CN-121613859-B - Multi-target cooperative control method and system for powder grinding

CN121613859BCN 121613859 BCN121613859 BCN 121613859BCN-121613859-B

Abstract

The invention discloses a powder grinding multi-target cooperative control method and system, which belong to the field of industrial process automation control, and comprise the steps of extracting power sequence data of a powder selecting machine load and a main motor of a mill, decomposing and separating an internal model component based on a variation mode and combining residence time distribution to generate a working condition quality mapping matrix, executing least square polynomial regression and numerical value branch matching to the working condition quality mapping matrix to construct candidate control parameters, calculating instantaneous frequency variance of the internal model component to construct dynamic weight, selecting a candidate vector with the minimum weight and executing amplitude truncation to generate a control instruction, collecting new data to calculate a prediction residual vector, executing least square iteration update to generate a correction regression equation, eliminating abnormal rows of the working condition quality mapping matrix by utilizing a quantile boundary on the prediction residual sequence, adding new data and executing secondary calibration. The invention adopts data-driven closed-loop control, and can improve the adaptability and control precision of the model to nonlinear working conditions.

Inventors

WANG TAO
ZHANG HUAIZHI
LI FENG
QIAN YEJI
LONG QIANG

Assignees

安徽国风矿业发展有限公司

Dates

Publication Date: 20260512
Application Date: 20251230

Claims (10)

1. The multi-target cooperative control method for powder grinding is characterized by comprising the following steps of: Extracting load time sequence data of a powder concentrator and power time sequence data of a main motor of a mill in a distributed control system, performing signal noise reduction processing based on variation modal decomposition, separating an internal model component, acquiring offline granularity detection data of an automatic laboratory, performing convolution mapping on the internal model component and the offline granularity detection data by using stay time distribution of materials in the grinding system, and generating a working condition quality mapping matrix; Executing least square polynomial regression on the working condition quality mapping matrix, taking the rotation speed of a powder concentrator, the rotation speed of a circulating fan and the total feeding quantity as independent variables, taking the square sum of unit yield power values and granularity difference as dependent variables to obtain a regression equation, sampling in a definition domain of the independent variables to generate a test vector set, substituting the test vector set into the regression equation to calculate a predicted value, executing the numerical value branch matching pair to remove the dominant vector, and extracting non-dominant vector to construct candidate control parameters; Calculating the instantaneous frequency variance of the internal model component, constructing dynamic weights, carrying out weighted summation on the unit yield power value and the granularity difference square sum, selecting a candidate control vector with the smallest weighted summation, calculating a differential increment through the candidate control vector and a current set value, and carrying out amplitude truncation to generate a frequency converter frequency and valve opening instruction to be issued and executed; Executing the frequency converter frequency and valve opening instruction to acquire a new unit yield power value and a new granularity detection value, calculating a prediction residual vector of the new unit yield power value and the new granularity detection value relative to the regression equation, and executing least square iteration update aiming at the regression equation to generate a modified regression equation; And extracting a predicted residual sequence of historical data by using the modified regression equation, calculating an upper quantile boundary of the predicted residual sequence in a sliding window, traversing the working condition quality mapping matrix, removing abnormal data lines with residual modular length exceeding the upper quantile boundary, adding the new unit yield power value and the new granularity detection value to the working condition quality mapping matrix, and executing least square regression to perform secondary calibration and archiving on the modified regression equation.
2. The method for multi-target cooperative control of powder grinding according to claim 1, wherein the generating the working condition quality mapping matrix comprises: performing variation modal decomposition on the powder concentrator load time sequence data and the mill main motor power time sequence data, outputting a modal component set, calculating cross-correlation coefficients of components in the modal component set and original time sequence data, and locking a component with a first cross-correlation coefficient as an internal model component; Extracting sampling time from the offline granularity detection data, calculating a cross-correlation function of the internal model component and the offline granularity detection data, taking the maximum time shift amount to generate average lag time, constructing a residence time distribution function by taking the average lag time as a hope, and performing discretization processing to obtain a time lag weight sequence; and taking the average lag time subtracted from the sampling time as an index, intercepting historical time window data with the same length as the time lag weight sequence in the internal model component, performing weighted summation operation on the historical time window data and the time lag weight sequence to obtain a weighted working condition characteristic value, and splicing the weighted working condition characteristic value with the offline granularity detection data to generate a working condition quality mapping matrix.
3. The powder grinding multi-target cooperative control method according to claim 1, wherein the extracting non-dominant vectors to construct candidate control parameters comprises: extracting data columns according to the rotating speed of the powder concentrator, the rotating speed of the circulating fan and the total feeding amount in the working condition quality mapping matrix, constructing a polynomial expansion design matrix, combining the unit yield power value and the particle size difference square sum to execute matrix operation, and calculating a regression coefficient vector to establish a regression equation; extracting a definition domain boundary value of the independent variable, executing equal step-length grid scanning in the definition domain boundary value to generate a test vector set, and substituting the test vector set into the regression equation to generate a predicted value set; and traversing the predicted value set to perform pairwise numerical comparison, and if the sum of the square sum of the unit yield power value and the granularity difference of the first vector is smaller than the second vector, marking the second vector as a dominant vector, removing the dominant vector, and reserving the unlabeled vector as a candidate control parameter.
4. A powder grinding multi-target cooperative control method according to claim 3, wherein the calculating the regression coefficient vector includes: calculating a self-square vector and a mutual product vector of the data column, and performing column vector splicing on the data column, the self-square vector and the mutual product vector to generate a polynomial expansion design matrix; and carrying out operation on the polynomial expansion design matrix to obtain an inverse matrix, and carrying out chain matrix multiplication on the inverse matrix, the transposed matrix of the polynomial expansion design matrix, the unit yield power value and the granularity difference square sum to generate a regression coefficient vector.
5. The method for multi-target cooperative control of powder grinding according to claim 1, wherein the generating the frequency converter frequency and the valve opening command issuing comprises: Counting the zero crossing number of the internal model component in sampling time, dividing the sampling time by the zero crossing number to obtain an average period, calculating the variance of each point amplitude of the internal model component relative to the average period as an instantaneous frequency variance, and mapping the instantaneous frequency variance into a normalized value by using a hyperbolic tangent function to serve as a dynamic weight for the square sum of the granularity difference; Calculating a unit yield power value mean value and a granularity difference square sum mean value by using the candidate control parameters, dividing the unit yield power value mean value and the granularity difference square sum mean value by the corresponding unit yield power value mean value and granularity difference square sum mean value respectively to generate a dimensionless ratio, performing weighted summation on the dimensionless ratio by using the dynamic weight, and taking a vector with the minimum summation result as a preferred vector; And calculating the data standard deviation of the rotating speed of the powder selecting machine by using the working condition quality mapping matrix as a safe fluctuation threshold, calculating the difference value between the optimal vector and the current set value, and generating a frequency converter frequency and valve opening instruction if the absolute value of the difference value is larger than the safe fluctuation threshold.
6. The powder grinding multi-target cooperative control method according to claim 4, further comprising: performing matrix multiplication operation by using the polynomial expansion design matrix and the regression coefficient vector to generate a theoretical reproduction value vector; And calculating a difference vector between the theoretical recovery value vector and the unit yield power value, and defining the difference vector as a static fitting deviation vector.
7. The multi-target cooperative control method for powder grinding according to claim 1, wherein generating the modified regression equation comprises: Substituting the frequency converter frequency and valve opening instructions into the regression equation to execute forward calculation to obtain a theoretical prediction value, and calculating the difference between the new unit yield power value and the new granularity detection value relative to the theoretical prediction value to obtain a prediction residual vector; Constructing a vector autocorrelation matrix by utilizing the frequency converter frequency and valve opening instruction, calculating an inverse matrix of the vector autocorrelation matrix to generate an orthogonal projection gain matrix, and performing chain multiplication on the frequency converter frequency and valve opening instruction, the prediction residual vector and the orthogonal projection gain matrix to generate a coefficient correction matrix; And extracting a current coefficient matrix aiming at the regression equation, performing matrix addition operation on the current coefficient matrix and the coefficient correction matrix, and reconstructing the regression equation to generate a corrected regression equation.
8. The method of claim 6, wherein the performing the secondary calibration and archiving of the modified regression equation comprises: Calculating coefficient difference values of the correction regression equation and the regression equation, multiplying the coefficient difference values by the polynomial expansion design matrix to obtain model drift vectors, and calculating difference values of the static fitting deviation vectors and the model drift vectors to generate a prediction residual sequence; Extracting the absolute upper quantile of the predicted residual sequence as a cleaning boundary, removing row vectors with residual modules larger than the cleaning boundary from the working condition quality mapping matrix, and defining the rest matrix data as a trusted sample matrix; And adding the new unit yield power value and the new granularity detection value to the trusted sample matrix, solving an inverse matrix for the trusted sample matrix, generating a regression coefficient after secondary calibration, and executing parameter coverage and archiving on the corrected regression equation.
9. The powder grinding multi-target cooperative control method of claim 8, wherein the archiving comprises: counting the number of lines of the trusted sample matrix to be used as the current sample size, and extracting the window length of the sliding window to be used as the upper limit of the capacity; and calculating the value obtained by subtracting the capacity upper limit from the current sample size as an overflow number, and sequentially removing row vectors equal to the overflow number from the initial row index of the trusted sample matrix if the overflow number is greater than zero.
10. A powder grinding multi-target cooperative control system applied to a powder grinding multi-target cooperative control method as defined in any one of claims 1 to 9, characterized in that the system comprises: The working condition quality mapping construction module is used for extracting load time sequence data of the powder concentrator and power time sequence data of a main motor of the mill in the distributed control system, performing signal noise reduction processing based on variation modal decomposition, separating out an internal model component, acquiring offline granularity detection data of an automatic laboratory, performing convolution mapping on the internal model component and the offline granularity detection data by using stay time distribution of materials in the grinding system, and generating a working condition quality mapping matrix; the regression prediction and non-dominated screening module is used for executing least square polynomial regression on the working condition quality mapping matrix, taking the rotational speed of a powder concentrator, the rotational speed of a circulating fan and the total feeding amount as independent variables, taking the square sum of unit yield power values and granularity difference as dependent variables to obtain a regression equation, sampling in a definition domain of the independent variables to generate a test vector set, substituting the test vector set into the regression equation to calculate a predicted value, executing numerical branch matching pair to remove dominated vectors, and extracting non-dominated vectors to construct candidate control parameters; the candidate vector sequencing and instruction generation module is used for calculating the instantaneous frequency variance of the internal model component, constructing dynamic weights, carrying out weighted summation on the unit yield power value and the granularity difference square sum, selecting a candidate control vector with the smallest weighted summation, calculating a differential increment through the candidate control vector and a current set value, and carrying out amplitude truncation to generate a frequency converter frequency and valve opening instruction for issuing and executing; The prediction residual calculation and regression updating module is used for executing the frequency converter frequency and valve opening instruction to acquire a new unit yield power value and a new granularity detection value, calculating a prediction residual vector of the new unit yield power value and the new granularity detection value relative to the regression equation, and executing least square iteration updating aiming at the regression equation to generate a corrected regression equation; And the residual threshold calculation and abnormal row rejection module is used for extracting a predicted residual sequence of historical data by utilizing the modified regression equation, calculating an upper quantile boundary of the predicted residual sequence in a sliding window, traversing the working condition quality mapping matrix, rejecting abnormal data rows with residual modular length exceeding the upper quantile boundary, adding the new unit yield power value and the new granularity detection value to the working condition quality mapping matrix, and executing least square regression to perform secondary calibration and archiving on the modified regression equation.

Description

Multi-target cooperative control method and system for powder grinding Technical Field The invention relates to the field of industrial process automation control, in particular to a powder grinding multi-target cooperative control method and system. Background The powder grinding process generally comprises the links of grinding, grading, material returning circulation and the like, and a production site is generally provided with a main motor of a mill, a powder selecting machine, a circulating fan, a feeding device, a regulating valve and other executive components. Under the distributed control architecture, data acquisition is carried out on the operation process, and the acquisition objects can comprise process variables such as powder concentrator load time sequence data, mill main motor power time sequence data, powder concentrator rotating speed, circulating fan rotating speed, total feeding quantity and the like. The product quality index can be represented by offline granularity detection data, and the energy consumption index can be represented by a unit yield power value. In the related art, china patent with the publication number CN121050232A discloses an operation regulation method of a pulverizer matched with a coal-fired boiler during peak shaving of thermal power, which specifically comprises the steps of determining a first typical particle size distribution interval when the coal-fired boiler is in a normal state and a second typical particle size distribution interval when the coal-fired boiler is in a peak shaving state, establishing a mathematical intelligent model between the coal-fired particle size and the performance of a thermal power generation system, configuring the number of the pulverizer and working parameters, and determining and starting the pulverizer with the corresponding number according to calculated coal-fired particle size requirements and combined with optimization results of real-time monitoring data and the mathematical intelligent model and dynamically regulating the working parameters of the pulverizer when the thermal power generation system is in the normal state or the peak shaving state. However, when the offline quality index and the online process variable are used for the collaborative regulation and control of the powder grinding process, there are constraints such as inconsistent time reference of the offline detection and the online process variable, noise interference of the process variable, difficulty in unified expression of multi-objective trade-off and the like. The off-line granularity detection obtains results through links such as sampling, conveying, sample preparation, detection and the like, has time delay between a detection value and sampling time, and meanwhile, the residence time of materials in a grinding loop is distributed, the time of different batches of samples reaching a grading loop has discreteness, and the direct corresponding relation between the off-line granularity detection value and an on-line process variable is unstable. The synchronous signals such as the powder concentrator load and the mill main motor power comprise trend items and fluctuation items, and the original signals are directly used for modeling, and regression parameters are sensitive to noise along with measurement noise or transient disturbance. The unit yield power value and the granularity deviation belong to different dimension targets, the weighting and the emphasis are weighed to change under different working condition fluctuation intensities, and the fixed weight or the fixed set value is difficult to cover the working condition change. In addition, raw material characteristic changes, equipment wear, sensor drift or intermittent faults cause mapping relation drift between process variables and quality and energy consumption indexes, and regression deviation is accumulated in a plurality of control periods when model updating based on latest data and abnormal sample cleaning are lacked, so that control parameter screening and instruction generation processes are affected. Disclosure of Invention In order to solve the problems, the invention provides a powder grinding multi-target cooperative control method and system, which adopt data-driven closed-loop control to improve the adaptability and control precision of a model to nonlinear working conditions. The above object can be achieved by the following scheme: The multi-target cooperative control method for powder grinding comprises the steps of extracting load time sequence data of a powder selecting machine and power time sequence data of a main motor of a grinding machine in a distributed control system, performing signal noise reduction processing based on variation modal decomposition, separating out an internal model component, acquiring offline granularity detection data of an automatic laboratory, performing convolution mapping on the internal model component and the offline granularity