Galvanized steel pipe surface characteristic prediction method and system based on process modeling

Abstract

The invention discloses a galvanized steel pipe surface characteristic prediction method based on process modeling, and relates to the technical field of intelligent manufacturing process optimization, and the galvanized steel pipe surface characteristic prediction method comprises the following steps: basic data acquisition, process state characteristic modeling, dynamic process diagram modeling, physical data dual-drive mixing and surface characteristic prediction. Adopting a derived feature construction method based on an embedded mechanism kernel equation to form process intermediate state variable enhanced data; by establishing a dynamic process diagram model, the state transfer and evolution process between steel pipe procedures is represented; a residual learning mechanism of a physical model and a data-driven model is combined, physical consistency constraint is introduced, a dual-drive fusion prediction framework is constructed, and multi-target surface characteristic hybrid prediction is realized; according to the method, high-precision and multi-target prediction of the characteristics such as the surface thickness, the adhesive force and the smoothness of the galvanized steel pipe is achieved, and the prediction stability and the physical interpretability under the complex galvanizing process are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing process optimization technology, specifically to a method and system for predicting the surface characteristics of galvanized steel pipes based on process modeling. Background Technology

[0002] The method and system for predicting the surface characteristics of galvanized steel pipes based on process modeling refers to establishing an intelligent prediction system that reflects the evolution of process states and physical reaction mechanisms by digitizing and modeling the data of each key process in the galvanizing production process (including degreasing, pickling, fluxing, galvanizing, and cooling). The core function of this method is to uniformly model multi-source sensor data, material characteristic data, and equipment operating parameters, and describe the state transfer logic between each process through dynamic process diagrams. Based on this, by introducing feature enhancement methods embedded with mechanistic equations and residual learning networks, the physical model and data model are synergistically optimized. Its application value is reflected in three aspects: First, it can significantly improve the accuracy and consistency of predicting the surface characteristics of galvanized steel pipes, avoiding the deviations of traditional empirical formulas or single machine learning models under complex working conditions; second, it can effectively realize the visualization and interpretability of the production process, providing a basis for decision-making for process optimization and quality control; third, it has scalability, enabling rapid migration and adaptive application under different production lines, different steel pipe materials, and multiple process conditions, thus having significant engineering promotion significance in the fields of intelligent manufacturing and quality prediction.

[0003] However, existing predictions of the surface characteristics of galvanized steel pipes suffer from technical problems, such as the inability to simultaneously characterize the complex coupling relationship between process parameters, material properties and equipment status, resulting in insufficient prediction accuracy and generalization ability.

[0004] Existing process state characteristic modeling methods suffer from technical problems, such as relying solely on data fitting, lacking mechanistic constraints, and failing to reflect the dynamic changes in pickling and fluxing processes.

[0005] Existing dynamic process diagram modeling methods suffer from technical problems such as insufficient modeling of inter-process state evolution relationships and a lack of time-varying transmission and physical feasibility constraint mechanisms.

[0006] Existing physical-data dual-drive hybrid methods suffer from technical problems such as the separation of physical and data models and the lack of consistency constraints in residual correction, leading to unstable prediction results. Summary of the Invention

[0007] To address the above issues and overcome the shortcomings of existing technologies, this invention provides a method and system for predicting the surface characteristics of galvanized steel pipes based on process modeling. Addressing the technical problem that existing surface characteristic prediction methods for galvanized steel pipes cannot simultaneously characterize the complex coupling relationships between process parameters, material properties, and equipment states, resulting in insufficient prediction accuracy and generalization ability, this solution creatively employs a comprehensive surface characteristic prediction process modeling enhancement method that combines process state feature modeling, dynamic process diagram modeling, and physical data dual-drive hybrid approach. This achieves accurate characterization of multi-dimensional dynamic characteristics and cross-process state transfer modeling throughout the galvanizing process. Furthermore, addressing the technical problem that existing process state feature modeling methods rely solely on data fitting and lack mechanistic constraints, failing to reflect the dynamic changes in pickling and fluxing processes, this solution creatively employs derived feature construction methods embedded with mechanistic kernel equations. This method constructs a model for process state characteristics, achieving enhanced dynamic feature representation of the pickling activation degree and surface coating thickness formation process. Addressing the technical problems of insufficient modeling of inter-process state evolution relationships and lack of time-varying transmission and physical feasibility constraints in existing dynamic process diagram modeling methods, this solution creatively adopts a dynamic process state transmission model oriented towards the galvanizing process flow to perform dynamic process diagram modeling, achieving spatiotemporal diagram modeling of the state evolution and reaction transmission of steel pipes across multiple processes. Furthermore, addressing the technical problems of the existing physical-data dual-drive hybrid method, which suffers from the disconnect between the physical model and the data model and the lack of consistency constraints in residual correction, leading to unstable prediction results, this solution creatively employs a residual learning network combined with the physical model for hybrid prediction of surface characteristics, achieving synergistic enhancement of physical consistency and data learning capabilities.

[0008] The technical solution adopted by this invention is as follows: The surface characteristic prediction method for galvanized steel pipes based on process modeling provided by this invention includes the following steps:

[0009] Step S1: Basic data collection;

[0010] Step S2: Modeling process state characteristics;

[0011] Step S3: Dynamic process diagram modeling;

[0012] Step S4: Physical data dual-drive hybridization.

[0013] Furthermore, in step S1, the basic data acquisition is used to build a data foundation, specifically by acquiring multi-source heterogeneous data of the galvanizing project, and performing time-series alignment and data preprocessing to obtain the original dataset of the galvanizing multidimensional process.

[0014] The multi-source heterogeneous data of the galvanizing process specifically includes real-time sensor data, material characteristic data, and equipment status parameters.

[0015] Further, in step S2, the process state feature modeling is used to calculate the intermediate process state features of pickling activation and flux film thickness. Specifically, based on the original dataset of the zinc plating multidimensional process, a derived feature construction method embedding the kernel equation of the mechanism is used to model the process state features and obtain enhanced data of intermediate process state variables, including the following steps:

[0016] Step S21: Mechanism feature construction, used to extract mechanistic feature quantities related to the pickling activation degree and flux film thickness formation process. Specifically, based on the original dataset of the zinc plating multidimensional process, pickling flux features are extracted and normalized to obtain mechanism feature data.

[0017] The pickling flux characteristics specifically include pickling Reynolds number, acid concentration saturation factor, and flux viscosity characteristic parameters;

[0018] Step S22: Mechanism core equation construction, used to establish core physical constraint equations corresponding to the pickling reaction kinetics and the coating formation process. Specifically, based on the mechanism characteristic data, the pickling activation degree change equation and the surface coating thickness formation equation are established respectively to obtain the mechanism core prediction value.

[0019] The acid washing activation change equation is specifically calculated based on temperature, acid concentration and flow state to determine the removal rate and obtain the activation change amount;

[0020] The surface coating thickness formation equation is specifically based on the flux concentration, temperature, desolventing rate, and drying parameters, and the film thickness formation rate is calculated.

[0021] Step S23: Derived feature construction, used to enhance the ability to express the dynamic change characteristics of the pickling and fluxing processes. Specifically, based on the mechanism feature data and the mechanism kernel prediction value, derived features reflecting the rate of change of pickling activation, the rate of increase of surface coating thickness and the fluctuation trend of operating conditions are extracted by performing time difference, sliding statistics and proportional function calculation. Combined with the basic inputs of temperature, acid concentration and flow parameters in the original dataset of the zinc plating multidimensional process, time series derivative features and relative change features are constructed to obtain a derived feature vector characterizing the dynamic response of the pickling and fluxing processes.

[0022] Step S24: Feature combination output, used to form a unified feature input for dynamic process diagram modeling and physical data dual-drive fusion. Specifically, based on the derived feature vector, feature splicing, normalization and correlation screening are performed to generate a comprehensive feature matrix containing pickling activation degree prediction value, surface coating thickness prediction value and dynamic change index. The comprehensive feature matrix is ​​then encapsulated as process intermediate state variable enhancement data, which serves as the input for the subsequent dynamic process diagram modeling in step S3 and physical data dual-drive fusion in step S4.

[0023] Further, in step S3, the dynamic process diagram modeling, used to characterize the transmission and evolution of the steel pipe's state between degreasing, pickling, fluxing, galvanizing, and cooling processes, specifically involves using a dynamic process state transmission model oriented towards the galvanizing process flow to perform dynamic process diagram modeling based on the original dataset of the galvanizing multidimensional process and the enhanced data of the intermediate process state variables, to obtain the steel pipe process chain state vector, including the following steps:

[0024] Step S31: Defining the topology of the process diagram, used to establish the topological relationships and node attribute representation structures between the main process nodes in the galvanizing production process. Specifically, based on the original dataset of the galvanizing multidimensional process and the enhanced data of intermediate state variables, a set of process nodes including degreasing, pickling, fluxing, galvanizing and cooling is defined, and a unidirectional directed edge relationship between adjacent processes is established. For each process node in the set of process nodes, a node attribute vector containing state variables such as pickling activation degree, surface coating thickness, surface temperature, viscosity parameters and cleanliness is constructed. By constructing physical feasible domain constraints, the basic topology data of the process diagram is obtained.

[0025] Step S32: Process association modeling, used to establish the time sequence correspondence between steel pipes at different process nodes. Specifically, based on the basic topology data of the process diagram, a time sequence index set is constructed to identify the node sequence relationship and state transmission direction of each steel pipe in the entire process from degreasing to cooling, thereby obtaining process constraint pipe fitting association mapping data.

[0026] Step S33: Dual-channel dynamic edge weight update, used to describe the time-varying transmission mechanism of state inheritance and processing action between processes. Specifically, based on the process constraint fitting association mapping data, dynamic update operators for the transportation channel and the reaction channel are constructed respectively, and the results of the two channels are weighted and fused to obtain the fused state data between processes.

[0027] The transport channel is used to characterize the inheritance ratio of state between adjacent processes;

[0028] The reaction channel is used to characterize the processing and rewriting effect of the current process on the state variables;

[0029] Step S34: Process state update, which is used to combine the transportation channel and the reaction channel to realize the dynamic evolution model of the steel pipe state between multiple processes. Specifically, based on the fused state data between processes, the state update calculation and physical boundary projection are performed to obtain the steel pipe process chain state vector.

[0030] Further, in step S4, the physical data dual-drive hybrid model is used to construct a residual learning model between the physical model and the data-driven model and to correct the surface characteristic data of the steel pipe. Specifically, based on the steel pipe process chain state vector and the process intermediate state variable enhancement data, a residual learning network combined with the physical model is used to perform hybrid prediction of surface characteristics, obtaining multi-objective hybrid prediction data of surface characteristics, including the following steps:

[0031] Step S41: Physical benchmark prediction, used to calculate the benchmark prediction results of physically solvable targets and form a multi-target physical benchmark prediction vector. Specifically, based on the steel pipe process chain state vector and the process intermediate state variable enhancement data, the analytical targets are calculated by using the heat conduction and reaction kinetic physical model. For targets that cannot be directly solved by physical equations, the prediction value is initialized to zero, and a multi-target physical benchmark prediction vector is constructed.

[0032] Step S42: Residual data construction, used to generate training targets required for residual learning and unify feature scale. Specifically, based on the multi-objective physical benchmark prediction vector and the steel pipe process chain state vector, a supervised residual target is constructed, the difference between the real surface characteristics and the physical benchmark prediction results is calculated, and a residual vector is obtained. The residual vector is then normalized and standardized to obtain standardized residual training data.

[0033] Step S43: Construct a multi-task residual learning network to model multi-objective residuals. Specifically, based on the standardized residual training data, a residual learning network containing an input encoding layer, a shared feature extraction layer, and a multi-task output layer is constructed to extract a comprehensive representation of the steel pipe process chain state and intermediate features. Through the multi-task structure, the surface characteristic residuals of thickness, adhesion, and smoothness are predicted to obtain the initial residual prediction vector.

[0034] Step S44: Physical constraint modeling, used to introduce physical consistency constraints in the residual correction process. Specifically, based on the initial residual prediction vector, a monotonicity constraint regularization term is introduced in the residual learning process to penalize gradient directions that violate physical laws, thereby obtaining a constraint fusion correction residual vector.

[0035] Step S45: Hybrid prediction output, used to realize the dual-drive fusion prediction of physical model and data-driven model and output multi-objective surface characteristic results. Specifically, based on the multi-objective physical benchmark prediction vector and the constraint fusion correction residual vector, the fusion calculation is performed by vector addition to obtain the hybrid prediction data of multi-objective surface characteristics.

[0036] The surface characteristic prediction system for galvanized steel pipes based on process modeling provided by this invention includes a basic data acquisition module, a process state feature modeling module, a dynamic process diagram modeling module, and a physical data dual-drive hybrid module.

[0037] The basic data acquisition module is used for basic data acquisition. Through basic data acquisition, the original dataset of the galvanizing multidimensional process is obtained, and the original dataset of the galvanizing multidimensional process is sent to the process state feature modeling module and the dynamic process diagram modeling module.

[0038] The process state feature modeling module is used for process state feature modeling. Through process state feature modeling, it obtains process intermediate state variable enhancement data and sends the process intermediate state variable enhancement data to the dynamic process diagram modeling module and the physical data dual-drive hybrid module.

[0039] The dynamic process diagram modeling module is used for dynamic process diagram modeling. Through dynamic process diagram modeling, the state vector of the steel pipe process chain is obtained, and the state vector of the steel pipe process chain is sent to the physical data dual-drive hybrid module.

[0040] The physical data dual-drive mixing module is used for physical data dual-drive mixing, and obtains multi-target surface characteristic mixed prediction data through physical data dual-drive mixing.

[0041] The beneficial effects achieved by adopting the above solution are as follows:

[0042] (1) In view of the technical problem that the existing surface characteristic prediction of galvanized steel pipes cannot simultaneously depict the complex coupling relationship between process parameters, material properties and equipment status, resulting in insufficient prediction accuracy and generalization ability, this solution creatively adopts a comprehensive surface characteristic prediction process modeling enhancement method that combines process status feature modeling, dynamic process diagram modeling and physical data dual-drive hybridization, which realizes accurate characterization of multi-dimensional dynamic features of the entire galvanizing process and cross-process status transfer modeling.

[0043] (2) In view of the technical problem that existing process state feature modeling methods rely solely on data fitting and lack mechanistic constraints, and cannot reflect the dynamic changes of pickling and fluxing processes, this solution creatively adopts a derived feature construction method that embeds mechanistic kernel equations to model process state features, thereby realizing the enhanced expression of dynamic features of pickling activation degree and surface coating thickness formation process.

[0044] (3) In view of the technical problems that existing dynamic process diagram modeling methods have insufficient modeling of the state evolution relationship between processes and lack of time-varying transmission and physical feasibility constraint mechanism, this solution creatively adopts a dynamic process state transmission model for galvanizing process flow to perform dynamic process diagram modeling, and realizes spatiotemporal diagram modeling of the state evolution and reaction transmission of steel pipes between multiple processes.

[0045] (4) In view of the technical problems in the existing physical data dual-drive hybrid method, there is a separation between the physical model and the data model and a lack of consistency constraints in residual correction, which leads to unstable prediction results. This solution creatively adopts a residual learning network that combines the physical model to perform hybrid prediction of surface characteristics, thereby achieving a synergistic enhancement of physical consistency and data learning ability. Attached Figure Description

[0046] Figure 1 A flowchart illustrating the surface characteristic prediction method for galvanized steel pipes based on process modeling provided by this invention.

[0047] Figure 2 A schematic diagram of the galvanized steel pipe surface characteristic prediction system based on process modeling provided by the present invention;

[0048] Figure 3 A flowchart illustrating the process state characteristics of step S2;

[0049] Figure 4 A schematic diagram of the process flow for modeling the dynamic process diagram in step S3;

[0050] Figure 5 This is a schematic diagram of the process for the dual-drive hybridization of physical data in step S4.

[0051] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0053] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0054] Example 1, see Figure 1 The present invention provides a method for predicting the surface characteristics of galvanized steel pipes based on process modeling, the method comprising the following steps:

[0055] Step S1: Basic data collection;

[0056] Step S2: Modeling process state characteristics;

[0057] Step S3: Dynamic process diagram modeling;

[0058] Step S4: Physical data dual-drive hybridization.

[0059] By performing the above operations, this solution addresses the technical problem that existing surface characteristic prediction methods for galvanized steel pipes cannot simultaneously depict the complex coupling relationship between process parameters, material properties, and equipment status, resulting in insufficient prediction accuracy and generalization ability. This solution creatively adopts a comprehensive surface characteristic prediction process modeling enhancement method that combines process status feature modeling, dynamic process diagram modeling, and physical data dual-drive hybrid approach. This method achieves accurate characterization of multi-dimensional dynamic features and cross-process status transfer modeling for the entire galvanizing process.

[0060] By constructing a process chain topology among the main processes such as pickling, fluxing, galvanizing and cooling, and introducing a joint optimization mechanism of physical constraints and data-driven approaches during the prediction process, the model can maintain physical consistency and prediction stability under different production conditions, significantly improving the comprehensive prediction accuracy of surface thickness, adhesion and smoothness.

[0061] Example 2, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. In step S1, the basic data acquisition is used to build a data foundation. Specifically, it involves acquiring multi-source heterogeneous data of the galvanizing project and performing time-series alignment and data preprocessing to obtain the original dataset of the galvanizing multidimensional process.

[0062] The multi-source heterogeneous data of the galvanizing project specifically includes real-time sensor data, material characteristic data, and equipment status parameters;

[0063] The real-time sensor data is specifically collected by sensors deployed in the process units of the production line, including the temperature and concentration of the degreasing tank, the temperature and concentration of the acid in the pickling tank, the concentration and temperature of the flux in the plating tank, the temperature and composition of the zinc in the galvanizing pot, the pressure and angle of the air knife, the temperature and flow rate of the cooling medium in the cooling section, and the production line conveying speed.

[0064] The material characteristic data is specifically collected and integrated through the manufacturing execution system, including the grade, chemical composition, pipe diameter and wall thickness of the steel substrate, and the chemical composition of the zinc ingot.

[0065] The equipment status parameters are collected through the equipment's own monitoring system and the production line monitoring data acquisition system, including equipment start-up and shutdown status, operating cycle, maintenance records and wear status of key components (including air knife nozzles and submerged rollers);

[0066] The timing alignment and data preprocessing include the following steps:

[0067] Step S11: Multi-source data association, specifically, based on the multi-source heterogeneous data of the galvanizing project, by performing unique index key matching and data association, an initial data set with a single steel pipe as the unit is obtained;

[0068] Step S12: Process timing alignment optimization, specifically, based on different timestamp data in the initial data set, and taking the absolute time of the steel pipe entering the production line as the benchmark, resampling and interpolation are performed at fixed time intervals to obtain a time-series aligned process data sequence;

[0069] Step S13: Abnormal data normalization, specifically, aligning the process data sequence according to the time series, manually detecting and removing outliers, and normalizing all numerical data by maximum and minimum values ​​to obtain the original dataset of the galvanizing multidimensional process.

[0070] Example 3, see Figure 1 , Figure 2 and Figure 3 This embodiment is based on the above embodiment. In step S2, the process state feature modeling is used to calculate the intermediate state features of the pickling activation degree and flux film thickness. Specifically, based on the original dataset of the zinc plating multidimensional process, a derived feature construction method with embedded mechanism kernel equations is used to model the process state features and obtain enhanced data of intermediate state variables. This includes the following steps:

[0071] Step S21: Mechanism feature construction, used to extract mechanistic feature quantities related to the pickling activation degree and flux film thickness formation process. Specifically, based on the original dataset of the zinc plating multidimensional process, pickling flux features are extracted and normalized to obtain mechanism feature data.

[0072] The pickling flux characteristics specifically include pickling Reynolds number, acid concentration saturation factor, and flux viscosity characteristic parameters;

[0073] Preferably, the formula for calculating the acid-washing Reynolds number is:

[0074] ;

[0075] In the formula, Re t It is the acid-washing Reynolds number. It is the density of the pickling solution. d is the linear velocity of the moving steel pipe, and d is the characteristic length of the steel pipe. It is the dynamic viscosity of the pickling solution;

[0076] The formula for calculating the acid concentration saturation factor is as follows:

[0077] ;

[0078] In the formula, It is the acid concentration saturation factor. It refers to the acid concentration of the pickling solution, C. 50 It is the acid concentration half-saturation value, which represents the acid concentration required to reach 50% of the maximum reaction rate;

[0079] The formula for calculating the viscosity characteristic parameter of the flux is:

[0080] ;

[0081] In the formula, It is a viscosity characteristic parameter of the flux, used to characterize the thickness of the liquid film formed by the flux on the surface of the steel pipe. It is the empirical coefficient of the plating solution. It's the viscosity of the flux. It is the speed at which the steel pipe moves out of the groove. is the density of the plating solution, and g is the acceleration due to gravity.

[0082] Step S22: Mechanism core equation construction, used to establish core physical constraint equations corresponding to the pickling reaction kinetics and the coating formation process. Specifically, based on the mechanism characteristic data, the pickling activation degree change equation and the surface coating thickness formation equation are established respectively to obtain the mechanism core prediction value.

[0083] The equation for the change in activation degree during acid washing is specifically based on the calculation of the removal rate and the change in activation degree according to temperature, acid concentration, and flow state. The calculation formula is as follows:

[0084] ;

[0085] In the formula, It is the rate of increase of acid pickling activation over time, k p It is the rate coefficient of the acid washing reaction, T pick,t It is the temperature of the pickling tank, f H It is a flow enhancement factor, used to represent the promoting effect of flow state on reaction rate, A act,t This is the percentage of pickling completed, 1-A act,t This represents the remaining percentage of the product that has not yet undergone pickling.

[0086] The surface coating thickness formation equation is specifically based on the flux concentration, temperature, desolventing rate, and drying parameters, and the film thickness formation rate is calculated using the following formula:

[0087] ;

[0088] In the formula, It is the rate at which the flux film thickness increases over time, k ads It is the adsorption rate coefficient, C flux,t It refers to the effective content of the flux, T flux,t It is the temperature of the flux, k. dry It is the film drying rate. It is the air knife wind speed, h flux,t It is the thickness of the surface coating layer;

[0089] Step S23: Derived feature construction, used to enhance the ability to express the dynamic change characteristics of the pickling and fluxing processes. Specifically, based on the mechanism feature data and the mechanism kernel prediction value, derived features reflecting the rate of change of pickling activation, the rate of increase of surface coating thickness and the fluctuation trend of operating conditions are extracted by performing time difference, sliding statistics and proportional function calculation. Combined with the basic inputs of temperature, acid concentration and flow parameters in the original dataset of the zinc plating multidimensional process, time series derivative features and relative change features are constructed to obtain a derived feature vector characterizing the dynamic response of the pickling and fluxing processes.

[0090] Preferably, the extraction of derived features reflecting the rate of change of pickling activation, the rate of increase of surface coating thickness, and the trend of operating conditions specifically includes the characteristics of the rate of change of pickling activation, the characteristics of the rate of increase of coating thickness, and the characteristics of the trend of operating conditions.

[0091] The characteristic of the rate of change of acid washing activation specifically includes the first derivative characteristic of activation and the relative rate of change of acid concentration.

[0092] The characteristic of the growth rate of the coating thickness specifically refers to the characteristic of the first derivative of the film thickness.

[0093] The operating condition fluctuation trend characteristics specifically include the relative volatility characteristics of the sliding window and the relative temperature change rate characteristics.

[0094] The first derivative characteristic of the activation degree is used to characterize the rate of change of the pickling reaction over consecutive time intervals, thereby reflecting the dynamic trend of the material surface being activated by pickling. The calculation formula is as follows:

[0095] ;

[0096] In the formula, It is a characteristic of the first derivative of activation, A act,t This is the percentage of pickling already completed, A. act,t-1 It is the proportion of acid pickling that was completed a moment ago. It is the discrete time step;

[0097] The relative change rate of acid concentration is used to reflect the effect of changes in acid concentration on the reaction rate. It characterizes the degree of fluctuation in acid washing conditions by calculating the ratio of the current acid concentration change to the previous time step. The calculation formula is as follows:

[0098] ;

[0099] In the formula, It is a characteristic of the relative change rate of acid concentration. It refers to the acid concentration of the pickling solution. It represents the acid concentration of the pickling solution at the previous moment;

[0100] The first derivative characteristic of the film thickness is used to represent the growth trend of the flux coating thickness over time, reflecting the dynamic process of film formation by the flux on the steel pipe surface. The calculation formula is as follows:

[0101] ;

[0102] In the formula, It is the first derivative characteristic of film thickness, h flux,t It is the thickness of the surface coating, h flux,t-1 It is the thickness of the surface coating at the previous moment;

[0103] The relative volatility characteristic of the sliding window is used to characterize the intensity of fluctuation of process parameters within a local time window, reflecting the influence of operating condition stability on the fluxing film formation process. The calculation formula is as follows:

[0104] ;

[0105] In the formula, F var It is the relative volatility characteristic of the sliding window, std is the standard deviation calculation function, x[tk:t] is the process parameter from the time interval tk to t, the process parameter specifically refers to the real-time sensor subset in the original dataset of the galvanizing multidimensional process that is related to the dynamic changes of pickling and fluxing processes, mean is the mean calculation function, and k is the length of the sliding window;

[0106] The relative temperature change rate characteristic is used to characterize the trend of fluxing or pickling temperature over time, reflecting the influence of temperature fluctuations on the reaction rate and film formation rate. The calculation formula is as follows:

[0107] ;

[0108] In the formula, It is a characteristic of the relative rate of change of temperature, T t It is the surface temperature parameter, T t-1 It is the surface temperature parameter from the previous moment;

[0109] Step S24: Feature combination output, used to form a unified feature input for the dual-drive fusion of dynamic process diagram modeling and physical data. Specifically, based on the derived feature vector, feature splicing, normalization, and correlation filtering are performed to generate a comprehensive feature matrix containing the predicted values ​​of pickling activation, surface coating thickness, and dynamic change indicators. The comprehensive feature matrix is ​​then encapsulated as enhanced data of intermediate process state variables, serving as the input for the subsequent steps S3 (dynamic process diagram modeling) and S4 (physical data dual-drive fusion).

[0110] As a further optimization of this embodiment, the formula for calculating the comprehensive feature matrix is ​​as follows:

[0111] ;

[0112] In the formula, Z is the comprehensive characteristic matrix. This is the predicted value of pickling activation. This is the predicted value of the surface coating thickness. It is a dynamically changing indicator;

[0113] Specifically, the predicted value of pickling activation is obtained by integrating the growth rate of pickling activation over time, and the calculation formula is as follows:

[0114] ;

[0115] In the formula, It is the rate of increase of acid pickling activation over time;

[0116] The predicted thickness of the surface coating is specifically obtained by integrating the growth rate of the flux film thickness over time, and the calculation formula is as follows:

[0117] ;

[0118] In the formula, It is the growth rate of the flux film thickness over time; as a further optimization of this embodiment, in the integration calculation of the predicted acid pickling activation value and the predicted surface coating thickness value, the current time t is taken as the starting point and the predicted time t+Δt is taken as the upper limit of integration, which is used to calculate the predicted value of the state variable in the next time step. The integration result reflects the evolution trend of the future state, thereby obtaining the predicted acid pickling activation value and the predicted surface coating thickness value.

[0119] The dynamic change indicators, i.e. the derived characteristics, specifically include the pickling activation rate change characteristic, the capping layer thickness growth rate characteristic, and the operating condition fluctuation trend characteristic, and the calculation formula is as follows:

[0120] ;

[0121] In the formula, It is a dynamically changing indicator. It is a characteristic of the first derivative of activation. It is a characteristic of the first derivative of film thickness. It is the relative volatility characteristic of the sliding window. It is a characteristic of the relative rate of change of temperature. It is a characteristic of the relative change rate of acid concentration.

[0122] By performing the above operations, this solution addresses the technical problem that existing process state feature modeling methods rely solely on data fitting, lack mechanistic constraints, and fail to reflect the dynamic changes in the pickling and fluxing processes. It creatively employs a derived feature construction method that embeds mechanistic kernel equations to model process state features, thereby achieving enhanced dynamic feature expression of the pickling activation degree and surface coating thickness formation process.

[0123] This method incorporates the kinetics of pickling reaction and the equation for film formation, embedding physical process characteristics into data-driven features. It then combines time difference, sliding statistics, and proportional function calculations to form time series derivatives and relative change features. This effectively solves the shortcomings of traditional feature construction, which lacks process correlation and dynamic sensitivity, making the prediction of intermediate process variables more consistent with the actual reaction mechanism.

[0124] Example 4, see Figure 1 , Figure 2 and Figure 4 This embodiment is based on the above embodiment. In step S3, the dynamic process diagram modeling is used to characterize the transmission and evolution of the steel pipe state between degreasing, pickling, fluxing, galvanizing, and cooling processes. Specifically, based on the original dataset of the galvanizing multidimensional process and the enhanced data of the intermediate state variables of the process, a dynamic process state transmission model oriented towards the galvanizing process flow is used to perform dynamic process diagram modeling to obtain the steel pipe process chain state vector, including the following steps:

[0125] Step S31: Defining the topology of the process diagram, used to establish the topological relationships and node attribute representation structures between the main process nodes in the galvanizing production process. Specifically, based on the original dataset of the galvanizing multidimensional process and the enhanced data of intermediate state variables, a set of process nodes including degreasing, pickling, fluxing, galvanizing and cooling is defined, and a unidirectional directed edge relationship between adjacent processes is established. For each process node in the set of process nodes, a node attribute vector containing state variables such as pickling activation degree, surface coating thickness, surface temperature, viscosity parameters and cleanliness is constructed. By constructing physical feasible domain constraints, the basic topology data of the process diagram is obtained.

[0126] Preferably, the basic topology data of the process diagram includes a set of process nodes, a state vector, a set of directed edge relationships, and a physically feasible region;

[0127] The formula for calculating the set of process nodes is:

[0128] ;

[0129] In the formula, Va is the set of process nodes, D is the degreasing process node, P is the pickling process node, F is the fluxing process node, Z is the galvanizing process node, and C is the cooling process node.

[0130] The formula for calculating the state vector is:

[0131] ;

[0132] In the formula, s v,t It is a state vector, where v is the node representation index, t is the time index, and A t It is an activation parameter, H t It is the surface coating thickness parameter, S t It is a cleanliness parameter, T t It is the surface temperature parameter, V t It is a viscosity parameter;

[0133] The formula for calculating the set of directed edge relations is:

[0134] ;

[0135] In the formula, It is a set of directed edge relations. It is a directed edge identifier;

[0136] The formula for calculating the physical feasible region is:

[0137] ;

[0138] In the formula, A is the physical feasible region, H is the physical feasible region of the activation parameter, S is the physical feasible region of the surface coating thickness parameter, T is the physical feasible region of the cleanliness parameter, and V is the physical feasible region of the surface temperature parameter.

[0139] Step S32: Process association modeling, used to establish the time sequence correspondence between steel pipes at different process nodes. Specifically, based on the basic topology data of the process diagram, a time sequence index set is constructed to identify the node sequence relationship and state transmission direction of each steel pipe in the entire process from degreasing to cooling, thereby obtaining process constraint pipe fitting association mapping data.

[0140] Step S33: Dual-channel dynamic edge weight update, used to describe the time-varying transmission mechanism of state inheritance and processing action between processes. Specifically, based on the process constraint fitting association mapping data, dynamic update operators for the transportation channel and the reaction channel are constructed respectively, and the results of the two channels are weighted and fused to obtain the fused state data between processes.

[0141] The transport channel is used to characterize the inheritance ratio of states between adjacent processes, and the calculation formula is as follows:

[0142] ;

[0143] In the formula, This is the process inheritance state vector of process node i at time t+1, obtained through the transportation channel, where i is the process node index and j is the index of the process's neighbor node. C is the edge weight of the transport channel between process node i and its neighboring node j. i,t It is the process continuity coefficient corresponding to process node i and time t, used to represent the process dwell time, s j,t It is the state vector of the neighboring node j of the process at time t;

[0144] Preferably, the edge weights of the transport channel satisfy the following conditions: and Specifically, it is determined by minimizing the error of the process inheritance state vector, and the calculation formula is as follows:

[0145] ;

[0146] In the formula, Here, represents the edge weights of the transportation corridor; `softmax` is the softmax function; and `Corr` is the similarity calculation function used to represent the state similarity between nodes. This is the edge weight adjustment coefficient, with a default value of 0.5. j,t s is the state vector of the neighboring node j at time t. i,t It is the state vector of process node i at time t;

[0147] Preferably, the formula for calculating the process continuity coefficient is:

[0148] ;

[0149] In the formula, C i,t It is the process continuity coefficient corresponding to process node i and time t. It is the discrete time step in dynamic process diagram modeling. This is the actual dwell time of the steel pipe at process node i. Its value is calculated from the effective length of the process segment and the running speed. It reflects the duration of the steel pipe being affected by processing in that process. The calculation formula is: , where L i It is the effective processing length. It is the linear velocity of the moving steel pipe;

[0150] The reaction channel is used to characterize the processing and modification effect of the current process on the state variable, and the calculation formula is as follows:

[0151] ;

[0152] In the formula, M is the vector representing the change in processing rewrite state of process node i at time t. i It is a 0 / 1 diagonal matrix used to characterize the range of influence of process node i on state variables, B i This is the processing gain matrix of the reaction process, representing the strength of the effect of the process on the state variables. It is the state processing function, u i,t These are process control variables, specifically including flux ratio, temperature, speed, and immersion time;

[0153] The weighting rule for weighted fusion of the two channel results is as follows:

[0154] ;

[0155] In the formula, This is the inter-process fusion status data of process node i at time t+1. It is the dual-channel fusion weight, with a value range of [0,1].

[0156] As a further optimization of this embodiment, the transport channel is used to express state inheritance, and the reaction channel is used to express state processing. The dynamic process state transfer model composed of the two makes the state evolution conform to the process mechanism rather than relying solely on statistical learning.

[0157] Step S34: Process state update, which is used to combine the transportation channel and the reaction channel to realize the dynamic evolution model of the steel pipe state between multiple processes. Specifically, based on the fused state data between processes, the state update calculation and physical boundary projection are performed to obtain the steel pipe process chain state vector.

[0158] Preferably, the calculation formula for the state update is:

[0159] ;

[0160] In the formula, s i,t+1 It is the state vector of the steel pipe process chain at process node i at time t+1. It is a physical feasible region projection operator used to project a state vector onto a physical constraint region. It is the inter-process fusion status data of process node i at time t+1.

[0161] This embodiment introduces a dual-channel time-varying transmission mechanism (transport channel and reaction channel) on the process diagram and combines it with the projection of the physical feasible region to achieve consistent and interpretable modeling of cross-process states. Compared with the time sequence diagram method based solely on statistical edge weights, this scheme ensures the physical rationality and reproducibility of state evolution, thereby improving the stability and traceability of the state transmission modeling of the entire process chain.

[0162] By performing the above operations, this solution addresses the technical problems of insufficient modeling of inter-process state evolution relationships and lack of time-varying transmission and physical feasibility constraint mechanisms in existing dynamic process diagram modeling methods. It creatively adopts a dynamic process state transmission model for galvanizing process flow to perform dynamic process diagram modeling, thereby realizing the spatiotemporal diagram modeling of the state evolution and reaction transmission of steel pipes between multiple processes.

[0163] By defining process nodes for degreasing, pickling, fluxing, galvanizing, and cooling, a dual-channel dynamic edge weight update operator for the transport and reaction channels is constructed. This enables the model to describe state inheritance and processing rewriting between nodes, and maintains the physical rationality of state variables through physical boundary projection. This achieves interpretable state evolution modeling for the entire process chain, improving the transparency and traceability of the prediction model.

[0164] Example 5, see Figure 1 , Figure 2 and Figure 5 This embodiment is based on the above embodiment. In step S4, the physical data dual-drive hybrid model is used to construct a residual learning model between the physical model and the data-driven model and to correct the surface characteristic data of the steel pipe. Specifically, based on the steel pipe process chain state vector and the process intermediate state variable enhancement data, a residual learning network combined with the physical model is used to perform surface characteristic hybrid prediction to obtain multi-objective surface characteristic hybrid prediction data, including the following steps:

[0165] Step S41: Physical benchmark prediction, used to calculate the benchmark prediction results of physically solvable targets and form a multi-target physical benchmark prediction vector. Specifically, based on the steel pipe process chain state vector and the process intermediate state variable enhancement data, the analytical targets are calculated by using the heat conduction and reaction kinetic physical model. For targets that cannot be directly solved by physical equations, the prediction value is initialized to zero, and a multi-target physical benchmark prediction vector is constructed.

[0166] The resolvable target specifically refers to the coating thickness, which is calculated using the following formula:

[0167] ;

[0168] In the formula, H(t) is the coating thickness, t is the time index, k0 is the reaction rate constant, and E a These are the surface reaction activation energy parameters, where R is the gas constant, T is the workpiece surface temperature, and C is the temperature at which the surface reacts. Zn It refers to the concentration of the zinc solution, C. s (t) is the interface concentration;

[0169] Step S42: Residual data construction, used to generate training targets required for residual learning and unify feature scale. Specifically, based on the multi-objective physical benchmark prediction vector and the steel pipe process chain state vector, a supervised residual target is constructed, the difference between the real surface characteristics and the physical benchmark prediction results is calculated, and a residual vector is obtained. The residual vector is then normalized and standardized to obtain standardized residual training data.

[0170] The formula for calculating the residual vector is:

[0171]

[0172] In the formula, R true It is the residual vector, Y true These are the actual measured surface characteristics, including coating thickness, adhesion, and smoothness. It is a multi-objective physical baseline prediction vector;

[0173] Step S43: Construct a multi-task residual learning network to model multi-objective residuals. Specifically, based on the standardized residual training data, a residual learning network containing an input encoding layer, a shared feature extraction layer, and a multi-task output layer is constructed to extract a comprehensive representation of the steel pipe process chain state and intermediate features. Through the multi-task structure, the surface characteristic residuals of thickness, adhesion, and smoothness are predicted to obtain the initial residual prediction vector.

[0174] The input encoding layer is used to map the process state feature vector into an intermediate representation;

[0175] The feature extraction layer is used to capture the implicit correlations and nonlinear relationships between different processes;

[0176] The multi-task output layer is used to output the corresponding residual components for each target.

[0177] Preferably, the initial residual prediction vector is calculated using the following formula:

[0178] ;

[0179] In the formula, It is the initial residual prediction vector, specifically in the form of [thickness residual, adhesion residual, surface finish residual], f multi It is a multi-task output layer operation function, f sharedIt is the feature extraction layer operation function, f enc is the input encoding layer operation function, s is the steel pipe process chain state vector, and z is the process intermediate state variable enhancement data;

[0180] Preferably, Table 1 is an example table of construction parameters for the residual learning network. As shown in the table, the model configuration is used to limit the structural form of each network module. For example, the input encoding layer adopts a combination structure of fully connected + normalization + activation function to ensure that process features from different sources are mapped to a unified feature space; the shared feature extraction layer adopts a stack of Transformer encoders to achieve cross-process feature dependency modeling through multi-head attention; the multi-task output layer outputs three residual prediction values ​​through a fully connected structure, corresponding to the three targets of thickness, adhesion and smoothness.

[0181] The model structure and parameter range of this invention are for illustrative purposes only and are not intended to be limiting.

[0182] The parameter name column corresponds to the core hyperparameters that can actually be adjusted during network training and inference, such as input mapping dimension, network depth, number of self-attention heads, and output dimension, which control feature representation ability, model complexity, and output structure, respectively.

[0183] The preferred parameter values ​​are used to provide the optimal configuration of the present invention in the actual galvanizing process prediction scenario. For example, the input mapping dimension is set to 128, the Transformer network depth is set to 3 layers, and the number of self-attention heads is set to 4 heads. This configuration can obtain stable residual prediction capability.

[0184] The adjustable parameter range provides upper and lower limits for model reproduction. For example, the input mapping dimension can be adjusted from 64 to 256, and the network depth can be adjusted from 2 to 5 layers to achieve effective modeling. Those skilled in the art can adjust the parameters within this range according to the data scale or computing resources of different production lines without affecting the implementation effect of the present invention.

[0185] Table 1. Example of construction parameters for residual learning networks

[0186]

[0187] Step S44: Physical constraint modeling, used to introduce physical consistency constraints in the residual correction process. Specifically, based on the initial residual prediction vector, a monotonicity constraint regularization term is introduced in the residual learning process to penalize gradient directions that violate physical laws, thereby obtaining a constraint fusion correction residual vector.

[0188] The formula for calculating the monotonicity constraint regularization term is as follows:

[0189] ;

[0190] In the formula, It is a monotonicity constraint regularization term, where n is the sample index. H is the adhesion value predicted by the model for the nth sample. n It is the coating thickness of the nth sample;

[0191] The formula for calculating the constraint fusion correction residual vector is as follows:

[0192] ;

[0193] In the formula, It is a constraint-fused correction residual vector. It is the initial residual prediction vector. This is the constraint weight coefficient, with a default value of 0.8;

[0194] Step S45: Hybrid prediction output, used to realize the dual-drive fusion prediction of physical model and data-driven model and output multi-objective surface characteristic results. Specifically, based on the multi-objective physical benchmark prediction vector and the constraint fusion correction residual vector, the fusion calculation is performed by vector addition to obtain the hybrid prediction data of multi-objective surface characteristics.

[0195] As a further optimization of this embodiment, the calculation formula for the multi-target surface characteristic mixed prediction data is as follows:

[0196] ;

[0197] In the formula, Y final It is multi-objective surface property mixed prediction data. It is a multi-objective physical baseline prediction vector. It is the constraint fusion correction residual vector.

[0198] By performing the above operations, this solution addresses the technical problems in existing physical-data dual-drive hybrid methods, such as the separation of physical and data models and the lack of consistency constraints in residual correction, which leads to unstable prediction results. It creatively adopts a residual learning network that combines the physical model to perform hybrid prediction of surface characteristics, thereby achieving a synergistic enhancement of physical consistency and data learning capabilities.

[0199] By using analytical targets calculated from physical models as a baseline, a residual network is used to model the unanalyzed parts in multiple tasks. A monotonic regularization term is introduced to suppress gradient directions that violate physical laws, ensuring the model maintains both physical interpretability and adaptive error correction capabilities. This dual-drive fusion mechanism effectively improves the prediction accuracy and stability of multi-objective properties such as surface thickness, adhesion, and smoothness, making it particularly suitable for surface performance prediction tasks under complex and nonlinear process conditions.

[0200] Example 6, see Figure 1 and Figure 2 This embodiment is based on the above embodiments. The galvanized steel pipe surface characteristic prediction system based on process modeling provided by the present invention includes a basic data acquisition module, a process state characteristic modeling module, a dynamic process diagram modeling module, and a physical data dual-drive hybrid module.

[0201] The basic data acquisition module is used for basic data acquisition. Through basic data acquisition, it obtains the original dataset of the galvanizing multidimensional process and sends the original dataset to the process state feature modeling module and the dynamic process diagram modeling module.

[0202] The process state feature modeling module is used for process state feature modeling. Through process state feature modeling, it obtains process intermediate state variable enhancement data and sends the process intermediate state variable enhancement data to the dynamic process diagram modeling module and the physical data dual-drive hybrid module.

[0203] The dynamic process diagram modeling module is used for dynamic process diagram modeling. Through dynamic process diagram modeling, the state vector of the steel pipe process chain is obtained, and the state vector of the steel pipe process chain is sent to the physical data dual-drive hybrid module.

[0204] The physical data dual-drive mixing module is used for physical data dual-drive mixing, and obtains multi-target surface characteristic mixed prediction data through physical data dual-drive mixing.

[0205] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0206] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0207] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A method for predicting the surface characteristics of galvanized steel pipes based on process modeling, characterized in that: The method includes the following steps: Step S1: Basic data collection to obtain the original dataset of the galvanizing multidimensional process; Step S2: Process state feature modeling. Based on the original dataset of the galvanizing multidimensional process, a derived feature construction method embedding the mechanism kernel equation is used to model the process state features and obtain enhanced data of intermediate process state variables. This includes the following steps: mechanism feature construction; mechanism kernel equation construction; derived feature construction; feature combination output. Step S3: Dynamic process diagram modeling. Based on the original dataset of the galvanizing multidimensional process and the enhanced data of the intermediate state variables of the process, a dynamic process state transfer model oriented towards the galvanizing process is adopted to perform dynamic process diagram modeling, and obtain the steel pipe process chain state vector. This includes the following steps: process diagram topology definition; process association modeling; dual-channel dynamic edge weight update; process state update. Step S4: Dual-drive physical data hybridization. Based on the steel pipe process chain state vector and the process intermediate state variable enhancement data, a residual learning network combined with a physical model is used to perform hybrid prediction of surface characteristics, resulting in multi-objective surface characteristic hybrid prediction data. This includes the following steps: physical baseline prediction; residual data construction; multi-task residual learning network construction; physical constraint modeling; and hybrid prediction output.

2. The method for predicting the surface characteristics of galvanized steel pipes based on process modeling according to claim 1, characterized in that: In step S1, the basic data acquisition is used to build the data foundation. Specifically, it involves acquiring multi-source heterogeneous data from the galvanizing project, performing time-series alignment and data preprocessing, and obtaining the original dataset of the multi-dimensional galvanizing process. The multi-source heterogeneous data of the galvanizing process specifically includes real-time sensor data, material characteristic data, and equipment status parameters.

3. The method for predicting the surface characteristics of galvanized steel pipes based on process modeling according to claim 2, characterized in that: In step S2, the process state feature modeling includes the following steps: Step S21: Mechanism feature construction, used to extract mechanistic feature quantities related to the pickling activation degree and flux film thickness formation process. Specifically, based on the original dataset of the zinc plating multidimensional process, pickling flux features are extracted and normalized to obtain mechanism feature data. The pickling flux characteristics specifically include pickling Reynolds number, acid concentration saturation factor, and flux viscosity characteristic parameters; Step S22: Mechanism core equation construction, used to establish core physical constraint equations corresponding to the pickling reaction kinetics and the coating formation process. Specifically, based on the mechanism characteristic data, the pickling activation degree change equation and the surface coating thickness formation equation are established respectively to obtain the mechanism core prediction value. The acid washing activation change equation is specifically calculated based on temperature, acid concentration and flow state to determine the removal rate and obtain the activation change amount; The surface coating thickness formation equation is specifically based on the flux concentration, temperature, desolventing rate, and drying parameters, and the film thickness formation rate is calculated. Step S23: Derived feature construction, used to enhance the ability to express the dynamic change characteristics of the pickling and fluxing processes. Specifically, based on the mechanism feature data and the mechanism kernel prediction value, derived features reflecting the rate of change of pickling activation, the rate of increase of surface coating thickness and the fluctuation trend of operating conditions are extracted by performing time difference, sliding statistics and proportional function calculation. Combined with the basic inputs of temperature, acid concentration and flow parameters in the original dataset of the zinc plating multidimensional process, time series derivative features and relative change features are constructed to obtain a derived feature vector characterizing the dynamic response of the pickling and fluxing processes. Step S24: Feature combination output, used to form a unified feature input for the dual-drive fusion of dynamic process diagram modeling and physical data. Specifically, based on the derived feature vector, feature splicing, normalization and correlation screening are performed to generate a comprehensive feature matrix containing the predicted value of pickling activation, the predicted value of surface coating thickness and dynamic change index.

4. The method for predicting the surface characteristics of galvanized steel pipes based on process modeling according to claim 3, characterized in that: In step S24, the comprehensive feature matrix will be encapsulated as process intermediate state variable enhancement data, which will serve as the input for the subsequent dynamic process diagram modeling in step S3 and the dual-drive fusion of physical data in step S4.

5. The method for predicting the surface characteristics of galvanized steel pipes based on process modeling according to claim 4, characterized in that: In step S3, the dynamic process diagram modeling includes the following steps: Step S31: Define the topology of the process diagram, which is used to establish the topological relationship and node attribute representation structure between the main process nodes in the galvanizing production process. Specifically, based on the original dataset of the galvanizing multidimensional process and the enhanced data of intermediate process state variables, define the process node set and establish the unidirectional directed edge relationship between adjacent processes. For each process node in the process node set, construct a node attribute vector and obtain the basic topology data of the process diagram by constructing physical feasible domain constraints. Step S32: Process association modeling, used to establish the time series correspondence between steel pipes at different process nodes. Specifically, based on the basic topology data of the process diagram, a time series index set is constructed to identify the node sequence relationship and state transmission direction of each steel pipe, thereby obtaining process constraint fitting association mapping data. Step S33: Dual-channel dynamic edge weight update, used to describe the time-varying transmission mechanism of state inheritance and processing action between processes. Specifically, based on the process constraint fitting association mapping data, dynamic update operators for the transportation channel and the reaction channel are constructed respectively, and the results of the two channels are weighted and fused to obtain the fused state data between processes. Step S34: Process state update, which is used to combine the transportation channel and the reaction channel to realize the dynamic evolution model of the steel pipe state between multiple processes. Specifically, based on the fused state data between processes, the state update calculation and physical boundary projection are performed to obtain the steel pipe process chain state vector.

6. The method for predicting the surface characteristics of galvanized steel pipes based on process modeling according to claim 5, characterized in that: In step S33, the transport channel is used to characterize the inheritance ratio of state between adjacent processes; the reaction channel is used to characterize the processing and rewriting effect of the current process on the state variable.

7. The method for predicting the surface characteristics of galvanized steel pipes based on process modeling according to claim 6, characterized in that: In step S4, the dual-drive mixing of physical data includes the following steps: Step S41: Physical benchmark prediction, used to calculate the benchmark prediction results of physically solvable targets and form a multi-target physical benchmark prediction vector. Specifically, based on the steel pipe process chain state vector and the process intermediate state variable enhancement data, the analytical targets are calculated by using the heat conduction and reaction kinetic physical model. For targets that cannot be directly solved by physical equations, the prediction value is initialized to zero, and a multi-target physical benchmark prediction vector is constructed. Step S42: Residual data construction, used to generate training targets required for residual learning and unify feature scale. Specifically, based on the multi-objective physical benchmark prediction vector and the steel pipe process chain state vector, a supervised residual target is constructed, the difference between the real surface characteristics and the physical benchmark prediction results is calculated, and a residual vector is obtained. The residual vector is then normalized and standardized to obtain standardized residual training data. Step S43: Construct a multi-task residual learning network to model multi-objective residuals. Specifically, based on the standardized residual training data, construct a residual learning network that includes an input encoding layer, a shared feature extraction layer, and a multi-task output layer to extract a comprehensive representation of the steel pipe process chain state and intermediate features. Through the multi-task structure, predict the surface characteristic residuals and obtain the initial residual prediction vector. Step S44: Physical constraint modeling, used to introduce physical consistency constraints in the residual correction process. Specifically, based on the initial residual prediction vector, a monotonicity constraint regularization term is introduced in the residual learning process to penalize gradient directions that violate physical laws, thereby obtaining a constraint fusion correction residual vector. Step S45: Hybrid prediction output, used to realize the dual-drive fusion prediction of physical model and data-driven model and output multi-objective surface characteristic results. Specifically, based on the multi-objective physical benchmark prediction vector and the constraint fusion correction residual vector, the fusion calculation is performed by vector addition to obtain the hybrid prediction data of multi-objective surface characteristics.

8. A galvanized steel pipe surface characteristic prediction system based on process modeling, used to implement the galvanized steel pipe surface characteristic prediction method based on process modeling as described in any one of claims 1-7, characterized in that: It includes a basic data acquisition module, a process state feature modeling module, a dynamic process diagram modeling module, and a physical data dual-drive hybrid module.

9. The galvanized steel pipe surface characteristic prediction system based on process modeling according to claim 8, characterized in that: The basic data acquisition module is used for basic data acquisition. Through basic data acquisition, the original dataset of the galvanizing multidimensional process is obtained, and the original dataset of the galvanizing multidimensional process is sent to the process state feature modeling module and the dynamic process diagram modeling module. The process state feature modeling module is used for process state feature modeling. Through process state feature modeling, it obtains process intermediate state variable enhancement data and sends the process intermediate state variable enhancement data to the dynamic process diagram modeling module and the physical data dual-drive hybrid module. The dynamic process diagram modeling module is used for dynamic process diagram modeling. Through dynamic process diagram modeling, the state vector of the steel pipe process chain is obtained, and the state vector of the steel pipe process chain is sent to the physical data dual-drive hybrid module. The physical data dual-drive mixing module is used for physical data dual-drive mixing, and obtains multi-target surface characteristic mixed prediction data through physical data dual-drive mixing.

Images