Dynamic modeling and prediction method of industrial big data combined with digital twinning

Abstract

This application relates to the field of industrial data processing and intelligent control technology, and discloses a method for dynamic modeling and prediction of industrial big data combined with digital twins. The method includes: acquiring operating status data, process execution data, control command records and prediction results, and constructing a sequence of control action events; segmenting the operating status data according to the sequence of control action events to form natural operating data segments and intervention operating data segments, extracting distribution characteristics, fluctuation characteristics and response characteristics, and generating a segmented behavioral feature set; calculating the control reflexive influence quantity and generating a reflexive drift marker, dividing the natural evolution dataset and the control feedback dataset, and predicting them separately; generating control risk constraint quantity, screening, delaying or limiting the control commands to be executed, and iteratively correcting the dataset according to the execution results; this method can distinguish between control feedback changes and natural evolution changes, reduce drift misjudgments, and suppress multi-round oscillations and parameter fluctuations.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial data processing and intelligent control, more specifically, the present application relates to an industrial big data dynamic modeling and prediction method combined with digital twinning. BACKGROUND

[0002] In the digital management and control scene of industrial production process, it is usually necessary to continuously collect running state data, process execution data and control execution data, map the current state of the industrial object using a digital twin model, predict future running trends, and generate control instructions based on the prediction results, thereby forming a closed-loop processing process combining state perception, prediction analysis and control execution.

[0003] In the prior art, there are technical solutions for applying digital twinning to industrial automation intelligent management and control. For example, CN121559984A discloses an automation intelligent management and control method for a factory and related devices. The method obtains real-time working condition data sets by collecting multi-source heterogeneous data from multiple information systems in real time and performing data fusion; obtains a digital twin model instance by synchronizing the state of a pre-constructed digital twin model based on the real-time working condition data sets; obtains an optimal control sequence by deducing and optimizing multiple control operations within a future preset time window based on the digital twin model instance; parses the optimal control sequence into a device control instruction set and executes it; then calculates the residual error between the actual running data and the predicted state trajectory according to the preset period, and performs adaptive correction on the digital twin model when the residual error meets the correction condition.

[0004] The above technical solution can realize the participation of digital twinning in the industrial control closed loop, and the processing focus is mainly on multi-source data fusion, state synchronization, control deduction, optimization and issuance, and model correction based on residual error; however, in the scene where digital twinning has directly participated in control tasks such as parameter adjustment, start and stop, load distribution, process switching, etc., once the control instructions are actually executed, the state quantity values, fluctuation modes and state response relationships in the subsequent time period will be changed; at this time, the running state data collected subsequently no longer only represents the natural evolution process of the industrial object itself, but also superimposes the feedback influence brought by the control action.

[0005] However, existing technologies typically incorporate newly acquired data after control execution directly into subsequent modeling samples for further state identification, trend prediction, or model correction. They lack a mechanism to distinguish the source of feedback data after control execution from naturally evolving data and to conduct subsequent prediction and control processing accordingly. Consequently, in subsequent closed-loop processing, data changes caused by control actions are easily misidentified as equipment malfunctions, operating condition drift, or model inaccuracies. This leads to subsequent predictions, control recommendations, and model corrections continuing to be based on mixed samples, resulting in a decreased correlation between control results and the actual operating state of the object.

[0006] For example, in scenarios involving energy consumption optimization, quality correction, or cycle time adjustment, the digital twin generates suggestions for air distribution adjustment, feeding speed correction, or cooling intensity adjustment based on the prediction results. After the relevant control commands are executed, local state variables may change significantly in a short period of time. If the subsequent processing does not distinguish whether this change originates from the industrial object's own operational changes or from the continuous effect of preceding control actions, and instead treats it uniformly as a new operating condition sample for subsequent prediction or model correction, a cyclical process may easily occur in the closed-loop operation where the prediction results affect the control actions, the control actions then change the subsequent data, and the subsequent data in turn affect the prediction results. This can lead to increased fluctuations in process parameters, repeated adjustments to control results, and decreased operational stability.

[0007] Therefore, existing technologies have at least the following problems: in the process of digital twins participating in industrial closed-loop control, there is a lack of processing methods to distinguish the sources of feedback data after control execution and natural evolution data, predict them separately, and further constrain subsequent control actions. This makes it easy for control feedback to be mixed into subsequent modeling, prediction, and control processing, making it difficult to ensure the pertinence, consistency, and stability of subsequent dynamic modeling and prediction processing.

[0008] Therefore, a dynamic modeling and prediction method for industrial big data that combines digital twins is needed to distinguish between the impact of control feedback and natural evolution information in the processing of operational data after control intervention, and to improve subsequent predictive analysis and control execution processes based on this. Summary of the Invention

[0009] To overcome the shortcomings of existing technologies in the process of digital twin participation in industrial closed-loop control, which lack the technical means to distinguish the source of feedback data after control execution from natural evolution data and to carry out subsequent prediction and control constraint processing accordingly, leading to the easy mixing of control feedback influences in subsequent modeling, prediction, and control processing, and making it difficult to ensure the pertinence, consistency, and stability of dynamic modeling and prediction processing, this application provides the following technical solution: This application discloses a method for dynamic modeling and prediction of industrial big data combining digital twins, characterized by including: Step S1: Obtain the operating status data, process execution data, control command records and prediction result time series table, and construct the control action event sequence according to the time correspondence; Step S2: The operational status data is segmented according to the sequence of control action events to form natural operational data segments and intervention operational data segments, and distribution characteristics, fluctuation characteristics and response characteristics are extracted to generate a segmented behavioral feature set; Step S3: Based on the segmented behavior feature set and prediction results, calculate the control reflexivity effect and generate a reflexivity drift label table according to the control reflexivity effect. Step S4: Based on the reflexive drift label table, the intervention operation data segment is divided into a natural evolution dataset and a control feedback dataset, and state prediction processing and feedback response prediction processing are performed respectively to generate natural evolution prediction results and control feedback prediction results. Step S5: Based on the natural evolution prediction results, control feedback prediction results, and control action event sequence, generate control risk constraints, and then filter, delay, or limit the control instructions to be executed based on the control risk constraints to generate constraint control instructions. Step S6: Write the execution results of the constraint control command into the control action event sequence, and update the segmented behavior feature set by combining the control reflexive influence quantity and the reflexive drift label table, and iteratively correct the natural evolution dataset and the control feedback dataset.

[0010] Compared with related technologies, this application has the following advantages: Compared with existing technologies, this application addresses the processing interruption point where the post-control execution operational data simultaneously contains both natural evolution information and control feedback information. It adds a continuous processing chain encompassing control action identification, feedback impact determination, data diversion prediction, and pre-control constraints. This improves the stability and consistency of subsequent dynamic modeling and prediction processing. Specifically: First, this application constructs a sequence of control action events by acquiring operational status data, process execution data, control command records, and prediction results. Based on this sequence, the operational status data is segmented into natural operational data segments and intervention operational data segments. Then, distribution characteristics, fluctuation characteristics, and response characteristics are extracted to generate a segmented behavioral feature set. This approach distinguishes the data segments affected by control actions from those unaffected by control actions based on time and interaction, providing a foundation for identifying the source of data changes. Compared to existing technologies that directly incorporate data after control execution into subsequent samples, this application improves the distinguishability of data sources in subsequent processing and reduces misjudgments caused by the mixing of control feedback data and natural evolution data.

[0011] Secondly, this application calculates the control reflexive influence based on the segmented behavioral feature set and prediction results, and generates a reflexive drift label table. Then, based on the reflexive drift label table, the intervention operation data segment is divided into a natural evolution dataset and a control feedback dataset, and state prediction processing and feedback response prediction processing are performed separately to generate natural evolution prediction results and control feedback prediction results. This allows for the separate processing of the industrial object's own state change trend and the feedback response trend caused by the control action, so that the subsequent prediction results correspond to data changes from different sources. As a result, the misidentification of data changes caused by control actions as equipment abnormalities, operating condition drift, or model inaccuracies can be reduced, and the ability of subsequent predictive analysis to distinguish the source of change can be improved.

[0012] Furthermore, this application generates control risk constraints based on natural evolution prediction results, control feedback prediction results, and control action event sequences. Based on these constraints, it performs screening, delay, or amplitude limiting on the control instructions to be executed, generating constrained control instructions. This allows for constraint processing of control actions that may cause feedback superposition, directional conflict, or deviation amplification, taking into account natural evolution trends, feedback response trends, and continuous action situations, before the control instructions enter the execution process. Compared to existing technologies that primarily perform model correction based on residuals after control execution, this application moves the processing position forward to before control execution, thus helping to reduce the continuous amplification risk under multiple rounds of closed-loop action and improving the stability of the control execution process.

[0013] Finally, this application also writes the execution results of constraint control instructions into the control action event sequence, and updates the segmented behavior feature set by combining the control reflexive influence quantity and reflexive drift marker, iteratively correcting the natural evolution dataset and control feedback dataset; this ensures that the data organization method, prediction basis, and control processing basis in subsequent cycles can maintain a continuous correspondence with the actual execution results, thus forming a complete scheme of: control action identification - data source differentiation - separate prediction - control constraints - post-execution iterative correction; thereby, it is beneficial to improve the pertinence of data organization in the dynamic modeling process of industrial big data, and enhance the consistency and operational reliability between subsequent prediction results and control processing.

[0014] In summary, this application introduces control action event identification, control reflexive effect determination, natural evolution data and control feedback data diversion prediction, and pre-control constraint processing into the processing of digital twins involved in industrial closed-loop control. This prevents the subsequent dynamic modeling and prediction processes from mixing the data changes caused by control actions with the natural evolution changes of industrial objects, thereby improving the data foundation upon which subsequent predictive analysis and control execution depend, and enhancing the stability, consistency, and reliability of the closed-loop operation process. Attached Figure Description

[0015] Figure 1A flowchart illustrating the steps of the industrial big data dynamic modeling and prediction method combining digital twins provided in this application; Figure 2 A schematic diagram illustrating the calculation of control reflexivity effects, generation of reflexivity drift label tables, and partitioning of the natural evolution dataset and control feedback dataset provided in this application; Figure 3 A schematic diagram illustrating the generation of control risk constraint quantities, processing of constraint control instructions, and iterative correction provided in this application. Detailed Implementation

[0016] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0017] Please see Figure 1 As shown, this embodiment provides a method for dynamic modeling and prediction of industrial big data combined with digital twins, including: Step S1 involves acquiring operating status data, process execution data, control command records, and prediction results, and constructing a control action event sequence according to the time correspondence. The purpose is to unify the state change process, process execution process, control action execution process, and prediction output process of industrial equipment during operation under the same time reference, forming a traceable control action event sequence, so as to identify the source of the influence of control actions on the data distribution evolution based on the control action event sequence in subsequent steps.

[0018] In practice, the operational status data generated by the industrial equipment during a continuous acquisition cycle is first read. This operational status data is a set of status data recorded sequentially by acquisition time. Each piece of operational status data includes an acquisition time stamp, a data source stamp, and a status value. Then, the operational status data is sorted according to the acquisition time stamp, and data records with missing times are deleted. For data records with duplicate times, one record is retained according to a preset retention rule, resulting in an operational status time series table arranged chronologically. Next, the process execution data for the time period corresponding to the operational status data is read. This process execution data characterizes the set values, adjustment values, and execution periods of various process parameters during the production process of the industrial equipment. Each piece of process execution data... Record the start time, end time, process parameter name, and process parameter value, and convert the process execution data into a process execution time series table corresponding to the operating status time series table. For process execution data that is continuously effective across multiple acquisition times, write the corresponding process parameter values ​​to each covered time position to obtain the process execution mapping result corresponding to each acquisition time. After obtaining the operating status time series table and the process execution time series table, read the control instruction records within the corresponding time period. The control instruction records are used to represent the action data set of the prediction result entering the control execution process. Each control instruction record includes at least the instruction generation time, instruction issuance time, instruction execution object, instruction adjustment direction, instruction action parameters, and instruction target value. Then, according to... The control command records are sorted by command issuance time and categorized according to the command execution object to obtain a control command timing table. For control command records lacking command generation or issuance time, the corresponding time can be supplemented based on the preceding and following related records in the control log. Records that fail to be supplemented are marked as invalid and removed from the control command timing table. Further, the prediction results corresponding to the control command records are read. The prediction results are the state assessment results or adjustment suggestion results output by the digital twin before the control command is generated. Each prediction result includes at least the prediction output time, prediction object, prediction variable, prediction value, and associated control object. The prediction results are sorted according to the prediction output time to form a prediction result timing table. The table is then used to match the associated control objects in the prediction result timing table with the instruction execution objects in the control instruction timing table. Successfully matched prediction results are paired with control instruction records to obtain a prediction-control pairing table. After obtaining the operating status timing table, process execution timing table, control instruction timing table, and prediction-control pairing table, the execution time is correspondingly processed. Specifically, using a unified time base as the alignment basis, operating status data, process execution data, control instruction records, and prediction results at the same time position are written into the same corresponding record. For control instruction records, the instruction issuance time is taken as the action start time, and the time when the parameter corresponding to the instruction execution object in the process execution data begins to change is taken as the action response time.When the time difference between the action response time and the command issuance time is within a preset allowable range, the interval corresponding to this time difference is determined as the control action identification interval. The preset allowable range can be set based on the correspondence between the equipment sampling period, the actuator response period, and the process parameter change period, used to limit the time correspondence between the control action and the process response. Finally, a control action event sequence is constructed based on the control action identification interval. Specifically, for each control command record corresponding to the completion time, the predicted output time, command issuance time, action response time, execution object, action parameters, action direction, and corresponding process execution interval are extracted to generate a control action event record. All control action event records are then sorted according to the command issuance time to form a control action event sequence. Each control action event record in the sequence represents the continuous action process of a prediction-driven control action from output to execution and then to process response. The constructed control action event sequence serves as the data input for step S2, used for subsequent segmented processing of the operating status data based on the control action event sequence.

[0019] In one embodiment, the industrial equipment is a continuous heat treatment device. The data acquisition cycle for the operating status is 10 seconds, and the acquisition period is from 08:00:00 to 08:10:00. The operating status data corresponding to 08:02:00 is: furnace temperature 865°C, feeding speed 1.20, motor current 48 kW, and outlet temperature 842°C. Process execution data is read according to the same time base. From 08:02:10, the target air distribution value is adjusted from 32 to 35, and from 08:02:20, the target feeding speed is adjusted from 1.20 to 1.15. Further, control command records are read. The command generation time for the combustion air distribution mechanism is 08:02:05, the command issuance time is 08:02:08, the command action parameter is air distribution opening, and the command target value is 35. The command generation time for the feeding mechanism is 08:02:16, the command issuance time is 08:02:18, and the command action parameter is feeding speed. The degree and the target value of the instruction are 1.15. Then, the prediction result time series table is read, where the associated control object corresponding to 08:02:05 is the combustion air distribution mechanism, and the associated control object corresponding to 08:02:16 is the feeding mechanism, and a prediction control pairing table is formed. Then, the time correspondence processing is performed with a 10-second acquisition cycle as a unified time base. The action response time corresponding to the combustion air distribution mechanism is found to be 08:02:10, and the action response time corresponding to the feeding mechanism is 08:02:20. Assuming the preset allowable range is 0 seconds to 30 seconds, the time difference between the two control instructions is 2 seconds. Therefore, 08:02:08 to 08:02:10 and 08:02:18 to 08:02:20 are respectively determined as the control action identification interval. Finally, two control action event records are constructed based on the control action identification interval and sorted according to the instruction issuance time to obtain the control action event sequence, which is used as the data input for step S2.

[0020] Step S2 involves segmenting the operational status data based on the control action event sequence to form natural operational data segments and intervention operational data segments. Distribution characteristics, fluctuation characteristics, and response characteristics are then extracted to generate a segmented behavioral feature set. The purpose of forming natural operational data segments and intervention operational data segments is to use the control action event sequence obtained in step S1 to distinguish between the natural evolution intervals unaffected by control actions and the feedback evolution intervals affected by control actions in the operational status data. This provides corresponding data carriers for subsequent extraction of distribution characteristics, fluctuation characteristics, and response characteristics, avoiding the mixing of behavioral changes caused by control actions with changes in the equipment's own operation.

[0021] In specific implementation, firstly, the operating status timing table and control action event sequence obtained in step S1 are read. Each control action event record in the control action event sequence includes the predicted output time, command issuance time, action response time, execution object, action parameters, action direction, and corresponding process execution interval. Then, the acquisition time marker in the operating status timing table is used as the segmented reference time axis. Subsequently, for each control action event record in the control action event sequence, the command issuance time and the start time of the corresponding process execution interval are extracted. The continuous time interval between the command issuance time and the end time of the corresponding process execution interval is marked as the candidate intervention interval. The corresponding process execution interval... The end time is taken from the time of instruction issuance of the next control event record that acts on the same execution object and has the same action parameter, or the time position when the corresponding action parameter in the process execution data is restored to the value before adjustment, so that each control event record corresponds to a traceable influence interval; then, interval merging processing is performed on all candidate intervention intervals. For candidate intervention intervals that intersect in time or are connected end to end, they are merged into a continuous intervention interval; for candidate intervention intervals that correspond to different execution objects but overlap in time, merging processing is also performed, and the source mark of each execution object is retained in the merged continuous intervention interval to obtain the intervention interval set; Then, the operational status data falling into the intervention interval set in the operational status time sequence table are extracted and formed into intervention operational data segments according to time continuity. These intervention operational data segments characterize operational segments affected by the controlled event sequence. The operational status data not falling into the intervention interval set in the operational status time sequence table are extracted and formed into natural operational data segments according to time continuity. These natural operational data segments characterize operational segments not affected by the controlled event sequence. Furthermore, to avoid execution propagation stagnation at segment boundaries, boundary transition intervals are set before the start position and after the end position of each intervention interval. The length of the boundary transition interval depends on the sampling period and the duration of the actuator's action. The correspondence between time and process parameter change duration is set; the boundary markers of the operating status data falling into the boundary transition range are recorded separately and are not included in the statistical subjects when forming natural operating data segments and intervention operating data segments, so as to reduce the impact of segment boundaries on subsequent feature extraction results; finally, all natural operating data segments and intervention operating data segments are numbered in chronological order, and the start time, end time, number of sampling points, source marker and associated control event record of each segment are written into the segment index table, and the natural operating data segments, intervention operating data segments and segment index table are output for the next step to extract distribution features, fluctuation features and response features based on natural operating data segments and intervention operating data segments.

[0022] Based on the embodiment of step S1, the data collection period for the running status timing table continues to be from 08:00:00 to 08:10:00, with a collection cycle of 10 seconds. Two control action event records are used, where the process execution interval corresponding to the first control action event record is from 08:02:10 to 08:04:50, and the process execution interval corresponding to the second control action event record is from 08:02:20 to 08:05:00. The data collection time marker is used as the dividing point. Based on the baseline timeline, the period from 08:02:08 to 08:04:50 is first marked as the first candidate intervention interval, and the period from 08:02:18 to 08:05:00 is marked as the second candidate intervention interval. Since the two times intersect, they are merged into a continuous intervention interval from 08:02:08 to 08:05:00, while retaining the source markers of the combustion air distribution mechanism and the feeding mechanism, resulting in a set of intervention intervals containing only one intervention interval. Subsequently, the operating status data falling within this intervention interval is extracted. An intervention operation data segment was formed, consisting of 18 sampling points. The operation status data not falling within this intervention interval was then divided into two natural operation data segments: the first natural operation data segment is from 08:00:00 to 08:02:00, with 13 sampling points; the second natural operation data segment is from 08:05:10 to 08:10:00, with 30 sampling points. Further, a front boundary transition interval of 08:01:50 to 08:02:00 and a rear boundary transition interval of 08:05 were established. From 08:05:20 to 08:10, the boundary markers of the four sampling points falling within the range are recorded separately and are not included in the statistical subjects. Finally, the first natural running data segment is numbered as segment 1, the intervention running data segment is numbered as segment 2, and the second natural running data segment is numbered as segment 3 in chronological order. The start time, end time, number of sampling points, source markers, and associated control events of each segment are written into the segment index table to output the natural running data segment, the intervention running data segment, and the segment index table.

[0023] The purpose of extracting distribution features, fluctuation features, and response features to generate a segmented behavioral feature set is to extract features from natural operation data segments and intervention operation data segments that can characterize the data numerical composition, time fluctuation patterns, and control action response relationships, and to generate a segmented behavioral feature set based on these features, so as to identify the control reflexive influence quantity and reflexive drift label table in subsequent steps.

[0024] In specific implementation, the natural operation data segment, the intervention operation data segment, and the segment index table are first read, and then processed segment by segment according to the segment number. For each natural operation data segment and each intervention operation data segment, the numerical sequence of each state variable within the corresponding time interval is extracted, and the distribution characteristics are calculated. The distribution characteristics include at least the starting value, ending value, mean, median, maximum value, minimum value, value range, and quantile interval width. The quantile interval width is obtained by sorting the numerical sequence according to its size and calculating the difference between the value corresponding to the preset upper quantile position and the value corresponding to the preset lower quantile position. The preset upper and lower quantile positions are designed to cover the main distribution of the sequence and exclude a small number of edge points. The numerical distribution pattern of the corresponding data segment is characterized. Next, the fluctuation characteristics of the numerical sequences of each state variable within the same data segment are calculated. These fluctuation characteristics include at least the sequence of differences between adjacent sampling points, the mean of the absolute values ​​of the differences between adjacent sampling points, the cumulative change per unit time, the length of continuous same-direction changes, and the number of local reverse changes. The sequence of differences between adjacent sampling points is obtained by subtracting the value of the previous sampling point from the current sampling point value. The cumulative change per unit time is obtained by summing the absolute values ​​of the differences between all adjacent sampling points within the corresponding data segment and dividing by the duration of the data segment. The length of continuous same-direction changes is obtained by counting the number of sampling points with consecutive differences of the same sign. The number of local reverse changes is obtained by counting the number of times the signs of adjacent differences change, thus characterizing the distribution pattern of the state variables within the same data segment. The data segment's fluctuation process is analyzed; subsequently, for each intervention operation data segment, response features are extracted by combining the associated control action event records recorded in the segment index table; among them, the response features include at least the response start offset, response duration, response direction consistency, response quantity change amplitude, and response recovery time; the response start offset is obtained by subtracting the action response time in the associated control action event record from the time position where the target state quantity first shows a continuous change within the intervention operation data segment; the response duration is obtained by the length of time the target state quantity changes continuously along the control action direction; the response direction consistency is obtained by comparing whether the action direction in the control action event record is consistent with the target state quantity change direction; and the response quantity change amplitude is... The amplitude is obtained by the difference between the value at the end of the response and the value at the beginning of the response. The response recovery time is obtained by the length of time it takes for the target state quantity to return to the interval near the beginning of the response. The idea behind setting the interval near the beginning of the response is to construct an allowable deviation range around the beginning of the response to identify the position where the state quantity returns from the controlled change to the original fluctuation band. For natural running data segments, since there are no records of associated control action events, response features are not extracted, but a no-response mark is written into the segment index table. Finally, the distribution features and fluctuation features corresponding to each natural running data segment are uniformly encoded with the distribution features, fluctuation features, and response features corresponding to each intervention running data segment to form a segmented behavioral feature record.Next, all segmented behavioral feature records are summarized according to segment number to generate a segmented behavioral feature set. Each record in the segmented behavioral feature set includes a segment number, segment type, time range, state variable name, distribution characteristics, fluctuation characteristics, response characteristics, or a no-response flag. The segmented behavioral feature set is then output to step S3 for subsequent calculation of the control reflexive effect.

[0025] After clarifying the above feature extraction rules, the following will further explain the process of extracting distribution features, fluctuation features, and response features and generating segmented behavior feature sets, based on the aforementioned continuous heat treatment equipment, segmented index table, and control action event record.

[0026] Based on the aforementioned segmented results, the data is further divided into three segments: Segment 1 (08:00:00 to 08:02:00), Segment 2 (08:02:08 to 08:05:00), and Segment 3 (08:05:10 to 08:10:00), with furnace temperature used as the state variable. The furnace temperature sequence for Segment 1 is 860, 861, 862, 861, 863, 864, 863, 864, 865, 864, 865, 866, 865. The furnace temperature sequence for Segment 2 is... The sequence of furnace temperature values ​​in the third segment is 865, 867, 869, 871, 872, 873, 873, 872, 871, 870, 869, 868, 867, 866, 866, 865, 865, 864. The values ​​are: 864, 864, 865, 865, 864, 864, 865, 865, 864, 864, 865, 865, 864, 865, 865, 864, 865, 865, 864, 865, 865, 864, 865, 865, 864, 865, 865, 864, 865, 865, 864, 864. First, follow... The data is processed segment by segment, numbered sequentially. For the first segment, the distribution characteristics are calculated, yielding the following values: initial value 860, ending value 865, mean 863.31, median 864, maximum 866, minimum 860, range 6, and quantile interval width 3. The quantile interval width is obtained by subtracting the value near the bottom of the sequence from the value near the top (865) after sorting. The fluctuation characteristics are then calculated, resulting in the sequence of differences between adjacent sampling points: 1, 1, -1, 2, 1, -1, 1, 1, -1, 1, 1, -1. The mean absolute value of the differences between adjacent sampling points is 1.08, and the cumulative change per unit time is 0. .108, the length of continuous unidirectional change is 2, the number of local reverse changes is 7, and a no-response mark is written in the segment index table. For the second segment, the distribution characteristics are calculated, and the following values ​​are obtained: starting value 865, ending value 864, mean 868.94, median 869, maximum value 873, minimum value 864, value range 9, quantile interval width 5. Then, the fluctuation characteristics are calculated, and the difference sequence between adjacent sampling points is 2, 2, 2, 1, 1, 0, -1, -1, -1, -1, -1, -1, -1, 0, -1, 0, -1, -1. The mean of the absolute value of the difference between adjacent sampling points is 1.06, and the cumulative change per unit time is 0.106. The length of continuous unidirectional change is 7, and the number of local reverse changes is 1. Then, the response features are extracted by combining the two associated control action event records recorded in the segmented index table. For the control action event record corresponding to the combustion air distribution mechanism, the action response time is 08:02:10. The furnace temperature in the second segment starts to rise continuously from 08:02:20, so the response starting point offset is 10 seconds. It changes continuously in the increasing direction until 08:03:00, so the response duration is 40 seconds. The consistency of the response direction is recorded as consistent. The response quantity change amplitude is 873 minus 865 equals 8. The response recovery time is calculated as 150 seconds based on the furnace temperature returning to the initial value of 865 around 08:04:40. The control event record for the feeding mechanism shows an action response time of 08:02:20. Since the furnace temperature in the second segment began to decrease continuously after 08:03:10, the corresponding response start offset was 50 seconds, continuously changing in the decreasing direction until 08:05:00. Therefore, the response duration was 110 seconds. The response direction consistency was recorded as consistent. The response quantity change amplitude was 864 minus 872, equal to -8. The response recovery time was recorded as not recovered because it did not return to the vicinity of 872 before the end of the second segment. For the third segment, the distribution characteristics were calculated, yielding the following values: initial value 864, ending value 864, mean 864.47, median 864, maximum value 865, minimum value 864, value range 1, and quantile interval width. 1. Next, calculate the fluctuation characteristics to obtain the alternating sequence of differences between adjacent sampling points, which are 0 and 1 or 0 and -1. The mean of the absolute values ​​of the differences between adjacent sampling points is 0.48, the cumulative change per unit time is 0.048, the length of continuous same-direction change is 1, and the number of local reverse changes is 14. Write a no-response mark in the segment index table. Finally, uniformly encode the distribution characteristics and fluctuation characteristics corresponding to the first segment, the distribution characteristics, fluctuation characteristics and response characteristics corresponding to the second segment, and the distribution characteristics and fluctuation characteristics corresponding to the third segment to form the first, second, and third segment behavior feature records, respectively. Then, summarize them according to the segment number to generate a segment behavior feature set, where the first record includes the segment number 1, the segment type natural running data segment, and the time. The first record includes the following parameters: range 08:00:00 to 08:02:00; state variable name: furnace temperature; corresponding distribution characteristics; corresponding fluctuation characteristics; and a no-response marker. The second record includes segment number 2, segment type: intervention operation data segment, time range 08:02:08 to 08:05:00; state variable name: furnace temperature; corresponding distribution characteristics; corresponding fluctuation characteristics; and two sets of response characteristics. The third record includes segment number 3, segment type: natural operation data segment, time range 08:05:10 to 08:10:00; state variable name: furnace temperature; corresponding distribution characteristics; corresponding fluctuation characteristics; and a no-response marker. The segmented behavior feature set is output to step S3 for subsequent calculation of the control reflexive influence quantity.

[0027] Step S3: Based on the segmented behavioral feature set and prediction results, calculate the control reflexivity influence quantity, and generate a reflexivity drift label table according to the control reflexivity influence quantity.

[0028] In this embodiment, based on the segmented behavior feature set and the prediction results, the control reflexivity influence is calculated, and a reflexivity drift label table is generated according to the control reflexivity influence. The purpose is to establish a correspondence between the segmented behavior feature set obtained in step S2 and the prediction results obtained in step S1, identify the degree of feedback influence of the prediction results on the distribution, fluctuation process, and response process of the operating state data after the control action, and distinguish the distribution changes caused by control closed-loop feedback from the distribution changes caused by equipment abnormalities, operating condition switching, or process disturbances, so as to perform hierarchical modeling of the data in subsequent steps based on the reflexivity drift label table. In specific implementation, the segmented behavior feature set output in step S2 and the output in step S1 are read first. The prediction result time series table is generated, wherein each segmented behavioral feature record in the segmented behavioral feature set includes at least the segment number, segment type, time range, state variable name, distribution characteristics, fluctuation characteristics, response characteristics, or no-response marker. Each prediction result in the prediction result time series table includes at least the prediction output time, prediction object, prediction variable, prediction value, and associated control object. Subsequently, based on the prediction output time, associated control object, and associated control action event records in the segment index table, the prediction results are paired with the corresponding intervention operation data segments to form a prediction segment correspondence table. Each record in the prediction segment correspondence table is used to characterize the time series correspondence between a prediction result and an intervention operation data segment. Next, for each record in the prediction segment correspondence table, the distribution characteristics, fluctuation characteristics, and response characteristics of the corresponding intervention operation data segment are extracted, and the distribution characteristics and fluctuation characteristics of the natural operation data segments that are temporally adjacent to the intervention operation data segment are extracted to form a comparison feature group. The comparison feature group is used to characterize the behavioral differences of the same state variable between the period affected by the control action and the period unaffected by the control action. Subsequently, based on the comparison feature group, the distribution offset, fluctuation offset, and response coupling amount are calculated. The distribution offset is obtained by summing the differences between the mean, median, range, and quantile interval width of the intervention operation data segment and the corresponding values ​​of the natural operation data segment. The fluctuation offset is obtained by summing the differences between the mean, cumulative change per unit time, length of continuous same-direction change, and number of local reverse changes of the absolute values ​​of the differences between adjacent sampling points of the intervention operation data segment and the corresponding values ​​of the natural operation data segment. The response coupling amount is obtained by summing the response starting point offset, response duration, response direction consistency, response quantity change amplitude, and response recovery time after unified conversion. The idea behind the unified conversion is to convert features with different dimensions into the same comparison scale so as to perform combined calculations on various features.

[0029] Further, prediction trend information is extracted based on the prediction variables and predicted values ​​in the prediction results, and the prediction trend information is compared with the consistency of the response direction and the magnitude of the response change in the response characteristics to obtain the prediction response consistency quantity. The prediction response consistency quantity characterizes the degree of correspondence between the prediction results and the actual behavior changes after the control action. When the prediction trend information is consistent with the response direction and the magnitude of the response change expands along the prediction trend direction, it is recorded as a positive correspondence; when the prediction trend information is inconsistent with the response direction or the magnitude of the response change does not expand along the prediction trend direction, it is recorded as a negative correspondence. In one embodiment, the prediction response... The consistency value should be determined based on at least three factors: the prediction direction, the response direction, and the response change amplitude. Specifically, when the prediction direction and response direction are consistent, and the response change amplitude is within a preset effective change range, it is recorded as positive consistency. When the prediction direction and response direction are opposite, or the response change amplitude is below the preset effective change range, it is recorded as negative consistency or weak consistency. For cases where the prediction direction is stable, whether the response change amplitude exceeds the natural fluctuation band is used as the consistency criterion. The natural fluctuation band is determined based on the fluctuation characteristics of adjacent natural running data segments. By introducing additional judgment rules for stable scenarios, the lack of consistency judgment due to relying solely on the direction field can be avoided.

[0030] Then, the control reflexive influence is calculated based on the distribution offset, fluctuation offset, response coupling, and predicted response consistency. The control reflexive influence is obtained by weighting the distribution offset, fluctuation offset, and response coupling, and then correcting the direction based on the predicted response consistency. The weighting is based on the weight of the state variable in the process, the influence of the controlled object on the production line cycle time, and the process sensitivity of the predicted variable. This allows the control reflexive influence to characterize the comprehensive impact of the prediction result on subsequent operation after control action feedback. In one embodiment, to ensure the repeatability of the control reflexive influence, the distribution offset, fluctuation offset, and response coupling are first scaled to obtain the distribution influence score, fluctuation influence score, and response influence score. The scaling process can employ interval mapping based on the statistical range of historical natural operation data segments, proportional conversion based on a reference baseline, or quantile mapping based on the historical sample distribution of the same state variable. Subsequently, based on the importance of the state variable, the sensitivity of the execution object, and the process correlation of the predictor variable, the distribution impact score, fluctuation impact score, and response impact score are weighted and synthesized to obtain the basic impact score. Then, the basic impact score is adjusted according to the predictive response consistency. When the predicted trend is consistent with the actual response direction, the basic impact score is maintained or enhanced; when the predicted trend is inconsistent with the actual response direction, the basic impact score is reduced to form a control reflexive impact. Through the above processing, different dimensional characteristics participate in the calculation on the same comparison scale, reducing the impact of implementer differences on the stability of the results.

[0031] Next, the control reflexive influence quantity corresponding to each prediction result is written into the control reflexive influence quantity table, and the prediction output time, segment number, state variable name, associated control object, and corresponding feature source are recorded synchronously in the control reflexive influence quantity table. Further, the control reflexive influence quantity table is read, and the control reflexive influence quantities are categorized according to the state variable name and segment number. The control reflexive influence quantities of the same state variable in multiple intervention operation data segments are sorted to form an influence quantity sequence. The segment judgment interval is determined based on the influence quantity sequence. The idea behind setting the segment judgment interval is to use the concentrated distribution interval of control reflexive influence quantities in historical intervention operation data segments as a reference, and to reserve an offset band outside the concentrated distribution interval to distinguish between control closed-loop feedback sources and non-control closed-loop feedback sources. Subsequently, the control reflexive influence quantity corresponding to the current intervention operation data segment is compared with the segment judgment interval. When the control reflexive influence quantity falls into the concentrated distribution interval and the corresponding predicted response consistency quantity is positively correlated, the corresponding intervention operation data segment is marked as a control feedback drift segment. When the control reflexive influence quantity deviates from the concentrated distribution interval, or the predicted response consistency quantity is negatively correlated, the corresponding intervention operation data segment is marked as a non-control feedback drift segment. When the same intervention operation data segment includes control reflexive influence quantities corresponding to multiple state quantities, the summation is performed according to the weight of the state quantity in the process to obtain the segment-level judgment result of the corresponding intervention operation data segment. Next, the segment-level determination results are verified against the source markers and associated control action event records in the segment index table. When the segment-level determination result is a control feedback drift segment, and there is an execution object in the source marker that is consistent with the control object associated with the prediction result, the control feedback drift segment marker is retained. When the segment-level determination result is a non-control feedback drift segment, or the source marker is inconsistent with the control object associated with the prediction result, a non-control feedback drift segment marker is written. For natural running data segments, since there are no response markers in the segment behavior feature records, the natural evolution segment marker is directly written. Finally, a reflexive drift marker table is generated according to the segment number, time range, state quantity name, control reflexive influence quantity value, prediction response consistency quantity result, and segment-level determination result. Each record in the reflexive drift marker table is used to characterize the source of behavioral change in the corresponding data segment. The reflexive drift marker table is output to step S4 so that the intervention running data segment can be divided into natural evolution dataset and control feedback dataset according to the reflexive drift markers.

[0032] It should be noted that the reflexive drift marker table is a collection of reflexive drift marker records summarized by segment number, where each record constitutes a reflexive drift marker.

[0033] After clarifying the above-mentioned rules for calculating the reflexive influence and generating the reflexive drift marker table, the following further explains the implementation process of step S3 in conjunction with the aforementioned embodiments of continuous heat treatment equipment, segmented behavior feature sets, prediction result time series tables, and segmented index tables. Based on the aforementioned embodiments, the intervention operation data segment from 08:02:08 to 08:05:00 in the second segment is continued as the analysis object, and two prediction results are used. The first prediction result has a prediction output time of 08:02:05, a predicted value that is too low, and the associated control object is the combustion air distribution mechanism. The second prediction result has a prediction output time of 08:02:16, a predicted value that is too high, and the associated control object is the feeding mechanism. The first and third segments are also used. The natural operating data segment of the segment is used, and the calculation process of the control reflexive influence is explained with furnace temperature as the state variable. First, a prediction segment correspondence table is established so that both prediction results are paired with the second segment. Then, the distribution characteristics, fluctuation characteristics, and response characteristics corresponding to the furnace temperature in the second segment are extracted, and the distribution characteristics and fluctuation characteristics corresponding to the first and third segments are also extracted. The mean values ​​of the first and third segments are taken as natural reference values. The distribution characteristics of the second segment are mean 868.94, median 869, range 9, and quantile interval width 5. The corresponding values ​​of the first segment are 863.31, 864, 6, and 3, and the corresponding values ​​of the third segment are 864.47, 864, 1, and 1. Therefore, the natural reference values ​​are 863.89, 864, 3.5, and 2, respectively. The distribution offset is 17.55. The fluctuation characteristics of the second segment are: mean absolute value of the difference between adjacent sampling points 1.06, cumulative change per unit time 0.106, continuous same-direction change length 7, and number of local reverse changes 1. The corresponding values ​​for the first segment are 1.08, 0.108, 2, and 7, and the corresponding values ​​for the third segment are 0.48, 0.048, 1, and 14. Therefore, the natural reference values ​​are 0.78, 0.078, 1.5, and 10.5, respectively. Thus, the absolute value of the fluctuation offset is calculated to be 3.692. The response characteristics corresponding to the combustion air distribution mechanism in the second segment are: response start offset 10 seconds, response duration 40 seconds, response direction consistency is consistent, response quantity change amplitude 8, and response recovery time 150 seconds. The response corresponding to the feeding mechanism... The characteristics are: response start-point offset of 50 seconds, response duration of 110 seconds, response direction consistency of "consistent", response quantity change amplitude of -8, and response recovery time recorded as "not recovered". Calculated according to a unified conversion rule, the response coupling quantity corresponding to the combustion air distribution mechanism is 8.83, and the response coupling quantity corresponding to the feeding mechanism is 15.67. Then, based on the predicted values, the predicted trend information is extracted and compared with the response direction consistency and response quantity change amplitude. The predicted response consistency quantities of the two prediction results are both recorded as positively correlated. Subsequently, the control reflexivity influence quantity is calculated as follows: distribution offset multiplied by 0.4, fluctuation offset multiplied by 0.2, response coupling quantity multiplied by 0.4, and then multiplied by the predicted response consistency quantity correction value. The control reflexivity influence quantity corresponding to the first prediction result is 11.3028, the control reflexivity influence quantity corresponding to the second prediction result is 14.0388, and is written into the control reflexivity influence quantity table. Further, influence quantity sequences are formed according to furnace temperature. Assuming the concentrated distribution interval is 9 to 15, both 11.3028 and 14.0388 fall into the concentrated distribution interval, and the predicted response consistency quantities are both positively correlated. Therefore, they are both marked as control feedback drift segments. Since the source marker in the second segment corresponds to the associated controlled object, the segment-level determination result of the second segment is retained as a control feedback drift segment. The first and third segments, due to the presence of no-response markers, are directly written into the natural evolution segment markers. Finally, a reflexivity drift marker table is generated.

[0034] Step S4: Based on the reflexive drift labeling table, the intervention operation data segment is divided into a natural evolution dataset and a control feedback dataset. State prediction processing and feedback response prediction processing are then performed separately to generate natural evolution prediction results and control feedback prediction results. In this embodiment, the purpose of dividing the intervention operation data segment into a natural evolution dataset and a control feedback dataset based on the reflexive drift labeling table, and performing state prediction processing and feedback response prediction processing separately to generate natural evolution prediction results and control feedback prediction results, is to use the reflexive drift labeling table generated in step S3 to distinguish the sources of behavioral changes in the intervention operation data segment, separating the operation segments caused by control loop feedback from those not belonging to control loop feedback, and establishing different prediction input and prediction output relationships for each. This allows the natural evolution prediction results to characterize the state change trend of industrial equipment under conditions unaffected by control loop feedback, and the control feedback prediction results to characterize the state response trend caused after the control action enters the execution process, thereby providing a separated prediction basis for the control risk constraint processing in subsequent steps.

[0035] In specific implementation, firstly, the intervention operation data segment and segment index table output in step S2, and the reflexive drift flag table output in step S3 are read; then, the reflexive drift flags are matched with the segment index table according to the segment number, and the segment-level judgment result corresponding to each intervention operation data segment is extracted; next, the intervention operation data segments whose segment-level judgment result is a control feedback drift segment are extracted, and the corresponding control reflexive influence quantity value, associated control object, associated control action event record and state quantity name are retained to form a control feedback candidate segment set; segments whose segment-level judgment result is not a control feedback drift segment are... The intervention operation data segments are extracted, and the corresponding segment numbers, time ranges, state variable names, and source markers are retained to form a set of natural evolution candidate segments. Subsequently, state variable decomposition processing is performed on the control feedback candidate segment set. When the same intervention operation data segment contains multiple state variables and the reflexive drift markers corresponding to different state variables are inconsistent, the corresponding state variable value sequences are extracted according to the state variable names. The state variable value sequences marked as control feedback drift segments are written into the control feedback candidate segment set, and the state variable value sequences not marked as control feedback drift segments are written into the natural evolution candidate segment set. The dataset is structured into sets to group different state variables within the same time interval according to their source of change. Further, a continuous splicing process is performed on the natural evolution candidate segment set. Candidate segments that are temporally adjacent and have the same state variable name are connected in chronological order to form natural evolution data segments. A source consistency splicing process is performed on the control feedback candidate segment set. Candidate segments with the same associated control object, the same action parameter, and are temporally continuous are connected in chronological order to form control feedback data segments. Then, dataset indexes are established for both the natural evolution data segments and the control feedback data segments. Each index record in the natural evolution dataset includes at least the data segment number, state variable name, time range, and source marker. Each index record in the control feedback dataset includes at least the data segment number, state variable name, time range, associated control object, associated control action event record, and control reflexive effect value. The corresponding state variable value sequence, process execution data, and prediction results are associated and written into the corresponding index record. Finally, the processed natural evolution data segments are aggregated to form a natural evolution dataset, and the processed control feedback data segments are aggregated to form a control feedback dataset.

[0036] Further, the natural evolution dataset and the control feedback dataset are read, and the input sequences are organized according to the data segment number and state variable name. For the natural evolution dataset, the state variable value sequence, the corresponding time position of the process execution data, and the operating state data in the continuous time interval before the data segment are extracted from each natural evolution data segment to form the natural evolution prediction input sequence. The natural evolution prediction input sequence is used to characterize the state continuity relationship when not affected by the control closed-loop feedback. Subsequently, state prediction processing is performed on each natural evolution prediction input sequence. The state variable values ​​in the input sequence are read in chronological order, and the results are calculated. The natural evolution prediction result is generated by extrapolating the state variable value range and change direction within the next prediction period based on the differences between adjacent sampling points, the mean of the time interval, and the most recent continuous change direction. This natural evolution prediction result includes at least the prediction time range, state variable name, prediction value range, and prediction change direction. For the control feedback dataset, the state variable value sequence, associated controlled object, associated control action event record, control reflexive influence value, and corresponding time position process execution data are extracted from each control feedback data segment to form the control feedback prediction input sequence. The control feedback prediction input sequence is used to characterize the response relationship of state variables to the control action after the control action enters the execution process. Then, feedback response prediction processing is performed on each control feedback prediction input sequence. First, the action start time, action response time, and action direction are determined based on the associated control action event records. Then, the state variable value sequence after the action response time is read, and the response start offset, response duration, response variable change amplitude, and recovery trend are calculated. The response start offset, response duration, response variable change amplitude, and recovery trend are then corrected based on the control reflexive influence. Finally, the response value range and response frequency for the next prediction period are extrapolated. In accordance with the direction, control feedback prediction results are generated. The larger the value of the control reflexive influence, the greater the correction ratio for the corresponding response duration and response change amplitude. The idea behind setting the correction ratio is to ensure that the feedback response prediction process reflects the continued effect of control closed-loop feedback on subsequent state changes. Finally, the natural evolution prediction results and control feedback prediction results are aligned according to the state variable name and prediction time range, and written into the prediction result table respectively. The natural evolution prediction results and control feedback prediction results are then output to step S5 for subsequent generation of control risk constraint quantities based on the natural evolution prediction results, control feedback prediction results, and control action event sequence.

[0037] Specifically, natural evolution extrapolation is performed based on the differences between adjacent sampling points, the mean of the time interval, and the most recent continuous change direction. First, the most recent continuous change direction is used as the initial change direction in the next prediction period. Then, the mean of the differences between adjacent sampling points is used as the unit sampling step size change, and the mean of the time interval is used as the central reference value to calculate the predicted value of the starting position of the next prediction period. Subsequently, according to the number of sampling points included in the prediction period, the unit sampling step size change is accumulated or decreased point by point to obtain the predicted value sequence. Then, the range of state variables is determined by the maximum and minimum values ​​in the predicted value sequence, and the most recent continuous change direction is used as the change direction to generate the natural evolution prediction result.

[0038] The response start-point offset, response duration, response magnitude, and recovery trend are corrected by incorporating the control reflexive influence. Specifically, the average response start-point offset, average response duration, and average response magnitude corresponding to the same associated controlled object in historical control feedback data segments are used as benchmark values. When the current control reflexive influence is higher than the interval corresponding to the benchmark value, the response duration and response magnitude are increased by a preset ratio, and the regression time corresponding to the recovery trend is extended by a preset ratio. When the current control reflexive influence is lower than the interval corresponding to the benchmark value, the response duration and response magnitude are decreased by a preset ratio, and the regression time corresponding to the recovery trend is shortened by a preset ratio. Subsequently, based on the corrected response start-point offset, response duration, and response magnitude, a response value sequence for the next prediction period is generated. The response value interval is determined by the maximum and minimum values ​​of the response value sequence, and the response direction is determined by the corrected response magnitude change direction, thus generating the control feedback prediction result.

[0039] Based on the aforementioned continuous heat treatment equipment embodiment, the second segment from 08:02:08 to 08:05:00 is used as the intervention operation data segment. This second segment was marked as the control feedback drift segment in step S3. The first segment from 08:00:00 to 08:02:00 and the third segment from 08:05:10 to 08:10:00 are used as natural evolution segments. First, based on the reflexive drift marking table and the segment index table, the second segment is written into the control feedback candidate segment set. The associated control objects—the combustion air distribution mechanism and the feeding mechanism—two associated control action event records, and the furnace temperature state variable names are retained. Simultaneously, the first and third segments are written into the natural evolution candidate segment set. Then, the state variables are split and spliced ​​to form the natural evolution... The dataset contains a natural evolution data segment A1 of furnace temperature and a control feedback data segment B1 of furnace temperature from the control feedback dataset. A1 includes the furnace temperature numerical sequences corresponding to segments 1 and 3, while B1 includes the furnace temperature numerical sequence corresponding to segment 2 and control reflexivity effects of 11.3028 and 14.0388. Further, a natural evolution prediction input sequence is constructed from A1. The furnace temperature values ​​(864, 864, 865, 865, 864, 864) of the last six consecutive sampling points at the end of segment 3 are taken. The differences between adjacent sampling points are calculated to be 0, 1, 0, -1, and 0, respectively. The mean of the differences between adjacent sampling points is 0, and the mean over the time interval is 864.33. The most recent continuous change direction is stationary. The next prediction period is assumed to include... With three sampling points, 864.33 is used as the starting predicted value, and the predicted values ​​are extrapolated point by point with a unit sampling step size of 0, resulting in the predicted value sequence 864.33, 864.33, and 864.33. Based on this, the prediction time range of the natural evolution prediction result is determined to be from 08:10:10 to 08:10:30, the state variable name is furnace temperature, the predicted value range is 864.33 to 864.33, and the predicted change direction is stationary. Then, a control feedback prediction input sequence is constructed for B1. The average response start offset of 20 seconds, the average response duration of 60 seconds, and the average response change amplitude of 6 in the historical control feedback data segment corresponding to the combustion air distribution mechanism are taken as the benchmark values, and the interval corresponding to the benchmark values ​​is taken as the control reflexive influence quantity. From 9 to 12, the current control reflexive influence quantity 11.3028 falls within the range corresponding to the benchmark value. Therefore, the original response starting point offset of 10 seconds, response duration of 40 seconds, response quantity change amplitude of 8, and recovery trend regression time of 150 seconds are retained unchanged. Assuming that the next prediction period contains 3 sampling points and the sampling period is 10 seconds, the response value sequence 866, 868, and 870 is generated according to the increasing direction of action and the rhythm of increasing by 2 every 10 seconds within the 40-second duration. The prediction time range of the control feedback prediction result is from 08:05:10 to 08:05:30. The state quantity name is furnace temperature, the response value range is 866 to 870, and the response direction is increasing. For the control reflexive influence quantity 14 corresponding to the feeding mechanism.0388, since it is higher than the corresponding range of the benchmark value, assuming a preset ratio of 20%, the original response duration of 110 seconds is corrected to 132 seconds, the original response value change amplitude of 8 is corrected to 9.6, and the regression time corresponding to the recovery trend is extended to 1.2 times the original value. In the next prediction period, the response value sequence 871, 869, and 867 is generated by decreasing in the direction of action. Based on this, the prediction time range of the control feedback prediction result is from 08:05:10 to 08:05:30, the state variable name is furnace temperature, the response value range is 867 to 871, and the response direction is decreasing. Finally, the above natural evolution prediction results and control feedback prediction results are written into the prediction result table according to the state variable name and prediction time range and output to step S5.

[0040] Step S5: Based on the natural evolution prediction results, control feedback prediction results, and control action event sequence, generate control risk constraint quantities, and then filter, delay, or limit the control instructions to be executed according to the control risk constraint quantities to generate constraint control instructions.

[0041] The purpose of generating control risk constraints is to analyze the correspondence between the natural evolution prediction results and control feedback prediction results obtained in step S4 and the control action event sequence obtained in step S1, identify the possible state deviations, feedback superposition, and cumulative effects caused by the continued execution of subsequent control actions, and generate control risk constraints accordingly. This allows for the selection, delay, or limitation of control instructions to be executed in subsequent steps based on the control risk constraints. In practice, the natural evolution prediction results and control feedback prediction results output in step S4, as well as the control action event sequence output in step S1, are read first. The natural evolution prediction results include at least the prediction time range, state variable name, prediction value range, and prediction change direction. The control feedback prediction results include at least the prediction time range, state variable name, response value range, and response direction. Each control action event record in the control action event sequence includes at least the prediction output time, instruction issuance time, action response time, execution object, action parameter, action direction, and corresponding process execution range. Subsequently, the natural evolution prediction results and control feedback prediction results are time-aligned according to the state variable name and prediction time range to form a prediction comparison record. Next, for each prediction comparison record, the predicted value range and predicted change direction are extracted from the natural evolution prediction results, and the response value range and response direction are extracted from the control feedback prediction results. The prediction deviation, directional conflict, and feedback superposition are calculated. The prediction deviation is obtained by subtracting the center value of the response value range from the center value of the prediction value range. The directional conflict is obtained by comparing whether the response direction and the predicted change direction are consistent. If they are consistent, it is recorded as the first state value; if they are inconsistent, it is recorded as the second state value. The feedback superposition is obtained by subtracting the width of the response value range from the width of the prediction value range, so as to characterize the degree of deviation and expansion of the control feedback from the natural evolution trend. Subsequently, the control action event records corresponding to the current prediction comparison record are read from the control action event sequence. The execution object, action parameters, action direction, action response time, and corresponding process execution interval are extracted. The cumulative continuous action amount is calculated by counting the number of control action event records that act on the same execution object and have the same action direction within a preset review period before the current prediction time range. The preset review period is designed to cover the time range within which continuous actions of the actuator may still affect the current state change, thus characterizing the cumulative effect of control actions over time. Further, the prediction deviation, direction conflict, feedback superposition, and cumulative continuous action amount are uniformly converted to form risk component values. The idea behind this uniform conversion is to convert deviation results and count results of different dimensions to the same comparison scale for combined calculation.Then, control risk constraints are generated based on the risk component values. Specifically, the risk component values ​​corresponding to the predicted deviation, directional conflict, feedback superposition, and continuous action accumulation are weighted and summed to obtain the control risk constraint corresponding to the current control instruction to be executed. During the weighted summation, the weights of each component are allocated based on the degree of influence of the state quantity on process quality, the degree of influence of the execution object on the production line cycle time, and the degree of influence of continuous control actions on equipment load. Finally, the control risk constraint, along with the corresponding state quantity name, prediction time range, execution object, action parameters, and associated control action event records, are written into the control risk constraint table, and the control risk constraint table is output for subsequent use in determining the control instruction to be executed based on the control risk constraint. The instructions are filtered, delayed, or limited. In one embodiment, the control risk constraint is composed of predicted deviation risk, directional conflict risk, feedback superposition risk, and continuous action cumulative risk. First, the predicted deviation, directional conflict, feedback superposition, and continuous action cumulative are converted into corresponding risk component values. Then, based on the impact of the execution object on process quality, equipment load sensitivity, and safety interlock proximity, each risk component value is weighted and synthesized to obtain the control risk constraint. The directional conflict can be represented by binary or multi-level discrete values, and the continuous action cumulative can be represented by at least one of the number of unidirectional control actions, duration, or cumulative amplitude within a preset review period. By decomposing the risk constraint into individually achievable component values, the engineering feasibility of the control constraint steps can be improved.

[0042] The purpose of generating constraint control instructions is to constrain subsequent control instructions to be issued based on control risk constraint quantities, thereby reducing parameter deviations and continuous action accumulation caused by the continued superposition of control actions under closed-loop feedback conditions, and generating constraint control instructions that can enter the execution process. In specific implementation, the control risk constraint table and the control instructions to be executed at the current time are read first. The control instructions to be executed include at least the instruction generation time, execution object, action parameters, action direction, and instruction target value. Then, the control instructions to be executed are matched with the corresponding records in the control risk constraint table according to the execution object and action parameters, and the control risk constraint quantity corresponding to the current control instruction to be executed is extracted. Next, risk grading is performed based on the control risk constraints to form instruction processing levels. The risk grading is based on dividing the historical control risk constraints into low-constraint, medium-constraint, and high-constraint zones, which are used to trigger screening, delay, and limiting processes respectively. Subsequently, when the control risk constraint falls into the low-constraint zone, screening is performed. Specifically, control instructions whose action direction is consistent with the predicted change direction in the natural evolution prediction results and does not overlap with the response direction in the control feedback prediction results are retained. Instructions whose action direction is opposite to the predicted change direction in the natural evolution prediction results and does not overlap with the response direction in the control feedback prediction results are removed. Execute control commands and output the filtered control commands to be executed; further, when the control risk constraint falls into the medium constraint zone, a delay process is performed, specifically: taking the command generation time of the current control command to be executed as the starting point, the command issuance time is postponed according to a preset delay duration, forming a delayed execution command; the preset delay duration is designed to cover a prediction period or a control action identification interval, used to re-evaluate the control action after the current feedback impact is released; then, when the control risk constraint falls into the high constraint zone, a limiting process is performed, specifically: reading the command target value of the control command to be executed, and reading the response value range width and control risk constraint from the corresponding control feedback prediction result, based on... Based on the change range of the target value of the control risk constraint relative to the current process execution value, a limited execution instruction is obtained. The idea behind setting the level ratio is that the larger the control risk constraint, the smaller the allowable change range, so as to reduce the amplification effect of continuous control actions on subsequent state changes. Furthermore, when there are multiple control instructions to be executed for the same execution object in the same processing cycle, they are first sorted by the instruction generation time, and then sorted by the size of the control risk constraint. The control instructions to be executed that are ranked earlier and have a lower processing level are retained, while the remaining control instructions to be executed are eliminated, delayed, or limited to form a unique constraint control instruction for the same execution object in the same processing cycle.Finally, the instructions that have undergone filtering, delaying, or limiting are recorded as constraint control instructions. The constraint control instructions include the execution object, action parameters, action direction, post-constraint target value, post-constraint issuance time, corresponding control risk constraint amount, and processing method marker. The constraint control instructions are then output to step S6 for subsequent writing of the control action event sequence and updating the segmented behavior feature set.

[0043] The implementation process of step S5 is further explained. Based on the aforementioned embodiment, the furnace temperature is used as the state variable for further explanation. The prediction time range for the natural evolution prediction result is from 08:05:10 to 08:05:30, the predicted value interval is 864.33 to 864.33, and the predicted change direction is stable. In the control feedback prediction result, the response value interval for the combustion air distribution mechanism is 866 to 870, and the response direction is increasing. The response value interval for the feeding mechanism is 867 to 871, and the response direction is... The control is reduced, and two control action event records corresponding to the current prediction comparison record are read. The action direction corresponding to the combustion air distribution mechanism is increasing, and the action direction corresponding to the feeding mechanism is decreasing. The preset review period is set to 5 minutes. The prediction deviation is calculated as 3.67, the directional conflict is 1, the feedback superposition is 4, and the continuous action accumulation is 2 for the corresponding record of the combustion air distribution mechanism. The risk items are calculated by multiplying the prediction deviation risk item value by 0.4, the directional conflict risk item value by 0.2, the feedback superposition risk item value by 0.2, and the continuous action risk item value by 0.2. The value is multiplied by 0.2 and weighted summed to obtain a control risk constraint of 3.068. The corresponding record for the feeding mechanism is calculated in the same way to obtain a control risk constraint of 3.268. Both are written into the control risk constraint table. Two control instructions to be executed are read, with the target value of the instruction corresponding to the combustion air distribution mechanism being 36 and the target value of the instruction corresponding to the feeding mechanism being 1.10. Assuming the low constraint zone is 0 to 2.5, the medium constraint zone is 2.5 to 3.2, and the high constraint zone is greater than 3.2, then the control risk constraint of the combustion air distribution mechanism is 3. 068 falls into the medium constraint zone, and a delay is performed. The preset delay time is 20 seconds, and the constraint is issued at 08:05:25. The control risk constraint value of 3.268 corresponding to the feeding mechanism falls into the high constraint zone, and a limit is performed. The current process execution value is 1.15, the original change range is 0.05, and after being compressed by 50%, the target value after constraint is 1.125. Finally, the delay execution command corresponding to the combustion air distribution mechanism and the limit execution command corresponding to the feeding mechanism are recorded as constraint control commands and output to step S6.

[0044] Step S6 involves writing the execution results of the constraint control instructions into the control action event sequence, updating the segmented behavior feature set in conjunction with the control reflexive influence and reflexive drift marker table, and iteratively correcting the natural evolution dataset and control feedback dataset. The purpose is to re-incorporate the results of the constraint control instructions generated in step S5 after actual execution into the control action event sequence, so that the constraint execution status of the control action, the state changes after execution, and the resulting changes in the source of behavior can continue to be recorded. Based on this, the segmented behavior feature set is updated in conjunction with the control reflexive influence and reflexive drift marker table, and the natural evolution dataset and control feedback dataset are further iteratively corrected so that the data partitioning and prediction processing in subsequent rounds are based on the latest execution results.

[0045] In specific implementation, firstly, the constraint control command output in step S5, the control action event sequence output in step S1, the control reflexive influence scale and reflexive drift marker table output in step S3, the segmented behavior feature set output in step S2, and the natural evolution dataset and control feedback dataset output in step S4 are read. Then, the running status data and process execution data after the constraint control command is executed are read; the actual issuance time, actual effective time, actual execution object, actual action parameter, actual target value, and actual action range corresponding to the constraint control command are extracted and compared with the constraint-post-constraint issuance time, execution object, action parameter, and constraint-post-constraint target value in the constraint control command; an execution result record is formed; wherein, the execution result record is used to characterize whether the constraint control command enters the execution process according to the constraint-post content and the actual action range after entering the execution process. Next, the execution result record is written into the control action event sequence. Specifically, new control action event records are added to the control action event sequence according to the actual issuance time; or the instruction issuance time, action parameters, target values, and corresponding process execution intervals in the original control action event records are replaced and updated. When the constraint control instruction corresponds to delayed processing, the delayed actual issuance time and corresponding actual action interval are written into the control action event sequence. When the constraint control instruction corresponds to amplitude limiting processing, the constrained target value and the actual action interval formed based on the constrained target value are written into the control action event sequence. When the constraint control instruction corresponds to filtering processing and has not entered the execution process, an unexecuted marker and the corresponding time position are written into the control action event sequence. Further, the intervention operation data segment related to the constraint control instruction is re-extracted based on the updated control action event sequence; and the state quantity value sequence within the time period corresponding to the execution result is read from the intervention operation data segment; combined with the control reflexive influence quantity and reflexive drift marker table in step S3, the corresponding segmented behavior feature records in the original segmented behavior feature set are updated. Specifically: When the updated control action event sequence shows that the constraint control command has not been executed; or the change in the state quantity after execution does not continue along the action direction corresponding to the original associated control object; the response feature in the corresponding segmented behavior feature record is rewritten as a no-response mark or a weak-response mark; and the association level of the corresponding control reflexive influence quantity in the segmented behavior feature record is reduced simultaneously; when the updated control action event sequence shows that the constraint control command has been executed; and the direction, duration, and magnitude of the change in the state quantity after execution are consistent with the original control feedback process; the response feature in the original segmented behavior feature record is retained; and the corresponding control reflexive influence quantity is written into the feature source field of the current round; when the updated control action event sequence shows that the source of the behavior change after execution is inconsistent with the original reflexive drift mark; the distribution feature, fluctuation feature, and response feature are recalculated with the updated execution result corresponding to the state quantity value sequence; and the corresponding content in the original segmented behavior feature record is replaced with the recalculated distribution feature, fluctuation feature, and response feature.Subsequently, the reflexive drift markers and dataset affiliation are re-examined based on the updated segmented behavioral feature records. Specifically: for segmented behavioral feature records originally marked as control feedback drift segments, if the updated response features weaken, the control reflexive influence decreases, and the state variable changes corresponding to the execution results revert to the natural evolution trend, then the corresponding data segment is removed from the control feedback dataset and written to the natural evolution dataset. For segmented behavioral feature records originally marked as natural evolution segments, if the updated execution results correspond to the state variable changes and the direction, duration, and magnitude of change of the associated control object, then the corresponding data segment is removed from the natural evolution dataset and written to the control feedback dataset. For data segments that have not undergone affiliation changes, only the index records, time ranges, feature sources, and associated execution result fields in the natural evolution dataset or control feedback dataset are updated.

[0046] In one implementation, whether a data segment corresponding to the execution result migrates between the natural evolution dataset and the control feedback dataset is determined by a joint assessment based on at least three conditions: change in response direction, change in response persistence, and trend of change in control reflexive influence. When the response direction of the data segment after execution no longer aligns with the direction of action of the associated control object, the response persistence decreases, and the control reflexive influence falls back to the natural evolution reference range, the data segment is migrated to the natural evolution dataset. When the data segment after execution and the associated control object form a correspondence in terms of action direction, start time, and magnitude of change, and the control reflexive influence increases and enters the control feedback reference range, the data segment is migrated to the control feedback dataset. For data segments that do not meet the migration conditions, only the index record and feature source field are updated. By setting a joint migration criterion, the subjectivity of data attribution adjustment can be reduced.

[0047] Furthermore, the corrected natural evolution dataset and control feedback dataset are subjected to iteration numbering processing to generate a version label for the current round dataset. The updated segmented behavior feature set, natural evolution dataset, and control feedback dataset are stored as input data for the next round of state prediction processing and feedback response prediction processing. Finally, the updated control action event sequence, segmented behavior feature set, natural evolution dataset, and control feedback dataset are output to enter the next round of data segmentation, impact identification, prediction processing, and control constraint processing based on the latest execution results.

[0048] In one embodiment, both natural evolution prediction processing and feedback response prediction processing can be implemented using rule extrapolation, statistical learning, machine learning, or digital twin simulation. When using rule extrapolation, the prediction interval for the next prediction period is formed based on the recent continuous change direction, mean change trend, and fluctuation bandwidth. When using statistical learning, state transition relationships are established based on historical fragments of similar state variables, and the prediction interval is output. When using digital twin simulation, the current process execution data, equipment status data, and constraint control commands are input into the digital twin model, and the natural evolution prediction result or control feedback prediction result is output.

[0049] Based on the aforementioned embodiments, the following explanation continues using two constraint control commands corresponding to the combustion air distribution mechanism and the feeding mechanism. The constraint control command for the combustion air distribution mechanism is a delayed processing, with a target value of 36 after constraint and a issuance time of 08:05:25. The constraint control command for the feeding mechanism is a limit processing, with a target value of 1.125 after constraint. After reading the execution results, the actual effective time of the combustion air distribution mechanism is 08:05:30, and the actual effective range is from 08:05:30 to 08:07:00. The actual effective time of the feeding mechanism is 08:05:10, and the actual effective range is from 08:05:10 to 08:06:40. The issuance time, target value, and process execution range of the corresponding commands are written into the control action event sequence. Then, based on the updated control action event sequence, the furnace temperature state value sequence is re-extracted, and the change in response value after the limit execution of the feeding mechanism is found to be 14.03. The original prediction level corresponding to 88 decreased to 3, and the response duration was reduced to 90 seconds. Therefore, the response features in the corresponding segmented behavior feature records were updated, and the control reflexive influence was reduced to 9.2000 while retaining the control feedback drift segment marker. At the same time, the response start offset after the delayed execution of the combustion air distribution mechanism was updated to 20 seconds, the response duration was updated to 80 seconds, and the response change amplitude was updated to 2. The control reflexive influence was reduced from 11.3028 to 8.1000. Since the updated state changes are close to the natural evolution trend, the corresponding data segment was moved out of the control feedback dataset and written into the natural evolution dataset. Finally, the corrected natural evolution dataset and control feedback dataset were renumbered to generate the current round dataset version marker. The updated control action event sequence, segmented behavior feature set, natural evolution dataset, and control feedback dataset were output as input data for the next round of state prediction processing and feedback response prediction processing.

[0050] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the aforementioned scope.

[0051] Finally: The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A dynamic modeling and prediction method for industrial big data combining digital twins, characterized in that: include: Step S1: Obtain the operating status data, process execution data, control command records and prediction result time series table, and construct the control action event sequence according to the time correspondence; Step S2: The operational status data is segmented according to the sequence of control action events to form natural operational data segments and intervention operational data segments, and distribution characteristics, fluctuation characteristics and response characteristics are extracted to generate a segmented behavioral feature set; Step S3: Based on the segmented behavior feature set and prediction results, calculate the control reflexivity effect and generate a reflexivity drift label table according to the control reflexivity effect. Step S4: Based on the reflexive drift label table, the intervention operation data segment is divided into a natural evolution dataset and a control feedback dataset, and state prediction processing and feedback response prediction processing are performed respectively to generate natural evolution prediction results and control feedback prediction results. Step S5: Based on the natural evolution prediction results, control feedback prediction results, and control action event sequence, generate control risk constraints, and then filter, delay, or limit the control instructions to be executed based on the control risk constraints to generate constraint control instructions. Step S6: Write the execution results of the constraint control command into the control action event sequence, and update the segmented behavior feature set by combining the control reflexive influence quantity and the reflexive drift label table, and iteratively correct the natural evolution dataset and the control feedback dataset.

2. The method for dynamic modeling and prediction of industrial big data combined with digital twins according to claim 1, characterized in that, Methods for constructing control action event sequences include: The system reads operating status data, process execution data, control command records, and prediction results to form operating status timing tables, process execution timing tables, control command timing tables, and prediction result timing tables, respectively. It matches the associated control objects in the prediction results with the command execution objects in the control command records to obtain a predictive control pairing table. It performs time-correspondence processing based on a unified time base and determines the control action identification interval based on the command issuance time and action response time. For the control command records that have completed time-correspondence processing, it extracts the predicted output time, command issuance time, action response time, execution object, action parameters, action direction, and corresponding process execution interval to generate control action event records. These records are then sorted by command issuance time to form a control action event sequence.

3. The method for dynamic modeling and prediction of industrial big data combined with digital twins according to claim 2, characterized in that, Methods for forming natural and interventional operational data segments include: Read the running status timing table and the control action event sequence, and use the acquisition time mark as the segmented reference time axis; determine the candidate intervention interval for each control action event record, and perform interval merging processing on all candidate intervention intervals to obtain the intervention interval set; form intervention running data segments according to time continuity for running status data that fall into the intervention interval set, and form natural running data segments according to time continuity for running status data that do not fall into the intervention interval set.

4. The method for dynamic modeling and prediction of industrial big data combined with digital twins according to claim 1, characterized in that, Methods for generating segmented behavioral feature sets include: Read the natural operation data segment, the intervention operation data segment, and the segment index table, and extract the numerical sequence of each state variable within the corresponding time interval according to the segment number; extract the distribution characteristics and fluctuation characteristics based on the numerical sequence; for each intervention operation data segment, extract the response characteristics by combining the associated control action event records recorded in the segment index table; for the natural operation data segment, write a no-response mark in the segment index table; uniformly encode the distribution characteristics and fluctuation characteristics corresponding to each natural operation data segment, as well as the distribution characteristics, fluctuation characteristics, and response characteristics corresponding to each intervention operation data segment, to form a segmented behavior characteristic record; summarize all segmented behavior characteristic records according to the segment number to generate a segmented behavior characteristic set.

5. The method for dynamic modeling and prediction of industrial big data combined with digital twins according to claim 2, characterized in that, Methods for generating reflexive drift marker tables include: Read the segmented behavioral feature set and the time series table of prediction results, and pair the prediction results with the corresponding intervention operation data segments to form a prediction segment correspondence table. For each record in the prediction segment correspondence table, extract the distribution characteristics, fluctuation characteristics, and response characteristics of the corresponding intervention operation data segment, and extract the distribution characteristics and fluctuation characteristics of the time-adjacent natural operation data segments to form a comparison feature group. Calculate the distribution offset, fluctuation offset, and response coupling based on the comparison feature group, and extract prediction trend information in conjunction with the prediction results to obtain the prediction response consistency quantity. Calculate the control reflexivity influence quantity based on the distribution offset, fluctuation offset, response coupling, and prediction response consistency quantity, and write it into the control reflexivity influence quantity table. Determine the segment-level judgment result based on the control reflexivity influence quantity, and generate a reflexivity drift label table in conjunction with the segment source marker.

6. The method for dynamic modeling and prediction of industrial big data combined with digital twins according to claim 1, characterized in that, Methods for dividing datasets into natural evolution datasets and control feedback datasets include: Read the intervention operation data segments, segment index table, and reflexive drift marker table, and match the reflexive drift markers with the segment index table according to the segment number. Extract the segment-level judgment result corresponding to each intervention operation data segment. Extract the intervention operation data segments whose segment-level judgment result is a control feedback drift segment to form a control feedback candidate segment set, and extract the intervention operation data segments whose segment-level judgment result is not a control feedback drift segment to form a natural evolution candidate segment set. Perform state variable splitting processing on the control feedback candidate segment set and the natural evolution candidate segment set according to the state variable name. Perform continuous splicing processing on the natural evolution candidate segment set and source consistency splicing processing on the control feedback candidate segment set. Establish natural evolution dataset index and control feedback dataset index respectively, and summarize the corresponding data segments to form natural evolution dataset and control feedback dataset.

7. The method for dynamic modeling and prediction of industrial big data combined with digital twins according to claim 6, characterized in that, Methods for generating natural evolution prediction results and control feedback prediction results include: Read the natural evolution dataset and the control feedback dataset, and organize the input sequence according to the data segment number and state variable name; extract the state variable numerical sequence, the process execution data at the corresponding time position, and the operating state data in the previous continuous time interval from the natural evolution dataset to form the natural evolution prediction input sequence, and perform state prediction processing on the natural evolution prediction input sequence to generate the natural evolution prediction result; extract the state variable numerical sequence, associated control object, associated control action event record, control reflexive influence value, and the process execution data at the corresponding time position from the control feedback dataset to form the control feedback prediction input sequence, and perform feedback response prediction processing on the control feedback prediction input sequence to generate the control feedback prediction result.

8. The method for dynamic modeling and prediction of industrial big data combined with digital twins according to claim 1, characterized in that, Methods for generating control risk constraints include: Read the natural evolution prediction results, control feedback prediction results, and control action event sequences, and align the natural evolution prediction results and control feedback prediction results in time according to the state variable name and prediction time range to form a prediction comparison record. For each prediction comparison record, extract the prediction value range, prediction change direction, response value range, and response direction, and calculate the prediction deviation, direction conflict, and feedback superposition. Read the control action event record corresponding to the current prediction comparison record, extract the execution object, action parameter, action direction, action response time, and corresponding process execution range, and calculate the cumulative amount of continuous action. Perform a unified conversion on the prediction deviation, direction conflict, feedback superposition, and cumulative amount of continuous action to form risk component values. Perform a weighted summation based on the risk component values ​​to generate the control risk constraint quantity.

9. The method for dynamic modeling and prediction of industrial big data combined with digital twins according to claim 8, characterized in that, Methods for generating constraint control instructions include: Read the control risk constraint table and the control instructions to be executed, and match the corresponding control risk constraint quantities according to the execution object and the action parameters; perform risk classification processing based on the control risk constraint quantities to form the instruction processing level; when the control risk constraint quantity falls into the low constraint zone, perform screening processing on the control instructions to be executed; when the control risk constraint quantity falls into the medium constraint zone, perform delay processing on the control instructions to be executed; when the control risk constraint quantity falls into the high constraint zone, perform amplitude limiting processing on the control instructions to be executed; when there are multiple control instructions to be executed for the same execution object in the same processing cycle, retain a unique control instruction to be executed after sorting by instruction generation time and control risk constraint quantity; record the instructions after screening, delay, or amplitude limiting processing as constraint control instructions.

10. The method for dynamic modeling and prediction of industrial big data combined with digital twins according to claim 8, characterized in that, Methods for iteratively refining natural evolution datasets and control feedback datasets include: Read constraint control commands, control action event sequences, control reflexivity impact scale, reflexivity drift labeling table, segmented behavior feature set, natural evolution dataset, and control feedback dataset; read the operational status data and process execution data after the constraint control commands are executed, extract the actual issuance time, actual execution object, actual action parameters, actual target value, and actual action range to form an execution result record; update the control action event sequence based on the execution result record; extract the corresponding intervention operation data segment based on the updated control action event sequence, and update the segmented behavior feature set in conjunction with the control reflexivity impact scale and reflexivity drift labeling table; correct the natural evolution dataset and control feedback dataset based on the updated segmented behavior feature set, and generate updated dataset version labels; output the updated control action event sequence, segmented behavior feature set, natural evolution dataset, and control feedback dataset.

Images