Process time prediction method and related apparatus

Abstract

This application discloses a process time prediction method and related apparatus, relating to the field of iron and steel smelting. This application obtains the steelmaking parameter set, refining parameter set, and continuous casting parameter set corresponding to the three processes of steelmaking, refining, and continuous casting for the target steel grade. The semantic vector of the target steel grade is obtained collaboratively based on the gradients of the steelmaking model, refining model, and continuous casting model, which can integrate the process semantic information of the three processes. The steelmaking parameter set, refining parameter set, and continuous casting parameter set, along with the semantic vector of the target steel grade, are respectively input into pre-constructed steelmaking, refining, and continuous casting time prediction models. This achieves accurate matching between the model corresponding to each process and its own process characteristics, allowing each model to focus on capturing the unique and most critical process features of the corresponding process without needing to learn complex hybrid mapping functions. This effectively avoids the internal logic confusion caused by a single model predicting the time of multiple processes, significantly improving the accuracy of time prediction for each process in steelmaking, refining, and continuous casting.

Description

Technical Field

[0001] This application relates to the field of iron and steel smelting technology, and in particular to a method and apparatus for predicting process time. Background Technology

[0002] Steelmaking, refining, and continuous casting are the core processes in steel production, and accurate prediction of their process times is a key prerequisite for improving production efficiency and achieving lean management. Specifically, the steelmaking process time refers to the continuous operation time from the moment molten iron is poured into the converter until the molten steel completes core smelting operations such as decarburization, heating, and preliminary dephosphorization and desulfurization, reaching the converter's final process requirements and being tapped. The refining process time refers to the continuous operation time from the moment molten steel from steelmaking enters the refining furnace until the molten steel completes all refining operations such as refined composition control, deep desulfurization and dephosphorization, vacuum degassing, temperature compensation, and alloying, reaching the target steel grade's purity and composition requirements and being tapped. The continuous casting process time refers to the continuous operation time from the moment qualified refined molten steel is poured into the continuous casting machine's crystallizer until all molten steel in that heat is poured, the billet undergoes crystallization cooling, straightening, and the casting process is completed.

[0003] Currently, the commonly used method in the industry for process time prediction is to use the same machine learning model to predict the process times for steelmaking, refining, and continuous casting processes separately. However, the technological mechanisms, response speeds, and influencing factors of steelmaking, refining, and continuous casting processes differ significantly. Steelmaking is highly nonlinear and involves violent reactions; refining is high-dimensional and has complex feature interactions. Using the same machine learning model to predict the process times for each of the three processes results in the machine learning model learning an overly complex hybrid mapping function. This makes it difficult for the machine learning model to capture the unique and most critical features of each process, leading to internal logical confusion within the machine learning model and resulting in low accuracy of the obtained process times. Summary of the Invention

[0004] In view of the above problems, this application provides a method and related apparatus for predicting process times to achieve the goal of obtaining process times with high accuracy. The specific solution is as follows: The first aspect of this application provides a method for predicting process time, including: Obtain the steelmaking parameter set for the steelmaking process, the refining parameter set for the refining process, and the continuous casting parameter set for the continuous casting process for the target steel grade; The target steel grade semantic vector is found from the preset steel grade and steel grade semantic vector. The target steel grade semantic vector is obtained based on the steelmaking model gradient representing the loss of the steelmaking time prediction model, the refining model gradient representing the loss of the refining time prediction model, and the continuous casting model gradient representing the loss of the continuous casting time prediction model. The steelmaking parameter set and the target steel grade semantic vector are input into the pre-constructed steelmaking time prediction model, and the steelmaking process time is obtained through the steelmaking time prediction model. The refining parameter set and the target steel grade semantic vector are input into the pre-constructed refining time prediction model, and the refining process time is obtained through the refining time prediction model. The continuous casting parameter set and the semantic vector of the target steel grade are input into the pre-constructed continuous casting time prediction model, and the continuous casting process time is obtained through the continuous casting time prediction model.

[0005] In one possible implementation, the step of searching for the target steel grade semantic vector corresponding to the target steel grade from the preset steel grade and steel grade semantic vector includes: Input the steel grade identifier of the target steel grade into the pre-built shared embedding layer; The target steel grade semantic vector corresponding to the steel grade identifier is obtained from the embedding weight matrix stored in the shared embedding layer. The embedding weight matrix contains steel grade semantic vectors corresponding to multiple steel grade identifiers. Obtain the semantic vector of the target steel grade output by the shared embedding layer.

[0006] In one possible implementation, the method for obtaining the embedding weight matrix includes: For each steel grade, the steelmaking parameter set corresponding to the steel grade is input into the steelmaking time prediction model, and the steelmaking model gradient is obtained through the steelmaking time prediction model; the refining parameter set corresponding to the steel grade is input into the refining time prediction model, and the refining model gradient is obtained through the refining time prediction model; the continuous casting time prediction model corresponding to the steel grade is input into the continuous casting time prediction model, and the continuous casting model gradient is obtained through the continuous casting time prediction model. For each steel grade, a semantic vector for the steel grade is obtained based on the gradient of the steelmaking model, the gradient of the refining model, and the gradient of the continuous casting model corresponding to the steel grade. Obtain the embedding weight matrix, which is composed of the semantic vectors of the steel grades corresponding to the multiple steel grades.

[0007] In one possible implementation, the step of obtaining the steel grade semantic vector based on the steelmaking model gradient, the refining model gradient, and the continuous casting model gradient corresponding to the steel grade includes: The gradient fusion value is calculated using the fusion formula α×▽1+β×▽2+γ×▽3, where α+β+γ=1, and α, β, and γ are preset weights; ▽1 is the gradient of the steelmaking model; ▽2 is the gradient of the refining model; and ▽3 is the gradient of the continuous casting model. The gradient fusion value is input into the Adam optimizer to obtain the optimized gradient; By updating formula W new =W old -η×Adam( The semantic vector of the steel type is calculated using the total), where Adam( (total) is the optimization gradient, W old W is the semantic vector of the steel type before the update. new This is the updated semantic vector for steel grades.

[0008] In one possible implementation, the target steel grade semantic vector includes: The alloying degree, representing the extent of alloying element addition; the cleanliness requirement, representing the purity index of inclusions and gas content in molten steel for the target steel grade; the temperature sensitivity, representing the sensitivity of the target steel grade to temperature fluctuations during smelting and casting; the process difficulty coefficient, representing the difficulty of process control in the entire production process of the target steel grade; the solidification characteristics, representing the solidification range width of the target steel grade in the continuous casting process; the cost level, representing the production cost of the target steel grade; the historical stability, representing the fluctuation of operating time and quality stability of the target steel grade in historical production processes; and the comprehensive quality level, representing the final product quality grade of the target steel grade.

[0009] In one possible implementation, the steelmaking time prediction model is a gated linear unit network; the refining time prediction model is a Transformer model; and the continuous casting time prediction model is a multilayer perceptron.

[0010] A second aspect of this application provides a process time prediction device, comprising: The first acquisition module is used to acquire the steelmaking parameter set for the steelmaking process, the refining parameter set for the refining process, and the continuous casting parameter set for the continuous casting process for the target steel grade. The first search module is used to search for the target steel grade semantic vector corresponding to the target steel grade from the preset steel grade and steel grade semantic vector. The target steel grade semantic vector is obtained based on the steelmaking model gradient representing the loss of the steelmaking time prediction model, the refining model gradient representing the loss of the refining time prediction model, and the continuous casting model gradient representing the loss of the continuous casting time prediction model. The second acquisition module is used to input the steelmaking parameter set and the target steel grade semantic vector into the pre-constructed steelmaking time prediction model, and obtain the steelmaking process time through the steelmaking time prediction model. The third acquisition module is used to input the refining parameter set and the target steel grade semantic vector into the pre-constructed refining time prediction model, and obtain the refining process time through the refining time prediction model. The fourth acquisition module is used to input the continuous casting parameter set and the target steel grade semantic vector into the pre-constructed continuous casting time prediction model, and obtain the continuous casting process time through the continuous casting time prediction model.

[0011] A third aspect of this application provides a computer program product including computer-readable instructions that, when executed on an electronic device, cause the electronic device to implement the process time prediction method of the first aspect or any implementation thereof.

[0012] A fourth aspect of this application provides an electronic device, including at least one processor and a memory connected to the processor, wherein: The memory is used to store computer programs; The processor is used to execute the computer program so that the electronic device can implement the process time prediction method of the first aspect or any implementation thereof.

[0013] The fifth aspect of this application provides a computer storage medium carrying one or more computer programs, which, when executed by an electronic device, enable the electronic device to perform the process time prediction method described in the first aspect or any implementation thereof.

[0014] By employing the above technical solution, this application provides a process time prediction method, which obtains the steelmaking parameter set, refining parameter set, and continuous casting parameter set corresponding to the three processes of steelmaking, refining, and continuous casting for the target steel grade. The semantic vector of the target steel grade is obtained collaboratively based on the gradients of the steelmaking model, refining model, and continuous casting model. It can integrate the process semantic information of the three processes, accurately characterize the process characteristics of the target steel grade in the entire production process, and provide unified and steel grade-specific shared feature support for the three process time prediction models, breaking down the information barriers between the prediction models of each process. The method also includes the process-specific parameters... The system inputs pre-built time prediction models for steelmaking, refining, and continuous casting processes with sets of parameters (such as the steelmaking parameter set for the steelmaking process, the refining parameter set for the refining process, and the continuous casting parameter set for the continuous casting process) and the semantic vector of the target steel grade. This achieves accurate matching between the model corresponding to each process and its own process characteristics. This allows each model to focus on capturing the unique and most critical process features of the corresponding process without having to learn complex hybrid mapping functions. This effectively avoids the internal logic confusion caused by a single model predicting the time of multiple processes in existing technologies, and significantly improves the accuracy of time prediction for each process in steelmaking, refining, and continuous casting. Attached Figure Description

[0015] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and the originals and elements are not necessarily drawn to scale.

[0016] Figure 1 A schematic diagram of a system architecture is provided for this application; Figure 2 A flowchart illustrating a process time prediction method provided in an embodiment of this application; Figure 3 A schematic diagram of the shared embedding layer provided in the embodiments of this application; Figure 4 This is a schematic diagram illustrating the process of obtaining the semantic vector of steel type provided in an embodiment of this application; Figure 5 A process time prediction architecture diagram provided for embodiments of this application; Figure 6 A detailed structural diagram of the steelmaking time prediction model 501 provided in the embodiments of this application; Figure 7 A detailed structural diagram of the refining time prediction model 502 provided in the embodiments of this application; Figure 8 A detailed structural diagram of the continuous casting time prediction model 503 provided in the embodiments of this application; Figure 9 A schematic diagram illustrating the training process of the continuous casting time prediction model provided in this application embodiment; Figure 10 A schematic diagram of a process time prediction device provided in an embodiment of this application; Figure 11 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0017] The embodiments of this application are described below with reference to the accompanying drawings. The terminology used in the implementation section of this application is for explaining specific embodiments only and is not intended to limit the scope of this application.

[0018] The embodiments of this application will now be described with reference to the accompanying drawings. Those skilled in the art will recognize that, with technological advancements and the emergence of new scenarios, the technical solutions provided in the embodiments of this application are equally applicable to similar technical problems.

[0019] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms are interchangeable where appropriate; this is merely a way of distinguishing objects with the same attributes in the embodiments of this application. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, so that a process, method, system, product, or apparatus that comprises a series of elements is not necessarily limited to those elements, but may include other elements not explicitly listed or inherent to those processes, methods, products, or apparatuses.

[0020] This application can be applied to the fields of intelligent steel manufacturing, metallurgical process optimization, industrial production scheduling, and metallurgical big data analysis. The following will introduce several application scenarios that can be applied to products, taking the prediction of the entire process of steelmaking, refining, and continuous casting as an example.

[0021] First, let's introduce the application scenarios of this application.

[0022] This application can be applied, but is not limited to, to applications with process time prediction methods or cloud services provided by cloud-side servers, which will be described in detail below: See Figure 1 , Figure 1 A schematic diagram of a system architecture is shown. The system may include a terminal 100 and a server 200. The server 200 can provide the methods provided in the embodiments of this application to one or more terminals.

[0023] The terminal 100 may be equipped with a process time prediction application. The application and webpage can provide an interface. The terminal 100 can receive relevant parameters (including but not limited to steel grade identification, steelmaking parameter set, refining parameter set, continuous casting parameter set, etc.) entered by the user on the steel grade parameter input interface, process condition setting interface, or prediction demand submission interface, and send the above parameters to the server 200. The server 200 can obtain the processing result based on the received parameters and return the processing result to the terminal 100.

[0024] It should be understood that in some optional implementations, the terminal 100 can also complete the action of obtaining the processing result based on the received parameters on its own, without the need for the server to cooperate. This application embodiment is not limited to this.

[0025] The following description Figure 1 The product form of the mid-terminal 100; The terminal 100 in this application embodiment can be a mobile phone, tablet computer, wearable device, vehicle device, augmented reality (AR) / virtual reality (VR) device, laptop computer, ultra-mobile personal computer (UMPC), netbook, personal digital assistant (PDA), etc., and this application embodiment does not impose any restrictions on it.

[0026] Terminal 100 may include a radio frequency unit, memory, input unit, display unit, camera (optional), audio circuitry (optional), speaker (optional), microphone (optional), headphone jack (optional), processor, external interface, power supply, and other components. Those skilled in the art will understand that the above-mentioned components are merely examples and do not constitute a limitation on the terminal or multifunctional device; it may include more or fewer components, or a combination of certain components, or different components.

[0027] The input unit can be used to receive input numeric or character information, and to generate key signal inputs related to user settings and function control of the portable multi-functional device. Specifically, the input unit may include a touchscreen (optional) and / or other input devices. Other input devices may include, but are not limited to, one or more of a physical keyboard, function keys (such as volume control buttons, power buttons, etc.), trackball, mouse, joystick, etc.

[0028] Among them, the input device can receive input data, etc.

[0029] The display unit can be used to display information input by the user or information provided to the user, various menus of the terminal, interactive interfaces, file display, and / or playback of any multimedia file. In the embodiments of this application, the display unit can be used to display parameter input interfaces, steel grade information selection interfaces, process condition setting interfaces, process time prediction result interfaces, prediction result comparison interfaces, abnormal parameter prompt interfaces, processing results, etc.

[0030] The memory can be used to store software code related to the process time prediction method, the processor can execute the steps of the process time prediction method, and can also schedule other units (such as the above-mentioned input unit and display unit) to achieve the corresponding functions.

[0031] This radio frequency unit (optional) can be used to receive and send signals during information transmission or calls.

[0032] In this embodiment of the application, the radio frequency unit can send data to the server 200 and receive the processing results sent by the server 200.

[0033] It should be understood that this radio frequency unit is optional and can be replaced with other communication interfaces, such as a network port.

[0034] Terminal 100 also includes a power source (such as a battery) for supplying power to the various components.

[0035] Terminal 100 also includes an external interface, which can be a standard Micro USB interface or a multi-pin connector, which can be used to connect terminal 100 to other devices for communication or to connect a charger to charge terminal 100.

[0036] Server 200 includes a bus, a processor, a communication interface, and memory. The processor, memory, and communication interface communicate with each other via the bus.

[0037] The memory can be used to store software code related to the process time prediction method, the processor can execute the steps of the chip's process time prediction method, and can also schedule other units to achieve the corresponding functions.

[0038] This application provides a method for predicting process time. The method for predicting process time according to this application will be described in detail below with reference to the accompanying drawings.

[0039] Reference Figure 2 , Figure 2 A flowchart illustrating a process time prediction method provided in this application embodiment is shown below. Figure 2 As shown in the figure, the process time prediction method provided in this application embodiment may include steps S201 to S205, which are described in detail below.

[0040] Step S201: Obtain the steelmaking parameter set for the steelmaking process, the refining parameter set for the refining process, and the continuous casting parameter set for the continuous casting process for the target steel grade.

[0041] The target steel grade refers to a specific steel grade for which the process time of the steelmaking process, refining process, and continuous casting process to be predicted is to be determined.

[0042] The steelmaking parameter set refers to the set of parameters that characterize all the core process conditions of the target steel grade in the steelmaking process, as detailed in Table 1.

[0043] Table 1

[0044] The refining parameter set refers to the set of parameters that characterize all the core process conditions of the target steel grade in the refining process, as detailed in Table 2.

[0045] Table 2

[0046] The continuous casting parameter set refers to the set of parameters that characterize all the core process conditions of the target steel grade in the continuous casting process, as detailed in Table 3.

[0047] Table 3

[0048] For example, steel production information systems (such as MES (Manufacturing Execution System)) can collect the full-process parameters of the target steel grade in the three processes of steelmaking, refining, and continuous casting, and organize them into dedicated parameter sets according to process categories to avoid mixing parameters from different processes.

[0049] Step S202: Find the target steel grade semantic vector corresponding to the target steel grade from the preset steel grade and steel grade semantic vector. The target steel grade semantic vector is obtained based on the steelmaking model gradient representing the loss of the steelmaking time prediction model, the refining model gradient representing the loss of the refining time prediction model, and the continuous casting model gradient representing the loss of the continuous casting time prediction model.

[0050] The correspondence between steel grades and their semantic vectors is obtained in advance through model training, where each steel grade corresponds to a unique semantic vector.

[0051] For example, the calculation process of the steelmaking model gradient is as follows: During the training of the steelmaking time prediction model, the difference between the predicted process time output by the steelmaking time prediction model and the labeled actual steelmaking time is compared to obtain a loss function, such as the MSE (Mean Squared Error Loss Function) loss function; the steelmaking model gradient is obtained by backpropagation of the loss function.

[0052] For example, the process of calculating the refinement model gradient is as follows: During the training of the refinement time prediction model, the difference between the predicted process time output by the refinement time prediction model and the labeled actual refinement time is compared to obtain the loss function, such as the Huber loss function; the refinement model gradient is calculated by backpropagation using this loss function.

[0053] For example, the calculation process of the continuous casting model gradient is as follows: During the training of the continuous casting time prediction model, the difference between the predicted process time output by the continuous casting time prediction model and the labeled actual continuous casting time is compared to obtain the loss function, such as the Huber loss function; the gradient of the continuous casting model is calculated by backpropagation using this loss function.

[0054] Step S203: Input the steelmaking parameter set and the target steel grade semantic vector into the pre-constructed steelmaking time prediction model, and obtain the steelmaking process time through the steelmaking time prediction model.

[0055] The construction process of the steelmaking time prediction model is as follows: obtain sample data corresponding to multiple steel grades, namely {steelmaking parameter set, steel grade semantic vector, labeled steelmaking process time}; input the sample data corresponding to multiple steel grades into the steelmaking time prediction model, compare the predicted process time output by the steelmaking time prediction model with the labeled steelmaking process time to obtain the loss function; update the parameters in the steelmaking time prediction model based on the loss function.

[0056] Step S204: Input the refining parameter set and the target steel grade semantic vector into the pre-constructed refining time prediction model, and obtain the refining process time through the refining time prediction model.

[0057] The construction process of the refining time prediction model is as follows: obtain sample data corresponding to multiple steel grades, namely {refining parameter set, steel grade semantic vector, and labeled refining process time}; input the sample data corresponding to multiple steel grades into the refining time prediction model, compare the predicted process time output by the refining time prediction model with the labeled refining process time to obtain the loss function; update the parameters in the refining time prediction model based on the loss function.

[0058] Step S205: Input the continuous casting parameter set and the target steel grade semantic vector into the pre-constructed continuous casting time prediction model, and obtain the continuous casting process time through the continuous casting time prediction model.

[0059] The construction process of the continuous casting time prediction model is as follows: obtain sample data corresponding to multiple steel grades, namely {continuous casting parameter set, steel grade semantic vector, and labeled continuous casting process time}; input the sample data corresponding to multiple steel grades into the continuous casting time prediction model, compare the predicted process time output by the continuous casting time prediction model with the labeled continuous casting process time to obtain the loss function; update the parameters in the continuous casting time prediction model based on the loss function.

[0060] In this application, the steelmaking time prediction model, refining time prediction model, and continuous casting time prediction model share a common target steel grade semantic vector.

[0061] This application provides a process time prediction method, which obtains the steelmaking parameter set, refining parameter set, and continuous casting parameter set corresponding to the three processes of steelmaking, refining, and continuous casting for the target steel grade. The semantic vector of the target steel grade is obtained collaboratively based on the gradients of the steelmaking model, refining model, and continuous casting model. It can integrate the process semantic information of the three processes, accurately represent the process characteristics of the target steel grade in the entire production process, and provide unified and steel grade-specific shared feature support for the three process time prediction models, breaking down the information barriers between the prediction models of each process. The method also includes the acquisition of the specific parameter sets for each process (such as steelmaking, refining, and continuous casting). The steelmaking parameter set corresponding to the steelmaking process, the refining parameter set corresponding to the refining process, and the continuous casting parameter set corresponding to the continuous casting process are input into the pre-constructed steelmaking, refining, and continuous casting time prediction models corresponding to the target steel grade semantic vector. This achieves accurate matching between the model corresponding to each process and its own process characteristics, so that each model does not need to learn complex hybrid mapping functions and can focus on capturing the unique and most critical process features of the corresponding process. This effectively avoids the internal logic confusion caused by a single model predicting the time of multiple processes in the existing technology, and significantly improves the accuracy of time prediction for each process of steelmaking, refining, and continuous casting.

[0062] In one optional implementation, step S202 is specifically implemented by steps A1 to A3.

[0063] Step A1: Input the steel grade identifier of the target steel grade into the pre-built shared embedding layer.

[0064] Step A2: Obtain the target steel grade semantic vector corresponding to the steel grade identifier from the embedding weight matrix stored in the shared embedding layer. The embedding weight matrix contains steel grade semantic vectors corresponding to multiple steel grade identifiers.

[0065] like Figure 3 The diagram shown is a schematic of the shared embedding layer provided in an embodiment of this application.

[0066] like Figure 3 The shared embedding layer shown includes: an input layer 301, an embedding matrix core layer 302, and an output layer 303.

[0067] The target steel grade identifier is input to the input layer 301; the input layer 301 inputs the target steel grade identifier to the embedding matrix core layer 302; the embedding matrix core layer 302 queries the embedding weight matrix to obtain the target steel grade semantic vector corresponding to the target steel grade identifier; the embedding matrix core layer 302 inputs the target steel grade semantic vector to the output layer 303; the output layer 303 outputs the target steel grade semantic vector.

[0068] For example, the steel grade identifier is an integer.

[0069] Figure 3In this example, we take 45 types of steel as an example. The 45 steel type identifiers are steel type 1, steel type 2, ..., steel type 45. Therefore, the embedding weight matrix includes 45 rows, and each row corresponds to the steel type semantic vector of a steel type identifier. For example, row 1 (i.e., the first row) is the steel type semantic vector of steel type 1, row 2 (i.e., the second row) is the steel type semantic vector of steel type 2, ..., row 45 (i.e., the forty-fifth row) is the steel type semantic vector of steel type 45.

[0070] Figure 3 Taking an 8-dimensional semantic vector for steel grades as an example, the semantic vector for steel grades is: Each dimension of the semantic vector corresponds to a specific process semantic. Through training, steel grades with similar process characteristics can be positioned close together in the vector space.

[0071] In summary, the embedding weight matrix is: That is, 45 rows × 8 columns. For example, the embedding weight matrix has 360 trainable parameters.

[0072] Step A3: Obtain the semantic vector of the target steel grade output by the shared embedding layer.

[0073] For example, such as Figure 3 As shown, the semantic vector of the target steel grade includes the following 8-dimensional semantic vectors: Dimension 1: Degree of alloying, which characterizes the extent to which alloying elements are added.

[0074] For example, the alloying degree can be set to [0.0, 1.0]. If the alloying degree is 0.0, the target steel grade corresponds to low-carbon steel, meaning the total content of alloying elements (such as manganese, silicon, chromium, nickel, molybdenum, etc.) in the steel is extremely low (usually ≤0.25%), containing only basic iron and carbon elements. No complex alloying operations are required, and the smelting process is simple. When the alloying degree is between 0.0 and 1.0, it corresponds to medium-carbon steel, low-alloy steel, and medium-alloy steel, with the total content of alloying elements increasing with the value, and the complexity of the alloying operation gradually increasing. If the alloying degree is 1.0, the target steel grade corresponds to high-alloy steel, meaning the total content of alloying elements in the steel is relatively high (usually ≥10%), requiring the addition of multiple and large amounts of alloying elements, making the alloying operation complex. In summary, the alloying degree directly affects the process time.

[0075] Dimension 2: The cleanliness requirements of the target steel grade for the purity of inclusions and gas content in molten steel.

[0076] For example, the cleanliness requirement ranges from [0.0, 1.0]. If the cleanliness requirement is 0.0, the target steel grade corresponds to ordinary steel, which has low requirements for the content of inclusions, gases (such as oxygen and nitrogen), and harmful impurities (such as sulfur and phosphorus) in the molten steel, and does not require complex purification processes. If the cleanliness requirement is 1.0, the target steel grade corresponds to ultra-pure steel, which has extremely strict control over the content of inclusions and harmful impurities, and requires special processes such as vacuum degassing and deep refining. The cleanliness requirement directly affects the process time.

[0077] Dimension 3: Temperature sensitivity of the target steel grade as stated in the table to temperature fluctuations during smelting and casting.

[0078] For example, the temperature sensitivity range is [0.0, 1.0]. If the temperature sensitivity is 0.0, the target steel grade corresponds to a temperature-insensitive steel grade, and temperature fluctuations have little impact on the quality of molten steel and the progress of the process, and the temperature control accuracy requirement is low. If the temperature sensitivity is 1.0, the target steel grade corresponds to a highly sensitive steel grade, and small temperature fluctuations can easily lead to defects such as component segregation and billet cracks, requiring strict control of the temperature of each process, and longer temperature compensation and waiting time.

[0079] Dimension 4: The process difficulty coefficient, which characterizes the difficulty of process control in the entire production process of the target steel grade.

[0080] For example, the range of the process difficulty coefficient is [0.0, 1.0]. If the process difficulty coefficient is 0.0, the target steel grade corresponds to easy-to-make steel, which is simple to control in terms of composition, easy to remove impurities, and has no special process requirements. The operation process of each step is simple. If the process difficulty coefficient is 1.0, the target steel grade corresponds to difficult-to-make steel, which requires precise control of composition uniformity and temperature stability, and involves special metallurgical processes (such as vacuum refining, controlled rolling and controlled cooling), and has high operation difficulty.

[0081] Dimension 5: Solidification characteristics, which are quantitative indicators representing the width of the solidification range of the target steel grade in the continuous casting process.

[0082] For example, the solidification characteristic value ranges from [0.0, 1.0]. If the solidification characteristic value is 0.0, the target steel grade corresponds to a narrow solidification range steel grade. This type of steel grade has a small temperature difference between the liquidus and solidus, resulting in a uniform and stable solidification process with a consistent solidification rate. Higher casting speeds can be used during continuous casting, leading to high casting efficiency and short continuous casting time. If the solidification characteristic value is between 0.0 and 1.0, the target steel grade corresponds to a medium solidification range steel grade. The temperature difference between the liquidus and solidus increases with the value, and the solidification rate gradually becomes uneven. If the solidification characteristic is set to 1.0, the target steel grade corresponds to a steel grade with a wide solidification range. This type of steel grade has a large temperature difference between the liquidus and solidus, and the solidification process is uneven. There is a significant temperature difference between the first solidification area and the last solidification area, which easily leads to defects such as central shrinkage cavities, compositional segregation, and hot cracks. During continuous casting, it is necessary to reduce the casting speed, optimize the secondary cooling regime, and extend the solidification cooling time, thereby increasing the continuous casting pouring time.

[0083] Dimension 6: Cost level that characterizes the production and manufacturing cost of the target steel grade.

[0084] For example, the cost level ranges from [0.0, 1.0]. If the cost level is 0.0, the target steel grade corresponds to a low-cost steel grade. In the production process of this type of steel grade, the raw materials (molten iron, scrap steel) are conventional and inexpensive, the amount of alloying elements added is minimal (or unnecessary), the smelting, refining, and continuous casting processes are simple, energy consumption is low, quality control requirements are relaxed, and the overall production cost is the lowest, without the need for additional complex processes or expensive materials. If the cost level is between 0.0 and 1.0, the target steel corresponds to a medium-cost steel grade. Its production cost gradually increases with the value. The raw materials can be conventional or slightly optimized, a small amount of alloying elements need to be added, the process complexity is moderate, energy consumption and quality control costs are moderate, and the overall cost is between low and high cost, suitable for conventional industrial use needs. If the cost grade is 1.0, the target steel corresponds to a high-cost steel grade. The production of this type of steel requires the use of high-quality, high-priced raw materials (such as high-purity molten iron and special scrap steel), and the addition of a variety of high-priced alloying elements (such as chromium, nickel, molybdenum, etc.). The smelting and refining processes are complex (such as adding special processes such as vacuum degassing and deep desulfurization), resulting in high energy consumption. Furthermore, it has stringent requirements for quality control throughout the entire process and high testing costs, making it the highest in overall production and manufacturing costs. It is used for high-end and special application scenarios.

[0085] Dimension 7: Characterizes the historical stability of the target steel grade in terms of operation time fluctuations and quality stability during historical production processes.

[0086] For example, the historical stability value ranges from [0.0, 1.0]. If the historical stability value is 0.0, the target steel grade corresponds to a steel grade with high historical stability. In the past batch production of this type of steel grade, the operation time of each process in steelmaking, refining, and continuous casting has a very small fluctuation range, the production process parameters are well adapted, the quality pass rate is high, the occurrence rate of abnormal conditions (such as excessive composition or billet defects) is extremely low, and the production process is highly controllable. If the historical stability value is between 0.0 and 1.0, the target steel grade corresponds to a steel grade with medium historical stability. In its past production, the operation time of each process has a certain fluctuation (the fluctuation range increases with the value), there are occasional minor abnormal conditions, the quality pass rate is at a medium level, and the production process is basically controllable. If the historical stability value is 1.0, the target steel grade corresponds to a steel grade with poor historical stability and large fluctuations. In the past production of this type of steel grade, the operation time of each process fluctuated greatly, the process parameters were poorly adapted, the occurrence rate of abnormal working conditions (such as temperature runaway, composition fluctuation, and production interruption caused by equipment failure) was high, the quality pass rate was unstable, and the controllability of the production process was weak.

[0087] Dimension 8: The overall quality grade that characterizes the final product quality grade of the target steel grade.

[0088] For example, the comprehensive quality grade ranges from [0.0, 1.0]. If the comprehensive quality grade is 0.0, the target steel grade corresponds to a common steel grade. This type of steel grade has lenient requirements for the core quality indicators of the final product, only needing to meet basic usage requirements, without strict control over compositional uniformity, microstructure, and surface quality. If the comprehensive quality grade is between 0.0 and 1.0, the target steel grade corresponds to a medium-quality steel grade, with its quality requirements gradually increasing as the value increases. If the comprehensive quality grade is 1.0, the target steel grade corresponds to a premium steel grade, which has extremely high comprehensive quality requirements for the final product.

[0089] For example, the semantic vector of the target steel grade is [0.23, 0.87, 0.15, 0.62, 0.41, 0.35, 0.78, 0.29].

[0090] The shared embedding layer is independent of the steelmaking time prediction model, refining time prediction model, and continuous casting time prediction model. The shared embedding layer maps steel grade identifiers to a low-dimensional dense vector. Through joint training of the three models, the shared embedding layer learns to generate a high-order steel grade semantic vector. This semantic vector integrates the comprehensive characteristics of the steel grade throughout the entire smelting process (such as metallurgical reaction intensity, composition control precision, and solidification quality sensitivity), forming a deep and compact numerical representation of the steel grade's essence. This representation not only significantly improves the prediction accuracy of each specialized model, but more importantly, it establishes a transferable steel grade process knowledge base. This allows the system to reliably perform analogical reasoning based on the proximity of the steel grade's position in this semantic vector space for sparse or entirely new steel grades, fundamentally enhancing the generalization ability and practical value of the prediction system. This fundamentally solves the problem of shallow steel grade feature representation.

[0091] The process of obtaining the embedding weight matrix in the shared embedding layer is described below. This process includes steps B1 to B3.

[0092] Step B1: For each steel grade, input the steelmaking parameter set corresponding to the steel grade into the steelmaking time prediction model, and obtain the steelmaking model gradient through the steelmaking time prediction model; input the refining parameter set corresponding to the steel grade into the refining time prediction model, and obtain the refining model gradient through the refining time prediction model; input the continuous casting time prediction model corresponding to the steel grade into the continuous casting time prediction model, and obtain the continuous casting model gradient through the continuous casting time prediction model.

[0093] Step B2: For each steel grade, obtain the steel grade semantic vector based on the gradient of the steelmaking model, the gradient of the refining model, and the gradient of the continuous casting model corresponding to the steel grade.

[0094] For example, there are multiple ways to implement step B2. This application provides, but is not limited to, the following method, which includes steps B21 to B23.

[0095] Please see Figure 4 This is a schematic diagram illustrating the process of obtaining the semantic vector of steel type provided in the embodiments of this application.

[0096] Step B21: Calculate the gradient fusion value ▽_total using the fusion formula α×▽1+β×▽2+γ×▽3, where α+β+γ=1, α, β, and γ are preset weights; ▽1 is the gradient of the steelmaking model; ▽2 is the gradient of the refining model; and ▽3 is the gradient of the continuous casting model.

[0097] For example, the preset weights for the steelmaking time prediction model are α=0.3, the preset weights for the refining time prediction model are β=0.5, and the preset weights for the continuous casting time prediction model are γ=0.2.

[0098] For example, the preset weights of the refining time prediction model are greater than those of the steelmaking time prediction model, which in turn are greater than those of the continuous casting time prediction model. This is because the refining process is most sensitive to the characteristics of the steel grade; the steelmaking process mainly affects the initial reaction process; and the continuous casting process is relatively stable and less affected by the steel grade.

[0099] Step B22: Input the gradient fusion value into the Adam optimizer to obtain the optimized gradient.

[0100] Step B23: Update formula W new =W old -η×Adam( The semantic vector of the steel type is calculated using the total), where Adam( (total) is the optimization gradient, W old W is the semantic vector of the steel type before the update. new This is the updated semantic vector for steel grades.

[0101] For example, η is the learning rate, for example, η=0.001, and the momentum parameters of the Adam optimizer include, but are not limited to: first moment decay rate β1=0.9, second moment decay rate β2=0.999.

[0102] Step B3: Obtain the embedding weight matrix, which is composed of the semantic vectors of the steel grades corresponding to the multiple steel grades.

[0103] It is understandable that as time goes by, the steelmaking time prediction model, refining time prediction model, and continuous casting time prediction model are constantly updated, so the gradients ▽1, ▽2, and ▽3 of the steelmaking model are constantly updated; the embedded weight matrix can be updated based on the updated gradients ▽1, ▽2, and ▽3 of the steelmaking model, refining model, and continuous casting model.

[0104] In this application, the steelmaking time prediction model, refining time prediction model, and continuous casting time prediction model share the same shared embedding layer, which reduces the number of parameters.

[0105] In one optional implementation, the steelmaking time prediction model is a gated linear unit network (GLU); the refining time prediction model is a Transformer model; and the continuous casting time prediction model is a multilayer perceptron.

[0106] To help those skilled in the art better understand the embodiments of this application, examples are provided below. Please refer to... Figure 5 This is a process time prediction architecture diagram provided in the embodiments of this application.

[0107] For example, the steelmaking time prediction model 501 includes: an input layer, a first GLU layer, a second GLU layer, and an output network.

[0108] For example, the refinement time prediction model 502 includes an input layer, a TabTransformer, a decoder, and an output network. The refinement time prediction model uses a TabTransformer based on tabular data. Among the features of the refinement process, the refinement parameter set has rich semantics, and the TabTransformer utilizes a self-attention mechanism to learn deep, context-dependent interactions between these refinement parameter sets.

[0109] For example, the continuous casting time prediction model 503 includes: an input layer, a first robust MLP layer, a second robust MLP layer, a third robust MLP layer, residual connections, and an output network. The continuous casting time prediction model in this application uses a robust MLP (Robust Multi-Layer Perceptron), which makes the continuous casting time prediction model simple in structure, fast in training, easy to understand, and easy to deploy.

[0110] For example, such as Figure 3 As shown, the shared embedding layer 504 includes an input layer, an embedding matrix core layer, and an output layer.

[0111] It is understandable that the steelmaking time prediction model 501, refining time prediction model 502, continuous casting time prediction model 503, and shared embedding layer 504 can be trained simultaneously. After training, the embedding weight matrix in the shared embedding layer is determined.

[0112] The following is about Figure 5 The detailed structure of the steelmaking time prediction model 501 shown in the figure is explained below. Figure 6 The diagram shown is a detailed structural diagram of the steelmaking time prediction model 501 provided in the embodiments of this application.

[0113] Based on Table 1, the steelmaking parameter set includes: equipment and material characteristics, i.e., the 4-dimensional converter number, which is obtained by one-hot encoding the converter identifier; and 7-dimensional process parameters.

[0114] The steelmaking parameter set and the 8-dimensional target steel grade semantic vector are concatenated to obtain a 19-dimensional input feature vector x=[x_device, x_embed, x_process], where x_device is the 4-dimensional converter number, x_embed is the 8-dimensional target steel grade semantic vector, and x_process is the 7-dimensional process parameter.

[0115] For example, the concatenation operation between the steelmaking parameter set and the 8-dimensional semantic vector of the target steel grade is as follows: .

[0116] from Figure 6 It can be seen that the input feature vector x = [x_device, x_embed, x_process] is input into the first GLU layer 601. The first GLU layer 601 includes a fully connected layer, a feature splitting layer, a gated computation layer, and a Dropout regularization layer.

[0117] The following describes the fully connected layer, feature splitting layer, gated computation layer, and Dropout regularization layer in the first GLU layer 601.

[0118] The fully connected layer in the first GLU layer 601 linearly transforms the 19-dimensional input feature vector into a 128-dimensional vector. Specifically, this can be achieved using the linear transformation formula Z1 = W1 × x + b1, where... , , .

[0119] The feature splitting layer can split the 128-dimensional vector Z1 into two 64-dimensional vectors A1 and B1: A1, B1 = split(Z1). , .

[0120] After the gated computation layer B1 is activated by the Sigmoid function, a gated weight vector h1 is generated, and the value range of h1 is [0,1]. ,in: For the sigmoid function, This represents element-wise multiplication (Hadamard product). A1 is the result of the original feature transformation. By multiplying element by element, B1 controls which features in A1 should be emphasized or suppressed.

[0121] Dropout regularization can be achieved through formulas. Regularize h1.

[0122] The first GLU layer 601 output Input to the second GLU layer 602.

[0123] The second GLU layer 602 includes a fully connected layer, a feature splitting layer, and a gated computation layer. The fully connected layer, feature splitting layer, and gated computation layer in the second GLU layer are described below.

[0124] Fully connected layers can handle 64-dimensional... A linear transformation is performed to obtain a 64-dimensional vector. For example, this can be achieved through the formula... Perform a linear transformation. , , .

[0125] Feature splitting layer through formula , , Z2 is split into A2 and B2, which are 32-dimensional units.

[0126] Gated computing layer through formula , The 32-dimensional feature h2 is calculated.

[0127] like Figure 6 As shown, the output network 603 includes a first fully connected layer, a second fully connected layer, and an output layer. The layers of the output network 603 are described below.

[0128] The first fully connected layer can transform a 32-dimensional vector into a 16-dimensional vector h3 using a formula. For example, the formula is: ,in: , , , .

[0129] For example, the first fully connected layer and the second fully connected layer can employ the ReLU activation function.

[0130] The second fully connected layer can transform a 16-dimensional vector into an 8-dimensional vector h4 using a formula. For example, the formula is: ,in: , , .

[0131] The output network can convert an 8-dimensional vector into a 1-dimensional vector using a formula. For example, the formula is: ,in: , , (Unit: minutes)

[0132] For example, in training a steelmaking time prediction model, the following loss function can be used: Total loss function: .

[0133] MSE loss: .

[0134] L2 regularization: ,in , (Regularization coefficient).

[0135] Gradient backpropagation: All parameters are updated by backpropagating layer by layer using the chain rule.

[0136] The following is about Figure 5 The detailed structure of the refining time prediction model 502 shown in the figure is explained below. For example... Figure 7 The diagram shown is a detailed structural diagram of the refining time prediction model 502 provided in the embodiments of this application.

[0137] For example, the refining time prediction model is the Transformer model.

[0138] As shown in Table 2, the input features of the refining time prediction model 502 include: data feature vector and category feature vector; where the data feature vector is... Among them, among them: This includes initial component characteristics, target component characteristics, and excipient addition characteristics. This includes derived features and the alloy content of molten steel (i.e., the difference between the target alloy composition content that needs to be achieved after the refining process of the target steel grade and the initial alloy composition content of the molten steel after the steelmaking process).

[0139] Derivative characteristics include, but are not limited to: carbon gap, silicon gap, manganese gap, phosphorus gap, sulfur gap, and temperature gap.

[0140] The carbon gap is calculated as: target carbon content - initial carbon content. The initial carbon content is the carbon content of the molten steel after the steelmaking process, and the target carbon content is the carbon content that the molten steel needs to reach after the refining process. The carbon gap reflects the adjustment requirements for carbon content during the refining process (positive values ​​indicate the need for carbon increase, negative values ​​indicate the need for carbon decarburization).

[0141] Silicon gap = Target silicon content - Initial silicon content. Wherein, initial silicon content is the silicon content in the molten steel after the steelmaking process, target silicon content is the silicon content in the molten steel after the refining process, and carbon gap is used to characterize the adjustment range of silicon content during the refining process, guiding the control of alloy addition.

[0142] Manganese gap = target manganese content - initial manganese content; where the initial manganese content is the manganese content in the molten steel after the steelmaking process, and the target manganese content is the manganese content in the molten steel after the refining process. The manganese gap reflects the need for the addition or removal of manganese alloys in the refining process and is directly related to the refining operation time.

[0143] Phosphorus gap = target phosphorus content - initial phosphorus content; where the initial phosphorus content is the phosphorus content in the molten steel after the steelmaking process, and the target phosphorus content is the phosphorus content that the molten steel needs to reach after the refining process. Since phosphorus is a harmful impurity, the absolute value of the phosphorus gap reflects the difficulty of deep dephosphorization in the refining process.

[0144] Sulfur gap = target sulfur content - initial sulfur content; where the initial sulfur content is the sulfur content in the molten steel after the steelmaking process, and the target sulfur content is the sulfur content in the molten steel after the refining process. The sulfur gap reflects the need and difficulty of deep desulfurization in the refining process and affects the refining operation time.

[0145] Temperature difference = target temperature - initial temperature; where the initial temperature is the temperature of the molten steel entering the station after the steelmaking process, and the target temperature is the temperature that the molten steel needs to reach when leaving the station after the refining process. The temperature difference reflects the need for temperature compensation or cooling of the molten steel in the refining process and directly affects the duration of temperature adjustment operations.

[0146] Category feature vector is ,in: Number the refining furnace. The semantic vector for the target steel grade.

[0147] For example, the data feature vector can be a 20-dimensional feature matrix; the category feature vector can be an 11-dimensional feature vector.

[0148] like Figure 7 As shown, before inputting the data feature vector into the refined time prediction model, it needs to be processed by MLP701. The MLP processing includes: transforming the 20-dimensional data feature vector into a 32-dimensional vector through a fully connected layer, i.e. ,in , Batch normalization is used to normalize 32-dimensional vectors, adjusting them to a uniform distribution range. Batch average, Batch variance , , Learnable parameters. Then, a 32-dimensional vector is output through the ReLU activation function, i.e. ,in , .

[0149] The categorical feature vector is input into TabTransformer702. TabTransformer includes, but is not limited to: linear projection, positional encoding, first Transformer encoder, second Transformer encoder, and global average pooling layer.

[0150] The linear projection layer can transform an 11-dimensional class feature vector into a 32-dimensional vector, as shown in the following formula: ,in , .

[0151] The position coding layer can be sine-cosine coded. An example is given below, assuming the position is... That is, refining furnace number = 0, target steel grade = 1. The position coding formula is as follows: , .in, This is the position coding matrix. The output of the position coding layer is: .

[0152] Both the first and second Transformer encoders consist of the following layers in sequence: a 4-head self-attention layer (8 dimensions / head), residual connections and a LayerNorm layer, a feedforward network layer (32-dimensional vector → 64-dimensional vector → 32-dimensional vector), residual connection layers and a LayerNorm layer, and Dropout: 0.1. Dropout: 0.1 means that during model training, in each iteration, 10% of the neurons in the current layer are randomly discarded, leaving only 90% of the neurons functioning normally.

[0153] Dimension of each head in the 4-head self-attention layer For the h-th head (h = 1, 2, 3, 4): , , ;in , , This represents the query weight matrix (Q), key weight matrix (K), and value weight matrix (V). Each head yields the following characteristics: , express The transpose of the matrix. The features from the four heads are concatenated to obtain: ,in .

[0154] Residual connections and LayerNorm layers are connected via Normalize.

[0155] The calculation process of the feedforward network layer is as follows: ,in , , , .

[0156] The GELU activation function is .

[0157] The output of the first Transformer encoder is: , .

[0158] The processing procedure of the second Transformer encoder is similar to that of the second Transformer encoder, and will not be repeated here. The output of the second Transformer encoder is: , .

[0159] The global average pooling layer is used to average the features at two locations. .

[0160] like Figure 7 As shown, the feature fusion layer 703 fuses the features output by MLP 701 and the features output by TabTransformer 702: .

[0161] like Figure 7 As shown, the refining time prediction model may include one or more decoders 704; multiple decoders 704 may be connected in series; the following explanation uses three decoders connected in series as an example.

[0162] The first decoder passes Convert a 64-dimensional vector to a 128-dimensional vector; , .

[0163] The second decoder passes through Convert a 128-dimensional vector to a 64-dimensional vector; , .

[0164] The third decoder passes through Convert a 64-dimensional vector to a 32-dimensional vector; .

[0165] like Figure 7 As shown, the refining time prediction model also includes an output network 705. The output network 705 consists of two fully connected layers and an output layer.

[0166] The first fully connected layer passes This transforms a 32-dimensional vector into a 16-dimensional vector. The second fully connected layer then... Transform a 16-dimensional vector into an 8-dimensional vector. The output network then... Convert an 8-dimensional vector into a 1-dimensional value. Wherein, ∈ The refining process time (in minutes).

[0167] During the training of the refined time prediction model, it can be trained using the Huber loss function (δ = 1.0).

[0168] The Huber loss function is: The threshold δ = 1.0 for the refining time prediction model.

[0169] The calculation process of the refining model gradient of the refining time prediction model is as follows: For : ,for : .

[0170] During the training of the refined time prediction model, the Adam optimizer (lr=0.001, wd=0.01) can be used for optimization and updates. The specific process is as follows: First moment estimation Second-order moment estimation , bias correction , Parameter update , where the parameters are: = 0.001 (learning rate), = 0.01 (weight decay), = 0.9, = 0.999, t = number of training steps.

[0171] The following is about Figure 5 The detailed structure of the continuous casting time prediction model 503 shown in the figure is explained below. For example... Figure 8 The diagram shown is a detailed structural diagram of the continuous casting time prediction model 503 provided in the embodiment of this application.

[0172] According to Table 3, the set of continuous casting parameters input to the continuous casting time prediction model 503 includes: 15-dimensional equipment specifications, 4-dimensional molten steel condition characteristics, and 5-dimensional process parameter characteristics.

[0173] Equipment specifications refer to the set of specifications and parameters of the core equipment involved in the continuous casting process, including but not limited to: Intermediate ladle number: refers to the unique identifier of the intermediate container used in the continuous casting process to receive refined qualified molten steel, stabilize the molten steel flow rate, and divert it to the crystallizer. Each intermediate ladle corresponds to a fixed capacity, material, and other specifications. Its identifier can be linked to query the core parameters of the intermediate ladle and affects the stability of molten steel pouring.

[0174] Crystallizer number: refers to the unique identifier of the core equipment (crystallizer) that realizes the solidification and forming of molten steel in the continuous casting process. The crystallizer is the key component for the transformation of molten steel from liquid to solid billet. Its identifier can be linked to query the core specifications of the crystallizer such as cross-sectional dimensions, length, and cooling method, which directly determines the billet forming efficiency.

[0175] Continuous casting machine number: This refers to the unique identifier of the entire continuous casting equipment used to complete the pouring of molten steel, straightening of the billet, and cooling. Each continuous casting machine corresponds to fixed parameters such as production capacity, casting speed range, and billet specification compatibility. Its identifier can be associated with the overall operating parameters of the continuous casting machine and affect the operating efficiency of the entire continuous casting process.

[0176] Slab specification characteristics: These refer to the core parameters that characterize the final slab produced by continuous casting, such as its size and shape.

[0177] The characteristics of molten steel refer to the state parameters of molten steel after it has been refined and is ready for continuous casting. These parameters directly affect the solidification rate of molten steel, the quality of the billet, and the difficulty of continuous casting operation, which in turn affects the continuous casting process time.

[0178] Process parameter characteristics refer to the core operating parameters set in the continuous casting process to achieve smooth pouring of molten steel and qualified forming of billets.

[0179] 15-dimensional equipment specifications 4-dimensional characteristics of molten steel conditions 5-dimensional process parameter characteristics 8-dimensional target steel grade semantic vector A 32-dimensional vector can be obtained by concatenation. .

[0180] It can handle 32-dimensional vectors Perform Z-score normalization, such as for Each feature conduct Processing, among which , Features on the training set The mean and standard deviation.

[0181] Combination Figure 8 As shown, the continuous casting time prediction model includes: a first robust MLP801, a second robust MLP802, a third robust MLP803, a residual connection 804, and an output network 805.

[0182] The execution process of the first robust MLP 801 is as follows: through the fully connected layer Convert the 32-dimensional vector to 64 dimensions; perform batch normalization using BatchNorm. (Batch average) (Batch variance) (Normalization) (Scaling and panning) , These are learnable parameters. Activated using the ReLU function. Processing is performed; via Dropout 0.15, Process it.

[0183] The execution process of the second robust MLP 802 is as follows: through the fully connected layer Convert a 64-dimensional vector to a 32-dimensional vector using BatchNorm. Perform batch normalization; activate using the ReLU function. Processing is performed; via Dropout 0.10 layer. Process it.

[0184] The execution process of the third robust MLP 803 is as follows: a 32-dimensional vector is transformed into a 16-dimensional vector through a fully connected layer. Activated by ReLU function Process it.

[0185] Execution via residual join ,in These are residual connection weights. =0.1.

[0186] The execution process of the output network 805 in the continuous casting time prediction model is as follows: The first fully connected layer passes through... This transforms a 16-dimensional vector into an 8-dimensional vector; the second fully connected layer... This transforms an 8-dimensional vector into a 4-dimensional vector; the output layer... This converts a 4-dimensional vector into a 1-dimensional numerical value, which represents the continuous casting process time.

[0187] During the training of the continuous casting time prediction model, it is possible to use methods such as... Figure 9 The training process is shown below.

[0188] like Figure 9 As shown, the formula for the Huber loss function for a single sample data is as follows: ,in (The threshold of the continuous casting model is smaller than that of the refining model.)

[0189] The gradient of the Huber loss is calculated as follows:

[0190] The formula for calculating the total loss function is as follows: ,in: (L2 regularization coefficient) (Weight decay coefficient) For all trainable parameters.

[0191] The AdamW optimizer update rules are as follows:

[0192] First-order moment estimation: .

[0193] Second-order moment estimation: .

[0194] Bias correction: .

[0195] Parameter update: .in: (learning rate) , , .

[0196] The early stopping mechanism refers to monitoring the validation set loss L_val; if L_val does not decrease beyond a threshold for P consecutive epochs, the early stopping mechanism will stop the test. Stop training and restore the best model, where P=20 (patience value). (Minimum improvement threshold).

[0197] In this application, the steelmaking time prediction model, refining time prediction model, and continuous casting time prediction model are trained completely independently. Each model uses its own loss function (such as MSE) and optimizer (such as Adam). When training a particular model, its error backpropagation simultaneously updates the model's own network weights and the corresponding parts of the shared embedding layers.

[0198] The above describes a process time prediction method provided by the embodiments of this application. The following describes the apparatus for performing the above process time prediction method.

[0199] Please see Figure 10 , Figure 10 This is a schematic diagram of a process time prediction device provided in an embodiment of this application. Figure 10 As shown, the process time prediction device includes: The first acquisition module 1001 is used to acquire the steelmaking parameter set for the steelmaking process, the refining parameter set for the refining process, and the continuous casting parameter set for the continuous casting process for the target steel grade. The first search module 1002 is used to search for the target steel grade semantic vector corresponding to the target steel grade from the preset steel grade and steel grade semantic vector. The target steel grade semantic vector is obtained based on the steelmaking model gradient representing the loss of the steelmaking time prediction model, the refining model gradient representing the loss of the refining time prediction model, and the continuous casting model gradient representing the loss of the continuous casting time prediction model. The second acquisition module 1003 is used to input the steelmaking parameter set and the target steel grade semantic vector into the pre-constructed steelmaking time prediction model, and obtain the steelmaking process time through the steelmaking time prediction model. The third acquisition module 1004 is used to input the refining parameter set and the target steel grade semantic vector into the pre-constructed refining time prediction model, and obtain the refining process time through the refining time prediction model. The fourth acquisition module 1005 is used to input the continuous casting parameter set and the target steel grade semantic vector into the pre-constructed continuous casting time prediction model, and obtain the continuous casting process time through the continuous casting time prediction model.

[0200] In one alternative implementation, the first lookup module includes: An input unit is used to input the steel grade identifier of the target steel grade into a pre-built shared embedding layer; The first acquisition unit is used to acquire the target steel grade semantic vector corresponding to the steel grade identifier from the embedding weight matrix stored in the shared embedding layer. The embedding weight matrix contains steel grade semantic vectors corresponding to multiple steel grade identifiers. The second acquisition unit is used to acquire the semantic vector of the target steel type output by the shared embedding layer.

[0201] In one alternative implementation, it also includes: The fifth acquisition module is used to, for each steel grade, input the steelmaking parameter set corresponding to the steel grade into the steelmaking time prediction model, and obtain the steelmaking model gradient through the steelmaking time prediction model; input the refining parameter set corresponding to the steel grade into the refining time prediction model, and obtain the refining model gradient through the refining time prediction model; input the continuous casting time prediction model corresponding to the steel grade into the continuous casting time prediction model, and obtain the continuous casting model gradient through the continuous casting time prediction model. The sixth acquisition module is used to obtain the semantic vector of the steel grade corresponding to each steel grade based on the gradient of the steelmaking model, the gradient of the refining model and the gradient of the continuous casting model corresponding to the steel grade. The seventh acquisition module is used to acquire the embedding weight matrix composed of the semantic vectors of steel grades corresponding to multiple steel grades.

[0202] In one optional implementation, the sixth acquisition module includes: The first calculation unit is used to calculate the gradient fusion value using the fusion formula α×▽1+β×▽2+γ×▽3, where α+β+γ=1, α, β, and γ are preset weights; ▽1 is the gradient of the steelmaking model; ▽2 is the gradient of the refining model; and ▽3 is the gradient of the continuous casting model. An optimization unit is used to input the gradient fusion value into the Adam optimizer to obtain the optimized gradient; The second calculation unit is used to update formula W. new =W old -η×Adam( The semantic vector of the steel type is calculated using the total), where Adam( (total) is the optimization gradient, W old W is the semantic vector of the steel type before the update. new This is the updated semantic vector for steel grades.

[0203] In one alternative implementation, the target steel grade semantic vector includes: The alloying degree, representing the extent of alloying element addition; the cleanliness requirement, representing the purity index of inclusions and gas content in molten steel for the target steel grade; the temperature sensitivity, representing the sensitivity of the target steel grade to temperature fluctuations during smelting and casting; the process difficulty coefficient, representing the difficulty of process control in the entire production process of the target steel grade; the solidification characteristics, representing the solidification range width of the target steel grade in the continuous casting process; the cost level, representing the production cost of the target steel grade; the historical stability, representing the fluctuation of operating time and quality stability of the target steel grade in historical production processes; and the comprehensive quality level, representing the final product quality grade of the target steel grade.

[0204] In one optional implementation, the steelmaking time prediction model is a gated linear unit network; the refining time prediction model is a Transformer model; and the continuous casting time prediction model is a multilayer perceptron.

[0205] This application also provides an electronic device in its embodiments. (See reference...) Figure 11 The diagram illustrates a structural schematic suitable for implementing the electronic device in the embodiments of this application. The electronic device in the embodiments of this application may include, but is not limited to, fixed terminals such as mobile phones, laptops, PDAs (personal digital assistants), PADs (tablet computers), desktop computers, etc. Figure 11The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0206] like Figure 11 As shown, the electronic device may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) 1101, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1102 or a program loaded from a storage device 1108 into a random access memory (RAM) 1103. When the electronic device is powered on, the RAM 1103 also stores various programs and data required for the operation of the electronic device. The processing unit 1101, ROM 1102, and RAM 1103 are interconnected via a bus 1104. An input / output (I / O) interface 1105 is also connected to the bus 1104.

[0207] Typically, the following devices can be connected to I / O interface 1105: input devices 1106 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 1107 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1108 including, for example, memory cards, hard drives, etc.; and communication devices 1109. Communication device 1109 allows electronic devices to exchange data via wireless or wired communication with other devices. Although Figure 11 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown. More or fewer devices may be implemented or have alternatively.

[0208] This application also provides a computer program product including computer-readable instructions, which, when executed on an electronic device, cause the electronic device to implement any of the process time prediction methods provided in this application.

[0209] This application also provides a computer-readable storage medium carrying one or more computer programs. When the one or more computer programs are executed by an electronic device, the electronic device can implement any of the process time prediction methods provided in this application.

[0210] It should also be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. In addition, in the device embodiment drawings provided in this application, the connection relationship between modules indicates that they have a communication connection, which can be implemented as one or more communication buses or signal lines.

[0211] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware, or it can be implemented by special-purpose hardware including application-specific integrated circuits, special-purpose CPUs, special-purpose memory, special-purpose components, etc. Generally, any function performed by a computer program can be easily implemented by corresponding hardware, and the specific hardware structure used to implement the same function can also be diverse, such as analog circuits, digital circuits, or special-purpose circuits. However, for this application, software program implementation is more often the preferred implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a readable storage medium, such as a computer floppy disk, USB flash drive, mobile hard disk, ROM, RAM, magnetic disk, or optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, training equipment, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0212] In the above embodiments, the implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, in the form of a computer program product.

[0213] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, training device, or data center to another website, computer, training device, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a training device or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives (SSDs)).

Claims

1. A method for predicting process time, characterized in that, include: Obtain the steelmaking parameter set for the steelmaking process, the refining parameter set for the refining process, and the continuous casting parameter set for the continuous casting process for the target steel grade; The target steel grade semantic vector is found from the preset steel grade and steel grade semantic vector. The target steel grade semantic vector is obtained based on the steelmaking model gradient representing the loss of the steelmaking time prediction model, the refining model gradient representing the loss of the refining time prediction model, and the continuous casting model gradient representing the loss of the continuous casting time prediction model. The steelmaking parameter set and the target steel grade semantic vector are input into the pre-constructed steelmaking time prediction model, and the steelmaking process time is obtained through the steelmaking time prediction model. The refining parameter set and the target steel grade semantic vector are input into the pre-constructed refining time prediction model, and the refining process time is obtained through the refining time prediction model. The continuous casting parameter set and the semantic vector of the target steel grade are input into the pre-constructed continuous casting time prediction model, and the continuous casting process time is obtained through the continuous casting time prediction model.

2. The process time prediction method according to claim 1, characterized in that, The step of searching for the target steel grade semantic vector corresponding to the target steel grade from the preset steel grade and steel grade semantic vector includes: Input the steel grade identifier of the target steel grade into the pre-built shared embedding layer; The target steel grade semantic vector corresponding to the steel grade identifier is obtained from the embedding weight matrix stored in the shared embedding layer. The embedding weight matrix contains steel grade semantic vectors corresponding to multiple steel grade identifiers. Obtain the semantic vector of the target steel grade output by the shared embedding layer.

3. The process time prediction method according to claim 2, characterized in that, The method for obtaining the embedding weight matrix includes: For each steel grade, the steelmaking parameter set corresponding to the steel grade is input into the steelmaking time prediction model, and the steelmaking model gradient is obtained through the steelmaking time prediction model; the refining parameter set corresponding to the steel grade is input into the refining time prediction model, and the refining model gradient is obtained through the refining time prediction model; the continuous casting time prediction model corresponding to the steel grade is input into the continuous casting time prediction model, and the continuous casting model gradient is obtained through the continuous casting time prediction model. For each steel grade, a semantic vector for the steel grade is obtained based on the gradient of the steelmaking model, the gradient of the refining model, and the gradient of the continuous casting model corresponding to the steel grade. Obtain the embedding weight matrix, which is composed of the semantic vectors of the steel grades corresponding to the multiple steel grades.

4. The process time prediction method according to claim 3, characterized in that, The step of obtaining the semantic vector of the steel grade based on the gradient of the steelmaking model, the gradient of the refining model, and the gradient of the continuous casting model corresponding to the steel grade includes: The gradient fusion value is calculated using the fusion formula α×▽1+β×▽2+γ×▽3, where α+β+γ=1, and α, β, and γ are preset weights; ▽1 is the gradient of the steelmaking model; ▽2 is the gradient of the refining model; and ▽3 is the gradient of the continuous casting model. The gradient fusion value is input into the Adam optimizer to obtain the optimized gradient; By updating formula W new =W old -η×Adam( The semantic vector of the steel type is calculated using the total), where Adam( (total) is the optimization gradient, W old W is the semantic vector of the steel type before the update. new This is the updated semantic vector for steel grades.

5. The process time prediction method according to any one of claims 1 to 4, characterized in that, The semantic vector of the target steel grade includes: The alloying degree, representing the extent of alloying element addition; the cleanliness requirement, representing the purity index of inclusions and gas content in molten steel for the target steel grade; the temperature sensitivity, representing the sensitivity of the target steel grade to temperature fluctuations during smelting and casting; the process difficulty coefficient, representing the difficulty of process control in the entire production process of the target steel grade; the solidification characteristics, representing the solidification range width of the target steel grade in the continuous casting process; the cost level, representing the production cost of the target steel grade; the historical stability, representing the fluctuation of operating time and quality stability of the target steel grade in historical production processes; and the comprehensive quality level, representing the final product quality grade of the target steel grade.

6. The process time prediction method according to any one of claims 1 to 4, characterized in that, The steelmaking time prediction model is a gated linear unit network; the refining time prediction model is a Transformer model; and the continuous casting time prediction model is a multilayer perceptron.

7. A process time prediction device, characterized in that, include: The first acquisition module is used to acquire the steelmaking parameter set for the steelmaking process, the refining parameter set for the refining process, and the continuous casting parameter set for the continuous casting process for the target steel grade. The first search module is used to search for the target steel grade semantic vector corresponding to the target steel grade from the preset steel grade and steel grade semantic vector. The target steel grade semantic vector is obtained based on the steelmaking model gradient representing the loss of the steelmaking time prediction model, the refining model gradient representing the loss of the refining time prediction model, and the continuous casting model gradient representing the loss of the continuous casting time prediction model. The second acquisition module is used to input the steelmaking parameter set and the target steel grade semantic vector into the pre-constructed steelmaking time prediction model, and obtain the steelmaking process time through the steelmaking time prediction model. The third acquisition module is used to input the refining parameter set and the target steel grade semantic vector into the pre-constructed refining time prediction model, and obtain the refining process time through the refining time prediction model. The fourth acquisition module is used to input the continuous casting parameter set and the target steel grade semantic vector into the pre-constructed continuous casting time prediction model, and obtain the continuous casting process time through the continuous casting time prediction model.

8. A computer program product, characterized in that, It includes computer-readable instructions that, when executed on an electronic device, cause the electronic device to implement the process time prediction method as described in any one of claims 1 to 6.

9. An electronic device, characterized in that, It includes at least one processor and a memory connected to the processor, wherein: The memory is used to store computer programs; The processor is used to execute the computer program to enable the electronic device to implement the process time prediction method as described in any one of claims 1 to 6.

10. A computer storage medium, characterized in that, The storage medium carries one or more computer programs that, when executed by an electronic device, enable the electronic device to implement the process time prediction method as described in any one of claims 1 to 6.

Images