A system and method for synergistic optimization of endpoint temperature and feedstock basicity for a sintering process

Abstract

This invention discloses a system and method for the synergistic optimization of endpoint temperature and raw material alkalinity in a sintering process. The system includes an independent auxiliary decision server and several functional modules integrated within it: a data acquisition module, an online raw material composition calculation module, a sintering endpoint prediction module, a causal decoupling and synergistic coefficient online identification module, a multivariable asymmetric time-delay decoupling module, an optimization suggestion module, and a reliable self-evolution module. The causal decoupling module constructs a structural causal model and calculates and outputs the decoupled synergistic optimization coefficients through online intervention, switching to a decoupling control mode when strong coupling occurs. The time-delay decoupling module constructs a multivariable time-delay matrix and deduces the state evolution across multiple time scales in a virtual reality space. The reliable self-evolution module introduces causal consistency verification to force model updates to not violate physical laws. This invention achieves causal decoupling, asymmetric time-delay compensation, and reliable self-evolution, solving the problems of strong coupling interference and control oscillations in the sintering process.

Description

Technical Field

[0001] This invention belongs to the field of intelligent auxiliary decision-making technology for sintering processes in iron and steel metallurgy, specifically relating to a system and method for synergistic optimization of the endpoint temperature and raw material alkalinity in the sintering process. Background Technology

[0002] During the sintering process, operators need to judge and adjust the sintering endpoint temperature and the basicity of the mixture based on the raw material composition and real-time process status. Currently, this process mainly relies on personal experience and lacks real-time, quantitative data prediction support, which often leads to delayed adjustments and makes it difficult to accurately and stably control the basicity near the target value, directly affecting the quality of sintered ore and production efficiency.

[0003] In existing technologies, raw material data is typically stored in a ClickHouse database, while process data originates from the on-site PLC system. These two types of data are independent and lack effective integration, making it impossible to obtain the true composition of the mixture and its dynamic impact on the sintering endpoint in real time. Furthermore, although some data acquisition or prediction algorithms exist, they are mostly isolated software programs that lack a unified hardware platform and standardized module connection methods, making them susceptible to interference and lacking in stability in complex industrial environments.

[0004] A search revealed several sintering process control schemes in the prior art. For example, prior art document 1 (D1: CN121254644A) discloses a multi-process collaborative control method and system for sintering, but its collaboration is essentially a shallow coupling based on rules (such as reducing machine speed when the ignition temperature is high), failing to address the technical pain point that "adjusting alkalinity will uncontrollably affect the endpoint, and adjusting the endpoint will accidentally destroy alkalinity." Prior art document 2 (D2: CN119763696A) discloses a sintering endpoint prediction method based on SVR, but it is limited to endpoint prediction and does not involve collaborative optimization with alkalinity. Prior art document 3 (D3: JPS5665943A) discloses an alkalinity control method based on CaO / SiO2 ratio calculation, but it only achieves single-objective feedback control and lacks dual-objective collaborative capability. Prior art document 4 (D4: CN103589862A) discloses a static batching adjustment method based on the alkalinity balance equation, but it does not consider the dynamic impact of batching adjustment on the endpoint.

[0005] Furthermore, the sintering process exhibits significant time delays (it typically takes 20-40 minutes from batch adjustment to the sintering endpoint response), and the response times of different input variables to different output variables are asymmetric (for example, flux adjustment affects basicity in 15 minutes but affects the endpoint in 25 minutes; machine speed adjustment affects the endpoint in 10 minutes but has virtually no effect on basicity). Existing technologies, such as the time delay compensation method disclosed in prior art (CN201110085227.7), only passively align single-variable time delays and cannot solve the coordinated control problem of multi-variable asymmetric time delay coupled systems.

[0006] Therefore, current sintering production still faces problems such as multi-source data separation, dispersed prediction functions, strong coupling interference, significant impact of asymmetric time delay, and lack of reliable self-evolution mechanism. There is an urgent need for an integrated hardware platform and supporting functional modules to achieve integrated collaborative optimization of data fusion, causal decoupling, asymmetric time delay compensation and reliable self-evolution. Summary of the Invention

[0007] This invention provides a system and method for co-optimizing the endpoint temperature and raw material alkalinity in the sintering process, aiming to solve the problems of strong coupling interference, significant asymmetric time delay, and lack of reliable self-evolution mechanism in the existing technology of sintering process, and to achieve true dual-objective decoupled control and multi-timescale co-optimization.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0009] A system for co-optimizing the endpoint temperature and raw material alkalinity in a sintering process includes a hardware carrier and multiple functional modules integrated on the hardware carrier.

[0010] Hardware platform: The hardware platform is an independent auxiliary decision-making server, equipped with a metal anti-interference shielding layer and heat dissipation module, and has a built-in database. The server is sized for installation in industrial control cabinets and is equipped with an industrial-grade processor and memory to ensure stable operation in complex environments.

[0011] Functional modules:

[0012] Data acquisition module: Integrated inside the server, including a first acquisition unit and a second acquisition unit. The first acquisition unit connects to the factory's ClickHouse database through a dedicated interface to collect chemical element analysis data of materials in the batching silo and instantaneous feeding data. The second acquisition unit connects to the sintering machine PLC system through an industrial Ethernet gateway to collect data on wind box temperature, negative pressure, material layer thickness, and machine speed.

[0013] Data integration module: Connected to the data acquisition module, it is used to align timestamps of multi-source data, clean the data, and store it uniformly in the built-in database.

[0014] Online raw material composition calculation module: Connected to the database, it dynamically calculates the instantaneous chemical composition of the mixture based on the real-time feed volume of each batching bin and the fixed test data of the materials in each bin.

[0015] Sintering endpoint prediction module: It has a built-in first prediction model (preferably XGBoost model, long short-term memory network or temporal attention network), whose inputs include chemical composition of mixture, wind box temperature, material layer thickness, machine speed and ignition temperature, and output the predicted value of sintering endpoint temperature.

[0016] The online identification module for causal decoupling and synergy coefficients is integrated within the server. It contains a structural causal model (SCM) of the sintering process, used to identify the dynamic coupling coefficient matrix K(t) between alkalinity and the final temperature in real time. It outputs the decoupled synergy optimization coefficients through online intervention calculation (Do-calculus). The specific implementation of this module includes:

[0017] Construct a structural cause-effect diagram of the sintering process to clarify the causal direction between variables (e.g., fuel ratio → final temperature, but the final temperature will not affect the fuel ratio in the reverse direction).

[0018] By calculating the expected change in endpoint temperature if alkalinity is adjusted alone through online intervention, a pure causal effect was obtained.

[0019] A dynamic coupling coefficient matrix K(t) is constructed based on pure causal effects, where the off-diagonal elements k ij (t) represents the causal effect strength of the j-th input variable on the i-th output variable;

[0020] When the absolute value of any off-diagonal element exceeds 0.7, the coupling coefficient is determined to exceed the preset threshold. The module automatically switches to decoupling control mode, prioritizing the adjustment of neutral variables (such as material layer thickness) that have the least impact on the coupled system, thus avoiding oscillations caused by directly adjusting strongly coupled variables.

[0021] Multivariable Asymmetric Time-Delay Decoupling Module: Integrated within the server, this module identifies the asymmetric time-delay response times of each input variable to each output variable, constructs a multivariable time-delay matrix, and, based on a model predictive control (MPC) framework, extrapolates future state evolution across multiple time scales in virtual reality space, generating corrected predicted values ​​after time-delay decoupling. The specific implementation of this module includes:

[0022] Based on historical operational data, cross-correlation analysis is used to identify the response delay time of each input variable to each output variable, and a multivariate time delay matrix is ​​constructed. ;

[0023] The time delay matrix diagonalization algorithm is used to decouple the strongly coupled multivariable time delay system into multiple independent single-variable time delay subsystems;

[0024] A digital twin (virtual reality space) of the sintering process is constructed in the server memory, and the process is projected forward in second-level time steps, simultaneously displaying the predicted evolution trajectory of the final temperature and alkalinity within the next 30 minutes.

[0025] For each subsystem, prediction and control are performed separately, and then the control variables are mapped back to the original variable space through inverse transformation.

[0026] The optimization suggestion module is connected to the causal decoupling and synergy coefficient online identification module, the multivariate asymmetric time delay decoupling module, and the target setting module. When the deviation between the predicted value and the target setting value exceeds the threshold, it generates a comprehensive adjustment suggestion based on the synergy optimization coefficient and the corrected predicted value after time delay decoupling. This suggestion includes a synergistic adjustment range for the feed volume of multiple batching bins, the fuel ratio adjustment amount, the engine speed adjustment amount, and the expected response time of each adjustment at multiple time scales (calculated based on the asymmetric time delay matrix).

[0027] Target setting module: It is equipped with a human-machine interface to receive the sintering endpoint target temperature and raw material target alkalinity set by the operator, and supports adaptive adjustment of the deviation threshold according to the working conditions.

[0028] Trusted self-evolution module: Integrated within the server, including:

[0029] Deviation monitoring unit is used to record and accumulate the deviation between predicted and actual values;

[0030] The trigger judgment unit is used to generate a model update instruction when the accumulated deviation exceeds a preset threshold;

[0031] Incremental training unit, used to incrementally train the first prediction model based on new data;

[0032] The causal consistency verification unit is used to force a verification after incremental training of the model to check whether the input-output relationship of the new model is consistent with the pre-set sintering physicochemical mechanism knowledge base. If the causal direction or significance changes in violation of physical laws (for example, the domain knowledge stipulates that "increasing the fuel ratio will increase the endpoint temperature", but the new model shows that "increasing the fuel ratio will decrease the endpoint temperature"), the model update will be prevented and manual review will be triggered.

[0033] The virtual sample generation unit generates virtual operating condition data that conforms to the constraints of material balance and heat balance based on generative adversarial network (GAN) to expand the incremental training dataset and prevent the model from overfitting in sparse data regions.

[0034] Deploy the verification unit to perform performance verification on the updated model, and deploy it online after successful verification.

[0035] Methods and solutions

[0036] A method for synergistic optimization of endpoint temperature and raw material basicity in a sintering process includes the following steps:

[0037] S1: Collect ingredient data and sintering process data, and store them after timestamp alignment and cleaning;

[0038] S2: Calculates the chemical composition of the mixture in real time based on the ingredient data;

[0039] S3: Input the chemical composition of the mixture and sintering process data into the first prediction model, and output the predicted value of the sintering endpoint temperature;

[0040] S4: Based on the structural causal model, the dynamic coupling coefficient matrix K(t) between alkalinity and endpoint temperature is identified in real time, and the decoupled synergistic optimization coefficient is output through online intervention calculation;

[0041] S5: Identify the asymmetric time delay response time of each input variable to each output variable, construct a multivariate time delay matrix, and based on the model predictive control framework, deduce the future state evolution at multiple time scales in virtual reality space, and generate corrected prediction values ​​after time delay decoupling.

[0042] S6: When the deviation between the predicted value and the target value exceeds the threshold, a comprehensive adjustment suggestion is generated based on the collaborative optimization coefficient and the corrected predicted value after time-delay decoupling.

[0043] S7: Record the deviation between the predicted value and the actual value. When the accumulated deviation exceeds the preset threshold, trigger the incremental training of the first prediction model and force the verification of whether the new model conforms to the physical laws through causal consistency verification. After the verification is passed, deploy it online.

[0044] Compared with the prior art, the beneficial effects of the present invention are:

[0045] 1. This invention achieves dynamic causal decoupling between alkalinity and endpoint temperature by constructing a structural causal model (SCM) and online intervention calculation. This solves the strong coupling interference problem in traditional methods, where "adjusting alkalinity will inevitably disturb the endpoint, and adjusting the endpoint will inevitably disturb alkalinity". When the coupling strength exceeds the limit, it automatically switches to decoupling control mode, prioritizes the adjustment of neutral variables, avoids control oscillations, and achieves causal decoupling and collaborative optimization.

[0046] 2. This invention addresses the multivariable asymmetric time delay characteristics of the sintering process (significant differences in response time between different inputs and different outputs). It constructs a multivariable time delay matrix and employs diagonal decoupling to deduce the state evolution across multiple time scales in a virtual reality space. This achieves a leap from "passive alignment" to "active prediction" and solves the control coordination problem caused by the inconsistent response speeds of multiple variables in industrial processes with large time delays.

[0047] 3. This invention introduces causal consistency verification, which forces model updates to not violate the sintering physicochemical mechanism, solving the problems of "black box risk" and "model mutation" in industrial AI deployment. At the same time, it uses virtual sample generation to combat data sparsity, ensuring that the model maintains prediction accuracy and reliability in long-term operation.

[0048] 4. This invention integrates all functional modules into an independent auxiliary decision server, with a built-in database, and is equipped with an anti-interference shielding layer and a heat dissipation module. It operates stably in complex industrial environments, forming a complete and closed-loop auxiliary decision hardware carrier.

[0049] 5. The system of the present invention operates independently and only needs to interface with existing equipment through reserved interfaces. No modification to the PLC control logic is required, ensuring the safety and stability of the existing production process. Attached Figure Description

[0050] Figure 1 This is a flowchart of the method of the present invention; Detailed Implementation

[0051] The present invention will be further described below with reference to the embodiments and accompanying drawings.

[0052] Example 1 System Hardware Configuration

[0053] In this embodiment, the hardware carrier is an independent auxiliary decision-making server. The server's size is adapted for installation in industrial control cabinets, facilitating on-site deployment. To adapt to the harsh environment of industrial sites, the server casing is equipped with a metal anti-interference shielding layer to isolate external electromagnetic interference; internally, a heat dissipation module consisting of cooling fans is integrated to form a heat dissipation loop, ensuring stable operation of the equipment in high-temperature environments. The server is equipped with an industrial-grade processor and at least 8GB of memory to ensure efficient data processing and model calculation capabilities. An Oracle database is embedded within the server as the data storage unit, and Ethernet and RS485 interfaces are reserved for connecting data acquisition modules, human-machine interfaces, and HMI interfaces.

[0054] Example 2: Specific Implementation of the Causal Decoupling and Online Coordination Coefficient Identification Module

[0055] The online identification module for causal decoupling and synergy coefficients is one of the core innovations of this invention, and its specific implementation is as follows:

[0056] (1) Construction of Structural Causal Model (SCM)

[0057] Construct a structural cause-effect diagram of the sintering process to clarify the causal direction between variables.

[0058] Let the set of variables be... ,include:

[0059] Input variables: feed rate of each batching bin (limestone bin, dolomite bin, fuel bin, etc.), machine speed, material layer thickness, ignition temperature, etc.

[0060] Intermediate variables: basicity of the mixture (B), sintering endpoint temperature (T);

[0061] Latent variables: the permeability index of the material layer (H) and the migration velocity of the combustion zone (V) cannot be directly measured. They are identified by integrating data-driven and mechanistic models.

[0062] The structural cause-effect graph is represented using a directed acyclic graph (DAG), where:

[0063] Fuel ratio → Combustion zone migration speed → Sintering endpoint temperature;

[0064] Flux ratio → Basicity of mixture;

[0065] Material layer thickness → Material layer permeability index → ​​Sintering endpoint temperature;

[0066] There is a two-way coupling edge between the basicity of the mixture and the sintering endpoint temperature, indicating that the two influence each other.

[0067] (2) Online intervention calculation (Do-calculus)

[0068] To obtain the expected change in endpoint temperature if alkalinity is adjusted alone, the intervention distribution needs to be calculated. nonconditional distribution Based on the structural causal model, the Do-calculus formula is adopted:

[0069] ;

[0070] Where Z represents the set of parent nodes of B (such as flux ratio) and possible mediating variables. Through this calculation, the pure causal effect coefficient can be obtained:

[0071] ;

[0072] Similarly, calculation .

[0073] (3) Construction of dynamic coupling coefficient matrix

[0074] Construct the dynamic coupling coefficient matrix K(t):

[0075] ;

[0076] in:

[0077] ;

[0078] ;

[0079] ;

[0080] .

[0081] When any non-diagonal element When the system is determined to be in a strongly coupled state, the module automatically switches to decoupling control mode.

[0082] (4) Decoupling control mode

[0083] In decoupled control mode, the system prioritizes adjusting "neutral variables." Neutral variables are those that have the least impact on the coupled system, and are determined by solving the following optimization problem:

[0084] ;

[0085] This involves selecting the variable that has the least impact on both alkalinity and the endpoint (such as bed thickness, machine speed, etc.), and indirectly influencing the target by adjusting the neutral variable, thus avoiding control oscillations caused by directly adjusting strongly coupled variables (such as flux or fuel).

[0086] In actual operation, the online identification module for causal decoupling and synergy coefficients updates the dynamic coupling coefficient matrix every 30 seconds to ensure that the optimization suggestions can adapt to rapid changes in working conditions.

[0087] Example 3: Specific Implementation of a Multivariable Asymmetric Time Delay Decoupling Module

[0088] The sintering process exhibits significant multivariable asymmetric time delay characteristics. The specific implementation of this module is as follows:

[0089] (1) Identification of multivariable time delay matrix

[0090] Based on historical operational data, cross-correlation analysis is used to identify the response delay time of each input variable to each output variable. Let the input variables be... With output variables The cross-correlation function is:

[0091] ;

[0092] Take The time τ at which the input variable reaches its maximum value is taken as the response delay time of the output variable to that input variable. .

[0093] Construct the multivariable time delay matrix τ:

[0094] ;

[0095] Where m is the number of input variables and there are two output variables (alkalinity and endpoint temperature). For example:

[0096] τ (flux adjustment → basicity) ≈ 15 minutes;

[0097] τ (flux adjustment → endpoint) ≈ 25 minutes;

[0098] τ(machine speed adjustment → alkalinity) ≈ ∞ (no significant effect);

[0099] τ (speed adjustment → endpoint) ≈ 10 minutes.

[0100] (2) Diagonalization of the time delay matrix and decoupling of the subsystem

[0101] A time-delay matrix diagonalization algorithm is employed to decouple a strongly coupled multivariable time-delay system into multiple independent single-variable time-delay subsystems. Let the system transfer function matrix be G(s), which contains time-delay terms. Find the decoupling matrix D(s) such that:

[0102] ;

[0103] This means achieving diagonalization between inputs and outputs, so that each output is only affected by its corresponding input.

[0104] (3) Virtual Reality Spatial Deduction

[0105] A digital twin (virtual reality space) of the sintering process is constructed in server memory, and it is projected forward in second-level time steps. The digital twin includes:

[0106] Current state vector X(t);

[0107] Queues of control commands that have been issued but have not yet taken effect;

[0108] The time delay matrix τ of each variable;

[0109] Predictive models.

[0110] The system performs a simulation every 10 seconds, simultaneously displaying the predicted evolution trajectory of the endpoint temperature and alkalinity within the next 30 minutes, forming a state prediction curve on the time axis.

[0111] (4) Predicted value correction and control quantity mapping

[0112] For each decoupled single-variable time-delay subsystem, a model predictive control (MPC) framework is used for prediction and control respectively to obtain the control quantity. Then, through inverse transformation... Mapping back to the original variable space yields the actual executable control instructions.

[0113] For example, virtual reality simulations show that the alkalinity will increase by 0.03 after 15 minutes due to the current flux adjustment, but the endpoint temperature will decrease by 8°C after 25 minutes due to the same adjustment. Based on this, the system generates a pre-adjustment instruction for "compensating the endpoint after 25 minutes" in advance, while ensuring that the alkalinity does not overshoot after 15 minutes.

[0114] Example 4: Specific Implementation of the Trusted Self-Evolution Module

[0115] The trusted self-evolution module ensures that the system maintains prediction accuracy and reliability during long-term operation. The specific implementation is as follows:

[0116] (1) Virtual Sample Generation

[0117] To address the issues of slow changes in sintering process conditions and potential data sparsity in incremental training, this module introduces a virtual sample generation unit. This unit generates virtual operating condition data that conforms to material and heat balance constraints based on a generative adversarial network (GAN).

[0118] The generator G(z) receives random noise z and generates virtual operating condition data; the discriminator D(x) determines whether the input data is real or generated. During training, a physical constraint loss function is introduced:

[0119] ;

[0120] The generated data is ensured to not only closely approximate the distribution of real data but also to satisfy the physicochemical conservation laws of the sintering process. The generated virtual data is used to augment the incremental training dataset, preventing the model from overfitting in sparse regions.

[0121] (2) Causal consistency check

[0122] After incremental training, the causal consistency verification unit forcibly verifies whether the input-output relationship of the new model is consistent with the pre-set sintering physicochemical mechanism knowledge base.

[0123] The knowledge base contains the following domain knowledge rules:

[0124] Rule 1: Increased fuel ratio → Increased sintering endpoint temperature (positive correlation);

[0125] Rule 2: Increased flux ratio → Increased basicity of the mixture (positive correlation);

[0126] Rule 3: Increased machine speed → Decreased sintering endpoint temperature (negative correlation);

[0127] Rule 4: Increased material layer thickness → Increased sintering endpoint temperature (positive correlation, within a certain range).

[0128] Validation method: Perform sensitivity analysis on each input variable in the new model and calculate... If the sign of the partial derivative is inconsistent with the rule, it is judged as a "violation of causal direction". For example, if the new model shows If the endpoint decreases due to increased fuel, an early warning is triggered, preventing the model from updating.

[0129] (3) Knowledge distillation maintains smoothness

[0130] To prevent abrupt changes in control commands due to model updates, knowledge distillation is employed. Let the old model be the teacher model M. teacher The new model is the student model M. student The distillation loss function is:

[0131]

[0132] The KL divergence term forces the output of the new model to be close to that of the old model under steady-state conditions, ensuring smooth output before and after the update and avoiding control abrupt changes.

[0133] (4) Deployment verification

[0134] After incremental training is completed and passes causal consistency verification, the validation unit is deployed to compare the performance of the new and old models on an offline dataset. If the prediction error (RMSE) of the new model is reduced by more than 5% compared to the old model, and all prediction accuracy metrics are no worse than the old model, then the validation is considered successful. After successful validation, the system automatically hot-deploys the new model to the sintering endpoint prediction module, achieving seamless switching.

[0135] Example 5: Example of generating comprehensive adjustment suggestions

[0136] At a certain point in operation, the system detected the following state:

[0137] Predicted sintering endpoint temperature: 365℃ (Target value: 380℃, deviation -15℃)

[0138] Predicted raw material alkalinity: 1.82 (Target value: 1.85, deviation -0.03)

[0139] The dynamic coupling coefficient matrix K(t) is: [1.0, 0.85; 0.65, 1.0] (the off-diagonal elements 0.85 and 0.65 are both relatively high, indicating a strong coupling state).

[0140] Current operating conditions: The material discharge rate in bin No. 3 (limestone bin) is too low, and the machine speed is too high;

[0141] The online identification module for causal decoupling and synergy coefficients determines that the coupling strength exceeds the limit (0.85>0.7), automatically switches to decoupling control mode, and selects the neutral variable "material layer thickness" as the priority adjustment variable.

[0142] The multivariable asymmetric time-delay decoupling module outputs time-delay information:

[0143] Flux adjustment affects basicity by 15 minutes and the endpoint by 25 minutes.

[0144] Speed ​​adjustment affects the endpoint by 10 minutes.

[0145] Adjusting the material layer thickness affects the endpoint by 20 minutes and alkalinity by 30 minutes.

[0146] Based on the above information, the optimization suggestion module generates comprehensive adjustment suggestions:

[0147] Recommended solution for decoupling control mode:

[0148] First priority (adjustment of neutral variables):

[0149] Increasing the material layer thickness by 5mm (the expected end temperature rises by 5-8℃ after 20 minutes, and the alkalinity rises by 0.01-0.02 after 30 minutes, minimizing the coupling effect).

[0150] Second priority (coordinated adjustment):

[0151] Increase the feed rate of No. 3 silo (limestone silo) by 40-60 kg / min (the alkalinity is expected to rise to 1.84-1.86 after 15 minutes).

[0152] To compensate for the drop in endpoint caused by flux adjustment after 25 minutes, the machine speed was simultaneously reduced by 0.05 m / min (the endpoint was expected to rise after 10 minutes, offsetting the flux adjustment effect after 25 minutes).

[0153] The suggestion is pushed to the operator through the HMI interface, and the operator can choose to adopt or adjust it according to the actual situation.

[0154] Example 6: Comparative Experiment with Existing Technologies

[0155] To verify the technical effectiveness of this invention, a 30-day comparative experiment was conducted at a sintering plant. During the experiment, the system provided collaborative optimization suggestions for operators' reference. The experimental data are as follows:

[0156] index The system was not used (same period in history). Using conventional forecasting systems Use this system Alkalinity hit rate (±0.03) 72.3% 78.5% 91.2% Endpoint temperature hit rate (±10℃) 68.1% 75.3% 89.6% Standard deviation of alkalinity fluctuation 0.041 0.035 0.022 Standard deviation of endpoint temperature fluctuation 12.3℃ 10.1℃ 6.5℃ Control the number of fluctuations (per month) - 8 times 0 times Fuel consumption per ton of ore 52.3 kg 51.1 kg 48.5 kg

[0157] Experimental results show that this system significantly improves the stability and economic benefits of sintering process control, especially solving the problem of "adjusting alkalinity leading to endpoint oscillation" in conventional prediction systems (reducing the number of oscillations from 8 to 0).

Claims

1. A system for synergistic optimization of endpoint temperature and raw material basicity in a sintering process, characterized in that, include: The hardware carrier is an independent auxiliary decision server, which is equipped with a metal anti-interference shielding layer and a heat dissipation module, and has a built-in database. The data acquisition module, integrated inside the server, is used to collect ingredient data and sintering process data; The online raw material composition calculation module is integrated into the server and connected to the database. It is used to calculate the chemical composition of the mixture in real time based on the ingredient data. The sintering endpoint prediction module is integrated into the server and has a built-in first prediction model. Its input end is connected to the online raw material composition calculation module and the data acquisition module to output the predicted value of the sintering endpoint temperature. The online identification module for causal decoupling and synergy coefficients is integrated into the server. It has a structural causal model of the sintering process built inside, which is used to identify the dynamic coupling coefficient matrix K(t) between alkalinity and the final temperature in real time, and to output the decoupled synergy optimization coefficients through online intervention calculation. When the coupling coefficient exceeds the preset threshold, the module automatically switches to decoupling control mode, prioritizing the adjustment of neutral variables that have the least impact on the coupled system; The multivariable asymmetric time delay decoupling module, integrated inside the server, is used to identify the asymmetric time delay response time of each input variable to each output variable, construct a multivariable time delay matrix, and, based on the model predictive control framework, extrapolate the future state evolution at multiple time scales in virtual reality space to generate corrected prediction values ​​after time delay decoupling. The optimization suggestion module is integrated inside the server. Its input end is connected to the causal decoupling and synergy coefficient online identification module, the multivariate asymmetric time delay decoupling module, and the target setting module. It is used to generate comprehensive adjustment suggestions containing multivariate synergy adjustment schemes based on the synergy optimization coefficient and the corrected prediction value after time delay decoupling. The trusted self-evolution module, integrated inside the server, includes a causal consistency verification unit. After incremental training of the model, it is used to forcibly verify whether the input-output relationship of the new model is consistent with the pre-set sintering physicochemical mechanism knowledge base. If the causal direction or significance changes in a way that violates physical laws, the model update is prevented and a manual review is triggered.

2. The system according to claim 1, characterized in that, The structural causal model constructed by the online identification module of causal decoupling and synergy coefficient includes latent variable nodes. These latent variable nodes represent physical quantities that cannot be directly measured, including the permeability index of the material layer or the migration speed of the combustion zone. They are identified through a data-driven approach that integrates with the mechanism model, and the causal effects of latent variables on the endpoint temperature and alkalinity are incorporated into the calculation of synergy optimization coefficients.

3. The system according to claim 1, characterized in that, The multivariable asymmetric time-delay decoupling module uses a time-delay matrix diagonalization algorithm to decouple the strongly coupled multivariable time-delay system into multiple independent single-variable time-delay subsystems, and performs prediction and control for each subsystem separately. The control quantity is then mapped back to the original variable space through inverse transformation.

4. The system according to claim 1, characterized in that, The trusted self-evolution module also includes a virtual sample generation unit, which generates virtual operating condition data that conforms to material balance and heat balance constraints based on generative adversarial networks. This data is used to expand the incremental training dataset and prevent the model from overfitting in sparse data regions.

5. The system according to claim 1, characterized in that, In the online identification module for causal decoupling and synergy coefficients, the off-diagonal element k of the dynamic coupling coefficient matrix K(t) ij (t) represents the causal effect strength of the j-th input variable on the i-th output variable. This matrix is ​​updated every 30 seconds. When the absolute value of any off-diagonal element exceeds 0.7, the coupling coefficient is determined to exceed the preset threshold.

6. The system according to claim 1, characterized in that, In the multivariable asymmetric time delay decoupling module, the virtual reality space is a digital twin of the sintering process constructed in the server memory. This digital twin extrapolates forward in second-level time steps, synchronously displaying the predicted evolution trajectory of the endpoint temperature and alkalinity within the next 30 minutes.

7. The system according to claim 1, characterized in that, The comprehensive adjustment suggestions generated by the optimization suggestion module include the coordinated adjustment range of the feed amount of multiple batching bins, the adjustment amount of fuel ratio, the adjustment amount of machine speed, and the expected response time of each adjustment at multiple time scales. The expected response time at multiple time scales is calculated based on an asymmetric time delay matrix.

8. The system according to claim 1, characterized in that, The target setting module is equipped with a human-machine interface for receiving the sintering endpoint target temperature and raw material target alkalinity set by the operator, and supports adaptive adjustment of the deviation threshold according to the working conditions.

9. A method for synergistic optimization of the endpoint temperature and raw material basicity in a sintering process, characterized in that, Includes the following steps: Step S1: Collect batching data and sintering process data through the data acquisition module, and store them in the built-in database after timestamp alignment and cleaning by the data integration module; Step S2: The online raw material composition calculation module calculates the chemical composition of the mixture in real time based on the ingredient data; Step S3: Input the chemical composition of the mixture and sintering process data into the first prediction model, and output the predicted value of the sintering endpoint temperature; Step S4: The online identification module for causal decoupling and synergy coefficients is based on the structural causal model, identifies the dynamic coupling coefficient matrix K(t) between alkalinity and endpoint temperature in real time, and outputs the decoupled synergy optimization coefficients through online intervention calculation. Step S5: The multivariable asymmetric time delay decoupling module identifies the asymmetric time delay response time of each input variable to each output variable, constructs a multivariable time delay matrix, and, based on the model predictive control framework, extrapolates the future state evolution at multiple time scales in virtual reality space, generating corrected prediction values ​​after time delay decoupling. Step S6: When the deviation between the predicted value and the target set value exceeds the threshold, the optimization suggestion module generates a comprehensive adjustment suggestion that includes a coordinated adjustment scheme for multiple variables based on the collaborative optimization coefficient and the corrected predicted value after time-delay decoupling. Step S7: The trusted self-evolution module records the deviation between the predicted value and the actual value. When the accumulated deviation exceeds the preset threshold, it triggers the incremental training of the first prediction model and forces the verification of whether the input-output relationship of the new model is consistent with the preset sintering physicochemical mechanism knowledge base through the causal consistency verification unit. After the verification is passed, it is deployed and updated online.

10. The method according to claim 9, characterized in that, The specific components of outputting the decoupled collaborative optimization coefficients in step S4 include: A directed acyclic graph of the sintering process is constructed using a structural causal model to clarify the causal direction between variables; By calculating "the expected change in endpoint temperature if alkalinity is adjusted alone" through online intervention, a pure causal effect was obtained. Construct a dynamic coupling coefficient matrix K(t) based on pure causal effects; When the off-diagonal elements of K(t) exceed the preset threshold, switch to decoupling control mode and prioritize the adjustment of the neutral variable that has the least impact on the coupled system.

Images