A real-time control method of combined grinding system fusing process mechanism and data model

Abstract

A kind of real-time control method of combined grinding system fusing process mechanism and data model, obtains the set value of controlled variable in current control loop of combined grinding system, determines the decision variable of real-time optimization control problem, constructs the objective function and constraint condition of combined grinding system, downloads the data to be analyzed, does smoothing processing, eliminates abnormal value, distinguishes stationary production data set and non-stationary production data, expands process mechanism, carries out data modeling by regression type machine learning algorithm or neural network model training and evaluation, expands abnormal condition research, generates abnormal condition response knowledge base, sends abnormal condition identification result to lower control system in real time, synchronizes response measure to control system, realizes closed-loop processing of abnormal condition, optimizes the objective function and constraint condition, and connects the optimization result with control system.

Description

Technical Field

[0001] This invention belongs to the field of cement production combined grinding technology, specifically involving a process mechanism and data model technology. Background Technology

[0002] Cement production mainly consists of processes such as raw material preparation, clinker calcination, and cement grinding. Cement grinding is a process in which mechanical external force is used to overcome the cohesive forces between solid molecules in the raw materials, causing them to break and form a certain particle size distribution.

[0003] While dual-closed-circuit combined cement grinding systems improve grinding efficiency and reduce unit power consumption through methods such as "replacing grinding with crushing" and "more crushing and less grinding," they still suffer from problems such as nonlinearity, large hysteresis, large inertia, and multiple disturbances. For example, when the grinding mill speed is too high or the mill tail blower speed is too low, the cement particles become too fine, leading to excessively rapid early hydration or poor crack resistance; cement output is affected by the grindability of the material, resulting in significant fluctuations; and unscientific batching leads to high grinding production costs. These problems have long plagued the stable production and control of combined cement grinding systems.

[0004] Combined grinding control involves controlling aspects such as the chemical composition of cement products, the fineness and specific surface area of ​​cement, the steady-flow bin weight, mill current, the current pumping up the mill bucket, and the differential pressure between the mill and the tube mill. In the specific control process, the effectiveness of each control loop is affected by material properties, differences in operator habits and adjustment methods, and fluctuations in operating conditions. Although cement plants both domestically and internationally have implemented automatic control methods, such as fuzzy control, PID control, and even advanced process control (APC), problems still exist, including unreasonable parameter tuning, poor performance, weak adaptability and handling capabilities, and excessive reliance on manual experience.

[0005] To address the above issues, this study adopts a real-time optimization control approach, comprehensively considering various influencing factors such as quality, cyclic load, unit power consumption, safety, and process parameters. An optimization problem combining mechanism and data is established. By solving the objective function with constraints, the target setpoints of the controlled variables in the control loop are determined. Special handling is implemented for abnormal operating conditions to achieve stability, economy, and safety of the combined grinding system. Summary of the Invention

[0006] To address the technical problem of excessive reliance on manual experience in setting controlled variables, which hinders the comprehensive consideration of quality, output, unit power consumption, safety, and cost, a real-time online optimization approach was adopted. This approach establishes a mapping relationship between the objective function and the setpoints of the controlled variables. By solving the objective function, the setpoints of the controlled variables are obtained. This technical solution fully leverages on-site production operation data, mechanistic models, and expert experience, resulting in key indicators that balance safety, efficiency, quality, low power consumption, and high output in production operations. It avoids excessive reliance on manual experience for the setpoints of controlled variables, enhances adaptability to operating conditions, and ensures that the optimized setpoints are reasonable and reliable. This improves the online commissioning rate of the APC control system, reduces the workload of on-site engineers, enhances system stability, reduces system failure rate, reduces system volatility, increases mill output, improves the stability of cement screening residue, improves the stability of mill main current and outlet hopper lifting current, and reduces unit power consumption.

[0007] The above method includes offline analysis and online execution. Offline analysis includes constructing a working condition identification model and a process model, smoothing data processing, removing outliers, and extracting steady-state data. Online execution includes online working condition identification, online selection and updating of the process model, optimization of the objective function, and output of calculated values. The specific implementation process is as follows: Figure 1 As shown. The detailed steps are as follows: Step 1: Obtain the setpoint of the controlled variable in the current control loop of the combined grinding system to clarify the decision variables for real-time optimization control.

[0008] Step 2: Construct the objective function and constraints of the combined grinding system. The objective function includes circulating load, material fineness sieve residue, specific surface area, unit consumption, and batching cost. The constraints include quality constraints, circulating load constraints, decision variable constraints, proportioning constraints, process constraints, and grinding efficiency constraints.

[0009] Step 3: Download the data to be analyzed, perform smoothing, and remove outliers.

[0010] Step 4: Differentiate between stationary production data sets and non-stationary production data sets.

[0011] Step 5: Expand the process mechanism, obtain the proportioning constraints and grinding efficiency of the material cost and constraints of the objective function constructed in Step 2, and perform data modeling by training and evaluating regression-type machine learning algorithms or neural network models to obtain material fineness and specific surface area, unit power consumption, cycle load, and unit consumption.

[0012] Step Six: Conduct an abnormal operating condition study, identify the non-stationary production data identified in Step Four that correspond to the following during the operation process: weighing bin overflow, roller press silo arching, roller press poor work performance, grinding mill over-grinding, grinding mill empty grinding, belt scale shutdown, and cyclone blockage. Combine the influencing factors of each abnormal operating condition to generate an abnormal operating condition response knowledge base.

[0013] Step 7: Send the abnormal operating condition identification results obtained in Step 6 to the lower-level control system in real time, and synchronize the countermeasures to the control system to achieve closed-loop processing of abnormal operating conditions.

[0014] Step 8: Optimize the objective function and constraints obtained in Step 5.

[0015] Step 9: Connect the optimization results obtained in Step 8 with the control system. Attached Figure Description

[0016] Figure 1 It's a flowchart. Figure 2 This is the stability analysis data of a certain No. 1 cement mill. Figure 3 This is the predicted trend of residue on a 45μm sieve from a certain No. 1 cement mill. Figure 4 It is the relationship between the objective function and the classifier current and the outlet mill bucket lifting current. Figure 5 This is a comparison of the sieve residue of a certain No. 1 cement mill before and after the 45μm sieve was put into use. Figure 6 This is a comparison of the main motor current and the outlet elevator current of a certain No. 1 cement mill before and after it was put into operation. Detailed Implementation

[0017] The following, in conjunction with the accompanying drawings, describes the implementation process based on the nine steps outlined in the aforementioned invention, as follows: Figure 1 As shown, the technical solution of the present invention will be described in detail below.

[0018] Step 1: Let the decision variables X include the following: mill outlet current X1, mill main unit current X2, steady flow bin weight X3, classifier current X4, mill head negative pressure X5, roller press current X6, and batching ratio X7 to X8. 12 These correspond to the proportions of clinker, gypsum, limestone, fly ash, slag, and externally added mineral powder, respectively.

[0019] Step 2: Let f1(X) to f5(X) correspond to the objective functions of circulating load, material fineness sieve residue, specific surface area, unit consumption, and batching cost, respectively. ω1 to ω5 correspond to the optimization weights of circulating load, material fineness sieve residue, specific surface area, unit consumption, and batching cost, respectively, set by the user according to control requirements. F and S represent the standard values ​​of material fineness sieve residue and specific surface area, respectively, set by the factory according to the production requirements of the cement type. f2(X)-F represents the degree of closeness between the predicted material fineness sieve residue and the factory's standard material fineness sieve residue; f3(X)-S represents the degree of closeness between the predicted specific surface area and the factory's standard specific surface area; p i Let φ represent the cost of the i-th ingredient and φ represent the grinding efficiency. The following formula relationship is obtained.

[0020]

[0021] Step 3: Download the data to be analyzed, perform smoothing, and remove outliers.

[0022] Step 4: Differentiate between stationary and non-stationary production data, such as... Figure 2 As shown.

[0023] Step 5: Considering the variable operating conditions of the combined grinding system, online model pooling is used for data modeling. A data segment with consistent distribution characteristics is selected to create a model. During online application, the corresponding model is selected for analysis and prediction based on the operating condition judgment results, and the model pool is optimized and updated. The input to the data model is the decision variable, i.e., the expected value of the controlled variable to be solved, and the output is the 45µm sieve residue. The trend of the predicted result of the 45µm sieve residue obtained through data modeling is as follows: Figure 3 As shown in the smooth curve, the actual values ​​of 45µm sieve residue within the same time period are as follows: Figure 3 As shown by the stepped broken line, the predicted value and the actual value have basically the same trend. The coefficient of determination is about 0.7 and the correlation coefficient is about 0.9. This indicates that the predicted value of 45um fineness sieve residue based on the data model can well reflect the trend change through the decision variable. By optimizing the closeness between the predicted value and the factory production standard value, the desired set value of the controlled variable can be obtained.

[0024] Step Six: Given the low frequency of abnormal operating conditions, data modeling is performed using small-sample learning. The occurrence of abnormal operating conditions is studied from data trends. Let T represent the abnormal operating condition to be identified, x represent known measurement data, and D... T D represents the available supervised information dataset. A Represent an auxiliary dataset independent of T, and construct the model p(y / x) = h(x, D). A D T), where p(y / x) represents the abnormal operating condition category y identified given the known measurement data x. Combining the influencing factors of each abnormal operating condition, an abnormal operating condition response knowledge base is generated.

[0025] Step 7: Send the abnormal operating condition identification results obtained in Step 6 to the lower-level control system in real time, and synchronize the countermeasures to the control system to achieve closed-loop processing of abnormal operating conditions.

[0026] Step 8: Calculate the decision variable values ​​corresponding to the minimum objective function using an optimization algorithm. The optimization methods used include, but are not limited to, stochastic gradient descent, Newton's method, conjugate gradient method, Lagrange multiplier method, and heuristic optimization algorithms. The concavity / convexity of the objective function is analyzed using a graphical approach, such as... Figure 4 The figure shows the relationship between the objective function and the classifier current and the outlet mill bucket lifting current. It is observed that the objective function that satisfies the constraints has multiple local optima, and the optimization problem is a multi-objective nonlinear optimization problem. In this embodiment, a heuristic optimization algorithm is selected to calculate and solve the optimization problem.

[0027] Step 9: The optimization results obtained in Step 8, i.e. the value of decision variable X, are sent to the APC system of the combined grinding system or other control systems via OPC communication protocol, Kafka message or HTTP interface, to be used as the set value of the controlled variable, so as to realize the automatic optimization of the set value of the controlled variable.

[0028] Following the above process, this invention achieves real-time optimization calculation of controlled variables in APC control loops or other control methods, avoiding excessive reliance on manual experience for controlled variable setpoints and improving system operational stability by 5%. When optimizing the setpoints of controlled variables, it comprehensively optimizes from the perspectives of cyclic load, quality, unit consumption, and batching cost, fully considering quality constraints, cyclic load constraints, decision variable constraints, proportioning constraints, process constraints, and grinding efficiency constraints. It also takes into account key indicators such as safety, efficiency, quality, low consumption, and high output in production operation, reducing unit power consumption by 2% and increasing mill output by 3%. During optimized control, when the system is in a stable state, it sends the real-time optimization results of the controlled variable setpoints. As a result, when the system is in a non-stationary state, it sends abnormal operating condition categories and corresponding handling solutions, increasing the online operation rate of the APC control system to over 98%, reducing the workload of field engineers, and significantly reducing the system failure rate. When constructing process functions, the process functions are updated through an online model pool, enhancing the adaptability to operating conditions and ensuring the stability of system operation. Statistical analysis of the system's operational volatility 10 days before and after commissioning shows that system volatility has decreased by 5%. When constructing process functions and objective functions, a dual-driven model of data and mechanism is adopted for modeling and analysis, fully leveraging the value of factory data while adhering to process mechanisms to ensure that the optimization results of the controlled variable setpoints are reasonable and reliable.

[0029] After implementing the aforementioned specific implementation methods, such as Figure 5 As shown, the stability of the cement mill residue on the 45µm sieve was significantly improved compared to before implementation; for example... Figure 6 As shown, the stability of the mill main current and the outlet current of the mill bucket are also significantly improved.

[0030] The above are embodiments of the present invention and do not limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention are included within the protection scope of the present invention.

Claims

1. A real-time control method for a combined grinding system integrating process mechanism and data model, characterized in that, include: Step 1: Obtain the setpoints of the controlled variables in the current control loop of the combined grinding system to identify the decision variables for the real-time optimization control problem; Step 2: Construct the objective function and constraints of the combined grinding system. The objective function includes circulating load, material fineness sieve residue, specific surface area, unit consumption, and batching cost. The constraints include quality constraints, circulating load constraints, decision variable constraints, proportioning constraints, process constraints, and grinding efficiency constraints. Step 3: Download the data to be analyzed, perform smoothing, and remove outliers; Step 4: Differentiate between stationary and non-stationary production data; Step 5: Expand the process mechanism, obtain the material cost and constraint conditions of the objective function constructed in Step 2, and the grinding efficiency. Use regression-based machine learning algorithms or neural network models for training and evaluation to perform data modeling, and obtain material fineness, specific surface area, unit power consumption, cycle load, and unit consumption. Considering the variable production conditions of the combined grinding system, use an online model pool for data modeling. Select a data segment with consistent data distribution characteristics to create a model. When applying it online, select the corresponding model for analysis and prediction based on the working condition judgment results, and optimize and update the model pool. Let the input of the data model be the decision variable, that is, the expected value of the controlled variable to be solved. By optimizing the closeness of the predicted value to the factory production standard value, the expected set value of the controlled variable is obtained. Step Six: Conduct an abnormal operating condition study, identifying the non-stationary production data identified in Step Four, corresponding to the following operational processes: weighing hopper overflow, roller press hopper arching, roller press poor work performance, mill over-grinding, mill empty grinding, belt scale shutdown, and cyclone blockage. Combine the influencing factors of each abnormal operating condition to generate an abnormal operating condition response knowledge base. Considering the low frequency of abnormal operating conditions, use small sample learning for data modeling, studying the occurrence of abnormal conditions from the perspective of data change trends. Let T represent the abnormal operating condition to be identified, x represent known measurement data, and D... T D represents the available supervised information dataset. A Represent an auxiliary dataset independent of T, and construct the model p(y / x) = h(x, D). A D T ), where p(y / x) represents the abnormal operating condition category y identified given the known measurement data x. Combining the influencing factors of each abnormal operating condition, an abnormal operating condition response knowledge base is generated. Step 7: Send the abnormal operating condition identification results obtained in Step 6 to the lower-level control system in real time, and synchronize the countermeasures to the control system to achieve closed-loop processing of abnormal operating conditions. Step 8: Optimize the solution to the objective function and constraints obtained in Step 5; Step 9: Send the optimization results obtained in Step 8, i.e. the values ​​of the decision variables, to the APC system of the combined grinding system or other control systems using the OPC communication protocol, Kafka message, or HTTP interface, so as to use them as the setpoints of the controlled variables and realize the automatic optimization of the setpoints of the controlled variables.

2. The real-time control method for a combined grinding system integrating process mechanism and data model according to claim 1, characterized in that, Step one further includes: setting the decision variable X to include the following: mill outlet current X1, mill main unit current X2, steady flow bin weight X3, classifier current X4, mill head negative pressure X5, roller press current X6, and batching ratio X7 to X. 12 These correspond to the proportions of clinker, gypsum, limestone, fly ash, slag, and externally added mineral powder, respectively.

3. The real-time control method for a combined grinding system integrating process mechanism and data model according to claim 2, characterized in that, Step two further includes: Let f1(X) to f5(X) correspond to the objective functions of circulating load, material fineness sieve residue, specific surface area, unit consumption, and batching cost, respectively; ω1 to ω5 correspond to the optimization weights of circulating load, material fineness sieve residue, specific surface area, unit consumption, and batching cost, respectively, set by the user according to control requirements; F and S represent the standard values ​​of material fineness sieve residue and specific surface area, respectively, set by the factory according to the production requirements of cement type; f2(X)-F represents the degree of closeness between the predicted material fineness sieve residue and the factory's standard material fineness sieve residue; f3(X)-S represents the degree of closeness between the predicted specific surface area and the factory's standard specific surface area; p i Let φ represent the cost of the i-th ingredient, and φ represent the grinding efficiency, thus yielding the formula. The relationship indicated.

4. The real-time control method for a combined grinding system integrating process mechanism and data model according to claim 1, characterized in that, Step eight further includes: calculating the decision variable values ​​corresponding to the minimum objective function through optimization algorithms, including stochastic gradient descent, Newton's method, conjugate gradient method, Lagrange multiplier method, and heuristic optimization algorithms, and analyzing the concavity and convexity of the objective function based on a graphical approach.

Images