Steel production line control system based on mechanism-ai coupled world model

Abstract

The application provides a steel production line control system based on a mechanism-AI coupled world model, comprising: collecting multi-source heterogeneous industrial data from a steel process industry physical production line, and preprocessing to obtain standardized multi-modal data; mapping the standardized multi-modal data to a unified hidden space to form a working condition hidden state capable of comprehensively representing the current production process; establishing a system of ordinary differential equations of production line macro state variables based on material conservation, energy conservation, mass conservation, momentum conservation, reaction kinetics, thermodynamic constraints and process dynamic characteristics as an ODE macro mechanism model; generating a spatial residual correction amount based on a PDE physical AI residual network according to the macro prediction result, the spatial coordinates and the working condition information; fusing the macro prediction result and the spatial residual correction amount according to a dynamic confidence weight to generate a final spatial-time process state field; and generating a control strategy according to the final spatial-time process state field and executing the control strategy to the production line equipment.

Description

Technical Field

[0001] This invention relates to the field of production line control technology, specifically to a steel production line control system based on a mechanism-AI coupled world model, and more specifically to a mechanism-AI coupled world model construction system driven by multimodal fusion perception for the steel process industry. Background Technology

[0002] The steel process industry typically includes continuous or semi-continuous production stages such as ironmaking, steelmaking, continuous casting, rolling, heat treatment, and finishing. These stages are characterized by strong coupling, long processes, high inertia, strong nonlinearity, and multivariable collaborative control. Existing steel production line control methods are usually centered on basic automation systems and process control systems. They collect field data through PLCs, DCSs, L1 basic automation systems, L2 process control systems, MES systems, and various online monitoring devices. Based on setpoint control, PID control, feedforward compensation, expert rules, lookup table models, process mechanism models, numerical simulation models, or local data-driven models, they control process parameters such as temperature, speed, pressure, flow rate, tension, reduction, cooling water volume, and energy medium distribution.

[0003] In existing production line control methods, the basic automation layer is typically responsible for real-time equipment control, such as motor speed control, valve opening control, roll gap control, cooling water control, and interlock protection. The process control layer usually calculates target setpoints based on steel grade, specifications, process procedures, historical experience, and online measurements, and then sends these setpoints to the basic automation layer for execution. The manufacturing execution system or production management system is mainly responsible for production planning, quality tracking, data recording, and process report generation. This type of control system can meet the basic automation needs in routine production, but its control effectiveness usually relies on pre-set process rules, human experience, local mechanism models, and single-point detection data. When raw material composition fluctuates, equipment status changes, production rhythm changes, new steel grades or specifications are switched to, or on-site disturbances increase, existing control methods are prone to problems such as decreased model accuracy, increased control setting deviations, and delayed response to abnormal operating conditions.

[0004] Specifically, existing steel production line control methods have the following shortcomings, including: First, existing control methods lack the ability to comprehensively utilize multi-source heterogeneous data. Steel production sites simultaneously generate one-dimensional time-series data such as temperature, pressure, flow rate, speed, current, and tension; discrete state data such as equipment start-up / shutdown, valve position, alarms, and interlocks; spatial data such as industrial images, infrared thermal images, point clouds, or defect images; and frequency domain data such as vibration, audio, current spectra, and ultrasonic signals. However, existing control systems typically process this data separately according to process, equipment, or data type, relying more on single sensors or a few key variables for feedback control. They lack a mechanism to uniformly map multi-modal data such as time series, state, images, and spectra into implicit operating states. Therefore, when sensor noise, modal missing data, local anomalies, or complex operating condition coupling occur, the system struggles to accurately identify the current production state and provides stable and comprehensive state input for subsequent mechanism modeling and control decisions.

[0005] Second, existing mechanistic or lumped parameter models struggle to simultaneously meet the demands for interpretability and spatially refined control. Traditional steel production line control widely utilizes mechanistic models, lookup table models, and lumped parameter models based on material conservation, energy conservation, momentum conservation, thermodynamics, reaction kinetics, or empirical formulas. These models possess a degree of physical interpretability and engineering usability, suitable for describing macroscopic variables such as average temperature, average composition, average pressure, average flow rate, average velocity, output, and energy consumption. However, these models typically only describe overall trends or average states, failing to express local distribution differences along the production line direction, width direction, thickness direction, equipment sections, or spatial grids. They also struggle to accurately compensate for spatial distribution deviations caused by complex boundary conditions, local disturbances, equipment aging, environmental changes, and operating condition drift. Therefore, relying solely on traditional macroscopic mechanistic models for production line control can easily lead to problems where macroscopic predictions are reasonable but local spatial state deviations are significant, thus affecting quality stability and control accuracy.

[0006] Third, traditional numerical simulation methods struggle to meet the real-time requirements of online control. While numerical calculation methods such as the finite element method, finite volume method, finite difference method, CFD, and FEM can perform high-precision calculations of spatially distributed states such as temperature, flow, stress, concentration, and velocity fields, they typically require fine mesh generation, complex boundary condition settings, and lengthy computation times. In the online control scenarios of the steel process industry, control cycles are usually short, and on-site conditions change rapidly. The control system needs to complete state estimation, prediction, strategy generation, and safety verification within a finite timeframe. Therefore, traditional high-precision numerical simulation methods are more suitable for offline analysis, process design, or localized condition verification, and are difficult to directly use as the core model for real-time production line control.

[0007] Fourth, purely data-driven AI prediction or control methods lack physical constraints and industrial credibility. In recent years, data-driven models such as machine learning, deep learning, CNN, RNN, LSTM, and Transformer have been increasingly used for tasks such as quality prediction, equipment status identification, process parameter prediction, and anomaly diagnosis. These methods can learn complex nonlinear relationships from historical data, but their prediction results are highly dependent on the distribution and quality of the training data. In scenarios involving new steel grades, new specifications, abnormal operating conditions, extreme operating conditions, or insufficient training samples, pure AI models are prone to problems such as unstable extrapolation, outputs that do not conform to physical laws, and uninterpretable control strategies. At the same time, pure AI models typically lack explicit connections with material conservation, energy conservation, momentum conservation, boundary conditions, safety constraints, and equipment capability constraints, making them difficult for industrial control systems to directly trust and use for closed-loop control.

[0008] Fifth, existing physical information neural networks or mechanism-AI hybrid methods are mostly geared towards single equipment, single processes, or single physical fields, making it difficult to form production line-level closed-loop control capabilities. Some existing methods embed physical equations into the neural network training process or use the output of the mechanism model as the input of the AI ​​model to improve prediction accuracy and generalization ability. However, such methods usually only build models for local objects, such as a single heating furnace, a single rolling process, a single temperature field, a single shape field, or a single quality prediction task. They lack an integrated framework for unified multimodal perception input, unified hidden state of operating conditions, macroscopic mechanism prediction, spatial residual correction, dynamic confidence fusion, safety constraint verification, control strategy generation, and data feedback iteration for the entire steel process industry production line. Therefore, although existing mechanism-AI hybrid methods can improve local prediction accuracy, they are difficult to support a production line-level control closed loop from state perception to control execution and model self-evolution.

[0009] Sixth, existing digital twin or industrial intelligent control platforms tend to focus more on monitoring, display, and decision support, lacking a safety closed-loop mechanism for real-time control. Existing digital twin platforms typically combine equipment models, process models, real-time data, and visualization systems for production monitoring, status display, offline simulation, report analysis, or decision support. However, in practical control applications, many platforms have not formed a complete closed loop consisting of multimodal perception, mechanism simulation, spatial correction, control strategy generation, safety constraint verification, equipment execution, and data feedback. Furthermore, existing platforms lack dynamic weighting and safety fallback mechanisms for situations such as unreliable AI model outputs, input conditions exceeding the training distribution, excessive physical residuals, communication anomalies, or inference latency exceeding limits, making it difficult to guarantee industrial safety when AI participates in control.

[0010] Therefore, existing steel production line control methods generally suffer from problems such as difficulty in unified perception of multi-source heterogeneous data, inability of macroscopic mechanism models to describe spatial distribution deviations, inability of traditional numerical simulation to meet real-time control requirements, lack of physical constraints and safety reliability in pure AI models, difficulty in forming production line-level closed loops in mechanism-AI hybrid models, and difficulty in continuous self-evolution of digital twin platforms. How to construct a steel production line control method and system that can unify and integrate multimodal industrial data, take into account both macroscopic mechanism explanation and spatial residual correction, support online control execution under safety constraints, and continuously iterate and optimize through real-time data feedback is a crucial technical problem that urgently needs to be solved for the intelligentization of the steel process industry and production line-level intelligent driving. Summary of the Invention

[0011] To address the shortcomings of existing technologies, the purpose of this invention is to provide a steel production line control system based on a mechanism-AI coupled world model.

[0012] A steel production line control system based on a mechanism-AI coupled world model, provided by the present invention, includes: The data acquisition and preprocessing module is used to collect multi-source heterogeneous industrial data from the physical production line of the steel process industry, and to preprocess the collected multi-source heterogeneous industrial data to obtain standardized multimodal data. The multimodal fusion sensing module is used to map standardized multimodal data to a unified latent space to form a latent state of the current production process that can comprehensively represent the current production process. The ODE macroscopic mechanism model module is used to establish a set of ordinary differential equations for the macroscopic state variables of the production line based on material conservation, energy conservation, mass conservation, momentum conservation, reaction kinetics, thermodynamic constraints, and process dynamic characteristics, as the ODE macroscopic mechanism model; the ODE macroscopic mechanism model generates macroscopic prediction results based on the hidden state of the current production process, control variables, and process parameters. The PDE Physical AI Residual Network Module is used to generate spatial residual corrections based on macroscopic prediction results, spatial coordinates, and operating condition information using the PDE Physical AI Residual Network. The mechanism-AI fusion output module is used to fuse macroscopic prediction results with spatial residual corrections according to dynamic confidence weights to generate the final spatial-temporal process state field. The online closed-loop control module is used to generate control strategies based on the final space-time process state field and execute them on the production line equipment.

[0013] Preferably, the data acquisition and preprocessing module includes: Module M1.1: Used to collect data from the physical production lines of the steel industry, including ironmaking, steelmaking, continuous casting, rolling, heat treatment, and finishing, including one-dimensional time-series numerical data, discrete switching state data, two-dimensional image and spatial data, and frequency domain spectrum data. The one-dimensional time-series numerical data includes: temperature, pressure, flow rate, speed, current, voltage, rolling force, tension, PLC analog quantity, valve opening, PID closed-loop curve, cooling water flow rate, and energy medium flow rate. The discrete data of the switch status includes: equipment start / stop signal, operating level, PLC switch quantity, IO point, valve position switch feedback, alarm status, interlock status, and process mode number. The image and spatial two-dimensional data include: visible light industrial images, infrared thermal imaging images, laser contour maps, point cloud projection maps, surface defect images, and equipment status images; The frequency domain spectrum data includes: on-site industrial audio, equipment vibration signals, motor current spectrum, ultrasonic flaw detection signals, envelope spectrum, and time spectrum. Module M1.2: Performs outlier removal, filtering, normalization, and sliding time window slicing on one-dimensional time-series numerical data to form a time-series input tensor; performs state encoding, missing data completion, time preservation, and event alignment on discrete switch-state data to form a discrete state input tensor; performs resolution normalization, region of interest extraction, coordinate calibration, and timestamp alignment on two-dimensional image and spatial data to form an image spatial input tensor; and performs time-frequency transformation, spectrogram generation, normalization, and equipment condition alignment on frequency-domain spectral data to form a spectral input tensor. Module M1.3: Using furnace number, coil number, slab number, material section number, equipment number, process section number or timestamp as indexes, it aligns data from different modalities to the same material object, the same equipment object or the same process time, thereby forming standardized multimodal data.

[0014] Preferably, the multimodal fusion sensing module includes: Module M2.1: Extracts the temporal dynamic features of the temporal input tensor to obtain one-dimensional temporal coding features; Module M2.2: Extracts the equipment state, process mode, and event change features from the discrete state input tensor to obtain discrete state encoding features; Module M2.3: Extracts the spatial distribution features of the image spatial input tensor to obtain the image spatial coding features; Module M2.4: Extracts device status and abnormal frequency features from the spectral input tensor to obtain frequency domain spectral coding features; Module M2.5: Maps one-dimensional temporal coding features, discrete state coding features, image spatial coding features, and frequency domain spectral coding features to a unified-dimensional working condition representation space;

[0015] in, The encoding features of mode m; This represents the mapping weight matrix for the m-th mode; This represents the bias term for the m-th mode; This represents the uniform dimension feature after the m-th modality mapping; Module M2.6: Establishes the correlation between different modalities and dynamically weights the importance of different modalities to obtain multimodal fusion features, including: A bidirectional cross-attention mechanism is employed, using one modality feature as the query vector and another modality feature as the key and value, to calculate intermodal attention enhancement features:

[0016] in, Let Q represent the attention enhancement result of mode n on mode m, where Q represents the query vector, K represents the key vector, and V represents the value vector; After obtaining the intermodal attention enhancement features, dynamic modal weights are calculated based on the current operating conditions and data quality scores:

[0017] in, Let g(·) represent the dynamic weights of the m-th mode, g(·) represent the weight generation network, and s represent the current operating state. This represents the data quality score for the m-th modality; The multimodal fusion feature is represented as follows:

[0018] in, This represents the multimodal fusion feature.

[0019] Module M2.7: Performs time-series modeling for multimodal fusion features to capture the evolution of production line conditions over time and obtain the hidden states of the conditions.

[0020] Preferably, the macroscopic mechanism model of ODE includes:

[0021] Where S(t) represents the set of macroscopic state variables at time t, U(t) represents the set of control variables at time t, and P(t) represents the set of process parameters at time t. This represents the hidden state of the operating condition output by the multimodal fusion sensing module. This represents the parameters of the macroscopic mechanism model of ODE. It represents the mechanism function composed of physical conservation relations, process dynamic relations, and empirical mechanism relations; The ODE macroscopic mechanism model module outputs a function. Transform macroeconomic state variables into target macroeconomic forecast results:

[0022] in, G(·) represents the macroscopic prediction result output by the ODE macroscopic mechanism model at time t, and G(·) represents the output function from the macroscopic state variable to the target predictor variable.

[0023] Preferably, the PDE physical AI residual network module includes: a time t, spatial coordinate x, and working condition hidden state. Macroscopic state variable S(t), control variable U(t), process parameter P(t), and macroscopic prediction results As input, the output spatial residual correction is:

[0024] in, This indicates that the PDE physical AI residual network is in time t and spatial location The residual correction amount output at that point. The parameter is PDE physical AI residual network, Indicates the parameters of the PDE physical AI residual network; When the spatial variable is a multidimensional spatial coordinate. Represented as:

[0025] in, to These represent different spatial directions, equipment sections, process locations, or discrete grid nodes, respectively. The PDE physical AI residual network learns the residual of the real spatial field relative to the ODE macroscopic prediction results based on time, spatial coordinates, hidden states of the operating conditions, macroscopic state variables, control variables, process parameters, and ODE macroscopic prediction results, thereby achieving physical constraint compensation for spatial distribution deviations.

[0026] Preferably, the PDE physical AI residual network includes: During the offline training phase, the target spatial field is obtained based on real observation data, offline simulation data, or computational data. ; This represents the true or high-confidence process state value that meets preset requirements at time t and spatial location x. Spatial residual is defined as:

[0027] Where R(t,x) represents the spatial residual of the real spatial field relative to the macroscopic prediction result of ODE; By introducing partial differential equation constraints corresponding to the target industrial process variables, their unified form is expressed as:

[0028] Where N(·) represents the partial differential equation operator corresponding to the target physical process, Y(t,x) represents the space-time process state field to be predicted, P(t) represents the process parameters, and U(t) represents the control variables; The total training loss function of the PDE physical AI residual network includes:

[0029] Where L represents the total loss function, This represents the data fitting loss. Indicates PDE physical loss. Indicates boundary condition loss. Represents the initial condition loss. This indicates ODE consistency loss. Indicates spatial smoothing loss, This represents the residual amplitude constraint loss. These represent the weighting coefficients of each loss term; The physical losses in PDE include:

[0030] in, Indicates the number of physical constraint sampling points. This indicates that after fusing the macroscopic prediction results of ODE with the output of the PDE physical AI residual network, the sampling points... The predicted state at the point; the PDE physical loss is used to constrain the network output to satisfy the PDE regularity loss corresponding to the target physical process; The data fitting loss includes:

[0031] in, Indicates the number of sampling points for the observation data; The ODE consistency loss includes:

[0032] in, Indicates a spatial region. The spatial average term represents the spatial-temporal process state field generated by the mechanism-AI fusion output module, which represents a measure of a spatial region. In spatial regions The average value on the above; the ODE consistency loss is used to ensure that the spatial field output by the PDE physical AI residual network is consistent with the output of the ODE macroscopic mechanism model in a macroscopic average sense; The residual magnitude constraint loss includes:

[0033] in, This indicates the maximum allowable residual correction magnitude.

[0034] Preferably, the mechanism-AI fusion output module includes:

[0035] in, This represents the final output space-time process state field. This indicates the results of macroeconomic forecasts. This represents the spatial residual correction amount output by the PDE physical AI residual network. Indicates dynamic confidence weights; The dynamic confidence weights are determined by the output confidence of the PDE physical AI residual network, the data quality score, the PDE physical residual, the operating condition distribution matching degree, and the communication health status.

[0036] in, This represents the Sigmoid function. This represents a weight generation network. Represents the physical residual of the PDE. This indicates the degree of deviation between the current operating conditions and the training distribution. Indicates the data quality score. Indicates the health status of communication.

[0037] Preferably, the online closed-loop control module includes: Module M6.1: Acquires real-time production data from systems including PLC, DCS, MES, L1 basic automation system, L2 process control system, sensor network, imaging system, and detection system; Module M6.2: Calls the multimodal fusion sensing module to calculate the real-time hidden state of the operating condition at online operation time t based on the current real-time production data. ; Module M6.3: Calls the ODE macroscopic mechanism model module to calculate the macroscopic prediction results for the next time step or future time window based on the current macroscopic state, control variables, process parameters, and implicit operating conditions. ;in, Indicates the prediction time step; Module M6.4: Calls the PDE Physical AI Residual Network module to calculate the spatial residual correction based on real-time operating condition hidden states, macroscopic prediction results, control variables, process parameters, and spatial coordinates. The final spatial-temporal process state field is generated through the mechanism-AI fusion output module. ; Module M6.5: Before generating the control strategy, it performs process boundary checks, equipment capability checks, control quantity upper and lower limit checks, interlock condition checks, communication status checks, confidence level checks, and abnormal rollback condition checks. When the current operating condition, model output, and equipment status meet the safety constraints, it enters the control strategy generation stage. When the safety constraints are not met, it refuses to generate or issue high-risk control strategies and triggers alarms, conservative control, or ODE safety baseline rollback. Module M6.6: Generates control strategies based on the final space-time process state field, target process indicators, equipment constraints, and production plan; the control strategies include temperature setting, speed setting, pressure reduction setting, tension setting, flow rate setting, pressure setting, valve opening setting, energy medium distribution strategy, production rhythm adjustment strategy, equipment maintenance strategy, and abnormal handling strategy; The control strategy is expressed as:

[0038] Where U*(t) represents the optimal control strategy, U represents the control variable to be optimized, and J(·) represents the control objective function. The target process state is represented by C, which represents equipment capacity constraints, safety boundary constraints, and process specification constraints. Module M6.6: Sends control policies that have passed security verification to the corresponding devices, controllers or actuators via the manufacturing API, equipment API, energy API, logistics API or security API.

[0039] Preferably, the system further includes: a real-time data feedback and model iteration module, used to receive real-time data, prediction results, control strategies, equipment execution results, and actual feedback results during online operation, and to continuously evaluate and iteratively update them, including: Module M7.1: Records the hidden states of operation conditions output by the multimodal fusion sensing module. Macroscopic prediction results output by the ODE macroscopic mechanism model module Spatial residual correction amount output by the PDE physical AI residual network module The fusion output of the mechanism-AI fusion output module And the control strategy U*(t) generated by the online closed-loop control module; Module M7.2: Collects the actual process status, quality indicators, equipment response, and abnormal information after the equipment is executed; Module M7.3: Simultaneously evaluates simulation tracking error and control strategy error; The simulation following error is used to compare the actual feedback result after the device executes with the simulation prediction result: ;in, This represents the simulation prediction results. This indicates the actual feedback result after the device executes the command; when When the value is consistently greater than the preset value, it indicates that the data is not closely following the actual working conditions and new residuals need to be learned or parameters need to be updated. The control strategy error is used to compare the deviation between the set value, the simulated value, and the actual value:

[0040]

[0041]

[0042] in, This represents the target setpoint given by the control strategy. This indicates the simulation results after the prediction is executed. Indicates the actual execution feedback value; if near but If the deviation is met but the preset requirements are still met, it indicates a problem with the equipment execution or on-site disturbance; if itself deviates If the preset requirements are met, it indicates that the control strategy generation effect is not good; if and If the long-term deviation is large but the preset requirements are met, it indicates that the residual needs to be updated. Module M7.4: Evaluate sample value based on multimodal fusion results, data quality, deviation from operating condition distribution, simulation error, control error, and operating condition status.

[0043] in, Indicates the sample value score. : Indicates the quality of multimodal data. This indicates the degree of deviation between the current operating conditions and the distribution of the training data. This represents the simulation tracking error. Indicates the control strategy error. Indicates the operating condition or abnormality type; Module M7.5: Writes high-value samples, abnormal samples, boundary condition samples, new steel grade samples, new specification samples, and samples with large model errors into the sample library; Module M7.6: Based on the sample library, periodically or event-triggered updates of the multimodal fusion perception model, ODE parameters, PDE physical AI residual network parameters, and control strategy model parameters; retraining trigger conditions include: simulation error continuously exceeding a threshold, control strategy error continuously exceeding a threshold, the appearance of new steel grades, new specifications, and new working conditions, the current working condition distribution significantly deviating from the training set and meeting preset requirements, the number of high-value samples in the sample library meeting preset requirements reaching a threshold, the periodic maintenance training time arriving, PDE residuals being consistently large and meeting preset requirements, and the safety rollback frequency increasing and meeting preset requirements.

[0044] Preferably, the system further includes: a safety monitoring and degradation rollback module, used to monitor the operating status of the PDE physical AI residual network, the mechanism-AI fusion output module, the online closed-loop control module and the communication link, and to ensure safe operation in abnormal situations, including: continuously monitoring the input data, model output, physical residual, model confidence, communication heartbeat and inference delay to determine whether the current model results meet the safety requirements of the industrial control scenario.

[0045] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention employs a multimodal fusion sensing module to encode one-dimensional time-series numerical data, discrete switching state data, two-dimensional image and spatial data, and frequency domain spectrum data respectively, and maps them to a unified-dimensional working condition latent space. This solves the problems of inconsistent multi-source heterogeneous data structure, inconsistent sampling frequency, and difficulty in data semantic fusion in the steel process industry. 2. This invention employs a bidirectional cross-attention mechanism, a dynamic modality weighting mechanism, and a global temporal modeling mechanism, enabling the system to adaptively adjust the contribution of each modality according to different working conditions and different data quality, thereby solving the problems of incomplete perception of a single modality, missing modalities, or unstable working condition recognition caused by modal noise. 3. This invention uses the ODE macroscopic mechanism model as the macroscopic mechanism framework of the production line, and calculates macroscopic prediction results based on material conservation, energy conservation, momentum conservation, mass conservation and process dynamic characteristics, thereby solving the problems of pure AI models lacking physical interpretability, difficulty in cold start and instability under abnormal operating conditions. 4. This invention uses a PDE physical AI residual network to compensate for spatial distribution deviations that are difficult to describe by the ODE macroscopic mechanism model, thereby solving the problem that traditional lumped parameter models or macroscopic mechanism models can only describe the overall trend and are difficult to describe the spatial field distribution characteristics. 6. This invention employs a fusion structure of "macroscopic mechanism prediction + spatial residual correction" to make the final output represent as follows: This solves the contradiction between the insufficient accuracy of a single mechanism model and the insufficient security of a single AI model; 7. This invention employs ODE consistency loss, PDE physical constraint loss, boundary condition loss, initial condition loss, spatial smoothing loss, and residual magnitude constraint loss to enable the PDE physical AI residual network to maintain physical consistency and macroscopic conservation consistency while learning spatial bias, thereby solving the problem that the output results of black box neural networks do not satisfy physical laws.

[0046] 8. This invention solves the problem of control risks that may result from AI model failure in industrial control scenarios by employing dynamic confidence weights and a security monitoring and degradation rollback module, which smoothly decays the residual correction weights to zero when the PDE physical AI residual network is abnormal or untrustworthy. Attached Figure Description

[0047] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a schematic diagram of a steel production line control system based on a mechanism-AI coupled world model. Detailed Implementation

[0048] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the scope of protection of the present invention.

[0049] Example 1 This invention provides a steel production line control system based on a mechanism-AI coupled world model. This system addresses continuous production processes such as ironmaking, steelmaking, continuous casting, rolling, heat treatment, and finishing. It collects multi-source heterogeneous data from the production line, including one-dimensional time-series numerical data, discrete switch-state data, two-dimensional image and spatial data, and frequency domain spectrum data. Through modal-specific encoding, unified dimension mapping, bidirectional cross-attention fusion, dynamic modal weight allocation, and global temporal modeling, it forms latent states representing the current operating status of the production line. Based on this, the system constructs a mechanism-AI coupled world model composed of an ODE macroscopic mechanism model and a PDE physical AI residual network. The ODE... The macroscopic mechanism model generates macroscopic prediction results based on material conservation, energy conservation, mass conservation, momentum conservation, and process dynamic characteristics. The PDE (Physical Derivative Architecture) residual network learns spatial distribution deviations that are difficult for the ODE macroscopic mechanism model to describe based on time and spatial coordinates, latent states of operating conditions, control variables, and process parameters. It then fuses the macroscopic mechanism prediction results with the spatial residual correction results through dynamic confidence weights to generate a space-time process state field. Finally, the system uses this space-time process state field for predictive control, optimized scheduling, anomaly diagnosis, predictive maintenance, and adaptive control, forming a long-term self-evolving closed loop through real-time data feedback, error assessment, sample accumulation, and model retraining. This invention upgrades the steel process industry from traditional single-point data monitoring, single-model prediction, and manual experience control to a production line-level intelligent driving system with multimodal perception, mechanism interpretation, spatial field prediction, safety rollback, and continuous evolution capabilities.

[0050] The steel production line control system based on the mechanism-AI coupled world model includes: a data acquisition and preprocessing module, a multimodal fusion sensing module, an ODE macroscopic mechanism model module, a PDE physical AI residual network module, a mechanism-AI fusion output module, an online closed-loop control module, a real-time data feedback and model iteration module, and a safety monitoring and degradation rollback module. These modules are connected in a logical sequence: "data acquisition—multimodal sensing—macroscopic mechanism prediction—spatial residual correction—fusion output—control execution—data feedback—model iteration," forming a closed-loop mechanism-AI coupled world model system for the steel process industry.

[0051] First, the data acquisition and preprocessing module obtains multi-source heterogeneous industrial data from the physical production lines of the steel process industry and organizes it into standardized multimodal samples. Subsequently, the multimodal fusion sensing module converts the standardized multimodal samples into latent operating states. This provides a unified input for mechanism modeling and residual correction; then, the ODE macroscopic mechanism model module generates macroscopic prediction results based on the implicit state of the operating condition, control variables, and process parameters. The PDE physical AI residual network module is based on Spatial coordinates and operating condition information generate spatial residual correction quantities. Then, the mechanism-AI fusion output module will... and The final space-time process state field is generated by fusing the data using dynamic confidence weights. Subsequently, the online closed-loop control module, based on... The system generates control strategies and executes them on production line equipment. Finally, the real-time data feedback and model iteration module records the actual execution results, prediction errors, and new operating condition samples for model retraining and parameter updates. Simultaneously, the safety monitoring and degradation rollback module runs throughout the entire online operation, ensuring the system can roll back to the ODE macroscopic mechanism safety baseline in abnormal situations. Through the connections between these modules, this system forms a complete technical framework of "multimodal perception—macroscopic mechanism prediction—spatial residual compensation—fusion output—online control—data feedback—model self-evolution."

[0052] Specifically, the data acquisition and preprocessing module is used to collect multi-source heterogeneous industrial data from the physical production lines of the steel process industry, and to clean, synchronize, align, standardize, and sample the collected data to form standardized multimodal samples that can be used by subsequent models. The physical production lines of the steel process industry include, but are not limited to, ironmaking, steelmaking, continuous casting, rolling, heat treatment, and finishing processes; the multi-source heterogeneous industrial data includes one-dimensional time-series numerical data, discrete data of switching states, two-dimensional image and spatial data, and frequency domain spectrum data. The data includes: one-dimensional time-series numerical data such as temperature, pressure, flow rate, speed, current, voltage, rolling force, tension, PLC analog signals, valve opening, PID closed-loop curves, cooling water flow rate, and energy medium flow rate; discrete switch-state data such as equipment start / stop signals, operating gears, PLC switch signals, I / O points, valve position switch feedback, alarm status, interlock status, and process mode number; two-dimensional image and spatial data such as visible light industrial images, infrared thermal imaging images, laser contour maps, point cloud projection maps, surface defect images, and equipment status images; and frequency domain spectrum data such as on-site industrial audio, equipment vibration signals, motor current spectrum, ultrasonic flaw detection signals, envelope spectrum, and time spectrum diagrams.

[0053] During data processing, this module first performs outlier removal, filtering, normalization, and sliding time window slicing on one-dimensional time-series numerical data to form a time-series input tensor. Then, it performs state encoding, missing data completion, time preservation, and event alignment on discrete switch-state data to form a discrete state input tensor. Simultaneously, it performs resolution normalization, region of interest extraction, coordinate calibration, and timestamp alignment on image and spatial two-dimensional data to form an image spatial input tensor. Finally, it performs time-frequency transformation, spectrogram generation, normalization, and equipment condition alignment on frequency-domain spectral data to form a spectral input tensor. After completing the above processing, the module uses furnace number, coil number, slab number, material section number, equipment number, process section number, or timestamp as indexes to align data from different modalities to the same material object, the same equipment object, or the same process time, thereby forming standardized multimodal samples.

[0054] The standardized multimodal samples output by this module are sent to the multimodal fusion perception module to generate the hidden state of the current production process. At the same time, the standardized multimodal samples can also be called by the real-time data feedback and model iteration module for offline training, sample accumulation and model retraining.

[0055] Specifically, the multimodal fusion sensing module is used to map industrial multimodal data with different data structures, sampling frequencies, and physical meanings to a unified latent space, forming a latent state of the current production process that can comprehensively characterize the current operating conditions. This module includes a one-dimensional temporal coding submodule, a discrete state coding submodule, an image spatial coding submodule, a frequency domain spectrum coding submodule, a unified dimension mapping submodule, a cross-modal fusion submodule, and a global temporal modeling submodule.

[0056] The one-dimensional temporal coding submodule receives the temporal input tensor output by the data acquisition and preprocessing module, and performs filtering enhancement, local temporal feature extraction, and long-term dependency modeling on the temporal input tensor to obtain one-dimensional temporal coding features. For the input temporal input tensor:

[0057] in, Indicates the number of samples in the batch. Indicates the length of the time window. Indicates the number of channels in the timing variable. Represents the real number field; this submodule can employ one or more combinations of filters, time-frequency analysis, 1D-CNN, GRU, TCN, or Transformer Encoder to output one-dimensional temporal coding features. ;in, This represents the encoding features of one-dimensional time-series numerical data. The discrete state encoding submodule receives the discrete state input tensor output by the data acquisition and preprocessing module and extracts equipment state, process mode, and event change features. For the input discrete state data: ;in, This indicates the number of discrete state channels. This submodule embeds and encodes discrete variables such as device start / stop, operating level, alarm, interlock, and I / O points, and extracts state transition features using a time-series coding model, outputting discrete state coding features. ;in, Encoding characteristics representing discrete data of switch states.

[0058] The image spatial coding submodule receives the image spatial input tensor output by the data acquisition and preprocessing module and extracts its spatial distribution features. For the input image spatial data: ;in, Indicates the image height. Indicates the image width. This indicates the number of image channels. This submodule can employ lightweight 2D-CNN, U-Net encoder, ViT-lite, or other spatial feature extraction networks to output image spatial coding features. ;in, The encoding features representing images and spatial two-dimensional data.

[0059] The frequency domain spectrum coding submodule takes the spectrum input tensor output by the data acquisition and preprocessing module and extracts equipment status and abnormal frequency features. This submodule first converts raw audio, vibration, current, and ultrasonic signals into spectrum diagrams, time-frequency diagrams, wavelet diagrams, or envelope spectra, and then uses a 2D-CNN or spectrum feature extraction network for encoding, outputting frequency domain spectrum coding features. ;in, This represents the coding characteristics of frequency domain spectrum data.

[0060] After obtaining the coding features of each modality, the unified dimension mapping submodule maps the coding features of different modalities to a unified dimension working condition representation space. For any modality m, its coding features are denoted as... This submodule transforms it into: through linear projection, multilayer perceptron, or normalized mapping. Where m represents the mode number, This represents the uniform dimension feature after the m-th modality mapping. This represents the dimension of the unified latent space. The mapping process can be represented as: ,in, This represents the mapping weight matrix for the m-th mode. This represents the bias term for the m-th modality. Through this process, data from different modalities are transformed into a unified feature representation with the same time and feature dimensions.

[0061] The cross-modal fusion submodule is used to establish the relationships between different modalities and dynamically weight the importance of different modalities. This submodule can employ a bidirectional cross-attention mechanism, using features from one modality as the query vector and features from another modality as the key and value, to calculate inter-modal attention enhancement features:

[0062] in, This represents the attention enhancement result of mode n on mode m, where Q represents the query vector, K represents the key vector, and V represents the value vector. After obtaining the inter-modal attention enhancement features, this submodule calculates the dynamic modal weights based on the current operating status and data quality score:

[0063] in, Let g(·) represent the dynamic weights of the m-th mode, g(·) represent the weight generation network, and s represent the current operating state. Let represent the data quality score for the m-th modality. The fused features are represented as follows:

[0064] in, This represents the multimodal fusion feature.

[0065] The global temporal modeling submodule further performs temporal modeling on the multimodal fusion features to capture the evolution of production line conditions over time. This submodule can use GRU, LSTM, TCN, Transformer Encoder, or state-space models to perform global temporal modeling on the multimodal fusion features and output the hidden states of the operating conditions. ;in, This represents the hidden state of the overall operating conditions at time t.

[0066] The multimodal fusion sensing module receives standardized multimodal samples from the data acquisition and preprocessing module as input, and its output... It is jointly invoked by the ODE macroscopic mechanism model module, the PDE physical AI residual network module, and the online closed-loop control module. In other words, the multimodal fusion perception module is the unified entry point for the entire system to enter "mechanism-AI coupled modeling" from "multi-source field data".

[0067] The ODE macroscopic mechanism model module is used to establish a set of ordinary differential equations for the macroscopic state variables of the production line based on material conservation, energy conservation, mass conservation, momentum conservation, reaction kinetics, thermodynamic constraints, and process dynamic characteristics. This serves as the macroscopic mechanism framework for the mechanism-AI coupled world model. This module is not limited to a specific process or physical variable, but rather describes the macroscopic process state evolving over time in the steel process industry, including average temperature, average composition, average thickness, average velocity, average pressure, average flow rate, material inventory, energy consumption, output, equipment load, and predicted quality indicators.

[0068] The macroscopic mechanism model of ODE can be expressed as follows:

[0069] Where S(t) represents the set of macroscopic state variables at time t, U(t) represents the set of control variables at time t, and P(t) represents the set of process parameters at time t. This represents the hidden state of the operating condition output by the multimodal fusion sensing module. Represents the parameters of the macroscopic mechanism model of ODE. This represents a mechanistic function composed of physical conservation relationships, process dynamic relationships, and empirical mechanistic relationships. The formula means that the rate of change of macroscopic state variables over time is jointly determined by the current macroscopic state, control inputs, process parameters, multimodal sensing results, and mechanistic parameters.

[0070] In actual operation, this module performs numerical solutions based on the current macroscopic state, real-time control variables, process parameters, and implicit operating conditions to obtain the macroscopic state prediction results for the next time step or future time window. Furthermore, the ODE macroscopic mechanism model module converts the macroscopic state variables into the target macroscopic prediction results through an output function.

[0071] in, G(·) represents the macroscopic prediction result output by the ODE macroscopic mechanism model at time t, and G(·) represents the output function from the macroscopic state variable to the target predictor variable.

[0072] The above set of ordinary differential equations is first solved to obtain the set of macroscopic state variables. Output function It is based on the control or prediction task from Functions that select, combine, or map target variables, such as extracting average temperature, average composition, average pressure, average flow rate, or quality indicators from macroscopic state variables, to form... .therefore, The result of solving ODE Task-oriented output.

[0073] The ODE macroscopic mechanism model module receives the output from the multimodal fusion sensing module. It combines the control variable U(t) and process parameter P(t) to predict the macroscopic mechanism; its output On the one hand, it serves as the macroscopic benchmark for the mechanism-AI fusion output module; on the other hand, it serves as the residual learning reference for the PDE physical AI residual network module. Therefore, the ODE macroscopic mechanism model module plays the role of "interpretable, safe, and stable mechanism backbone" in the entire system.

[0074] The PDE physical AI residual network module is used to describe and compensate for spatial distribution deviations that are difficult to express by the ODE macroscopic mechanism model, and to learn the local corrections of process conditions in time and space. This module can serve as a general-purpose spatial field residual modeling module, applicable to industrial process variables with spatial distribution attributes such as temperature field, concentration field, composition field, pressure field, velocity field, stress field, strain field, thickness field, plate shape field, defect risk field, and equipment state field.

[0075] This module uses time t, spatial coordinate x, and implicit working state as parameters. Macroscopic state variable S(t), control variable U(t), process parameter P(t), and macroscopic prediction results of ODE As input, the output spatial residual correction is:

[0076] in, This represents the residual correction amount output by the AI ​​network at time t and spatial location x. The parameter is PDE physical AI residual network, This represents the parameters of the PDE physical AI residual network. When the spatial variable is a multidimensional spatial coordinate, x can be represented as:

[0077] in, to These represent different spatial directions, equipment sections, process locations, or discrete grid nodes. The PDE physical AI residual network can employ PINN, PI-CNN, U-Net, GNN, graph attention network, Transformer, coordinate network, or a combination of the above networks. Specifically, PINN or PI-CNN is used to embed partial differential equation constraints into the training process, U-Net or CNN is used to handle regular grid spatial fields, and GNN or graph attention network is used to handle irregular spatial nodes, complex topological relationships, or non-uniform equipment sections.

[0078] The PDE physical AI residual network is based on time and spatial coordinates, latent states of operating conditions, macroscopic state variables, control variables, process parameters, and ODE macroscopic prediction results. It learns the residual of the real spatial field relative to the ODE macroscopic prediction results, thereby achieving physical constraint compensation for spatial distribution deviations.

[0079] During the offline training phase, this module obtains the target spatial field based on real observation data, offline simulation data, or high-precision computational data. ; in, This represents the true or high-confidence process state value at time t and spatial location x. The spatial residual is defined as:

[0080] Here, R(t,x) represents the spatial residual of the real spatial field relative to the macroscopic prediction result of ODE. Through this definition, the learning objective of the PDE physical AI residual network is not to directly replace the mechanistic model, but to learn the spatial bias that the macroscopic mechanistic model of ODE cannot express, thereby reducing the learning difficulty of the AI ​​model and improving its physical interpretability.

[0081] This module also introduces partial differential equation constraints corresponding to the target industrial process variables, which can be expressed in a unified form as follows:

[0082] in, (·) denotes the partial differential equation operator corresponding to the target physical process, Y(t,x) represents the space-time process state field to be predicted, P(t) represents the process parameters, and U(t) represents the control variables. The partial differential equation operator can be set as a heat transfer equation, convection-diffusion equation, mass conservation equation, momentum conservation equation, reaction kinetics equation, elastoplastic deformation equation, fluid-structure interaction equation, or equipment state evolution equation according to different processes.

[0083] The corresponding PDE physical loss is defined as:

[0084] in, Indicates the number of physical constraint sampling points. This indicates that after fusing the macroscopic prediction results of ODE with the output of the PDE physical AI residual network, the sampling points... The predicted state at that point. This loss is used to constrain the network output to satisfy the PDE law loss corresponding to the target physical process.

[0085] The training loss of the PDE (Physical Engineering Dependency) physical AI residual network includes data fitting loss, PDE physical loss, boundary condition loss, initial condition loss, ODE consistency loss, spatial smoothing loss, and residual magnitude constraint loss. The total loss function is expressed as:

[0086] Where L represents the total loss function, This represents the data fitting loss. Indicates PDE physical loss, Indicates boundary condition loss. Represents the initial condition loss. This indicates ODE consistency loss. Indicates spatial smoothing loss, This represents the residual amplitude constraint loss. These represent the weighting coefficients of each loss term.

[0087] The data fitting loss is expressed as:

[0088] in, This indicates the number of sampling points for the observation data.

[0089] The ODE consistency loss is expressed as:

[0090] in, Indicates a spatial region. The spatial average term represents the spatial-temporal process state field generated by the mechanism-AI fusion output module, which represents a measure of a spatial region. In spatial regions The average value over the period. This loss term is used to ensure that the spatial field output by the PDE physical AI residual network is consistent with the output of the ODE macroscopic mechanism model in a macroscopic average sense.

[0091] The residual amplitude constraint loss is expressed as:

[0092] in, This indicates the maximum allowable residual correction magnitude.

[0093] The PDE physical AI residual network module receives the output from the multimodal fusion sensing module. and the output of the ODE macroscopic mechanism model module And generate spatial residual correction amount. The residual correction is then sent to the mechanism-AI fusion output module for fusion with the ODE macroscopic prediction results. Therefore, this module plays the role of "spatial bias compensation" and "physical constraint AI correction".

[0094] The mechanism-AI fusion output module is used to fuse the output of the ODE macroscopic mechanism model with the output of the PDE physical AI residual network to generate the final space-time process state field. This module does not directly replace the ODE macroscopic mechanism model, nor does it directly rely on pure AI prediction results. Instead, it adopts a "macroscopic mechanism prediction + spatial residual correction" approach, combining the stability and interpretability of ODE with the spatial representation capability of the PDE physical AI residual network.

[0095] The fusion output can be expressed as:

[0096] in, This represents the final output space-time process state field. This represents the macroscopic prediction result output by the ODE macroscopic mechanism model. This represents the spatial residual correction amount output by the PDE physical AI residual network. This represents the dynamic confidence weight. When only time-level confidence control is required... It can be simplified to When different degrees of residual correction are required for different spatial regions, For spatial distribution weights.

[0097] The dynamic confidence weights can be jointly determined by the output confidence of the PDE physical AI residual network, the data quality score, the PDE physical residuals, the operating condition distribution matching degree, and the communication health status.

[0098] in, This represents the Sigmoid function. This represents a weight generation network. Represents the physical residual of the PDE. This indicates the degree of deviation between the current operating conditions and the training distribution. Indicates the data quality score. This indicates the health status of communication. Through this weight design, when the PDE physical AI residual network is in a reliable state, the system enhances the spatial residual correction effect; when the model is unreliable or the input conditions exceed the training distribution, the system reduces the residual correction weight.

[0099] The mechanism-AI fusion output module receives the output from the ODE macroscopic mechanism model module. and the output of the PDE physical AI residual network module and output The final space-time process state field is sent to the online closed-loop control module, anomaly diagnosis module, predictive maintenance module, and visualization module for subsequent control decisions and state assessment. Therefore, this module is the core interface connecting "mechanism-AI modeling" and "online control execution".

[0100] The online closed-loop control module is used to generate control strategies based on the final space-time process state field, and then distribute them to production line equipment or control systems after safety verification. This module includes a real-time data access submodule, a real-time sensing submodule, a real-time simulation submodule, a space correction submodule, a safety constraint verification submodule, a control strategy generation submodule, and an equipment execution submodule.

[0101] The real-time data access submodule acquires current production data in real time from PLC, DCS, MES, L1 basic automation system, L2 process control system, sensor network, image system and detection system.

[0102] The real-time sensing submodule calls the multimodal fusion sensing module to calculate the real-time hidden state of the operating condition at online operation time t based on the current real-time production data. ; in, This indicates the real-time hidden status of the operating condition at time t during online operation.

[0103] The real-time simulation submodule calls the ODE macroscopic mechanism model module to calculate the macroscopic prediction results for the next time step or future time window based on the current macroscopic state, control variables, process parameters, and implicit operating conditions.

[0104] in, This indicates the prediction time step.

[0105] The spatial correction submodule calls the PDE physical AI residual network module to calculate the spatial residual correction amount based on the real-time hidden state, macroscopic prediction results, control variables, process parameters, and spatial coordinates.

[0106] The final spatial-temporal process state field is generated through the mechanism-AI fusion output module.

[0107] Before generating the control strategy, the safety constraint verification submodule performs checks on process boundaries, equipment capabilities, upper and lower limits of control variables, interlocking conditions, communication status, confidence levels, and abnormal rollback conditions. When the current operating conditions, model output, and equipment status meet the safety constraints, the system enters the control strategy generation stage; when the safety constraints are not met, the system refuses to generate or issue high-risk control strategies and triggers alarms, conservative control, or ODE safety baseline rollback.

[0108] The control strategy generation submodule generates control strategies based on the final space-time process state field, target process parameters, equipment constraints, and production plan. These control strategies include temperature settings, speed settings, reduction settings, tension settings, flow rate settings, pressure settings, valve opening settings, energy medium distribution strategies, production rhythm adjustment strategies, equipment maintenance strategies, and anomaly handling strategies. The control strategy can be represented as:

[0109] Where U*(t) represents the optimal control strategy, U represents the control variable to be optimized, and J(·) represents the control objective function. This represents the target process state, and C represents equipment capability constraints, safety boundary constraints, and process specification constraints. The equipment execution submodule sends the control strategy, which has undergone safety verification, to the corresponding equipment, controller, or actuator through the manufacturing API, equipment API, energy API, logistics API, or safety API.

[0110] The online closed-loop control module receives the mechanism - AI fusion output module generated by the module The data is then transformed into an executable control strategy U*(t). The actual results after the control strategy is executed are fed back to the real-time data feedback and model iteration module. Therefore, this module realizes a closed-loop implementation from "model prediction" to "safety decision-making" and then to "equipment execution". The real-time data feedback and model iteration module is used to receive real-time data, model prediction results, control strategies, equipment execution results and actual feedback results during online operation, and to continuously evaluate and iteratively update the model. This module includes an online prediction recording submodule, a result feedback submodule, an error evaluation submodule, a sample evaluation submodule, a sample accumulation submodule and a model retraining submodule.

[0111] The online prediction and recording submodule records the output of the multimodal fusion perception module. The output of the ODE macroscopic mechanism model module The output of the PDE physical AI residual network module The output of the mechanism-AI fusion output module The control strategy U*(t) generated by the online closed-loop control module is also included. The result feedback submodule collects the actual process status, quality indicators, equipment response, and abnormal information after the equipment is executed.

[0112] The error assessment submodule performs simultaneous simulation tracking error and control strategy error assessment. The simulation tracking error is used to compare the actual feedback results after device execution with the simulation prediction results.

[0113] in, This represents the model simulation prediction results. This represents the actual feedback result after the device executes the command. When A consistently large error indicates that the model is not accurately reflecting real-world conditions, requiring the model to learn new residuals or update its parameters. The control strategy error is used to compare the deviations between the setpoint, simulated values, and actual values.

[0114]

[0115]

[0116] in, This represents the target setpoint given by the control strategy. This represents the simulation results after the model predictions are executed. This represents the actual execution feedback value. If near but Deviation indicates a problem with equipment execution or on-site disturbance; if itself deviates This indicates that the control strategy generation effect is not good; if and If the long-term deviation is large, it indicates that the model residuals need to be updated.

[0117] The sample evaluation submodule assesses the sample value based on multimodal fusion results, data quality, deviation from operating condition distribution, simulation error, control error, and operating condition status.

[0118] in, Indicates the sample value score. Indicates the quality of multimodal data. This indicates the degree of deviation between the current operating conditions and the distribution of the training data. This represents the simulation tracking error. Indicates the control strategy error. This indicates the operating condition or anomaly type. The sample sedimentation submodule writes high-value samples, abnormal samples, boundary condition samples, new steel grade samples, new specification samples, and samples with large model errors into the sample library.

[0119] The model retraining submodule updates the multimodal fusion sensing model, ODE parameters, PDE physical AI residual network parameters, and control strategy model parameters periodically or event-triggered based on the aforementioned sample library. Retraining trigger conditions include: simulation error continuously exceeding a threshold, control strategy error continuously exceeding a threshold, the appearance of new steel grades, new specifications, or new working conditions, a significant deviation of the current working condition distribution from the training set, the number of high-value samples in the sample library reaching a threshold, the expiration of periodic maintenance training time, persistently large PDE residuals, and an increased frequency of safety rollbacks.

[0120] The real-time data feedback and model iteration module receives the execution results from the online closed-loop control module and converts these results into samples required for model training and parameter updates. The updated model is then invoked again by the multimodal fusion perception module, the ODE macroscopic mechanism model module, the PDE physical AI residual network module, and the online closed-loop control module. Therefore, this module implements a closed-loop self-evolutionary process of "online prediction—safety constraint verification and control strategy generation—equipment execution—result feedback—error assessment—sample assessment—sample accumulation—model retraining—model update".

[0121] The security monitoring and degradation rollback module is used to monitor the operating status of the PDE physical AI residual network, the mechanism-AI fusion output module, the online closed-loop control module, and the communication link, and to ensure the safe operation of the system in abnormal situations. This module continuously monitors input data, model output, physical residuals, model confidence, communication heartbeat, and inference latency to determine whether the current model results meet the security requirements of the industrial control scenario.

[0122] In actual operation, this module performs input anomaly detection to check whether the current input data exceeds the training distribution or the range allowed by the process; or The system performs the following checks: whether the output exceeds the physically feasible range; whether the model output meets the set PDE physical constraints; whether the dynamic confidence weight is lower than the safety threshold; whether the communication between modules is normal; and whether the online inference of the model meets the control cycle requirements.

[0123] When the security monitoring and degradation rollback module detects an anomaly, the system executes a degradation strategy, causing the dynamic confidence weight to gradually decay:

[0124] And make the final output return to:

[0125] The system smoothly switches from a hybrid mode of "ODE macroscopic mechanism model + PDE physical AI residual network" to a safety baseline mode supported only by the ODE macroscopic mechanism model. The rollback process uses a smooth decay method to avoid abrupt changes in control setpoints, and records abnormal information, input and output data before and after the abnormality, and model status in the fault analysis library.

[0126] The security monitoring and degradation rollback module runs through the entire process of online inference, fusion output, and control execution. It provides security constraints on the PDE physical AI residual network and the mechanism-AI fusion output module, and switches the system to the ODE macroscopic mechanism model output in abnormal situations. Thus, while pursuing high-precision spatial field prediction and intelligent control, the system retains the security benchmark and reliable rollback capability required for industrial control.

[0127] This invention encodes one-dimensional time-series data, discrete state data, image spatial data, and frequency domain spectrum data in the steel process industry into the same working condition latent space, avoiding the problem of using different modal data in a fragmented manner. By employing a bidirectional cross-attention and dynamic modal weighting mechanism, the system can adaptively select effective modalities based on current operating conditions and data quality, enhancing the robustness of multimodal perception. The ODE macroscopic mechanism model serves as the macroscopic mechanism framework, preserving the physical interpretability and safety benchmark capabilities of traditional mechanism models. Instead of directly replacing the mechanism model, the PDE physical AI residual network learns spatial residuals, thereby improving spatial field prediction accuracy while reducing the risk of AI model runaway. Through ODE consistency constraints, the spatial field output by the PDE physical AI residual network is kept consistent with the ODE macroscopic mechanism model in an overall average sense, thus preventing AI correction from disrupting macroscopic conservation relationships. Dynamic confidence weights and a safety degradation fallback mechanism are employed to automatically revert to the ODE macroscopic mechanism model output when the AI ​​model becomes abnormal or unreliable, meeting the safety requirements of industrial control systems. A closed-loop self-evolutionary process of "offline modeling—online inference—control execution—result feedback—sample accumulation—model retraining" is constructed, enabling the model to adapt to long-term changes in operating conditions.

[0128] The steel production line control method based on a mechanism-AI coupled world model includes: Step S1: Collect multi-source heterogeneous industrial data from the physical production line of the steel process industry through the data acquisition and preprocessing module, and preprocess the collected multi-source heterogeneous industrial data to obtain standardized multimodal data. Step S2: The standardized multimodal data is mapped to a unified latent space through the multimodal fusion sensing module to form a latent state of the working condition that can comprehensively represent the current production process. Step S3: The ODE macroscopic mechanism model module establishes a set of ordinary differential equations for the macroscopic state variables of the production line based on material conservation, energy conservation, mass conservation, momentum conservation, reaction kinetics, thermodynamic constraints, and process dynamic characteristics, which serves as the ODE macroscopic mechanism model. The ODE macroscopic mechanism model generates macroscopic prediction results based on the hidden state of the current production process, control variables, and process parameters. Step S4: Generate spatial residual correction based on macroscopic prediction results, spatial coordinates, and working condition information using the PDE Physical AI Residual Network module; Step S5: The macroscopic prediction results and spatial residual corrections are fused with dynamic confidence weights through the mechanism-AI fusion output module to generate the final spatial-temporal process state field; Step S6: The online closed-loop control module generates a control strategy based on the final space-time process state field and executes it on the production line equipment; Step S7: Receive real-time data, prediction results, control strategies, equipment execution results, and actual feedback results during online operation through the real-time data feedback and model iteration module, and perform continuous evaluation and iterative updates; Step S8: Monitor the operating status of the PDE physical AI residual network, mechanism-AI fusion output module, online closed-loop control module and communication link through the safety monitoring and degradation rollback module, and ensure safe operation in abnormal situations, including: continuously monitoring input data, model output, physical residual, model confidence, communication heartbeat and inference delay to determine whether the current model results meet the safety requirements of industrial control scenarios.

[0129] Those skilled in the art will understand that, in addition to implementing the system, apparatus, and their modules provided by this invention in purely computer-readable program code, the same program can be implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system, apparatus, and their modules provided by this invention can be considered a hardware component, and the modules included therein for implementing various programs can also be considered structures within the hardware component; alternatively, modules for implementing various functions can be considered both software programs implementing the method and structures within the hardware component.

[0130] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A steel production line control system based on a mechanism-AI coupled world model, characterized in that, include: The data acquisition and preprocessing module is used to collect multi-source heterogeneous industrial data from the physical production line of the steel process industry, and to preprocess the collected multi-source heterogeneous industrial data to obtain standardized multimodal data. The multimodal fusion sensing module is used to map standardized multimodal data to a unified latent space to form a latent state of the current production process that can comprehensively represent the current production process. The ODE macroscopic mechanism model module is used to establish a set of ordinary differential equations for the macroscopic state variables of the production line based on material conservation, energy conservation, mass conservation, momentum conservation, reaction kinetics, thermodynamic constraints, and process dynamic characteristics, as the ODE macroscopic mechanism model; the ODE macroscopic mechanism model generates macroscopic prediction results based on the hidden state of the current production process, control variables, and process parameters. The PDE Physical AI Residual Network Module is used to generate spatial residual corrections based on macroscopic prediction results, spatial coordinates, and operating condition information using the PDE Physical AI Residual Network. The mechanism-AI fusion output module is used to fuse macroscopic prediction results with spatial residual corrections according to dynamic confidence weights to generate the final spatial-temporal process state field. The online closed-loop control module is used to generate control strategies based on the final space-time process state field and execute them on the production line equipment.

2. The steel production line control system based on a mechanism-AI coupled world model according to claim 1, characterized in that, The data acquisition and preprocessing module includes: Module M1.1: Used to collect data from the physical production lines of the steel industry, including ironmaking, steelmaking, continuous casting, rolling, heat treatment, and finishing, including one-dimensional time-series numerical data, discrete switching state data, two-dimensional image and spatial data, and frequency domain spectrum data. The one-dimensional time-series numerical data includes: temperature, pressure, flow rate, speed, current, voltage, rolling force, tension, PLC analog quantity, valve opening, PID closed-loop curve, cooling water flow rate, and energy medium flow rate. The discrete data of the switch status includes: equipment start / stop signal, operating level, PLC switch quantity, IO point, valve position switch feedback, alarm status, interlock status, and process mode number. The image and spatial two-dimensional data include: visible light industrial images, infrared thermal imaging images, laser contour maps, point cloud projection maps, surface defect images, and equipment status images; The frequency domain spectrum data includes: on-site industrial audio, equipment vibration signals, motor current spectrum, ultrasonic flaw detection signals, envelope spectrum, and time spectrum. Module M1.2: Performs outlier removal, filtering, normalization, and sliding time window slicing on one-dimensional time-series numerical data to form a time-series input tensor; performs state encoding, missing data completion, time preservation, and event alignment on discrete switch-state data to form a discrete state input tensor; performs resolution normalization, region of interest extraction, coordinate calibration, and timestamp alignment on two-dimensional image and spatial data to form an image spatial input tensor; and performs time-frequency transformation, spectrogram generation, normalization, and equipment condition alignment on frequency-domain spectral data to form a spectral input tensor. Module M1.3: Using furnace number, coil number, slab number, material section number, equipment number, process section number or timestamp as indexes, it aligns data of different modalities to the same material object, the same equipment object or the same process time, thereby forming standardized multimodal data.

3. The steel production line control system based on a mechanism-AI coupled world model according to claim 1, characterized in that, The multimodal fusion sensing module includes: Module M2.1: Extracts the temporal dynamic features of the temporal input tensor to obtain one-dimensional temporal coding features; Module M2.2: Extracts the equipment state, process mode, and event change features from the discrete state input tensor to obtain discrete state encoding features; Module M2.3: Extracts the spatial distribution features of the image spatial input tensor to obtain the image spatial coding features; Module M2.4: Extracts device status and abnormal frequency features from the spectral input tensor to obtain frequency domain spectral coding features; Module M2.5: Maps one-dimensional temporal coding features, discrete state coding features, image spatial coding features, and frequency domain spectral coding features to a unified-dimensional working condition representation space; in, The encoding features of mode m; This represents the mapping weight matrix for the m-th mode; This represents the bias term for the m-th mode; This represents the uniform dimension feature after the m-th modality mapping; Module M2.6: Establishes the correlation between different modalities and dynamically weights the importance of different modalities to obtain multimodal fusion features, including: A bidirectional cross-attention mechanism is employed, using one modality feature as the query vector and another modality feature as the key and value, to calculate intermodal attention enhancement features: in, Let Q represent the attention enhancement result of mode n on mode m, where Q represents the query vector, K represents the key vector, and V represents the value vector; After obtaining the intermodal attention enhancement features, dynamic modal weights are calculated based on the current operating conditions and data quality scores: in, Let g(·) represent the dynamic weights of the m-th mode, g(·) represent the weight generation network, and s represent the current operating state. This represents the data quality score for the m-th modality; The multimodal fusion feature is represented as follows: in, This represents the multimodal fusion feature. Module M2.7: Performs time-series modeling for multimodal fusion features to capture the evolution of production line conditions over time and obtain the hidden states of the conditions.

4. The steel production line control system based on a mechanism-AI coupled world model according to claim 1, characterized in that, The macroscopic mechanism model of ODE includes: Where S(t) represents the set of macroscopic state variables at time t, U(t) represents the set of control variables at time t, and P(t) represents the set of process parameters at time t. This represents the hidden state of the operating condition output by the multimodal fusion sensing module. This represents the parameters of the macroscopic mechanism model of ODE. It represents the mechanism function composed of physical conservation relations, process dynamic relations, and empirical mechanism relations; The ODE macroscopic mechanism model module outputs a function. Transform macroeconomic state variables into target macroeconomic forecast results: in, G(·) represents the macroscopic prediction result output by the ODE macroscopic mechanism model at time t, and G(·) represents the output function from the macroscopic state variable to the target predictor variable.

5. The steel production line control system based on a mechanism-AI coupled world model according to claim 1, characterized in that, The PDE physical AI residual network module includes: time t, spatial coordinate x, and hidden state of operation. Macroscopic state variable S(t), control variable U(t), process parameter P(t), and macroscopic prediction results As input, the output spatial residual correction is: in, This indicates that the PDE physical AI residual network is in time t and spatial location The residual correction amount output at that point. The parameter is PDE physical AI residual network, Indicates the parameters of the PDE physical AI residual network; When the spatial variable is a multidimensional spatial coordinate. Represented as: in, to These represent different spatial directions, equipment sections, process locations, or discrete grid nodes, respectively. The PDE physical AI residual network learns the residual of the real spatial field relative to the ODE macroscopic prediction results based on time, spatial coordinates, hidden states of the operating conditions, macroscopic state variables, control variables, process parameters, and ODE macroscopic prediction results, thereby achieving physical constraint compensation for spatial distribution deviations.

6. The steel production line control system based on a mechanism-AI coupled world model according to claim 5, characterized in that, The PDE physical AI residual network includes: During the offline training phase, the target spatial field is obtained based on real observation data, offline simulation data, or computational data. ; This represents the true or high-confidence process state value that meets preset requirements at time t and spatial location x. Spatial residual is defined as: Where R(t,x) represents the spatial residual of the real spatial field relative to the macroscopic prediction result of ODE; By introducing partial differential equation constraints corresponding to the target industrial process variables, their unified form is expressed as: Where N(·) represents the partial differential equation operator corresponding to the target physical process, Y(t,x) represents the space-time process state field to be predicted, P(t) represents the process parameters, and U(t) represents the control variables; The total training loss function of the PDE physical AI residual network includes: Where L represents the total loss function, This represents the data fitting loss. Indicates PDE physical loss. Indicates boundary condition loss. Represents the initial condition loss. This indicates ODE consistency loss. Indicates spatial smoothing loss, This represents the residual amplitude constraint loss. These represent the weighting coefficients of each loss term; The physical losses in PDE include: in, Indicates the number of physical constraint sampling points. This indicates that after fusing the macroscopic prediction results of ODE with the output of the PDE physical AI residual network, the sampling points... The predicted state at the point; the PDE physical loss is used to constrain the network output to satisfy the PDE regularity loss corresponding to the target physical process; The data fitting loss includes: in, Indicates the number of sampling points for the observation data; The ODE consistency loss includes: in, Indicates a spatial region. The spatial average term represents the spatial-temporal process state field generated by the mechanism-AI fusion output module, which represents a measure of a spatial region. In spatial regions The average value on the above; the ODE consistency loss is used to ensure that the spatial field output by the PDE physical AI residual network is consistent with the output of the ODE macroscopic mechanism model in a macroscopic average sense; The residual magnitude constraint loss includes: in, This indicates the maximum allowable residual correction magnitude.

7. The steel production line control system based on a mechanism-AI coupled world model according to claim 1, characterized in that, The mechanism-AI fusion output module includes: in, This represents the final output space-time process state field. This indicates the results of macroeconomic forecasts. This represents the spatial residual correction amount output by the PDE physical AI residual network. Indicates dynamic confidence weights; The dynamic confidence weights are determined by the output confidence of the PDE physical AI residual network, the data quality score, the PDE physical residual, the operating condition distribution matching degree, and the communication health status. in, This represents the Sigmoid function. This represents a weight generation network. Represents the physical residual of the PDE. This indicates the degree of deviation between the current operating conditions and the training distribution. Indicates the data quality score. Indicates the health status of communication.

8. The steel production line control system based on a mechanism-AI coupled world model according to claim 1, characterized in that, The online closed-loop control module includes: Module M6.1: Acquires real-time production data from systems including PLC, DCS, MES, L1 basic automation system, L2 process control system, sensor network, imaging system, and detection system; Module M6.2: Calls the multimodal fusion sensing module to calculate the real-time hidden state of the operating condition at online operation time t based on the current real-time production data. ; Module M6.3: Calls the ODE macroscopic mechanism model module to calculate the macroscopic prediction results for the next time step or future time window based on the current macroscopic state, control variables, process parameters, and implicit operating conditions. ;in, Indicates the prediction time step; Module M6.4: Calls the PDE Physical AI Residual Network module to calculate the spatial residual correction based on real-time operating condition hidden states, macroscopic prediction results, control variables, process parameters, and spatial coordinates. The final spatial-temporal process state field is generated through the mechanism-AI fusion output module. ; Module M6.5: Before generating the control strategy, it performs process boundary checks, equipment capability checks, control quantity upper and lower limit checks, interlock condition checks, communication status checks, confidence level checks, and abnormal rollback condition checks. When the current operating condition, model output, and equipment status meet the safety constraints, it enters the control strategy generation stage. When the safety constraints are not met, it refuses to generate or issue high-risk control strategies and triggers alarms, conservative control, or ODE safety baseline rollback. Module M6.6: Generates control strategies based on the final space-time process state field, target process indicators, equipment constraints, and production plan; the control strategies include temperature setting, speed setting, pressure reduction setting, tension setting, flow rate setting, pressure setting, valve opening setting, energy medium distribution strategy, production rhythm adjustment strategy, equipment maintenance strategy, and abnormal handling strategy; The control strategy is expressed as: Where U*(t) represents the optimal control strategy, U represents the control variable to be optimized, and J(·) represents the control objective function. The target process state is represented by C, which represents equipment capacity constraints, safety boundary constraints, and process specification constraints. Module M6.6: Sends control policies that have passed security verification to the corresponding devices, controllers or actuators via the manufacturing API, equipment API, energy API, logistics API or security API.

9. The steel production line control system based on a mechanism-AI coupled world model according to claim 1, characterized in that, The system also includes a real-time data feedback and model iteration module, used to receive real-time data, prediction results, control strategies, equipment execution results, and actual feedback results during online operation, and to continuously evaluate and iteratively update them, including: Module M7.1: Records the hidden states of operation conditions output by the multimodal fusion sensing module. Macroscopic prediction results output by the ODE macroscopic mechanism model module Spatial residual correction amount output by the PDE physical AI residual network module The fusion output of the mechanism-AI fusion output module And the control strategy U*(t) generated by the online closed-loop control module; Module M7.2: Collects the actual process status, quality indicators, equipment response, and abnormal information after the equipment is executed; Module M7.3: Simultaneously evaluates simulation tracking error and control strategy error; The simulation following error is used to compare the actual feedback result after the device executes with the simulation prediction result: ;in, This represents the simulation prediction results. This indicates the actual feedback result after the device executes the command; when When the value is consistently greater than the preset value, it indicates that the data is not closely following the actual working conditions and new residuals need to be learned or parameters need to be updated. The control strategy error is used to compare the deviation between the set value, the simulated value, and the actual value: in, This represents the target setpoint given by the control strategy. This indicates the simulation results after the prediction is executed. Indicates the actual execution feedback value; if near but If the deviation is met but the preset requirements are still met, it indicates a problem with the equipment execution or on-site disturbance; if itself deviates If the preset requirements are met, it indicates that the control strategy generation effect is not good; if and If the long-term deviation is large but the preset requirements are met, it indicates that the residual needs to be updated. Module M7.4: Evaluate sample value based on multimodal fusion results, data quality, deviation from operating condition distribution, simulation error, control error, and operating condition status. in, Indicates the sample value score. : Indicates the quality of multimodal data. This indicates the degree of deviation between the current operating conditions and the distribution of the training data. This represents the simulation tracking error. Indicates the control strategy error. Indicates the operating condition or abnormality type; Module M7.5: Writes high-value samples, abnormal samples, boundary condition samples, new steel grade samples, new specification samples, and samples with large model errors into the sample library; Module M7.6: Based on the sample library, periodically or event-triggered updates of the multimodal fusion perception model, ODE parameters, PDE physical AI residual network parameters, and control strategy model parameters; retraining trigger conditions include: simulation error continuously exceeding a threshold, control strategy error continuously exceeding a threshold, the appearance of new steel grades, new specifications, and new working conditions, the current working condition distribution significantly deviating from the training set and meeting preset requirements, the number of high-value samples in the sample library meeting preset requirements reaching a threshold, the periodic maintenance training time arriving, PDE residuals being consistently large and meeting preset requirements, and the safety rollback frequency increasing and meeting preset requirements.

10. The steel production line control system based on a mechanism-AI coupled world model according to claim 1, characterized in that, The system also includes a safety monitoring and degradation rollback module, which monitors the operating status of the PDE physical AI residual network, the mechanism-AI fusion output module, the online closed-loop control module, and the communication link, and ensures safe operation in abnormal situations. This includes continuously monitoring the input data, model output, physical residuals, model confidence, communication heartbeat, and inference latency to determine whether the current model results meet the safety requirements of the industrial control scenario.

Images