Sintering endpoint closed-loop control system and method based on multi-field fusion network

By employing a multi-field fusion network-based closed-loop control method for the sintering endpoint, and utilizing matrix temperature measurement and two-dimensional deviation decoupling, combined with feedforward prediction and feedback closed loop, the problems of detection accuracy and control lag in the sintering endpoint control are solved, enabling timely and precise control of the endpoint and improving the stability and efficiency of the sintering process.

CN123065255APending Publication Date: 2026-11-10NANJING HUIXIANG AUTOMATION SYST ENG CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611592549.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-15
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing sintering endpoint control technologies suffer from insufficient detection accuracy and control lag, making it difficult to achieve timely and precise control. This is mainly because single-point or small-scale temperature detection is susceptible to interference from gas leakage and air crosstalk, and the operation relies on manual labor, resulting in delayed and inconsistent endpoint judgment.

Method used

A closed-loop control method for the sintering endpoint using a multi-field fusion network is adopted. By comprehensively characterizing the multi-physics field through matrix temperature measurement, and combining two-dimensional deviation decoupling and feedforward prediction and feedback closed loop, control commands are generated to achieve precise control of the sintering endpoint.

Benefits of technology

It enables timely and precise control of the sintering endpoint, reduces control lag, improves the accuracy and stability of endpoint judgment, and ensures the efficient operation of the sintering process.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123065255A_ABST
    Figure CN123065255A_ABST
Patent Text Reader

Abstract

The application relates to the technical field of automatic control, and discloses a sintering endpoint closed-loop control system and method based on a multi-field fusion network, which comprises the following steps: sequentially numbering sintering air boxes, screening endpoint observation sections, and arranging a temperature measurement matrix; according to the temperature measurement matrix, the temperature of each temperature measurement point is collected in real time to form temperature measurement time sequence data, and associated process parameters are synchronously collected; through pretreatment, a time sequence database is constructed; the temperature of a transverse temperature measurement point of each air box is fitted, a coarse sintering endpoint is located through curve inflection point identification and weighted fusion; a space-time temperature field is constructed for the endpoint observation section, a predicted sintering endpoint and endpoint temperature are generated through endpoint prediction; a two-dimensional deviation parameter is generated through deviation comparison of a target air box number and the predicted sintering endpoint; and through feedforward prediction and feedback closed loop, a regulation and control instruction is generated and issued.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of automatic control technology, and more specifically to a closed-loop control system and method for the sintering endpoint based on a multi-field fusion network. Background Technology

[0002] Sintering is a crucial step in the steel production process, and the quality and yield of its product, sintered ore, directly affect the stable operation and overall energy consumption of subsequent blast furnace ironmaking. The sintering endpoint, as a comprehensive parameter reflecting the completeness of the sintered material, has a significant impact on the yield, quality, and energy consumption of sintered ore due to its accurate determination and stable control. With the increasing demand for refined control and intelligent upgrading in the steel industry, sintering endpoint control technology is evolving from traditional manual experience-based judgment towards automation and intelligence. At the detection level, soft measurement methods based on the temperature of the exhaust gas from the bellows are widely used. At the control level, researchers have successively proposed fuzzy control strategies based on the temperature rise point of the exhaust gas and endpoint prediction methods combined with neural network models, aiming to improve the timeliness and accuracy of sintering endpoint control.

[0003] However, existing technologies still have shortcomings in practical applications. Specifically, in the detection stage, existing methods mostly rely on the exhaust gas temperature of a single or a few air boxes to estimate the endpoint. However, during the sintering process, phenomena such as air leakage on the material surface of the trolley and air leakage at the tail of the machine often occur, which can easily cause the measured exhaust gas temperature value to deviate from the actual working conditions. Moreover, a single point or a few temperature points cannot fully characterize the coupling relationship of multiple physical fields such as temperature field, airflow field and material field, which limits the accuracy of endpoint judgment. In the control stage, most sintering production sites are still mainly operated manually. Operators make adjustments based on the detection data. The judgment process has a certain lag and the operating standards are difficult to be completely unified. Even if some systems introduce feedback adjustment mechanisms, there is still a time delay between the occurrence of deviation and the execution of adjustment, making it difficult to achieve timely and accurate control of the sintering endpoint. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the present invention aims to provide a closed-loop control system and method for the sintering endpoint based on a multi-field fusion network. This system comprehensively characterizes multiple physical fields and suppresses interference from gas leakage through matrix temperature measurement. It improves the accuracy of endpoint judgment by combining coarse positioning with endpoint prediction. Furthermore, it shortens the control lag by combining two-dimensional deviation decoupling, feedforward prediction, and feedback closed loop, thereby achieving the regulation of the sintering endpoint.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A closed-loop control method for the sintering endpoint based on a multi-field fusion network includes: The sintering air boxes were numbered sequentially, and the final observation section was selected and the temperature measurement matrix was arranged. Based on the temperature measurement matrix, the temperature of each temperature measurement point is collected in real time to form temperature measurement time series data. Related process parameters are collected synchronously, and a time series database is constructed through preprocessing. Temperature fitting is performed on the transverse temperature measurement points of each wind box, and the endpoint of coarse sintering is located by curve inflection point identification and weighted fusion. A spatiotemporal temperature field is constructed for the endpoint observation segment, and the predicted sintering endpoint and endpoint temperature are generated through endpoint prediction. By comparing the deviation between the target windbox number and the predicted sintering endpoint, a two-dimensional deviation parameter is generated. Using feedforward prediction and feedback closed loop, a control command is generated and issued.

[0006] Preferably, the screening endpoint observation segment includes: All air boxes are sequentially numbered along the direction of travel of the sintering machine trolley; Retrieve historical temperature time series data of all wind boxes, and divide the configuration period and positioning period into the historical temperature time series data according to the sintering cycle, and generate configuration data source and positioning data source. The configured data source is called to calculate the average temperature of each wind box, thus obtaining the average temperature sequence of the entire wind box. The uniform temperature sequence of the entire wind box is subjected to first-order forward difference processing to calculate the longitudinal temperature gradient between adjacent wind boxes, thereby obtaining the gradient sequence of the configuration period, and the standard deviation of the gradient sequence is calculated. Configure the confidence coefficient and calculate the gradient threshold based on the standard deviation; A positioning gradient sequence is constructed for the positioning data source, and rising and falling edges are filtered in combination with a gradient threshold to obtain the gradient positioning interval.

[0007] Preferably, the configured confidence coefficient includes: For the gradient sequence of the configuration period, it is divided into two equal segments according to time, denoted as follows: , And calculate the standard deviation respectively. , ; Construct a set of candidate confidence coefficients. For each candidate confidence coefficient, calculate the candidate threshold by multiplying the standard deviation by the candidate confidence coefficient. , ; use position The rising and falling edges, using position The rising and falling edges; based on , The absolute difference between the rising edge and the falling edge is calculated separately, and the positioning difference of each candidate confidence coefficient is calculated by summing the absolute differences. The candidate confidence coefficient with the smallest positioning difference is selected as the confidence coefficient.

[0008] Preferably, the screening endpoint observation segment further includes: Iterate through the location data sources and identify the bellows number of the temperature peak at each sampling time; The number of times each windbox is recorded as a temperature peak during the statistical positioning cycle is counted, and the peak frequency of each windbox is calculated to obtain the peak frequency sequence. By accumulating the peak frequencies, the cumulative frequency from the first bellows to the i-th bellows is calculated, and the cumulative frequency sequence is obtained. Based on the peak frequency sequence, the position standard deviation is calculated with the peak frequency as the weight of the bellows number, and the confidence level is adaptively configured based on the position standard deviation. In the cumulative frequency sequence, by traversing forward, the bellows numbers at which the cumulative frequency first reaches the lower and upper limits of the cumulative frequency are located respectively, thereby obtaining the frequency positioning interval; The endpoint observation segment is determined by taking the union of the gradient positioning interval and the frequency positioning interval.

[0009] Preferably, the arrangement of the temperature measurement matrix includes: A temperature data source is constructed based on historical temperature time-series data and encrypted measured cross-sectional distribution data; In the transverse direction of the bellows cross-section, the hysteresis distance is set with the transverse densification spacing as the step size; The experimental variability function value is calculated for each lag distance using the average temperature at each transverse measuring point. The experimental variability function values ​​of each bellows at the same lag distance are taken, and the lateral variability curve of the endpoint observation segment is obtained by averaging. The lateral sample variance is calculated using the temperature measurement data source. The lateral variation curve is fitted with least squares to obtain the sill value and lateral range. The sill value is compared with the lateral sample variance to calculate the relative deviation. The fit is considered valid only if the relative deviation is within the set range. Based on the spatial sampling theorem, the largest integer value that satisfies the condition is taken as the lateral measurement point spacing, and the number of lateral measurement points is calculated in combination with the cross-sectional width. ; Calculate the number of vertical measuring points in the vertical direction of the bellows. .

[0010] Preferably, the arrangement of the temperature measurement matrix further includes: The flow field distribution data inside the bellows is obtained. The gradient magnitudes of the transverse and vertical flow velocities are calculated by using the absolute values ​​of the first-order partial derivatives, and the velocity gradient field on the cross-section of the bellows is obtained. Calculate the gradient integral of the transverse velocity within the cross-sectional width; The theoretical integral for each transverse interval is obtained by using the ratio of the gradient integral to the number of transverse measurement points. The integral is then formed by integrating from the leftmost side of the cross-sectional width towards the right. One horizontal interval; Within the effective height, divide into One vertical interval; The temperature measurement point is installed at the center of the interval, thus forming a temperature measurement matrix inside each air box.

[0011] Preferably, the positioning of the coarse sintering endpoint includes: Based on the temperature measurement matrix of the i-th wind box, all temperature measurement points are grouped according to their horizontal positions, and the representative temperature of each horizontal position is obtained by averaging. Each representative temperature is numbered to form the lateral temperature sequence of the i-th wind box, and the temperature function of the i-th wind box is generated by least squares polynomial fitting. For each temperature function, the first derivative is taken to find the extreme points. The maximum points are selected by the sign of the second derivative value and used as inflection points. Inside the i-th windbox, the second derivative value of the inflection point is called, and the sharpness of the inflection point is calculated through absolute value operation. The number of inflection points is counted, and the lateral unimodality is determined, including flat, unimodal, and multimodal.

[0012] Preferably, the positioning of the coarse sintering endpoint further includes: For all bellows identified as single-peak within the final observation segment, the arithmetic mean of the sharpness of the inflection point is taken to obtain the comparison benchmark. If the sharpness of a single bellows is less than the comparison benchmark, it is judged as medium confidence; otherwise, it is judged as high confidence; for flat and multi-peaked bellows, it is judged as low confidence. Based on the confidence level, a confidence coefficient is set. For a single bellows, the temperature at the inflection point is called, and the initial weight of the corresponding bellows is obtained by multiplying the temperature and the confidence coefficient. The initial weights of all bellows are proportionally normalized to obtain the fusion weights of each bellows. Based on the fusion weight, the bellows numbers in the endpoint observation section are weighted and fused to obtain the coarse sintering endpoint.

[0013] Preferably, the step of generating and issuing control commands includes: Select high-quality production periods and extract actual sintering endpoints to form a sample set; The target bellows number is obtained by performing an arithmetic mean operation and rounding on the actual sintering endpoints in the sample set. Based on the difference between the predicted sintering endpoint and the target wind box number, the longitudinal endpoint deviation is calculated. ; Calculate the horizontal partition deviation based on the maximum and minimum temperature values ​​of the horizontal interval. ; Set longitudinal deviation threshold Lateral deviation threshold ; Only when or This triggers the adjustment mechanism; For the aforementioned adjustment mechanism, sample pairs are extracted, gain coefficients are obtained through least squares linear regression, and feedforward pre-adjustment amounts are calculated, including feedforward speed pre-adjustment amounts and feedforward opening pre-adjustment amounts. Using a PID controller, calculate the feedback correction, including the feedback machine speed correction and the feedback opening correction; Control commands are generated by superimposing the feedforward pre-adjustment amount and the feedback correction amount.

[0014] A closed-loop control system for the sintering endpoint based on a multi-field fusion network includes: a temperature acquisition module, a positioning module, and a deviation correction module. The temperature acquisition module is used to number the air boxes deployed under the sintering machine trolley, select the end observation section, arrange the temperature measurement matrix, and collect the temperature and related process parameters of each temperature measurement point in the temperature measurement matrix in real time. The data is then preprocessed and stored in the time series database. The positioning module is used to fit the temperature of the transverse temperature measurement point of each wind box, locate the coarse sintering endpoint by positioning the curve inflection point and weighted fusion, construct a spatiotemporal temperature field for the endpoint observation segment, and generate the predicted sintering endpoint by endpoint prediction. The deviation correction module is used to set the target windbox number, compare the deviation with the predicted sintering endpoint, generate two-dimensional deviation parameters, and generate control commands by using feedforward prediction and feedback closed loop.

[0015] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: By organically linking the selection of endpoint observation sections, matrix-style temperature measurement arrangement, and multi-physics parameter acquisition, a complete temperature measurement system covering the sintering machine's running direction and the trolley's horizontal and vertical directions was constructed, enabling a comprehensive characterization of the coupling state of multiple physical fields such as temperature, airflow, and materials, providing an accurate data foundation for subsequent prediction and control. Simultaneously, the lateral temperature of each windbox is fitted, and the coarse sintering endpoint is located by curve inflection point identification and weighted fusion. Then, the predicted sintering endpoint and endpoint temperature are generated by combining the spatiotemporal temperature field and endpoint prediction, forming a progressive endpoint perception mechanism from coarse to fine. Even if individual windboxes are affected by local operating condition disturbances and cause temperature abnormalities, the influence of abnormal windboxes on endpoint positioning can be effectively suppressed by inflection point credibility discrimination and weighted fusion, making the endpoint prediction results closer to the actual operating conditions. In the control phase, the deviation between the target bellows number determined by data-driven determination and the predicted sintering endpoint is compared to generate two-dimensional deviation parameters: longitudinal endpoint deviation and lateral zone temperature difference deviation of the trolley. This decouples the two issues of overall sintering speed and lateral sintering uniformity, corresponding to independent adjustment of trolley speed and damper, respectively. This ensures that the control objective is clear and the adjustment means correspond one-to-one with the problem. Using a control strategy that combines feedforward prediction and feedback closed loop, the feedforward loop outputs the adjustment direction in advance based on the prediction results, shortening the response time from the occurrence of deviation to the execution of adjustment. The feedback loop corrects the adjustment amount based on the actual deviation, ensuring the accuracy and stability of the control. The two complement each other, enabling the sintering endpoint to be timely and accurately controlled to the target position. Attached Figure Description

[0016] Figure 1 The flowchart shows a closed-loop control method for the sintering endpoint based on a multi-field fusion network. Figure 2 This is a flowchart of the process for selecting the endpoint observation segment in this invention; Figure 3 This is a flowchart illustrating the process of locating the endpoint of coarse sintering in this invention; Figure 4 This is a structural diagram of a sintering endpoint closed-loop control system based on a multi-field fusion network. Detailed Implementation

[0017] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof.

[0018] Example 1 like Figures 1-3 As shown, this embodiment provides a closed-loop control method for the sintering endpoint based on a multi-field fusion network, including the following steps: The air boxes deployed under the sintering machine trolley are sequentially numbered from 1 to n. By selecting the endpoint observation section, a temperature measurement matrix is ​​arranged, and the temperature of each temperature measurement point in the temperature measurement matrix is ​​collected in real time to form temperature measurement time series data. At the same time, the associated process parameters related to the sintering endpoint are collected in real time, such as the negative pressure of each air box, the material layer thickness of the trolley (obtained by a high-temperature layer thickness detector), the trolley speed, the ignition furnace temperature, the oxygen content of the air box, and the opening degree of the air box valve. The data is preprocessed and stored in the time series database. Based on the temperature measurement time series data, the temperature of the transverse temperature measurement point of each wind box is fitted to generate a temperature curve function, and the derivative of each temperature curve function is solved to locate the curve inflection point. By weighted fusion of the wind box numbers at all curve inflection points in the endpoint observation segment, the endpoint of coarse sintering is located. Based on spatial and temporal dimensions, a spatiotemporal temperature field is constructed for the endpoint observation segment to characterize the continuous distribution of temperature in space and its dynamic evolution over time. At the same time, the coarse sintering endpoint is marked, and the process parameters and the coarse sintering endpoint are used as prior combustion information. Through endpoint prediction, the predicted sintering endpoint and endpoint temperature are generated. Based on the preset target bellows number, the deviation is compared with the predicted sintering endpoint. Combined with the lateral temperature difference of the spatiotemporal temperature field, the deviation is judged to generate two-dimensional deviation parameters, including longitudinal endpoint deviation and trolley lateral zone temperature difference deviation, which correspond to longitudinal machine speed adjustment and lateral damper independent adjustment, respectively. Using feedforward prediction and feedback closed loop, control commands are generated and issued to the underlying execution equipment according to safety priority.

[0019] Furthermore, the steps for selecting the endpoint observation segment include: Along the direction of travel of the trolley, from the head to the tail, all the bellows under the trolley are numbered sequentially as 1, 2, ..., n according to their physical location, and the numbering is mapped one-to-one with the spatial location of the bellows; Based on the existing temperature sensors in each wind box, historical temperature time series data of n wind boxes in at least two consecutive sintering cycles are retrieved as the basic data source for the final observation segment selection. This basic data source is the original single-point temperature detection data of the wind box. Historical temperature time series data is divided into configuration period and positioning period according to sintering cycle. For example, the latest sintering cycle in the historical temperature time series data is set as the positioning period, and the sintering cycle that is not the positioning period is the configuration period. Configuration data source and positioning data source are generated. The configuration data source is used to calculate the gradient threshold, and the positioning data source is used to locate the endpoint observation segment. Call the configured data source to calculate the average temperature of the time-series temperature data for the i-th bellows (i=1,2,…,n) within the configured period. The average temperature sequence of the entire air box during the configuration period is obtained. ; This represents the average temperature of the nth bellows; The average temperature sequence of all wind boxes is processed by first-order forward difference, that is, the longitudinal temperature gradient between adjacent wind boxes is calculated by the difference in average temperature between adjacent wind boxes. The gradient sequence of the configuration period is obtained. ; Let represent the longitudinal temperature gradient between the nth bellows and the previous bellows, where Since the mixture in the trolley area corresponding to the first wind box has just been fed and ignited, it is mainly in the drying and preheating stage and has not yet entered the intense combustion stage. Therefore, the case where the first wind box is the rising edge is not considered in this application. For the gradient sequence of the configuration period, it is divided into two equal segments according to time, denoted as follows: , And calculate the standard deviation respectively. , In this implementation, the standard deviation is calculated using the sample standard deviation, i.e., by dividing by the degrees of freedom. The unbiased estimation calculation method This represents the number of data points in the gradient sequence within the corresponding segment. Since the product of the confidence coefficient and the standard deviation is the gradient threshold, the absolute value of the longitudinal temperature gradient in the gradient sequence of the configuration period is calculated, and the maximum and minimum absolute values ​​are taken. The ratio of the maximum value to the overall standard deviation of the gradient sequence of the configuration period is used as the upper limit of the set, and the ratio of the minimum value to the overall standard deviation is used as the lower limit of the set. A set of candidate confidence coefficients is constructed with a step size of 0.2. For each candidate confidence coefficient, the candidate threshold is calculated by multiplying the standard deviation by the candidate confidence coefficient. ,include The corresponding first candidate threshold , The corresponding second candidate threshold ,use position The rising and falling edges, using position The rising and falling edges; where, the gradient sequence is traversed in the positive direction from the nose to the tail, locating the first edge that satisfies... The bellows number is the rising edge, which is the starting position of a significant temperature increase in the combustion zone. The gradient sequence is traversed in reverse from the tail to the nose to locate the first position that satisfies this condition. The bellows number is the falling edge, which is the end position where the combustion zone temperature drops back to the baseline; based on , The absolute difference between the two rising edges and the absolute difference between the two falling edges are calculated separately. The location difference of each candidate confidence coefficient is calculated by summing the two absolute differences. The candidate confidence coefficient with the smallest location difference is selected as the final confidence coefficient. For example, the sintering machine trolley has 10 air boxes, i.e., n=10, and the candidate confidence coefficient set K={1.0,1.5,2.0,2.5,3.0} is configured in the gradient sequence of the cycle. Corresponding to {2,3,5,6,24,35,10,-15,-32} Corresponding to {2,2,4,6,2,38,35,-15,-35}; but gradient mean ,at this time corresponding standard deviation , gradient mean ,at this time corresponding standard deviation ;in, for The Middle There are gradients here. , for The Middle There are gradients here. ; When the candidate confidence coefficient is 1.0, , ;use Forward traversal of {2,3,5,6,24,35,10,-15,-32}, the corresponding rising edge That is, the bellows numbered 6, through reverse traversal, corresponds to the falling edge That is, bellows numbered 10; utilizing Forward traversal of {2,2,4,6,2,38,35,-15,-35}, the corresponding rising edge That is, the bellows numbered 7, through reverse traversal, corresponds to the falling edge That is, the bellows numbered 10; pass The location discrepancy was found to be 1 when the candidate confidence coefficient was 1.0. Similarly, when the candidate confidence coefficient is 1.5, the positioning difference is 0. When the candidate confidence coefficients are 2.0 / 2.5 / 3.0, there is no rising edge, so the corresponding positioning difference is equivalent to infinity. At this time, the positioning difference is the minimum of 0, and the corresponding final confidence coefficient is 1.5. Calculate the configuration standard deviation for the gradient sequence with configuration period. The gradient threshold is obtained by configuring the product of the standard deviation and the confidence coefficient. ; To locate the data source, a gradient sequence is constructed, denoted as the location gradient sequence. This is then processed by performing a forward traversal of the location gradient sequence (and...). Comparison) and reverse traversal (and) (Comparison) Locating the rising edge from the localization gradient sequence and falling edge Thus, the gradient positioning range is obtained. ; Iterate through the location data sources. For each sampling time, identify the bellows number with the highest temperature peak, i.e., the bellows number with the highest exhaust gas temperature at each sampling time. After iterating through all sampling times, count the number of times the i-th bellows is recorded as the temperature peak during the positioning cycle. The peak frequency of the i-th bellows is calculated by the ratio of the number of sampling times to the total number of sampling moments. Thus, the peak frequency sequence is obtained. ; Based on the cumulative distribution function, the cumulative frequency from the first bellows to the i-th bellows is calculated by accumulating the peak frequencies. The cumulative frequency sequence is obtained. ;in, ; Based on the peak frequency sequence, the bellows numbers are weighted and summed using the peak frequency as the weight. The weighted sum is then divided by the number of bellows to obtain the weighted mean. The standard deviation of each bellows number is calculated using this weighted mean, yielding the positional standard deviation. The confidence level is then adaptively configured based on this positional standard deviation, and the cumulative lower and upper limits are derived from the confidence level. , , The confidence levels are as follows: when the standard deviation of the location is less than 1, the endpoint dispersion is small and the operating conditions are stable, so a 90% confidence level is selected; when the standard deviation of the location is between 1 and 2, it indicates that the endpoint dispersion is moderate, so a 95% confidence level is selected; when the standard deviation of the location is greater than 2, it indicates that the endpoint dispersion is large and the operating conditions fluctuate greatly, so a 99% confidence level is selected. In the cumulative frequency sequence, the bellows number where the cumulative frequency first reaches the lower limit is located by forward traversal, and this number serves as the statistical starting position for the sintering endpoint activity range. The bellows number at which the cumulative frequency first reaches the cumulative upper limit is located is used as the statistical termination point of the sintering endpoint activity range. Thus, the frequency positioning range is obtained. ; When using the gradient method, the boundary of significant temperature changes in the combustion zone is located from a spatial perspective based on the temperature change rate of adjacent bellows. However, this method is insensitive to conditions with gentle temperature gradients. When using the frequency method, the range of the endpoint activity is located from a temporal perspective based on the probability distribution of temperature peaks. However, this method lacks robustness to single abnormal shifts (such as a large drift of the endpoint at a certain moment). Each method has its own limitations. This application cross-validates from both spatial and temporal dimensions to improve the reliability and comprehensiveness of the observation segment location. The union of the gradient location interval and the frequency location interval is taken to finally determine the endpoint observation segment. ,in, , , This indicates taking the minimum value. This indicates taking the maximum value.

[0020] Furthermore, the steps for setting up the temperature measurement matrix include: A temperature data source is constructed based on historical temperature time-series data of the entire windbox and temporarily encrypted measured cross-sectional distribution data. Specifically, in each windbox within the final observation section, temporarily encrypted temperature measurement points are arranged in the transverse and vertical directions of the cross-section to collect temperature time-series data for at least one complete sintering cycle. The transverse encryption spacing is as follows: That is, 10 temperature measuring points are evenly and densely arranged in the transverse direction along the cross-section of the bellows, and the vertical spacing is increased. That is, six temperature measuring points are evenly and densely arranged along the vertical direction of the cross-section. The width of the cross-section. For effective height; In the transverse direction of the bellows cross-section, a series of hysteresis distances are set with the transverse densification spacing as the step size; the maximum value of the hysteresis distance does not exceed half the width of the bellows cross-section, ensuring that each hysteresis level has a sufficient number of location point data pairs, and each hysteresis distance represents the transverse spatial interval between two temperature measurement positions. For each bellows in the final observation section, the temperature of each lateral measuring point within the acquisition period is averaged over time as the spatial representative temperature of the corresponding lateral measuring point. The experimental variability function value is calculated for each lag distance, as shown in the following expression:

[0021] In the formula, The first observation segment at the end point Experimental variability function values ​​of the typhoon box at a lag distance h. This refers to the sequential numbering of each bellows in the terminal observation section. , This represents the total number of bellows in the terminal observation section, and , bellows The total number of data pairs of points that are exactly h apart. Horizontal position The space represents temperature. For the corresponding bellows One horizontal temperature measurement point, To and The space at a distance h represents temperature; The experimental variability function values ​​of each bellows at the same lag distance are taken and averaged to obtain the transverse variability curve of the final observation section, in order to eliminate the influence of accidental operating conditions of a single bellows. The expression is as follows:

[0022] In the formula, The average experimental variability function of the endpoint observation segment at the lag distance h; Using temperature data sources, for the same horizontal measuring point containing vertical measuring points, the spatial representative temperature of each horizontal measuring point is calculated through temperature mean calculation. Based on the calculation logic of sample standard deviation, the horizontal sample variance of the endpoint observation segment is calculated as a theoretical reference value for the sill value. Based on the spherical model in the theory of variograms, least squares fitting is performed on the discrete data point set in the transverse variogram, as shown in the following expression:

[0023] in, The lag distance, In lag distance The experimental variogram values ​​are as follows. The nugget value, i.e., the variogram value when h approaches 0, characterizes measurement error and microscale variation. For structural variance, For sill values, For the variable range, the nugget value, structural variance, and variable range are unknown parameters at this time, and need to be solved by least squares fitting; Fitting and solving , , Three parameters are used, and the solved sill value is compared with the lateral sample variance to calculate the relative deviation. That is, the absolute difference between the sill value and the lateral sample variance is used as the numerator and the lateral sample variance is used as the denominator. A very small division constant is added to the denominator to prevent the denominator from being 0 when the lateral sample variance approaches 0. If the relative deviation is within the set range, such as less than 15%, it indicates that the fit is effective; otherwise, use an exponential model or a Gaussian model to refit until the fit is effective. The range obtained through fitting is the lateral range. This represents the largest scale of spatial correlation of temperature in the lateral direction. According to the spatial sampling theorem, the interval between lateral measurement points should not exceed [a certain value]. Then, the largest integer value that satisfies the condition is taken as the horizontal measurement point spacing. The number of transverse measuring points is obtained by rounding up the ratio of the cross-sectional width to the spacing between transverse measuring points. ; Similarly, in the vertical direction of the bellows, after setting the vertical densification interval for the sequence, the vertical range is obtained by calculating the experimental variogram value, averaging the endpoint observation segment, least squares fitting, and verifying with the sill value. According to the spatial sampling theorem, the interval between vertical measuring points should not exceed [a certain value]. Then, the largest integer value that satisfies the condition is taken as the vertical measuring point spacing. The number of vertical measuring points is obtained by rounding up the ratio of the effective height to the distance between vertical measuring points. ; Because the velocity and temperature distribution within the bellows are spatially non-uniform across its cross-section—specifically, near the bellows sidewalls and at airflow bends, the velocity gradient is large and temperature changes drastically, while the central region has a gentler velocity and more uniform temperature—uniform sampling would lead to insufficient sampling in areas of intense gradients and redundant sampling in areas of gentler gradients. Therefore, after determining the number of measuring points, their locations are not uniformly distributed but rather optimized based on the velocity gradient within the bellows. This prioritizes the placement of measuring points in areas of dramatic flow and temperature changes. Flow field distribution data within the bellows is obtained through actual measurements using a hot-wire anemometer. The gradient magnitude of the lateral velocity is calculated using the absolute value of the first partial derivative of the velocity with respect to the lateral coordinate. The gradient magnitude of the vertical flow velocity is calculated by using the absolute value of the first-order partial derivative of the flow velocity with respect to the vertical coordinate. , thus obtaining the velocity gradient field on the cross-section of the bellows; where, a local coordinate system is constructed for the cross-section, with the leftmost inner wall of the bellows cross-section as the origin and the positive x-axis direction to the right; Within the cross-sectional width, the gradient magnitude of the transverse velocity is integrated along the transverse coordinate to calculate the gradient integral of the transverse velocity. Using the same theoretical integral for each interval, the ratio of the gradient integral to the number of transverse measurement points is obtained, i.e. We obtain the theoretical integral for each horizontal interval, and integrate from the leftmost side of the width to the right, forming... One horizontal interval; Similarly, within the effective height, divide into One vertical interval; For each horizontal section, the center of the section is used as the horizontal installation position of the temperature measuring point. For each vertical section, the center of the section is used as the vertical installation position of the temperature measuring point, thus forming a temperature measuring matrix inside each air box. Each temperature measuring point is equipped with a thermocouple with a corresponding range and response time, which is connected to a multi-channel signal acquisition device through a compensating wire, and transmitted to the PLC via fiber optic cable through a convergence box to achieve millisecond-level synchronous acquisition.

[0024] Furthermore, the steps for locating the endpoint of coarse sintering include: Based on the temperature measurement matrix of the i-th wind box, all temperature measurement points are grouped according to their horizontal positions, and the representative temperature of each horizontal position is obtained by averaging the temperature within a single horizontal position. The vertical deployment of temperature measurement points is to improve the robustness of the representative temperature and avoid over-reliance on a single measurement position. Once the measurement point is disturbed by local air leakage, dust accumulation, or local air leakage, the horizontal representative temperature will be distorted, thereby contaminating the entire horizontal fitting curve and peak location. Based on the representative temperatures at each lateral position within the i-th bellows, each representative temperature is assigned a number using the bellows number and the point number at each lateral position. That is, the representative temperature at the j-th lateral position of the i-th bellows is... This forms the lateral temperature sequence of the i-th bellows. The points corresponding to the locations are numbered sequentially from left to right based on the number of horizontal measurement points, i.e., 1, ..., ... , ; Based on the horizontal temperature series, with the point number as the independent variable and the representative temperature as the dependent variable, the temperature function of the i-th bellows is generated by least squares polynomial fitting. The expression is as follows:

[0025] In the formula, The point numbering for the transverse measuring points. , Let be the order of the polynomial fitting, and In this embodiment, we take , ... The coefficients of the polynomial to be determined are obtained by least squares polynomial fitting. Take the first derivative of each temperature function and set the first derivative to zero to find the extreme points; Since the minimum point represents the lowest horizontal temperature, while the peak inflection point represents the horizontal position with the highest exhaust gas temperature, corresponding to the local maximum point of the temperature function, the second derivative of the temperature function of the i-th fan is taken. Substituting the coordinates of the extreme point into the second derivative function, the second derivative value of the extreme point is obtained. The maximum point is selected by the sign of the second derivative value, that is, when the second derivative value is negative, this extreme point is the maximum point, and the maximum point is taken as the inflection point. ; The second derivative value of the inflection point is used to calculate its sharpness through absolute value operations. The greater the sharpness, the sharper the peak value, the more concentrated the lateral distribution of the combustion zone, and the more reliable the peak temperature represents the combustion intensity of the bellows. The number of inflection points is counted within the i-th bellows. And perform lateral unimodal determination; like This indicates that the temperature curve is monotonous or plateau-shaped in the horizontal direction, with no temperature peaks, and is therefore judged to be flat. The bellows cannot extract an effective representative of the combustion intensity. like This indicates that the transverse temperature distribution has a single-peak shape, which is a normal operating condition and is therefore determined to be single-peak. like This indicates that there are multiple independent high temperature zones in the transverse temperature distribution, which is a direct representation of abnormal working conditions such as material surface air leakage, uneven material distribution, and local air leakage. It is determined to be multi-peak, and the peak temperature of the wind box loses its unique representativeness, so its peak value is deemed unreliable. For the sharpness confidence assessment, for all bellows determined to be single-peak within the final observation segment, the arithmetic mean of the sharpness at the inflection point is taken to obtain the comparison benchmark. If the sharpness of a single bellows is less than the comparison benchmark, it indicates that the peak value is relatively flat and the combustion distribution is diffuse, and it is judged as medium confidence, and the corresponding confidence coefficient is set to 0.5. If the sharpness of a single bellows is not less than the comparison benchmark, it indicates that the peak value is sharp and the combustion is concentrated, and it is judged as high confidence, and the corresponding confidence coefficient is set to 1. For flat and multi-peak bellows, it is judged as low confidence, and the corresponding confidence coefficient is set to 0, that is, the bellows peak value is removed and does not participate in the fusion. For a single bellows, the temperature at the inflection point is called, and the initial weight of the corresponding bellows is obtained by multiplying the temperature with the confidence coefficient. The initial weights of all bellows are proportionally normalized to obtain the fusion weight of each bellows, so that the sum of the fusion weights is 1. Based on the fusion weight, the bellows numbers in the endpoint observation section are weighted and fused to obtain the coarse sintering endpoint.

[0026] Furthermore, the steps for generating the predicted sintering endpoint and endpoint temperature include: A three-dimensional spatial coordinate system is established with the trolley's travel direction as the longitudinal direction (mapped to continuous longitudinal coordinates by the bellows number), the trolley's width direction as the transverse direction, and the bellows' depth direction as the vertical direction. Each temperature measurement point in the final observation section is mapped to the three-dimensional spatial coordinate system to obtain the three-dimensional spatial coordinates of each temperature measurement point. The temperature time series data of each temperature measurement point is then bound to the corresponding three-dimensional spatial coordinates to form matrix temperature measurement data with three-dimensional coordinate labels. Based on matrix temperature measurement data, three-dimensional spatial interpolation is used to reconstruct the temperature of discrete temperature measurement points into a continuous temperature field, generating a three-dimensional temperature field snapshot covering the space of the final observation segment, which characterizes the continuous temperature distribution within the space of the final observation segment. Using a sliding time window, the three-dimensional temperature field snapshot is converted into a time series along the time axis, and the time series is smoothed using an exponentially weighted moving average (EWMA) to generate a spatiotemporal temperature field; the expression is shown below:

[0027] In the formula, Let be the smoothed spatiotemporal temperature field at time t. This is a snapshot of the three-dimensional temperature field input at time t. For smoothing weighting coefficients, In this embodiment, we take ; The coarse sintering endpoint is used as the combustion peak mark, which is embedded at the corresponding longitudinal coordinate of the spatiotemporal temperature field to mark the current combustion zone center position in the temperature field. The spatiotemporal features of the temperature field, which characterize the spatial distribution of temperature and the evolution of combustion state, are extracted from the spatiotemporal temperature field. These features include the temperature peak, the longitudinal position of the peak, the transverse temperature gradient, the longitudinal temperature gradient, and the position of the combustion zone front. The position of the combustion zone front is the spatial boundary position in the direction of the trolley's movement (from the front of the trolley to the rear of the trolley) where the combustion reaction has not yet fully developed and the temperature begins to rise from a low temperature. The associated process parameters are used as multi-physics field combustion information, and the coarse sintering endpoint is used as combustion prior location information based on the exhaust gas temperature curve. The associated process parameters and the coarse sintering endpoint together constitute the combustion prior information for endpoint prediction. Using the spatiotemporal characteristics of the temperature field, associated process parameters, and the rough sintering endpoint as common inputs, a pre-built prediction model is invoked to predict the endpoint and generate the predicted sintering endpoint and predicted endpoint temperature under the current operating conditions. The prediction model includes an input layer, multiple hidden layers, and an output layer. The input layer is used to receive and normalize the input feature vector consisting of the spatiotemporal characteristics of the temperature field, associated process parameters and the rough sintering endpoint, so as to eliminate the influence of data with different dimensions on the model. Hidden layers are used to extract high-dimensional coupling and correlation features between input features and sintering endpoint through nonlinear activation mapping, enabling the model to learn the complex nonlinear relationship between multi-physics parameters and sintering endpoint. The output layer is used to output the predicted sintering endpoint and the predicted endpoint temperature; Historical operating condition data are extracted from the time series database to construct a training sample set. The sample inputs are the spatiotemporal characteristics of the temperature field at each historical moment, the associated process parameters and the rough sintering endpoint, and the sample labels are the historical actual sintering endpoint and endpoint temperature. The loss function is constructed using the mean squared error between the model's predicted output and the actual label, and serves as the objective for training optimization. With the goal of minimizing the loss function, the weights and bias parameters of each layer of the network are updated iteratively through backpropagation and gradient descent. The Adam optimizer is used, and the weights and bias parameters of each layer of the network are updated iteratively through backpropagation and mini-batch gradient descent. The learning rate decay strategy is adopted, the initial learning rate is determined by optimization based on the historical validation set, the training batch size is set, and mini-batch samples are used to participate in parameter updates in each iteration. Repeat the iteration until the loss function converges to obtain the trained prediction model. The convergence criterion is that either of the following two conditions is met: the decrease in the loss function on the validation set for multiple consecutive iterations is less than a preset convergence threshold (e.g., 10). -4 ); The number of iterations reaches the maximum iteration limit (e.g., 500 iterations).

[0028] Furthermore, the steps for generating control instructions include: The high-quality production periods that meet the process qualification requirements for sintered ore quality indicators are selected from historical data, such as stable FeO content within the process allowable range and finished product rate meeting the standard. The actual sintering endpoints at each moment within the high-quality production period are extracted, that is, the historical actual endpoint data represented by the bellows number, to form a sample set. The arithmetic mean of the actual sintering endpoints in the sample set is calculated to obtain the statistical center value. Since the bellows number is a discrete integer, the statistical center value is rounded to obtain the target bellows number, which represents the target state that the controlled variable should reach. This is used to approximate the actual endpoint (i.e. the predicted sintering endpoint output by the prediction model) to the set target position. Based on the difference between the predicted sintering endpoint and the target wind box number, the longitudinal endpoint deviation is calculated. , This indicates that the predicted endpoint is lagging behind the target; at this point, the combustion zone has not yet reached the target bellows, and sintering is slow. This indicates that the predicted endpoint is ahead of the target, at which point the combustion zone has already crossed the target bellows, and sintering is too fast; Based on the horizontal intervals in the temperature measurement matrix deployment, the average temperature of each horizontal interval in the final observation segment is extracted from the spatiotemporal temperature field. The maximum and minimum temperature values ​​are also extracted. The difference between the maximum and minimum temperature values ​​is used to obtain the horizontal partitioning deviation. This reflects the difference in sintering speed between the left and right sides caused by uneven material distribution or uneven airflow on the trolley. Set deviation thresholds, including longitudinal deviation thresholds. Lateral deviation threshold Among them, the lateral deviation samples under historical normal working conditions are obtained, and the lateral deviation threshold is set using the 3σ criterion, that is, the sample standard deviation of the lateral deviation samples is calculated and three times the sample standard deviation is used as the lateral deviation threshold. The longitudinal deviation samples under historical normal working conditions are obtained, and the longitudinal deviation threshold is set using the 3σ criterion, that is, the sample standard deviation of the longitudinal deviation samples is calculated and three times the sample standard deviation is used as the longitudinal deviation threshold. like and The current operating condition is determined to be normal, and no adjustment is required. like or If a longitudinal overall offset or lateral sintering unevenness is detected, an adjustment mechanism is triggered. For the adjustment mechanism, historical sample pairs of changes in the endpoint position caused by trolley speed adjustment are extracted from the time-series database. Each sample represents a speed adjustment amount and the corresponding longitudinal endpoint deviation. Using the longitudinal endpoint deviation as the independent variable and the speed adjustment amount as the dependent variable, a linear relationship is fitted using least-squares linear regression. The slope obtained from the regression is the longitudinal feedforward gain coefficient. This coefficient is only applied when there is an overall longitudinal offset. By using the longitudinal feedforward gain coefficient and the current longitudinal endpoint deviation The product operation is used to calculate the feedforward speed pre-adjustment; Sample pairs of historical data on lateral temperature difference changes caused by damper opening adjustments were extracted from the time-series database. Using lateral partitioning deviation as the independent variable and damper opening adjustment amount as the dependent variable, a linear relationship was fitted using least-squares linear regression. The slope obtained from the regression is the lateral feedforward gain coefficient. This coefficient is only applied when lateral sintering inhomogeneity exists. The current deviation between the lateral gain coefficient and the lateral partition. The product operation is used to calculate the feedforward opening pre-adjustment amount; Using PID controllers—a longitudinal PID controller with longitudinal endpoint deviation as input and a lateral PID controller with lateral partition deviation as input—feedback corrections are calculated, including feedback machine speed corrections (only when lateral sintering unevenness exists) and feedback damper opening corrections (only when lateral sintering unevenness exists). The longitudinal PID controller is used for machine speed regulation, and the lateral PID controller is used for damper opening regulation. The proportional, integral, and derivative coefficients of both PID controllers are obtained through offline optimization tuning using historical operating data. Historical production datasets are extracted, containing timing records of sintering endpoint longitudinal deviation, lateral partition deviation, machine speed output, damper opening output, and system response output. Sintering cycle segments with no equipment faults and in closed-loop regulation are selected, while abnormal data segments involving start-up / shutdown, forced human intervention, and equipment alarm faults are removed to obtain the original PID tuning dataset. The original dataset is divided into a training subset and a validation subset. The training subset is used for simulation of candidate PID parameters. The loop operation uses a validation subset to evaluate the control effect of each set of candidate parameters. Overshoot, settling time, and cumulative absolute error are used as control evaluation indicators. Overshoot is the maximum deviation of the control output from the target value, settling time is the time required for the deviation to converge to within the allowable error band, and cumulative absolute error is the time integral of the absolute value of the deviation during the entire adjustment process. By superimposing the control evaluation indicators, a comprehensive evaluation function is constructed. A grid traversal is performed on the proportional, integral, and derivative parameters of the PID controller. For each set of candidate PID parameters, closed-loop simulation is performed based on historical time-series data from the training subset to calculate the comprehensive evaluation function corresponding to that set of parameters. The set of parameters that minimizes the comprehensive evaluation function is selected as the initial PID parameters for tuning. The initial PID parameters are substituted into the validation subset for closed-loop simulation verification. If both overshoot and settling time meet the process constraints, the set of parameters is determined as the officially used PID parameters. If the process constraints are not met, the parameter search interval is narrowed, and the optimization is re-tested. By superimposing the feedforward pre-adjustment amount and the feedback correction amount, control commands are generated and sent to the corresponding underlying execution devices through the industrial communication network. The trolley speed command is sent to the trolley frequency converter for execution, and the damper opening command is sent to the damper actuators of each air box for execution.

[0029] Example 2 Another embodiment provided by the present invention: See Figure 4 The sintering endpoint closed-loop control system based on multi-field fusion network includes: a temperature acquisition module, a positioning module, and a deviation correction module. The temperature acquisition module is used to number the air boxes deployed under the sintering machine trolley, select the endpoint observation section, arrange the temperature measurement matrix, collect the temperature of each temperature measurement point in the temperature measurement matrix in real time, form temperature measurement time series data, and collect the associated process parameters related to the sintering endpoint in real time, and store them into the time series database after data preprocessing. The positioning module is used to fit the temperature of the transverse temperature measurement point of each wind box. By solving the derivative, the inflection point of the curve is located. Based on the weighted fusion of all curve inflection points in the endpoint observation segment, the coarse sintering endpoint is located. A spatiotemporal temperature field is constructed for the endpoint observation segment. The process parameters and the coarse sintering endpoint are used as the combustion prior information. Through endpoint prediction, the predicted sintering endpoint and endpoint temperature are generated. The deviation correction module is used to set the target windbox number, compare the deviation with the predicted sintering endpoint, generate two-dimensional deviation parameters, and generate control commands by using feedforward prediction and feedback closed loop.

[0030] In summary, this invention sequentially numbers and selects endpoint observation sections for the sintering air boxes, arranges a matrix of temperature measurement points covering both horizontal and vertical directions, and collects the temperature of each measurement point in real time to form temperature measurement time-series data. Simultaneously, it collects and stores related process parameters such as negative pressure, material layer thickness, and trolley speed into a time-series database. This multi-point temperature measurement comprehensively characterizes the coupling state of the temperature field, airflow field, and material field, avoiding the limitation of endpoint judgment accuracy caused by interference from single or a few temperature points due to material surface leakage or tail-end air leakage. Temperature fitting is performed on the horizontal temperature measurement points of each air box, and the coarse sintering endpoint is located through curve inflection point identification and weighted fusion. Then, a spatiotemporal temperature field is constructed for the endpoint observation section, and the predicted sintering endpoint and endpoint temperature are generated through endpoint prediction, achieving a coarse-to-fine process. The progressive endpoint perception effectively suppresses the impact of abnormal bellows on positioning. The target bellows number determined by data-driven determination is compared with the predicted sintering endpoint to generate two-dimensional deviation parameters: longitudinal endpoint deviation and lateral zone temperature difference deviation of the trolley. These parameters correspond to the independent adjustment of the trolley speed and the damper, respectively. Control commands are generated through feedforward prediction and feedback closed loop and distributed to the underlying execution equipment in stages. The feedforward link outputs the adjustment direction in advance based on the prediction results, shortening the response time from the occurrence of deviation to the execution of adjustment. The feedback link corrects the adjustment amount based on the actual deviation, ensuring control accuracy. This avoids the lag in manual operation judgment, the difficulty in unifying operation standards, and the time delay in feedback adjustment, thus achieving timely and accurate control of the sintering endpoint.

[0031] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A closed-loop control method for sintering endpoint based on multi-field fusion network, characterized in that, include: The sintering air boxes were numbered sequentially, and the final observation section was selected and the temperature measurement matrix was arranged. Based on the temperature measurement matrix, the temperature of each temperature measurement point is collected in real time to form temperature measurement time series data. Related process parameters are collected synchronously, and a time series database is constructed through preprocessing. Temperature fitting is performed on the transverse temperature measurement points of each wind box, and the endpoint of coarse sintering is located by curve inflection point identification and weighted fusion. A spatiotemporal temperature field is constructed for the endpoint observation segment, and the predicted sintering endpoint and endpoint temperature are generated through endpoint prediction. By comparing the deviation between the target windbox number and the predicted sintering endpoint, a two-dimensional deviation parameter is generated. Using feedforward prediction and feedback closed loop, a control command is generated and issued.

2. The sintering endpoint closed-loop control method based on a multi-field fusion network according to claim 1, characterized in that, The screening endpoint observation segment includes: All air boxes are sequentially numbered along the direction of travel of the sintering machine trolley; Retrieve historical temperature time series data of all wind boxes, and divide the configuration period and positioning period into the historical temperature time series data according to the sintering cycle, and generate configuration data source and positioning data source. The configured data source is called to calculate the average temperature of each wind box, thus obtaining the average temperature sequence of the entire wind box. The uniform temperature sequence of the entire wind box is subjected to first-order forward difference processing to calculate the longitudinal temperature gradient between adjacent wind boxes, thereby obtaining the gradient sequence of the configuration period, and the standard deviation of the gradient sequence is calculated. Configure the confidence coefficient and, in conjunction with the standard deviation, calculate the gradient threshold; A positioning gradient sequence is constructed for the positioning data source, and rising and falling edges are filtered in combination with a gradient threshold to obtain the gradient positioning interval.

3. The sintering endpoint closed-loop control method based on a multi-field fusion network according to claim 2, characterized in that, The configured confidence coefficients include: For the gradient sequence of the configuration period, it is divided into two equal segments according to time, denoted as follows: , And calculate the standard deviation respectively. , ; Construct a set of candidate confidence coefficients. For each candidate confidence coefficient, calculate the candidate threshold by multiplying the standard deviation by the candidate confidence coefficient. , ; use position The rising and falling edges, using position The rising and falling edges; based on , The absolute difference between the rising edge and the falling edge is calculated separately, and the positioning difference of each candidate confidence coefficient is calculated by summing the absolute differences. The candidate confidence coefficient with the smallest positioning difference is selected as the confidence coefficient.

4. The sintering endpoint closed-loop control method based on a multi-field fusion network according to claim 2, characterized in that, The screening endpoint observation segment also includes: Iterate through the location data sources and identify the bellows number of the temperature peak at each sampling time; The number of times each windbox is recorded as a temperature peak during the statistical positioning cycle is counted, and the peak frequency of each windbox is calculated to obtain the peak frequency sequence. By accumulating the peak frequencies, the cumulative frequency from the first bellows to the i-th bellows is calculated, and the cumulative frequency sequence is obtained. Based on the peak frequency sequence, the position standard deviation is calculated with the peak frequency as the weight of the bellows number, and the confidence level is adaptively configured based on the position standard deviation. In the cumulative frequency sequence, by traversing forward, the bellows numbers at which the cumulative frequency first reaches the lower and upper limits of the cumulative frequency are located respectively, thereby obtaining the frequency positioning interval; The endpoint observation segment is determined by taking the union of the gradient positioning interval and the frequency positioning interval.

5. The sintering endpoint closed-loop control method based on a multi-field fusion network according to claim 1, characterized in that, The temperature measurement matrix arrangement includes: A temperature data source is constructed based on historical temperature time-series data and encrypted measured cross-sectional distribution data; In the transverse direction of the bellows cross-section, the hysteresis distance is set with the transverse densification spacing as the step size; The experimental variability function value is calculated for each lag distance using the average temperature at each transverse measuring point. The experimental variability function values ​​of each bellows at the same lag distance are taken, and the lateral variability curve of the endpoint observation segment is obtained by averaging. The lateral sample variance is calculated using the temperature measurement data source. The lateral variation curve is fitted with least squares to obtain the sill value and lateral range. The sill value is compared with the lateral sample variance to calculate the relative deviation. The fit is considered valid only if the relative deviation is within the set range. Based on the spatial sampling theorem, the largest integer value that satisfies the condition is taken as the lateral measurement point spacing, and the number of lateral measurement points is calculated in combination with the cross-sectional width. ; Calculate the number of vertical measuring points in the vertical direction of the bellows. .

6. The sintering endpoint closed-loop control method based on a multi-field fusion network according to claim 5, characterized in that, The temperature measurement matrix arrangement also includes: The flow field distribution data inside the bellows is obtained. The gradient magnitudes of the transverse and vertical flow velocities are calculated by using the absolute values ​​of the first-order partial derivatives, and the velocity gradient field on the cross-section of the bellows is obtained. Calculate the gradient integral of the transverse velocity within the cross-sectional width; The theoretical integral for each transverse interval is obtained by using the ratio of the gradient integral to the number of transverse measurement points. The integral is then formed by integrating from the leftmost side of the cross-sectional width towards the right. One horizontal interval; Within the effective height, divide into One vertical interval; The temperature measurement point is installed at the center of the interval, thus forming a temperature measurement matrix inside each air box.

7. The sintering endpoint closed-loop control method based on a multi-field fusion network according to claim 1, characterized in that, The positioning of the coarse sintering endpoint includes: Based on the temperature measurement matrix of the i-th wind box, all temperature measurement points are grouped according to their horizontal positions, and the representative temperature of each horizontal position is obtained by averaging. Each representative temperature is numbered to form the lateral temperature sequence of the i-th wind box, and the temperature function of the i-th wind box is generated by least squares polynomial fitting. For each temperature function, the first derivative is taken to find the extreme points. The maximum points are selected by the sign of the second derivative value and used as inflection points. Inside the i-th windbox, the second derivative value of the inflection point is called, and the sharpness of the inflection point is calculated through absolute value operation. The number of inflection points is counted, and the lateral unimodality is determined, including flat, unimodal, and multimodal.

8. The sintering endpoint closed-loop control method based on a multi-field fusion network according to claim 7, characterized in that, The positioning of the coarse sintering endpoint also includes: For all bellows identified as single-peak within the final observation segment, the arithmetic mean of the sharpness of the inflection point is taken to obtain the comparison benchmark. If the sharpness of a single bellows is less than the comparison benchmark, it is judged as medium confidence; otherwise, it is judged as high confidence; for flat and multi-peaked bellows, it is judged as low confidence. Based on the confidence level, a confidence coefficient is set. For a single bellows, the temperature at the inflection point is called, and the initial weight of the corresponding bellows is obtained by multiplying the temperature and the confidence coefficient. The initial weights of all bellows are proportionally normalized to obtain the fusion weights of each bellows. Based on the fusion weight, the bellows numbers in the endpoint observation section are weighted and fused to obtain the coarse sintering endpoint.

9. The sintering endpoint closed-loop control method based on a multi-field fusion network according to claim 1, characterized in that, The steps for generating and issuing control instructions include: Select high-quality production periods and extract actual sintering endpoints to form a sample set; The target bellows number is obtained by performing an arithmetic mean operation and rounding on the actual sintering endpoints in the sample set. Based on the difference between the predicted sintering endpoint and the target wind box number, the longitudinal endpoint deviation is calculated. ; Calculate the horizontal partition deviation based on the maximum and minimum temperature values ​​of the horizontal interval. ; Set longitudinal deviation threshold Lateral deviation threshold ; Only when or This triggers the adjustment mechanism; For the aforementioned adjustment mechanism, sample pairs are extracted, gain coefficients are obtained through least squares linear regression, and feedforward pre-adjustment amounts are calculated, including feedforward speed pre-adjustment amounts and feedforward opening pre-adjustment amounts. Using a PID controller, calculate the feedback correction, including the feedback machine speed correction and the feedback opening correction; Control commands are generated by superimposing the feedforward pre-adjustment amount and the feedback correction amount.

10. A sintering endpoint closed-loop control system based on a multi-field fusion network, used to implement the sintering endpoint closed-loop control method based on a multi-field fusion network as described in any one of claims 1-9, characterized in that, include: Temperature acquisition module, positioning module, deviation correction module; The temperature acquisition module is used to number the air boxes deployed under the sintering machine trolley, select the end observation section, arrange the temperature measurement matrix, and collect the temperature and related process parameters of each temperature measurement point in the temperature measurement matrix in real time. The data is then preprocessed and stored in the time series database. The positioning module is used to fit the temperature of the transverse temperature measurement point of each wind box, locate the coarse sintering endpoint by positioning the curve inflection point and weighted fusion, construct a spatiotemporal temperature field for the endpoint observation segment, and generate the predicted sintering endpoint by endpoint prediction. The deviation correction module is used to set the target windbox number, compare the deviation with the predicted sintering endpoint, generate two-dimensional deviation parameters, and generate control commands by using feedforward prediction and feedback closed loop.