: firstly constructing a knowledge graph based on the thickness, composition, process and performance standard data of non-oriented silicon steel, and obtaining a reference production path from the knowledge graph through an interpolation algorithm after inputting target iron loss; thirdly, training the mechanism model and the machine learning model in sequence by using industrial big data to form a prediction model fusing the mechanism and the data; and then, defining an iron loss optimization objective function, converting the prediction model into a product development model, and carrying out optimization solution on a production path by adopting a gradient descent method. And finally, the obtained optimal production path is put into industrial trial production, and the trial production result is fed back to the product development model for continuous optimization, so that the prediction precision and practicability of the product development model are continuously improved.

Description

Technical Field

[0001] This invention relates to an intelligent recommendation method for the production process of non-oriented silicon steel, belonging to the technical field of silicon steel production. Background Technology

[0002] Traditional product development and design rely heavily on human experience and laboratory research, requiring repeated trial and error to verify solutions. This not only leads to long experimental cycles but also incurs high trial and error costs, severely restricting the development efficiency of the target product.

[0003] The R&D model for non-oriented silicon steel follows a similar pattern, relying heavily on a trial-and-error cycle of "laboratory research – small-batch trial production – performance testing." This process not only requires repeated adjustments to the composition and annealing process but also involves the preparation and testing of numerous physical samples. This results in a long development cycle for the target product (new product) and high costs for materials, energy, and labor, which has become a key bottleneck restricting rapid product delivery and market responsiveness.

[0004] Therefore, building a digital R&D platform for silicon steel products has become an inevitable choice for promoting technological upgrades. The platform integrates production data, materials knowledge, and artificial intelligence algorithms to establish a digital mapping model of composition, process, and performance, enabling virtual prediction of performance and intelligent recommendation of process parameters, thereby significantly reducing reliance on physical experiments.

[0005] Patent application number 2023115891845 discloses a method for controlling the performance stability of non-oriented silicon steel based on a prediction model. This method employs a process optimization approach based on a performance prediction model, using preceding process parameters as known inputs and iteratively optimizing subsequent process parameters within the performance prediction model. This method involves dynamic process optimization for each coil of steel to achieve the target performance, but the specific model algorithm is not described in the paper. Summary of the Invention

[0006] To address the aforementioned problems, this invention proposes an intelligent recommendation method for non-oriented silicon steel production processes, the specific technical solution of which is as follows: A smart recommendation method for non-oriented silicon steel production processes includes the following steps: Step (1): Collect standard data of existing non-oriented silicon steel products and construct a silicon steel knowledge graph; the standard data refers to the target thickness + composition parameters + process parameters + performance parameters corresponding to different grades of non-oriented silicon steel. Step (2): Input the target thickness and target iron loss of the target product, and obtain the baseline composition and process path based on the knowledge graph and interpolation algorithm; Step (3): Collect industrial big data of existing non-oriented silicon steel products. The industrial big data refers to the actual thickness, composition parameters, process parameters and performance parameters of each coil of steel; train the mechanism model and the machine learning model, and fuse the mechanism model and the machine learning model to obtain the prediction model; Step (4): Based on the prediction model in step (3), construct the product development model. The specific process is as follows: Define the objective function for iron loss optimization, and use the baseline composition + process path obtained in step (2) as the initial value. Perform iterative calculation through gradient descent method to output the final "composition parameters + process parameters". Step (5): Apply the target product's "composition parameters + process parameters" to the on-site industrial trial production, and compare the measured iron loss value with the target iron loss value. If the average deviation value is ≥ the threshold, then feed the industrial trial production data back to step (3) to retrain the prediction model, thereby improving the accuracy of the product development model constructed in step (4). Step (6): Repeat step (5) until the average deviation between the measured performance value and the target value is less than the threshold.

[0007] Furthermore, in step 1, the thickness of the non-oriented silicon steel includes 0.50 mm and 0.35 mm; the composition parameters include Si, Al, and Mn alloying elements; the process parameters include hot rolling, normalizing, and annealing processes; and the performance parameter is the silicon steel loss value. The knowledge graph is classified according to thicknesses of 0.50 mm and 0.35 mm.

[0008] Furthermore, in step 2, after determining the target product's target thickness and target iron loss, the closest products A and B are searched from the knowledge graph of the corresponding thicknesses. It is assumed that product A has a higher iron loss value than the target iron loss, and the distance is... d A Product B's iron loss value is lower than the target iron loss, and the distance is... d B The steelmaking composition and process parameters of the target product are determined through an interpolation algorithm. The specific formula is as follows: , in and These represent the component parameters or process parameters of product A and product B, respectively. and These represent the weighting coefficients for product A and product B, respectively. This indicates the baseline composition or process parameters of the target product.

[0009] Furthermore, in step 3, firstly, industrial big data on the thickness, composition parameters, process parameters, and performance parameters of non-oriented silicon steel is obtained to train the mechanism model and the machine learning model. The calculation formula structure of the mechanism model is as follows: , In the formula, This is the iron loss value. For hysteresis loss, For eddy current losses, For abnormal losses, For operating frequency, For material thickness, The saturation magnetic flux density For material density, The resistivity of the material These are the model coefficients; The input parameters for the mechanistic model are the Si and Al contents and the material thickness. The output parameter is the iron loss value. Based on industrial big data of non-oriented silicon steel thickness, composition parameters, process parameters, and performance parameters, model coefficients were obtained. .

[0010] Furthermore, after determining the mechanism model, industrial big data on thickness, Si, and Al parameters are input into the mechanism model to obtain calculated values ​​for the iron loss mechanism. Due to the complex on-site environment and the variability of industrial big data on process parameters, the mechanistic model cannot fully capture all the characteristics of the production site, resulting in certain deviations in the mechanistic model, such as the iron loss deviation value. ; To compensate for the shortcomings of the mechanistic model, a machine learning model is introduced, combining industrial big data to learn the complex patterns of the field environment. Specifically, the machine learning algorithm used is the Bayesian optimization algorithm. The output of the Bayesian optimization algorithm is the predicted value and the 95% confidence interval. The input parameters of the Bayesian optimization algorithm are all industrial big data corresponding to the components and process parameters, and the output parameter is the iron loss deviation value. A complete Bayesian optimization model is obtained through training with industrial big data; After obtaining the mechanistic model and the Bayesian optimization model, a prediction module is constructed. First, the industrial big data of process parameters is broken down, and the thickness, Si, and Al parameters are input into the mechanistic model to obtain the calculated values ​​of the iron loss mechanism. Industrial big data, including thickness, composition parameters, process parameters, and performance parameters, is input into a Bayesian optimization model to obtain bias prediction values. The two are then combined to output the final iron loss prediction value. .

[0011] Furthermore, in step 4, the prediction model is transformed into a product development model through reverse calculation. The specific process is as follows: First, the exploration function is defined as... ,in This represents the predicted iron loss value. This represents the range of values ​​within a 95% confidence interval. , These represent the weight coefficients, ranging from 0 to 1, and their sum is 1. The exploration function means that the predicted value and the confidence interval value are used to calculate a model score based on the corresponding weight coefficients, and the maximum model score is used as the model optimization target. For product development, the model score needs to be close to the target value for iron loss, rather than the higher the model score, the better. Therefore, the exploration function needs to be transformed into: , In the formula, The target value for iron loss of the target product. The specific formula for the predicted value of the prediction model is as follows: , , in, These are values ​​calculated using a mechanistic model, based on explicit known formulas. The input parameters are Si, Al, and material thickness. ; These are the predictions from the Bayesian model. The input parameters of the Bayesian model include Si, Al, and All component parameters and process parameters, including those included; Indicates temperature.

[0012] Furthermore, by adding a negative sign before the exploration function, we construct the objective function for optimizing iron loss. Sorce The minimum value of the objective function for optimizing iron loss is found through iterative calculation using the gradient descent method. The formula for calculating the gradient of the objective function for optimizing iron loss is: , Here, X represents all composition parameters and process parameters. Starting from X0, a new X is calculated in each iteration until the gradient condition is met, at which point the final composition parameters and process parameters are output. Parameter X = ( x 0, x 1, x 2, x 3, x 4, x 5....., x n ), n is the total number of all component parameters and process parameters. Take the partial derivative for each component parameter and process parameter, and iterate continuously through formula (1). X is the input and the gradient is the output until the gradient satisfies the condition. The mechanism model is an explicit formula, which directly solves the multidimensional spatial gradient. The Bayesian model is a black box model, which approximates the gradient in a very small range near the current value using the difference method. Then, the baseline component and process path obtained by interpolation through the knowledge graph are used as the initial value X0 and substituted into formula (1) to calculate the gradient at position X0. ,judge If the value is ≤0.01, the final production process composition and process parameters will be output; otherwise, the output will be... As input values ​​for the next iteration The learning rate determines the model's learning speed, ranging from 0 to 0.5. After multiple iterations, the gradient judgment condition of the iron loss optimization objective function is met, and the final composition parameters and process parameters of the production process are output.

[0013] Furthermore, in step 5, the final production path output by the product development model is applied to industrial trial production. The threshold for the average deviation between the average iron loss value of the trial production product and the target iron loss value is 0.2 W / kg. If the condition is met, the process is solidified; if the condition is not met, the industrial trial production data is embedded into the industrial big data database as known data and used as new sample data to further train the mechanism model and Bayesian model to improve the accuracy of the product development model.

[0014] The working principle of this invention is: (1) This invention first constructs a knowledge graph based on standard data of composition, process, and performance of non-oriented silicon steel. After inputting the target iron loss, a benchmark production path is obtained from the knowledge graph through an interpolation algorithm. Next, a mechanism model and a machine learning model are trained sequentially using industrial big data, and the mechanism and data are integrated to form a prediction model. Subsequently, an objective function for iron loss optimization is defined, the prediction model is transformed into a product development model, and the gradient descent method is used to optimize the production path. Finally, the obtained optimal process path is put into industrial trial production, and the trial production results are fed back to the product development model for continuous optimization, thereby continuously improving the prediction accuracy and practicality of the product development model.

[0015] (2) Current digital R&D technologies for silicon steel products mostly rely on purely data-driven machine learning models. Due to the lack of constraints and compensation from metallurgical mechanisms, the product development models have significant prediction errors, especially with insufficient generalization ability under fluctuating operating conditions, making them difficult to apply directly to actual industrial scenarios. To address this, this invention proposes a fusion modeling method based on deviation prediction. By constructing a dual-driven architecture of "mechanism-data", the accuracy and engineering applicability of the product development model are effectively improved.

[0016] (3) In the product development model optimization iteration process, the setting of initial values ​​has a crucial impact on the optimization effect. If the initial values ​​are not chosen reasonably, the product development model is prone to getting trapped in local optima, and the number of iteration steps will increase significantly, thereby reducing computational efficiency. To address this problem, this invention proposes an initial value generation method based on knowledge graphs: first, a knowledge graph is constructed using existing product data, and then a set of benchmark process parameters is calculated using a similar product interpolation algorithm, which serves as the initial values ​​for product development model iteration. This method effectively improves the convergence speed and optimization efficiency of the product development model, and solves the problems of slow iteration and low efficiency caused by traditional initialization strategies.

[0017] The beneficial effects of this invention are: Compared to existing technologies, the focus of this invention is on the digital R&D design of silicon steel products. Specifically, for the target product, a set of solidifiable composition and process standard paths is designed by integrating mechanistic models and machine learning models, thereby meeting the needs of subsequent stable mass production of the product. Furthermore, by integrating the mechanistic model into the machine learning algorithm, the large deviations caused by the machine learning model in predicting unknown products can be compensated to a certain extent.

[0018] This invention integrates mechanistic models and machine learning models. On one hand, it overcomes the shortcomings of pure mechanistic models in adapting to complex and variable process parameters in real-world environments. On the other hand, it improves the problem of significantly reduced prediction accuracy of purely data-driven machine learning models when they fall outside the range of data features. Especially during product development, when the target component is not within the existing intermediate range but exceeds the established boundaries, relying solely on machine learning leads to a substantial decrease in prediction accuracy. By introducing a mechanistic model, even in such "out-of-scope" situations, the model can still provide a reasonable baseline prediction value based on physical mechanisms, thereby effectively controlling the error range and significantly improving the model's generalization ability.

[0019] This invention deeply integrates knowledge graphs, mechanistic models, and machine learning algorithms to construct a data-driven digital design system. Users only need to input the key target parameters of the product (such as thickness and iron loss) to automatically generate recommended solutions covering the entire process parameters from steelmaking composition, hot rolling, normalizing to annealing.

[0020] This invention completely changes the traditional trial-and-error development model, shifting process design from experience-driven to scientific decision-making. It can significantly shorten the product design cycle, reduce development costs, and improve the accuracy and reliability of process parameters. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall process of the present invention. Figure 2 This is a schematic diagram of the machine learning model training of the present invention. Figure 3 This is a schematic diagram of the prediction model of the present invention. Figure 4 This is a schematic diagram of the product development model of the present invention. Figure 5 This is a comparison chart of the calculated and measured values ​​of the iron loss mechanism according to an embodiment of the present invention. Figure 6 These are graphs showing the measured iron loss data of each batch in the industrial trial production of this invention. Detailed Implementation

[0022] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0023] Combined with appendix Figure 1 As can be seen, the overall process of this invention is as follows: (1) Collect standard data on the target thickness (0.50mm and 0.35mm) + composition parameters (alloying elements such as Si, Al, and Mn) + process parameters + performance parameters (silicon steel loss value) of existing non-oriented silicon steel products, and construct a silicon steel knowledge graph. Process parameters include hot rolling process (heating furnace temperature, roughing temperature, finishing temperature, coiling temperature, etc.), normalizing process (normalizing speed, normalizing temperature, etc.), and annealing process (annealing speed, annealing temperature, etc.).

[0024] (2) Input the thickness specification and target iron loss of the target product (new product). Based on the knowledge graph, obtain the baseline composition and process path through interpolation algorithm. Find similar products A and B (products one above and one below the target iron loss value, respectively) from the knowledge graph of the corresponding thickness. Assume that the iron loss value of product A is higher than the target iron loss, and the distance is... d A Product B's iron loss value is lower than the target iron loss, and the distance is... d B The steelmaking composition and process parameters of the target product are determined using an interpolation algorithm. The specific formula is as follows: , in and These represent the component parameters or process parameters of product A and product B, respectively. and These represent the weighting coefficients for product A and product B, respectively. This indicates the baseline composition or process parameters of the target product.

[0025] (3) Collect industrial big data on the thickness, composition, process and performance of existing non-oriented silicon steel products (actual data of each coil of steel), train the mechanism model and machine learning model, and then obtain a prediction model that integrates "mechanism + AI".

[0026] The calculation formula structure for the mechanism model is as follows: , In the formula, This is the iron loss value. For hysteresis loss, For eddy current losses, For abnormal losses, For operating frequency, For material thickness, The saturation magnetic flux density For material density, The resistivity of the material These are the model coefficients; The input parameters for the mechanistic model are the Si and Al contents and the material thickness. The output parameter is the iron loss value. Based on industrial big data of non-oriented silicon steel thickness, composition parameters, process parameters, and performance parameters, model coefficients were obtained. .

[0027] The input parameters for the mechanistic model are the Si and Al contents and the material thickness. d The output parameter is the iron loss value. Formulas can be used Simplified explanation. Based on big data from the non-oriented silicon steel industry, the model coefficients were obtained through least squares regression. Since the least squares method is existing technology, it will not be described in detail in this invention.

[0028] After determining the mechanism model, inputting the thickness, Si, and Al parameters into the mechanism model will yield the calculated values ​​of the iron loss mechanism. Due to the complex on-site environment and variable process parameters, the mechanistic model cannot fully capture all the characteristics of the production site, resulting in certain deviations in the mechanistic model, and thus, the iron loss deviation value. .

[0029] To overcome the shortcomings of existing mechanistic models, a machine learning model is introduced, combining industrial big data to learn the complex patterns of the field environment. For details, please refer to [link / reference needed]. Figure 2 The machine learning algorithm used is the Bayesian optimization algorithm. This algorithm outputs a predicted value and a 95% confidence interval (compared to traditional models, it not only outputs a single predicted value but also includes a fluctuation range within which the measured value has a 95% probability of falling). The input parameters for the Bayesian optimization algorithm are all composition and process parameters (including composition-hot rolling-normalizing-annealing), and the output parameter is the iron loss deviation value. A complete Bayesian optimization model can be trained using industrial big data (the training process is based on existing technology and will not be elaborated here). For example... Figure 2 As shown.

[0030] After obtaining the mechanistic model and the Bayesian optimization model, a prediction model is constructed, such as... Figure 3 As shown, the process parameters are first broken down, with thickness, Si, and Al parameters input into the mechanistic model to obtain mechanistic calculation values, and all parameters input into the Bayesian model to obtain deviation prediction values. The two are then combined to output the final iron loss prediction value. .like Figure 3 As shown.

[0031] (4) Based on the prediction model, a product development model is constructed, an objective function for optimizing iron loss is defined, and the baseline composition + process path is used as the initial value of the gradient function of the objective function. Iterative calculation is performed using the gradient descent method to output the final production path of "composition parameters + process parameters". The specific process is as follows: The "mechanism + AI" integrated prediction model is transformed into a product development model through reverse computation. The specific process is as follows: First, the exploration function is defined as... ,in Indicates the predicted value. This indicates the 95% confidence interval range. , These represent the weighting coefficients. This function means that the predicted value and confidence interval are calculated based on the corresponding weighting coefficients (the weighting coefficients are determined according to actual needs, especially when the data volume is small and unknown intervals need to be explored). Take the larger value, especially when the data volume is large. Calculate a model score by taking the largest value, and use the maximum model score as the model optimization target.

[0032] For product development, the model score needs to be close to the iron loss target value, rather than the higher the score, the better. Therefore, the exploration function needs to be transformed into: , In the formula, The target value for iron loss of the target product. The specific formula for the predicted value of the prediction model is as follows: , , in, These are values ​​calculated using a mechanistic model, based on explicit known formulas. The input parameters are Si, Al, and material thickness. ; These are the predictions from the Bayesian model. The input parameters of the Bayesian model include Si, Al, and All component parameters and process parameters, including the formula The brackets in the text contain all the components and process parameters, specifically the average furnace temperature, minimum furnace temperature, maximum furnace temperature, furnace time, average roughing temperature, average finishing rolling temperature, minimum finishing rolling temperature, maximum finishing rolling temperature, average coiling temperature, average normalizing rate, average normalizing temperature, average annealing process rate, average annealing temperature, minimum annealing temperature, maximum annealing temperature, annealing furnace tension, C, Si, Mn, P, S, Cr, Ni, Cu, Mo, Ti, N, Al, Sn; It indicates temperature, and the temperature is further subdivided into heating furnace temperature, roughing rolling temperature, finishing rolling temperature, etc.

[0033] Based on the aforementioned exploration function Sorce, an objective function for optimizing iron loss is constructed, and iterative calculation is performed using the gradient descent method. Since the principle of gradient descent is to find the minimum value, while the exploration function Sorce needs to obtain the maximum value, a negative sign needs to be added before Sorce, converting it to -Sorce. The gradient calculation formula for the objective function for optimizing iron loss is: , The mechanistic model uses an explicit, known formula, allowing direct calculation of the multidimensional gradient. The Bayesian model, however, is a black-box model with an unknown formula, thus its gradient can only be approximated within a very small range around the current value using the finite difference method. Then, the production path obtained through knowledge graph interpolation is used as the initial value X0, substituted into the iron loss optimization objective function -Sorce, to calculate the gradient. ,judge If the value is ≤0.01, the final production path will be output if the condition is met; otherwise, the output path will be... As the input value for the next iteration, after multiple iterations, until the gradient determination condition of the iron loss optimization objective function is met, the final composition parameters and process parameters of the production process are output. For example... Figure 4 As shown.

[0034] (5) Apply the “composition parameters + process parameters” of the target product to the on-site industrial trial production, and compare the measured iron loss value with the target iron loss value. If the average iron loss value deviates from the target iron loss value by more than 0.2 W / kg, the industrial trial production data is fed back to the product development model to improve the accuracy.

[0035] The following is a specific application example of the present invention: (1) Collect non-oriented silicon steel data to construct a knowledge graph. The thickness is divided into 0.50mm and 0.35mm. The alloy composition includes elements such as Si and Al. The process parameters include hot rolling process, normalizing process and annealing process. The performance parameter is the power frequency iron loss value of silicon steel.

[0036] Table 1 Knowledge Graph of Non-oriented Silicon Steel

[0037] (2) Input the target product thickness (0.50mm) and target iron loss value (3.1W / kg). First, find similar product numbers 5 and 6 from the knowledge graph, and calculate the distance of the target product's iron loss to numbers 5 and 6: d 5 = 3.3 - 3.1 = 0.2 d 6 = 3.1 - 3 = 0.1. Then, interpolation is performed based on the distance weights to obtain the baseline production path for the target product: Si element: 1.42% * 0.1 / (0.1 + 0.2) + 2.08% * 0.2 / (0.1 + 0.2) = 1.86% Furnace temperature: 1130*0.1 / (0.1+0.2) + 1120*0.2 / (0.1+0.2) = 1123℃ Normalizing temperature: 940*0.1 / (0.1+0.2) + 920*0.2 / (0.1+0.2) = 927℃ Annealing temperature: 910*0.1 / (0.1+0.2) + 940*0.2 / (0.1+0.2) = 930℃ The baseline production path obtained through interpolation is used as the initial value for subsequent model iteration calculations.

[0038] (3) Collect industrial big data on non-oriented silicon steel, about 10,000 volumes, covering thickness specifications of 0.35mm and 0.50mm, and composition systems with Si content ranging from 0.3% to 2.6%. Substitute Si, Al and finished product thickness as inputs and silicon steel loss as output into the mechanism model, and perform regression fitting on the mechanism model using the least squares method.

[0039] ,

[0040] By fitting the coefficients of the computer theoretical model, respectively This leads to the mechanistic model. Among them, the calculated and measured values ​​of iron loss mechanism are, for example, Figure 5 As shown.

[0041] (4) Take all parameters of steelmaking composition, normalizing process and annealing process as input, and the deviation between the calculated value of iron loss mechanism and the target value as output, and substitute them into the Bayesian model for training (the training process is completed automatically by the script program, which will not be described in detail here) to obtain the fitted iron loss deviation prediction model. The final prediction model is obtained by fusing the mechanistic model and the Bayesian model. .

[0042] (5) Define the objective function for optimizing iron loss. Considering that the model training data is sufficient (around 10,000 data points), the weight coefficients... The value is 0.8. The value is 0.2, with a 95% confidence interval. This can be automatically solved using a Bayesian model. Using the baseline production path as the initial value, iterative calculations are performed using the gradient descent method to find the minimum value of the objective function for optimizing iron loss. The gradient calculation formula is as follows:

[0043] Where X is a multidimensional process parameter, X = ( x 1, x 2, x 3, x 4, x 5....), requiring partial derivatives for each process parameter, mechanism model f It is an explicit formula and can be solved directly using a Bayesian model. g This is a black-box model, calculated using a difference approximation. After each iteration, the output gradient is determined. If the condition is met, the iteration terminates; otherwise, let... To begin the next iteration ( (This represents the learning rate, with a value of 0.1).

[0044] (6) After further optimizing the production path through the predictive model, the final composition and process parameters were obtained: Si element 1.825%, furnace temperature 1132℃, normalizing temperature 935℃; annealing temperature 925℃. The process scheme was applied to the production site for a trial production of one batch, totaling 50 coils. The iron loss data is as follows: Figure 6 As shown, the average value is 3.101 W / kg and the standard deviation is 0.003 W / kg, which meets the design requirements.

[0045] Those skilled in the art will understand that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as herein.

[0046] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. An intelligent recommendation method for non-oriented silicon steel production processes, characterized in that, Includes the following steps: Step (1): Collect standard data of existing non-oriented silicon steel products and construct a silicon steel knowledge graph; the standard data refers to the target thickness + composition parameters + process parameters + performance parameters corresponding to different grades of non-oriented silicon steel. Step (2): Input the target thickness and target iron loss of the target product, and obtain the baseline composition and process path based on the knowledge graph through interpolation algorithm; Step (3): Collect industrial big data of existing non-oriented silicon steel products. The industrial big data refers to the actual thickness, composition parameters, process parameters and performance parameters of each coil of steel; train the mechanism model and the machine learning model, and fuse the mechanism model and the machine learning model to obtain the prediction model; Step (4): Based on the prediction model in step (3), construct the product development model. The specific process is as follows: Define the objective function for iron loss optimization, and use the baseline composition + process path obtained in step (2) as the initial value. Perform iterative calculation through gradient descent method to output the final "composition parameters + process parameters". Step (5): Apply the target product's "composition parameters + process parameters" to the on-site industrial trial production, and compare the measured iron loss value with the target iron loss value. If the average deviation value is ≥ the threshold, then feed the industrial trial production data back to step (3) to retrain the prediction model, thereby improving the accuracy of the product development model constructed in step (4). Step (6): Repeat step (5) until the average deviation between the measured performance value and the target value is less than the threshold.

2. The intelligent recommendation method for non-oriented silicon steel production process according to claim 1, characterized in that, In step 1, the thickness of the non-oriented silicon steel includes 0.50 mm and 0.35 mm; the composition parameters include Si, Al, and Mn alloying elements; the process parameters include hot rolling, normalizing, and annealing processes; and the performance parameter is the silicon steel loss value. The knowledge graph is classified according to thicknesses of 0.50 mm and 0.35 mm.

3. The intelligent recommendation method for non-oriented silicon steel production process according to claim 1, characterized in that, In step 2, after determining the target thickness and target iron loss of the target product, the closest products A and B are found from the knowledge graph of the corresponding thicknesses. It is assumed that the iron loss value of product A is higher than the target iron loss, and the distance is... d A Product B's iron loss value is lower than the target iron loss, and the distance is... d B The steelmaking composition and process parameters of the target product are determined through an interpolation algorithm. The specific formula is as follows: ， in and These represent the component parameters or process parameters of product A and product B, respectively. and These represent the weighting coefficients for product A and product B, respectively. This indicates the baseline composition or process parameters of the target product.

4. The intelligent recommendation method for non-oriented silicon steel production process according to claim 1, characterized in that, In step 3, firstly, industrial big data on the thickness, composition parameters, process parameters, and performance parameters of non-oriented silicon steel is obtained to train the mechanism model and the machine learning model. The calculation formula structure of the mechanism model is as follows: ， In the formula, This is the iron loss value. For hysteresis loss, For eddy current losses, For abnormal losses, For operating frequency, For material thickness, The saturation magnetic flux density For material density, The resistivity of the material These are the model coefficients; The input parameters for the mechanistic model are the Si and Al contents and the material thickness. The output parameter is the iron loss value. Based on industrial big data of non-oriented silicon steel thickness, composition parameters, process parameters, and performance parameters, model coefficients were obtained. .

5. The intelligent recommendation method for non-oriented silicon steel production process according to claim 1, characterized in that, After determining the mechanism model, industrial big data on thickness, Si, and Al parameters are input into the mechanism model to obtain the calculated values ​​of the iron loss mechanism. Due to the complex on-site environment and the variability of industrial big data on process parameters, the mechanistic model cannot fully capture all the characteristics of the production site, resulting in certain deviations in the mechanistic model, such as the iron loss deviation value. ; To compensate for the shortcomings of the mechanistic model, a machine learning model is introduced, combining industrial big data to learn the complex patterns of the field environment. Specifically, the machine learning algorithm used is the Bayesian optimization algorithm. The output of the Bayesian optimization algorithm is the predicted value and the 95% confidence interval. The input parameters of the Bayesian optimization algorithm are all industrial big data corresponding to the components and process parameters, and the output parameter is the iron loss deviation value. A complete Bayesian optimization model is obtained through training with industrial big data; After obtaining the mechanistic model and the Bayesian optimization model, a prediction module is constructed. First, the industrial big data of process parameters is broken down, and the thickness, Si, and Al parameters are input into the mechanistic model to obtain the calculated values ​​of the iron loss mechanism. Industrial big data, including thickness, composition parameters, process parameters, and performance parameters, is input into a Bayesian optimization model to obtain bias prediction values. The two are then combined to output the final iron loss prediction value. .

6. The intelligent recommendation method for non-oriented silicon steel production process according to claim 1, characterized in that, In step 4, the prediction model is transformed into a product development model through reverse calculation. The specific process is as follows: First, the exploration function is defined as... ,in This represents the predicted iron loss value. This represents the range of values ​​within a 95% confidence interval. , These represent weighting coefficients, with values ​​ranging from 0 to 1, and their sum is 1. The meaning of the exploration function is to calculate a model score based on the predicted value and the confidence interval value according to the corresponding weight coefficients, and use the maximum model score as the model optimization target. For product development, the model score needs to be close to the target value for iron loss, rather than the higher the model score, the better. Therefore, the exploration function needs to be transformed into: ， In the formula, The target value for iron loss of the target product. The specific formula for the predicted value of the prediction model is as follows: ， ， in, These are values ​​calculated using a mechanistic model, based on explicit known formulas. The input parameters are Si, Al, and material thickness. ; These are the predictions from the Bayesian model. The input parameters of the Bayesian model include Si, Al, and All component parameters and process parameters, including Indicates temperature.

7. The intelligent recommendation method for non-oriented silicon steel production process according to claim 6, characterized in that, To construct the iron loss optimization objective function, add a negative sign before the exploration function. Sorce The minimum value of the objective function for optimizing iron loss is found by iterative calculation using the gradient descent method. The formula for calculating the gradient of the objective function for iron loss optimization is: ， Here, X represents all composition parameters and process parameters. Starting from X0, a new X is calculated in each iteration until the gradient condition is met, at which point the final composition parameters and process parameters are output. Parameter X = ( x 0, x 1, x 2, x 3, x 4, x 5....., x n ), n is the total number of all component parameters and process parameters. Take the partial derivative for each component parameter and process parameter, and iterate continuously through formula (1). X is the input and the gradient is the output until the gradient satisfies the condition. The mechanism model is an explicit formula, which directly solves the multidimensional spatial gradient. The Bayesian model is a black box model, which approximates the gradient in a very small range near the current value using the difference method. Then, the baseline component and process path obtained by interpolation through the knowledge graph are used as the initial value X0 and substituted into formula (1) to calculate the gradient at position X0. ,judge If the value is ≤0.01, the final production process composition and process parameters will be output; otherwise, the output will be... As input values ​​for the next iteration The learning rate determines the model's learning speed, ranging from 0 to 0.

5. After multiple iterations, the gradient judgment condition of the iron loss optimization objective function is met, and the final composition parameters and process parameters of the production process are output.

8. The intelligent recommendation method for non-oriented silicon steel production process according to claim 1, characterized in that, In step 5, the final production path output by the product development model is applied to industrial trial production. The threshold for the average deviation between the average iron loss value of the trial production product and the target iron loss value is 0.2 W / kg. If the condition is met, the process is solidified; if the condition is not met, the industrial trial production data is embedded into the industrial big data database as known data and used as new sample data to further train the mechanism model and Bayesian model to improve the accuracy of the product development model.

Images