Vacuum furnace AI pressure control method based on multi-modal data fusion

Abstract

The application discloses a kind of vacuum furnace AI pressure control methods based on multi-modal data fusion, it is related to vacuum furnace control technical field, including, acquisition vacuum furnace current control period in multi-modal data, execution time alignment, abnormality is excluded, dimension is normalized and stage slice, form synchronous control frame;According to synchronous control frame, determine the pressure control phase corresponding to current control period, write the stage pressure control constraint corresponding to current pressure control phase into stage control account book, form stage control account book entry;Execute compensation control instruction set, compare and revise actual pressure response with prediction result, rewrite revision result.The application realizes continuous representation to current control period pressure evolution trend, orderly constraint to compensation action release opportunity and unified scheduling to pump group, valve, temperature rise and backfilling collaborative relationship by executing pressure prediction and compensation control.

Description

Technical Field

[0001] This invention relates to the field of vacuum furnace control technology, and in particular to an AI-based pressure control method for vacuum furnaces based on multimodal data fusion. Background Technology

[0002] Vacuum furnaces are widely used in heat treatment, brazing, sintering, and functional material preparation. Their pressure control process typically operates in conjunction with heating regimes, evacuation paths, valve regulation, and backfilling strategies. With advancements in sensor configuration, actuator linkage, and process digitization, technologies for multi-source sensing, staged modeling, predictive control, and closed-loop compensation for vacuum furnace operation are continuously developing. Pressure control is evolving from single-point feedback regulation to intelligent control methods that combine process mechanisms and data analysis.

[0003] In actual operation of a vacuum furnace, pressure changes are simultaneously affected by factors such as pressure control stage switching, pump and valve action lag, temperature coupling, and backfill disturbances. The control process exhibits significant stage dependence and propagation lag, making it difficult to establish a stable correlation between control information from different sources within a unified temporal context. Therefore, how to achieve coordinated organization of multimodal states, stage constraints, disturbance characterization, and time-delay propagation characteristics around the same control cycle has become a core technical problem in the precise pressure control of vacuum furnaces. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an AI-based pressure control method for vacuum furnaces based on multimodal data fusion to solve the problems of timing mismatch and inaccurate compensation in pressure control under multiple disturbances and stage switching conditions.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides an AI-based pressure control method for a vacuum furnace based on multimodal data fusion. The method includes: acquiring multimodal data within the current control cycle of the vacuum furnace; performing time alignment, anomaly removal, dimension normalization, and stage slicing to form a synchronous control frame; determining the pressure control stage corresponding to the current control cycle based on the synchronous control frame; writing the stage pressure control constraints corresponding to the current pressure control stage into a stage control ledger to form a stage control ledger entry; and combining the synchronous control frame and the stage control ledger entry to identify the disturbance type, disturbance intensity, and disturbance reliability corresponding to the current pressure deviation to form a disturbance... The system controls the ledger entries; based on the synchronous control frame, stage control ledger entries, and disturbance control ledger entries, it extracts the pressure response hysteresis corresponding to the pump group, valve, temperature rise, and backfill, forming a time-delay feature set; it fuses and encodes the time-delay feature set to form a pressure propagation time-delay fingerprint vector; it executes pressure prediction and compensation control based on the stage control ledger entries, disturbance control ledger entries, and pressure propagation time-delay fingerprint vector, generating prediction results and a compensation control instruction set; it executes the compensation control instruction set, compares the actual pressure response with the prediction results to identify and correct the deviation, and writes the correction results back.

[0007] As a preferred embodiment of the vacuum furnace AI pressure control method based on multimodal data fusion described in this invention, the step of forming a synchronization control frame is as follows: The multimodal data includes pressure time-series data, temperature time-series data, valve opening data, pump unit operating status data, process formula data, furnace charging parameter data, historical anomaly records, and historical control result data. Time offset correction and unified time base projection are performed on the multimodal data to obtain the alignment control sequence; The alignment control sequence is combined with historical control result data and historical anomaly records to remove abnormal data. Robust dimensional normalization is then performed based on process formula data and furnace charging parameter data to obtain the normalized control sequence. The stage slice with hysteresis constraint is executed based on the normalized control sequence, and the stage slice results are spliced ​​with the normalized state variables, trend variables, process formula data, furnace loading parameter data, historical anomaly records and historical control result data corresponding to the normalized control sequence to form a synchronous control frame.

[0008] As a preferred embodiment of the vacuum furnace AI pressure control method based on multimodal data fusion described in this invention, the stage pressure control constraints include the target pressure curve, the allowable fluctuation band, the stage switching window, and the control constraint priority.

[0009] As a preferred embodiment of the vacuum furnace AI pressure control method based on multimodal data fusion described in this invention, the steps for forming the ledger entries in the formation stage are as follows: Based on the stage slice results, normalized state variables, and trend variables in the synchronous control frame, and combined with the stage start and end conditions in the process formula data, the pressure control stage corresponding to the current control cycle is determined. The target pressure curve, allowable fluctuation range, stage switching window, and control constraint priority corresponding to the pressure control stage are written into the preset stage control ledger to form stage control ledger entries.

[0010] As a preferred embodiment of the vacuum furnace AI pressure control method based on multimodal data fusion described in this invention, the step of forming disturbance control ledger entries is as follows: By combining the actual pressure data in the synchronous control frame with the target pressure curve and allowable fluctuation band in the stage control ledger entries, a stage-normalized pressure deviation flow is constructed to determine the disturbance window corresponding to the current pressure deviation. Based on the disturbance window, corresponding disturbance hypotheses are constructed around the pump group, valve, temperature rise and backfill respectively. Combined with historical control result data, counterfactual repair pressure trajectories corresponding to each disturbance hypothesis are generated. By combining the counterfactual repair pressure trajectory corresponding to each perturbation hypothesis with the actual pressure data, target pressure curve and allowable fluctuation band, the perturbation evidence quality value corresponding to each perturbation hypothesis is calculated. Conflict suppression fusion is performed on the perturbation evidence quality values ​​corresponding to each perturbation hypothesis to determine the perturbation type, perturbation intensity, and perturbation confidence level corresponding to the current pressure deviation; The disturbance type, disturbance intensity, and disturbance confidence level, along with the current control cycle number, current control stage, and current disturbance window, are written into the preset disturbance control ledger to form disturbance control ledger entries.

[0011] As a preferred embodiment of the vacuum furnace AI pressure control method based on multimodal data fusion described in this invention, the steps for forming the time-delay feature set are as follows: By combining synchronous control frames, stage control ledger entries, and disturbance control ledger entries, the stage-gated excitation flow and pressure response flow corresponding to pump sets, valves, heating, and backfilling are constructed. Based on the stage-gated excitation flow and pressure response flow, the pressure response hysteresis corresponding to the pump group, valve, temperature rise and backfill is determined respectively; Based on the pressure response hysteresis corresponding to the pump set, valve, temperature rise and backfill, the strength of time hysteresis evidence, the peak width of time hysteresis and the order difference between objects are extracted to form a time hysteresis feature set.

[0012] As a preferred embodiment of the vacuum furnace AI pressure control method based on multimodal data fusion described in this invention, the steps for forming the pressure propagation time-delay fingerprint vector are as follows: A continuous time-delay propagation graph is constructed using a time-delay feature set, and a memory update is performed on the continuous time-delay propagation graph to form the state of the time-delay propagation graph; Extract the root time-delay subgraph centered on the pressure node from the state of the time-delay propagation graph, and perform whole-graph embedding encoding on the root time-delay subgraph to form the pressure propagation time-delay fingerprint vector.

[0013] As a preferred embodiment of the vacuum furnace AI pressure control method based on multimodal data fusion described in this invention, the steps for generating the prediction results and compensation control instruction set are as follows: By combining the synchronization control frame, stage control ledger entries, disturbance control ledger entries, and pressure propagation delay fingerprint vector, the compensation prediction context packet and timing release constraints corresponding to the current control cycle are constructed. Perform rolling pressure prediction on the compensation prediction context packet and timing release constraints to obtain the prediction result corresponding to the current control cycle; Based on the prediction results, multiple sets of compensation candidate sequences are generated. Combined with stage control pressure constraints, disturbance type, disturbance intensity, disturbance confidence and pressure propagation time delay fingerprint vector, each compensation candidate sequence is screened to determine the target compensation candidate sequence. Based on the target compensation candidate sequence, target pressure curve, allowable fluctuation band, and stage switching window, safety constraint correction is performed to obtain the compensation control instruction set.

[0014] As a preferred embodiment of the vacuum furnace AI pressure control method based on multimodal data fusion described in this invention, the prediction result includes at least the next time-domain pressure trajectory; The compensation control command set includes pump start / stop adjustment amount, pump speed adjustment amount, throttle valve opening adjustment amount, vacuum valve switching amount, bypass valve switching amount, air charging valve opening adjustment amount, and heating rate correction amount.

[0015] As a preferred embodiment of the vacuum furnace AI pressure control method based on multimodal data fusion described in this invention, the step of writing back the correction result is as follows: The system executes the compensation control instruction set and collects the actual pressure response. It then compares the actual pressure response with the predicted results to determine the deviation. Based on the deviation results, it performs online correction on the predicted results for the next control cycle to generate the correction results. The compensation control instruction set, actual pressure response, prediction results, deviation results, and correction results are written and written back.

[0016] The beneficial effects of this invention are as follows: by performing pressure prediction and compensation control based on stage control ledger entries, disturbance control ledger entries, and pressure propagation time delay fingerprint vectors, it achieves continuous characterization of the pressure evolution trend in the current control cycle, orderly constraint on the timing of compensation action release, and unified scheduling of the coordinated relationship between pump groups, valves, heating, and backfilling; it enables the output compensation control instruction set to have clear stage specificity, disturbance adaptability, and time delay matching, and can provide a stable basis for the deviation correction of actual pressure response and prediction results and online write-back. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of an AI-based pressure control method for vacuum furnaces based on multimodal data fusion.

[0019] Figure 2 This is a schematic diagram illustrating the formation of entries in the phased control ledger.

[0020] Figure 3 A schematic diagram of the formation of disturbance control ledger entries.

[0021] Figure 4 This is a schematic diagram illustrating the linkage between the pressure propagation time delay fingerprint vector and the compensation control.

[0022] Figure 5 This is a data graph for the stage switching window.

[0023] Figure 6 This is a data graph showing the distribution of pressure response hysteresis for each object.

[0024] Figure 7 This is a comparison chart of the three curves of the disturbance window. Detailed Implementation

[0025] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0026] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0027] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0028] Reference Figures 1-7 This is one embodiment of the present invention, which provides an AI pressure control method for a vacuum furnace based on multimodal data fusion, comprising the following steps: S1. Collect multimodal data within the current control cycle of the vacuum furnace, perform time alignment, anomaly removal, dimension normalization, and stage slicing to form a synchronous control frame.

[0029] S1.1: Multimodal data includes pressure time series data, temperature time series data, valve opening data, pump unit operating status data, process formula data, furnace charging parameter data, historical anomaly records, and historical control result data; It should be noted that the pressure time series data refers to the sequence of chamber pressure values ​​that change continuously over time and are collected in real time by the capacitive pressure sensor inside the vacuum furnace chamber. Temperature time series data refers to the sequence of temperature values ​​recorded by the thermocouple temperature measuring element inside the furnace, reflecting the evolution of the current heating or cooling stage over time; Valve opening data refers to real-time status data fed back by electric or pneumatic actuators, characterizing the current physical opening angle or percentage of throttling valves, gas charging valves, and vacuum valves in the vacuum furnace pipeline; Pump unit operating status data refers to real-time electrical signal data output by the vacuum pump unit PLC controller, reflecting the current start / stop status, speed frequency, and operating load of the mechanical pump, Roots pump, or diffusion pump. Process formulation data refers to the set of process control program parameters that are pre-entered by the host computer human-machine interface (HMI) and include the target pressure curve, stage temperature setpoint, heat preservation time and stage switching logic. Furnace loading parameter data refers to static process auxiliary parameters that are manually entered by operators or identified by scanning codes, and that characterize the material, weight, bulk density, and loading method of the current batch of materials. Historical anomaly records refer to the time-series log data of abnormal operating conditions such as sensor failures, valve jamming, or pressure over-limits that occurred during past production processes, stored by the equipment alarm module. Historical control result data refers to the mapping relationship data between control commands executed in previous control cycles and the actual pressure / temperature trajectories achieved, which are archived in the database.

[0030] S1.2: Perform time offset correction and unified time reference projection on the multimodal data to obtain the alignment control sequence; Specifically, time offset correction is performed on each data sequence in the multimodal data to eliminate minor differences in the acquisition time of different sensors, and a reference clock source for unified time reference projection is determined. The reference clock source is the master clock in the vacuum furnace control element. Data with sampling frequency higher than the reference clock source in the multimodal data is downsampled to retain characteristic peaks. Data with sampling frequency lower than the reference clock source in the multimodal data is interpolated using linear interpolation or spline interpolation algorithms. The processed data sequence is arranged according to the timestamp of the reference clock source to form an alignment control sequence.

[0031] S1.3: Combine the alignment control sequence with historical control result data and historical anomaly records, remove abnormal data, and perform robust dimensional normalization based on process formula data and furnace charging parameter data to obtain the normalized control sequence; Specifically, the alignment control sequence is timestamped and associated with historical control result data and historical anomaly records to construct a complete dataset containing both current and historical moments. Based on the fault type and occurrence time in the historical anomaly records, abnormal data in the complete dataset is marked and removed. Abnormal data includes data on sensor malfunctions, valve jamming, or pressure exceeding limits. Based on the target pressure curve in the process formula data and the material material and weight in the furnace charging parameter data, the pressure control stage identifier in the process formula data and the material characteristic identifier in the furnace charging parameter data are analyzed. This allows the data in the complete dataset to be divided into positive and negative indicators. For positive indicators, the dimensions are determined using minimization or averaging methods; for negative indicators, the dimensions are determined using a reverse method. Based on the material weight and material type in the furnace loading parameter data, combined with the historical control effect in the historical control result data, the weight of each data item is calculated using the entropy method. The determined dimensions and weights are applied to the corresponding data items to determine the dimensions and weights of each data item. Robust dimension normalization is performed on the aligned control sequence after removing outlier data, converting data with different dimensions into a unified dimensionless numerical range. The normalized data is then combined with the process formula data and furnace loading parameter data to form a normalized control sequence.

[0032] The weight of each data item is calculated using the entropy method, expressed as: ; ; In the formula, Indicates the first Information entropy of each data item; Indicates the index of the data item; The normalization coefficient represents the entropy value calculation. This represents the total number of sampling times; Indicates the sampling time index; Indicates the first Item data item in the first The feature weight at each sampling time, that is, the proportion of the value at that time to the total value of the data at all times; Indicates the first Item data item in the first The feature weights at each sampling time are calculated using the natural logarithm. Indicates the first The weight of each data item; This indicates the total number of data items involved in the weight calculation.

[0033] It should be noted that the normalization coefficient for entropy calculation comes from the correction factor set in information entropy theory to eliminate the influence of sample size on the magnitude of entropy. Its mathematical essence is the reciprocal of the natural logarithm of the total number of samples. It is used to map the calculated information entropy value to a standardized interval of 0 to 1, ensuring that the entropy values ​​of datasets with different sample sizes are comparable. The exemplary value range is usually between 0 and 1. The value is strictly determined by the total number of sampling moments in the alignment control sequence, that is, by calculating the natural logarithm of the total number of sampling moments and taking its reciprocal.

[0034] S1.4: Execute stage slices with hysteresis constraints based on the normalized control sequence, and splice the stage slice results with the normalized state variables, trend variables, process formula data, furnace loading parameter data, historical anomaly records, and historical control result data corresponding to the normalized control sequence to form a synchronous control frame.

[0035] Specifically, based on the preset stage division threshold, stage slicing with hysteresis constraints is performed on the normalized control sequence. By setting the rising and falling thresholds of the stage division threshold, a hysteresis interval is formed to eliminate frequent jumps in the normalized control sequence at the stage critical point, determining the pressure control stage slice result to which the current moment belongs. The normalized state quantity at the current moment is obtained from the normalized control sequence, which is a dimensionless pressure value and temperature value. Differentiation is performed on the normalized control sequence at the current moment to obtain the change trend quantity, which represents the pressure change rate and temperature change rate. Historical prior features matching the current pressure control stage slice result are retrieved from historical anomaly records and historical control result data. Historical prior features include historical disturbance patterns and historical control gains. The pressure control stage slice result, normalized state quantity, change trend quantity, process formula data, furnace loading parameter data, and historical prior features are spliced ​​together along the feature dimension to form a synchronous control frame.

[0036] It should be noted that the stage division threshold is set comprehensively based on the theoretical parameter values ​​of different pressure control stages in the process formula data, combined with the stable operating range in the historical control result data; the exemplary value range is between 0 and 1 after normalization, such as 0.2 to 0.9; the value is based on the ratio of the target parameters (such as temperature and pressure) of each stage in the process formula data to the equipment limit parameters, and a certain safety margin needs to be reserved to avoid false triggering; The rising threshold of the stage division threshold is set based on the stage division threshold, plus a hysteresis bias value (determined based on the statistical results of the historical fluctuation amplitude of the normalized control sequence in the neighborhood of the stage division threshold); the exemplary value range is higher than the stage division threshold, for example, when the stage division threshold is 0.5, the rising threshold can be set to 0.55; the value is based on the noise amplitude of the normalized control sequence during stage switching, and it is necessary to ensure that the rising threshold is higher than the noise peak value to prevent jitter; The drop threshold for the stage division threshold is set based on the stage division threshold minus a hysteresis bias value; the exemplary value range is lower than the stage division threshold, for example, when the stage division threshold is 0.5, the drop threshold can be set to 0.45; the value is based on the noise amplitude of the normalized control sequence during stage switching, and it is necessary to ensure that the drop threshold is lower than the noise valley value to prevent jitter.

[0037] S2. Determine the pressure control stage corresponding to the current control cycle based on the synchronization control frame, and write the pressure control constraints corresponding to the current pressure control stage into the stage control ledger to form a stage control ledger entry.

[0038] S2.1: Stage pressure control constraints include target pressure curve, allowable fluctuation band, stage switching window, and control constraint priority.

[0039] It should be noted that the target pressure curve is based on the ideal pressure change trajectory set in each stage of the process formula data; its function is to provide a reference path for vacuum furnace control and guide the pressure timing data to change according to the predetermined process requirements. The allowable fluctuation range is the upper and lower limit deviation range set around the target pressure curve; its function is to define the qualified range of pressure control. When the pressure timing data is within this range, it is determined that the process is stable; if it exceeds this range, adjustment is triggered. The stage switching window is the control stage transition judgment area determined by combining the rising and falling thresholds of the stage division threshold; its function is to use the hysteresis constraint mechanism to eliminate data jumps at critical points and ensure that the control stage slice results only switch smoothly when specific conditions are met. Control constraint priority is the execution order of various control indicators set based on the material characteristics and historical anomaly records in the furnace loading parameter data; its function is to determine the priority of ensuring the control effect of key indicators (such as safety or core process parameters) when multimodal data conflicts or resources are limited.

[0040] S2.2: Based on the stage slice results, normalized state variables and trend variables in the synchronous control frame, and combined with the stage start and end conditions in the process formula data, determine the pressure control stage corresponding to the current control cycle; Specifically, the stage slice results in the synchronous control frame are analyzed to obtain the candidate pressure control stage identifier for the current moment. The normalized state quantity in the synchronous control frame is compared with the stage start and end conditions in the process formula data. The stage start and end conditions include the starting pressure value and the ending pressure value of the target pressure curve. If the normalized state quantity is between the starting pressure value and the ending pressure value, it is determined that the static access condition is met. The directionality of pressure change is determined based on the change trend quantity in the synchronous control frame. The change trend quantity represents the pressure change rate. If the direction of the change trend quantity is consistent with the evolution direction of the target pressure curve, it is determined that the dynamic evolution condition is met. Only when both the static access condition and the dynamic evolution condition are met simultaneously, the candidate pressure control stage identifier is determined as the pressure control stage corresponding to the current control cycle.

[0041] S2.3: Write the target pressure curve, allowable fluctuation range, stage switching window and control constraint priority corresponding to the pressure control stage into the preset stage control ledger to form stage control ledger entries.

[0042] Specifically, based on the process formula data corresponding to the pressure control stage, the path parameters of the target pressure curve are extracted. The path parameters include the trajectory shape and rate of change of pressure. Based on the path parameters of the target pressure curve, the upper and lower limits of the allowable fluctuation band are determined. The upper and lower limits define the qualified range of pressure control. Combined with the rising and falling thresholds of the stage division threshold, the judgment area of ​​the stage switching window is determined. The judgment area is used to eliminate data jumps at the stage critical point. Based on the material characteristics in the furnace loading parameter data, the execution order of the control constraint priority is determined. The execution order specifies the priority order of pressure control and temperature control. The current control cycle number, the current pressure control stage, the path parameters of the target pressure curve, the upper and lower limits of the allowable fluctuation band, the judgment area of ​​the stage switching window, and the execution order of the control constraint priority are packaged and written into the preset stage control ledger to form a stage control ledger entry.

[0043] It should be noted that the stage control ledger refers to a logical recording unit or data storage structure constructed based on process formula data, used for the structured storage of key control parameters such as target pressure curves, allowable fluctuation bands, stage switching windows, and control constraint priorities corresponding to the pressure control stage.

[0044] Figure 5This diagram illustrates the local changes between the target pressure curve, the actual pressure response, and the next time-domain pressure trajectory within the phase switching window. The blue curve represents the target pressure curve, the orange curve represents the actual pressure response, the green curve represents the next time-domain pressure trajectory, and the light blue shaded area represents the phase switching window. As shown in the diagram, during the phase switching process, the actual pressure response experiences a significant trough around 900s, while the next time-domain pressure trajectory lies between the target pressure curve and the actual pressure response. This allows the diagram to reflect the downward and upward trends of the actual pressure in advance, demonstrating good process tracking and transition adaptation capabilities during the phase switching process.

[0045] S3. By combining the synchronization control frame and the stage control ledger entry, identify the disturbance type, disturbance intensity and disturbance confidence corresponding to the current pressure deviation, and form a disturbance control ledger entry.

[0046] S3.1: Combining the actual pressure data in the synchronous control frame with the target pressure curve and allowable fluctuation band in the stage control ledger entries, construct the stage normalized pressure deviation flow and determine the disturbance window corresponding to the current pressure deviation. Specifically, the process involves reading the timestamp from the synchronization control frame and the actual pressure data collected by the pressure sensor. Based on the timestamp in the synchronization control frame, linear interpolation is performed on the target pressure curve of the stage control ledger entry to extract the baseline pressure value of the target pressure curve at the corresponding moment. The difference between the actual pressure data and the baseline pressure value of the target pressure curve is used as the original pressure deviation, which represents the instantaneous difference between the actual pressure and the target trajectory. The upper and lower limits of the allowable fluctuation band in the stage control ledger entry are obtained, and a normalized denominator interval is constructed. If the original pressure deviation is positive, the ratio of the original pressure deviation to the upper limit of the allowable fluctuation band is calculated. If the original pressure deviation is negative, the ratio of the original pressure deviation to the lower limit of the allowable fluctuation band is calculated. This ratio is used to transform the original pressure deviation into a dimensionless stage-normalized pressure deviation flow, which represents the relative degree of deviation of the actual pressure from the target trajectory. Based on the absolute value of the stage-normalized pressure deviation flow, it is mapped to a preset disturbance level interval to determine the disturbance window corresponding to the current pressure deviation. The disturbance window is used to define the severity level of the pressure fluctuation.

[0047] It should be noted that the specific steps for setting the disturbance level range include: extracting the upper and lower limit deviation values ​​of the allowable fluctuation band as the benchmark range; classifying the fluctuation level based on the degree of process quality impact corresponding to different fluctuation amplitudes in historical control result data; and mapping the fluctuation level to the benchmark range to form the disturbance level range. The exemplary value range is usually normalized 0 to 1.5, where 0 to 0.5 corresponds to slight fluctuation, 0.5 to 1.0 corresponds to moderate fluctuation, and 1.0 to 1.5 corresponds to severe fluctuation. The value is based on the qualified pressure fluctuation range defined by the upper and lower limit deviation values ​​of the allowable fluctuation band, and the product defect rate or process abnormality probability caused by exceeding this range in historical control result data, to ensure that the disturbance level range can accurately reflect the actual impact of pressure fluctuation on process stability.

[0048] S3.2: Based on the disturbance window, construct corresponding disturbance assumptions around the pump group, valve, temperature rise and backfill respectively, and generate the counterfactual repair pressure trajectory corresponding to each disturbance assumption by combining historical control result data; Specifically, the severity level of pressure fluctuations defined by the disturbance window is analyzed. The severity level of pressure fluctuations represents the degree to which the current pressure deviates from the target trajectory. Based on the severity level of pressure fluctuations, corresponding disturbance hypotheses are constructed for the four execution links: pump set, valve, heating, and backfilling. The disturbance hypotheses include the abnormal pumping speed hypothesis, valve opening deviation hypothesis, heating rate lag hypothesis, and insufficient backfilling flow hypothesis. Historical control result data is retrieved, and the abnormal pumping speed hypothesis, valve opening deviation hypothesis, heating rate lag hypothesis, and insufficient backfilling flow hypothesis are matched with the pressure response records in the historical control result data. The historical corrected pressure change trend when the hypothesis conditions are met is extracted. Based on the historical corrected pressure change trend and the severity level of pressure fluctuations defined by the disturbance window, the ideal pressure evolution path of the execution link under the state of no abnormality is deduced, thereby generating the counterfactual repair pressure trajectory corresponding to each disturbance hypothesis. The counterfactual repair pressure trajectory represents the expected pressure recovery curve after eliminating the disturbance factors.

[0049] Figure 7 This is used to illustrate the relationship between the target pressure curve, actual pressure data, and counterfactual repair pressure trajectory within the perturbation window. The blue curve represents the target pressure curve, the orange curve represents the actual pressure data, and the green curve represents the counterfactual repair pressure trajectory. As can be seen from the figure, the actual pressure data is generally lower than the target pressure curve, while the counterfactual repair pressure trajectory is located between the two and is closer to the target pressure curve. It can form a reasonable repair trend after the perturbation occurs, thus demonstrating good perturbation adaptation and deviation correction capabilities.

[0050] S3.3: Combine the counterfactual repair pressure trajectory corresponding to each perturbation hypothesis with the actual pressure data, target pressure curve, and allowable fluctuation band to calculate the perturbation evidence quality value corresponding to each perturbation hypothesis. The expression is as follows: ; In the formula, Indicates the first The quality value of the perturbation evidence corresponding to each perturbation hypothesis; Indicates the perturbation hypothesis index; This indicates the total number of sampling points within the calculation window; This represents the reciprocal of the total number of sampling points within the calculation window; Indicates the index of the time sampling moment; Indicates the first Actual pressure data at any given moment; Indicates the first The counterfactual repair pressure trajectory corresponding to the perturbation hypothesis is in the first... The pressure value at any given moment; This indicates the half-width value of the allowed fluctuation band; Indicates the first The pressure value of the target pressure curve at any given time.

[0051] S3.4: Perform conflict suppression fusion on the perturbation evidence quality values ​​corresponding to each perturbation hypothesis to determine the perturbation type, perturbation intensity, and perturbation confidence level corresponding to the current pressure deviation; Specifically, based on the quality value of the disturbance evidence corresponding to each disturbance hypothesis, the orthogonal sum rule in Dempster-Shafer evidence theory is used to perform conflict suppression fusion on the quality value of the disturbance evidence corresponding to each disturbance hypothesis. The composite confidence assignment value of each of the pump group disturbance hypothesis, valve disturbance hypothesis, temperature rise disturbance hypothesis, and backfilling disturbance hypothesis is calculated. The hypothesis with the largest composite confidence assignment value is selected as the disturbance type corresponding to the current pressure deviation. The composite confidence assignment value corresponding to the disturbance type is determined as the disturbance confidence. The disturbance intensity is determined based on the difference between the counterfactual repair pressure trajectory and the actual pressure data according to the disturbance type corresponding to the current pressure deviation.

[0052] ; ; In the formula, Indicates the first The composite confidence assignment values ​​corresponding to each perturbation hypothesis; Indicates the perturbation hypothesis index; Indicates the coefficient of conflict of evidence; Indicates the normalization factor; This indicates the perturbation hypothesis supported by the first set of evidence (e.g., calculations based on pressure trajectory repair errors); Indicates the index of the hypothesis supported by the first set of evidence; This indicates the perturbation hypothesis supported by the second set of evidence (e.g., calculations based on the degree of deviation from the fluctuation band); Indicates the index of the hypothesis supported by the second set of evidence; The identification framework contains a set of all mutually exclusive and exhaustive perturbation hypotheses (i.e., the set of four hypotheses: pump set, valve, heating, and backfill). Indicates the first A set of evidence (e.g., based on pressure trajectory repair error characteristics) supports the hypothesis. The basic probability allocation value (i.e., the normalized form of the perturbation evidence quality value); Indicates the index of the first set of evidence; Indicates the first A set of evidence (e.g., based on fluctuation band deviation characteristics) supports the hypothesis. The basic probability allocation value (i.e., the normalized form of the perturbation evidence quality value); Indicates the index of the second set of evidence; Representing the empty set means and This points to different perturbation hypotheses, meaning the two sets of evidence conflict.

[0053] S3.5: Write the disturbance type, disturbance intensity, and disturbance confidence level, along with the current control cycle number, current control stage, and current disturbance window, into the preset disturbance control ledger to form a disturbance control ledger entry.

[0054] Specifically, the disturbance type, disturbance intensity, and disturbance confidence level, along with the current control cycle number, current control stage, and current disturbance window, are written into a preset disturbance control ledger. The preset disturbance control ledger integrates the received disturbance type, disturbance intensity, and disturbance confidence level, along with the current control cycle number, current control stage, and current disturbance window, into a complete record, which is the disturbance control ledger entry.

[0055] It should be noted that the preset disturbance control ledger refers to a record carrier used to store control process data such as disturbance type, disturbance intensity, disturbance confidence, current control cycle number, current control pressure stage, and current disturbance window.

[0056] S4. Based on the synchronous control frame, stage control ledger entries and disturbance control ledger entries, extract the pressure response hysteresis corresponding to the pump group, valve, temperature rise and backfill to form a time delay feature set; perform fusion encoding on the time delay feature set to form a pressure propagation time delay fingerprint vector.

[0057] S4.1: Combining synchronous control frames, stage control ledger entries, and disturbance control ledger entries, construct the stage-gated excitation flow and pressure response flow corresponding to the pump group, valves, heating, and backfilling; Specifically, the system reads pump operation status data, valve opening data, temperature timing data, and valve opening change information corresponding to the charging valve from the synchronous control frame, respectively, as the basic excitation sequences for the pump, valve, temperature rise, and backfilling. It then reads the current pressure control stage, target pressure curve, allowable fluctuation band, stage switching window, and control constraint priority from the stage control ledger entries, performs stage gating on each basic excitation sequence, and forms the stage-gated basic excitation sequence. Finally, it reads the disturbance type, disturbance intensity, disturbance confidence level, and current disturbance window from the disturbance control ledger entries, performs disturbance correction on the stage-gated basic excitation sequence, and forms the stage-gated excitation flow corresponding to the pump, valve, temperature rise, and backfilling. Finally, based on the actual pressure data in the synchronous control frame, and combined with the current pressure control stage and the current disturbance window to extract the corresponding time period, a pressure response flow is formed.

[0058] S4.2: Based on the stage-gated excitation flow and pressure response flow, determine the pressure response hysteresis corresponding to the pump group, valve, temperature rise and backfill respectively; Specifically, the corresponding excitation start time is extracted from the stage gating excitation flow corresponding to the pump group, valve, heating and backfilling respectively, and the time when the preset pressure response threshold is reached is extracted from the pressure response flow corresponding to the pump group, valve, heating and backfilling respectively. The time difference between the time when the preset pressure response threshold is reached and the corresponding excitation start time for the pump group, valve, heating and backfilling respectively is determined as the pressure response hysteresis for the pump group, valve, heating and backfilling respectively.

[0059] It should be noted that the pressure response threshold is set based on the steady-state pressure value of the pressure response flow by multiplying the steady-state pressure value by a preset proportional coefficient (e.g., 90% to 95%). An exemplary value range is 90% to 95% of the steady-state pressure value. The basis for this value is to accurately capture the moment when the pressure response flow reaches and stabilizes in its final state, thereby accurately obtaining the pressure response hysteresis.

[0060] Figure 6 This diagram illustrates the pressure response hysteresis distribution characteristics of four types of objects: pump sets, valves, heating, and backfill. As shown in the figure, the overall pressure response hysteresis is the highest for heating, the lowest for valves, and the middle for pump sets and backfill. This indicates that there are significant time lag differences in the effect of different objects on pressure changes, and the time lag characteristics of different objects can be identified. This provides a basis for constructing pressure propagation time lag fingerprint vectors and for multi-object collaborative scheduling.

[0061] S4.3: Based on the pressure response lag of the pump set, valve, heating and backfill, extract the time lag evidence strength, time lag peak width information and the order difference information between objects to form a time lag feature set.

[0062] Specifically, the pressure response hysteresis corresponding to the pump set, valve, heating, and backfill is obtained, and the values ​​of the pressure response hysteresis corresponding to the pump set, valve, heating, and backfill are directly used as the strength of time-delay evidence. The fluctuation range of the pressure response hysteresis corresponding to the pump set, valve, heating, and backfill within a preset time window is extracted, and this fluctuation range is used as the time-delay peak width information. The pressure response hysteresis corresponding to the pump set, valve, heating, and backfill is arranged according to the preset order of action of the objects, and the order difference between the pressure response hysteresis of two adjacent objects is extracted, and this order difference is used as the order difference information between objects. The strength of time-delay evidence, the time-delay peak width information, and the order difference information between objects are combined to form a time-delay feature set.

[0063] It should be noted that the time window is based on the historical fluctuation cycle of the pressure response lag and is set by selecting the shortest time interval that covers the complete fluctuation process of the pressure response lag. The exemplary value range is 1 to 3 times the fluctuation cycle of the pressure response lag. The basis for the value is to fully capture the dynamic change characteristics of the pressure response lag, while avoiding data redundancy due to an excessively long window or feature omission due to an excessively short window. The sequence of actions of objects is based on the physical order of the pump set, valve, heating and backfilling in the process flow. It is set by sorting out the logical order of the triggering instructions of each object in the control flow. An exemplary sequence is usually "pump set - valve - heating - backfilling".

[0064] S4.4: Construct a continuous time-delay propagation graph using the time-delay feature set, and perform memory update on the continuous time-delay propagation graph to form the state of the time-delay propagation graph; Specifically, the time delay evidence strength, time delay peak width information, and order difference information between objects are extracted from the time delay feature set. Pump groups, valves, heating, backfilling, and pressure monitoring points are used as nodes, and time delay peak width information is used as the connection edge attribute value to construct a continuous time delay propagation graph. The node attribute value and connection edge attribute value corresponding to the current control cycle are fused with the historical attribute value to complete the memory update of the continuous time delay propagation graph and form the time delay propagation graph state.

[0065] S4.5: Extract the root time-delay subgraph centered on the pressure node from the time-delay propagation graph state, and perform whole-graph embedding encoding on the root time-delay subgraph to form the pressure propagation time-delay fingerprint vector.

[0066] Specifically, all nodes in the time-delay propagation graph are traversed, and nodes marked as pressure monitoring points in the identification information are selected and identified as pressure nodes. Starting from the pressure node, the process extends outward along the connecting edge to a preset number of hops, extracting the pressure node and its adjacent nodes and connecting edges to form a root time-delay subgraph. The node attribute values ​​and connecting edge attribute values ​​in the root time-delay subgraph are aggregated and arranged according to a preset vector dimension to form a pressure propagation time-delay fingerprint vector.

[0067] It should be noted that the hop count range is based on the influence radius of the pressure node in the time-delay propagation graph state. It is set by selecting the minimum number of hops that can cover the main time-delay propagation path of the pressure node. The example value is 2 hops or 3 hops. The basis for the value is to ensure that the root time-delay subgraph can completely contain the core associated nodes of the pressure node, while avoiding redundancy of the subgraph due to an excessively large range or loss of key features due to an excessively small range. The vector dimension is set based on the information capacity of the node attribute values ​​and connection edge attribute values ​​in the root time-delay subgraph. It is set by selecting the minimum dimension value that can fully carry the aggregated attribute value features. Example values ​​are 128 or 256. The basis for this value is to ensure that the pressure propagation time-delay fingerprint vector can fully express the time-delay features of the root time-delay subgraph, while avoiding computational redundancy due to excessively high dimension or feature compression loss due to excessively low dimension.

[0068] S5. Perform pressure prediction and compensation control based on the stage control ledger entries, disturbance control ledger entries, and pressure propagation time delay fingerprint vector, generate prediction results and compensation control instruction set; execute the compensation control instruction set, compare the actual pressure response with the prediction results and correct the deviation, and write back the corrected results.

[0069] S5.1: Combine the synchronization control frame, stage control ledger entry, disturbance control ledger entry, and pressure propagation delay fingerprint vector to construct the compensation prediction context packet and timing release constraint corresponding to the current control cycle; Specifically, the time reference information, normalized state quantity, and trend quantity of the current control cycle are extracted from the synchronization control frame; the current control cycle number, current pressure control stage, target pressure curve, allowable fluctuation band, stage switching window, and control constraint priority are extracted from the stage control ledger entries; and the disturbance type, disturbance intensity, disturbance confidence, and current disturbance window are extracted from the disturbance control ledger entries. The time reference information, normalized state quantity, trend quantity, current pressure control stage, target pressure curve, allowable fluctuation band, stage switching window, control constraint priority, disturbance type, disturbance intensity, disturbance confidence, current disturbance window, and pressure propagation time delay fingerprint vector are encapsulated to form the compensation prediction context packet corresponding to the current control cycle; the allowable release interval is determined based on the stage switching window; the disturbance action interval is determined based on the current disturbance window; and the ordering relationship of each compensation action is determined based on the control constraint priority to form the timing release constraint corresponding to the current control cycle.

[0070] S5.2: Perform rolling pressure prediction on the compensation prediction context packet and timing release constraints to obtain the prediction result corresponding to the current control cycle; Specifically, the time reference information, normalized state quantity, trend quantity, target pressure curve, allowable fluctuation band, stage switching window, control constraint priority, disturbance type, disturbance intensity, disturbance confidence, current disturbance window, and pressure propagation delay fingerprint vector are extracted from the compensation prediction context package. The time reference information is used as the prediction time axis, the normalized state quantity and trend quantity are used as the current state input, the target pressure curve, allowable fluctuation band, stage switching window, and current disturbance window are used as prediction boundary conditions, and the disturbance type, disturbance intensity, disturbance confidence, and pressure propagation delay fingerprint vector are used as prediction correction parameters. Combined with the allowable release time point determined by the timing release constraint, the pressure change trend is rolled forward on the prediction time axis according to a preset step size to generate the prediction result corresponding to the current control cycle.

[0071] Specifically, the preset step size is set based on the minimum change period of the time delay coefficient in the pressure propagation time delay fingerprint vector. The example value is 1 / 5 to 1 / 10 of the minimum change period of pressure propagation. The basis for this value is to ensure that the subtle changes of the time delay coefficient can be captured when rolling the pressure change trend, while avoiding the decrease in prediction accuracy due to an excessively large step size or the computational redundancy due to an excessively small step size.

[0072] It should be noted that the prediction results include at least the next time-domain pressure trajectory; the next time-domain pressure trajectory is used to guide the coordinated action of pump sets, valves and heating equipment in the next time domain to ensure that the pressure change trend is consistent with the target state.

[0073] S5.3: Based on the prediction results, generate multiple sets of compensation candidate sequences, and combine them with stage control pressure constraints, disturbance type, disturbance intensity, disturbance confidence and pressure propagation time delay fingerprint vector to screen each compensation candidate sequence and determine the target compensation candidate sequence; Specifically, the process involves reading the next time-domain pressure trajectory from the prediction results to determine the direction, segment, and duration of deviation of the predicted pressure relative to the target pressure curve within the current control cycle. Based on the direction, segment, and duration of deviation, multiple candidate compensation sequences are generated, differing in compensation magnitude, release time, and sequence of actions, focusing on the correction direction, timing, and magnitude of the current pressure deviation. The process also involves reading the stage control pressure constraints, verifying each candidate compensation sequence, and eliminating those that would cause the predicted pressure to exceed the allowable fluctuation range, enter the stage switching window conflict zone, or violate control constraint priority after compensation. Furthermore, the process involves reading the disturbance type and intensity, eliminating candidate compensation sequences that are inconsistent with the current disturbance direction or whose compensation intensity is inappropriate. Finally, based on the disturbance reliability, the remaining candidate compensation sequences are further refined. The ranking process is as follows: when the perturbation confidence level is within a preset high confidence range, priority is given to retaining the compensation candidate sequences that directly correct the current perturbation type; when the perturbation confidence level is within a preset low confidence range, priority is given to retaining the compensation candidate sequences that have the least overlap with the stage switching window and the fewest conflicts with control constraint priorities. The pressure propagation delay fingerprint vector is read, and the release time, the combination relationship between candidate actions, and the order of actions of each compensation candidate sequence are checked to see if they are consistent with the current pressure propagation delay relationship. Compensation candidate sequences that would cause release disorder, action conflict, or delay mismatch are eliminated. The remaining compensation candidate sequences are ranked in order of compliance with the allowable fluctuation band and stage switching window, correspondence with the perturbation type and intensity, and matching with the pressure propagation delay relationship. The group with the highest ranking is determined as the target compensation candidate sequence.

[0074] It should be noted that the high confidence interval and low confidence interval are set based on the statistical distribution of disturbance confidence in historical control result data, combined with experience in the vacuum furnace pressure control process, and are used to characterize whether the current disturbance judgment result is in a high confidence state or a low confidence state, respectively; for example, the high confidence interval is 0.75 to 1.00, and the low confidence interval is 0 to 0.40; the basis for the value is that the disturbance confidence is a normalized confidence value, the larger the value, the more reliable the current disturbance judgment result, and the smaller the value, the higher the uncertainty of the current disturbance judgment result.

[0075] S5.4: Based on the target compensation candidate sequence, target pressure curve, allowable fluctuation band and stage switching window, perform safety constraint correction to obtain the compensation control instruction set.

[0076] Specifically, the initial compensation magnitude and initial compensation timing are extracted from the target compensation candidate sequence. The initial compensation magnitude is compared with the target pressure curve, and the initial compensation magnitude is adjusted according to the pressure peak of the target pressure curve to obtain the corrected compensation magnitude. The corrected compensation magnitude is compared with the upper and lower limits of the allowable fluctuation band. If the corrected compensation magnitude exceeds the upper and lower limits of the allowable fluctuation band, the corrected compensation magnitude is trimmed to the boundary value of the allowable fluctuation band to obtain the boundary corrected compensation magnitude. The time range of the stage switching window is read, and the execution time of the boundary corrected compensation magnitude is compared with the time range of the stage switching window. If the execution time of the boundary corrected compensation magnitude is within the time range of the stage switching window, the execution time of the boundary corrected compensation magnitude is shifted outside the time range of the stage switching window to obtain the final compensation magnitude and final compensation timing, which are then encapsulated into control instructions to form a compensation control instruction set.

[0077] S5.5: The compensation control instruction set includes pump start / stop adjustment amount, pump speed adjustment amount, throttle valve opening adjustment amount, vacuum valve switching amount, bypass valve switching amount, charging valve opening adjustment amount, and heating rate correction amount.

[0078] It should be noted that the pump start-stop adjustment range is used to change the number of pumps involved in the operation to provide macroscopic pressure and flow support, and to quickly respond to large pressure fluctuations. The pump speed adjustment is used to fine-tune the output flow of the pump to smooth pressure fluctuations and achieve precise and continuous control of pressure changes. The throttle valve opening adjustment is used to change the cross-sectional area of ​​the pipeline to generate local resistance, thereby precisely limiting or releasing local pressure and preventing pressure overshoot; Vacuum valve switching is used to establish or disconnect a vacuum path to quickly reduce internal pressure and meet the process requirements of low-pressure or vacuum environments. The bypass valve switching function is used to transfer the flow of the medium through the bypass pipeline when the pressure in the main passage is abnormal or when diversion is required, so as to play a protective role of diverting pressure relief or maintaining minimum flow circulation. The opening adjustment of the inflation valve is used to replenish the gas medium to increase the pressure, to compensate for the pressure drop caused by leakage or disturbance, and to maintain pressure stability. The heating rate correction is used to address the thermodynamic coupling relationship between temperature and pressure. By controlling the rate of temperature change, it indirectly corrects the pressure change trend and eliminates control errors caused by thermal hysteresis.

[0079] S5.6: Execute the compensation control instruction set and collect the actual pressure response, compare the deviation between the actual pressure response and the prediction result, and perform online correction on the prediction result corresponding to the next control cycle based on the deviation result to form the correction result; Specifically, the final compensation amplitude and timing sequence in the compensation control instruction set are analyzed. Based on the final compensation timing sequence, the control action corresponding to the final compensation amplitude is executed at the corresponding time point. The real-time pressure value after the execution of the control action is collected by the pressure sensor to form the actual pressure response. The prediction result corresponding to the current control cycle is obtained, and the real-time pressure value of the actual pressure response is compared with the pressure value at the same time point in the prediction result corresponding to the current control cycle to obtain the deviation result. The prediction result corresponding to the next control cycle is obtained, and the correction direction is determined based on the deviation result. The pressure value at each time point in the prediction result corresponding to the next control cycle is adjusted according to the correction direction to obtain the corrected pressure value. The corrected pressure values ​​are arranged in chronological order to form the correction result.

[0080] S5.7: Write and write back the compensation control instruction set, actual pressure response, prediction results, deviation results, and correction results.

[0081] Specifically, the system acquires the compensation control instruction set, actual pressure response, prediction results, deviation results, and correction results. It then creates a control record data structure and fills the instruction, response, prediction, deviation, and correction fields of the control record data structure with the final compensation magnitude and timing, the real-time pressure value of the actual pressure response, the pressure value of the prediction result corresponding to the current control cycle, the judgment value of the deviation result, and the corrected pressure value of the correction result. Finally, it writes the instruction field back to the status register of the actuator, the response field back to the display cache of the monitoring interface, the prediction field back to the data source of the trend analysis chart, the deviation field back to the record list of the abnormal alarm module, and the correction field back to the initial value of the prediction result corresponding to the next control cycle.

[0082] In summary, this invention achieves continuous characterization of the pressure evolution trend in the current control cycle, orderly constraints on the timing of compensation action release, and unified scheduling of the coordinated relationship between pumps, valves, heating, and backfilling by executing pressure prediction and compensation control based on stage control ledger entries, disturbance control ledger entries, and pressure propagation time delay fingerprint vectors. This results in an output compensation control command set with clear stage-specificity, disturbance adaptability, and time delay matching, and provides a stable basis for correcting deviations between actual pressure response and prediction results, as well as for online write-back.

[0083] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A vacuum furnace AI pressure control method based on multimodal data fusion, characterized in that, include: Collect multimodal data within the current control cycle of the vacuum furnace, perform time alignment, anomaly removal, dimension normalization, and stage slicing to form a synchronous control frame; Based on the synchronous control frame, the pressure control stage corresponding to the current control cycle is determined, and the stage pressure control constraints corresponding to the current pressure control stage are written into the stage control ledger to form a stage control ledger entry; the stage pressure control constraints include the target pressure curve, the allowable fluctuation band, the stage switching window, and the control constraint priority. By combining synchronization control frames and stage control ledger entries, the disturbance type, disturbance intensity, and disturbance confidence level corresponding to the current pressure deviation are identified, and disturbance control ledger entries are formed. The steps are as follows: By combining the actual pressure data in the synchronous control frame with the target pressure curve and allowable fluctuation band in the stage control ledger entries, a stage-normalized pressure deviation flow is constructed to determine the disturbance window corresponding to the current pressure deviation. Based on the disturbance window, corresponding disturbance hypotheses are constructed around the pump group, valve, temperature rise and backfill respectively. Combined with historical control result data, counterfactual repair pressure trajectories corresponding to each disturbance hypothesis are generated. By combining the counterfactual repair pressure trajectory corresponding to each perturbation hypothesis with the actual pressure data, target pressure curve and allowable fluctuation band, the perturbation evidence quality value corresponding to each perturbation hypothesis is calculated. Conflict suppression fusion is performed on the perturbation evidence quality values ​​corresponding to each perturbation hypothesis to determine the perturbation type, perturbation intensity, and perturbation confidence level corresponding to the current pressure deviation; The disturbance type, disturbance intensity, and disturbance confidence level, along with the current control cycle number, current control stage, and current disturbance window, are written into the preset disturbance control ledger to form disturbance control ledger entries; Based on the synchronous control frame, stage control ledger entries, and disturbance control ledger entries, the pressure response hysteresis corresponding to the pump group, valve, temperature rise, and backfill is extracted to form a time delay feature set; the time delay feature set is fused and encoded to form a pressure propagation time delay fingerprint vector. Pressure prediction and compensation control are performed based on stage control ledger entries, disturbance control ledger entries, and pressure propagation time delay fingerprint vectors, generating prediction results and compensation control instruction sets; the compensation control instruction sets are executed, the actual pressure response is compared with the prediction results and corrected, and the corrected results are written back.

2. The vacuum furnace AI pressure control method based on multimodal data fusion as described in claim 1, characterized in that, The steps for forming the synchronization control frame are as follows: The multimodal data includes pressure time-series data, temperature time-series data, valve opening data, pump unit operating status data, process formula data, furnace charging parameter data, historical anomaly records, and historical control result data. Time offset correction and unified time base projection are performed on the multimodal data to obtain the alignment control sequence; The alignment control sequence is combined with historical control result data and historical anomaly records to remove abnormal data. Robust dimensional normalization is then performed based on process formula data and furnace charging parameter data to obtain the normalized control sequence. The stage slice with hysteresis constraint is executed based on the normalized control sequence, and the stage slice results are spliced ​​with the normalized state variables, trend variables, process formula data, furnace loading parameter data, historical anomaly records and historical control result data corresponding to the normalized control sequence to form a synchronous control frame.

3. The vacuum furnace AI pressure control method based on multimodal data fusion as described in claim 1, characterized in that, The steps for controlling ledger entries during the formation phase are as follows: Based on the stage slice results, normalized state variables, and trend variables in the synchronous control frame, and combined with the stage start and end conditions in the process formula data, the pressure control stage corresponding to the current control cycle is determined. The target pressure curve, allowable fluctuation range, stage switching window, and control constraint priority corresponding to the pressure control stage are written into the preset stage control ledger to form stage control ledger entries.

4. The vacuum furnace AI pressure control method based on multimodal data fusion as described in claim 1, characterized in that, The steps for forming the time-delay feature set are as follows: By combining synchronous control frames, stage control ledger entries, and disturbance control ledger entries, the stage-gated excitation flow and pressure response flow corresponding to pump sets, valves, heating, and backfilling are constructed. Based on the stage-gated excitation flow and pressure response flow, the pressure response hysteresis corresponding to the pump group, valve, temperature rise and backfill is determined respectively; Based on the pressure response hysteresis corresponding to the pump set, valve, temperature rise and backfill, the strength of time hysteresis evidence, the peak width of time hysteresis and the order difference between objects are extracted to form a time hysteresis feature set.

5. The vacuum furnace AI pressure control method based on multimodal data fusion as described in claim 1 or 4, characterized in that, The steps for forming the pressure propagation time-delay fingerprint vector are as follows: A continuous time-delay propagation graph is constructed using a time-delay feature set, and a memory update is performed on the continuous time-delay propagation graph to form the state of the time-delay propagation graph; Extract the root time-delay subgraph centered on the pressure node from the state of the time-delay propagation graph, and perform whole-graph embedding encoding on the root time-delay subgraph to form the pressure propagation time-delay fingerprint vector.

6. The vacuum furnace AI pressure control method based on multimodal data fusion as described in claim 5, characterized in that, The steps for generating the prediction results and compensation control instruction set are as follows: By combining the synchronization control frame, stage control ledger entries, disturbance control ledger entries, and pressure propagation delay fingerprint vector, the compensation prediction context packet and timing release constraints corresponding to the current control cycle are constructed. Perform rolling pressure prediction on the compensation prediction context packet and timing release constraints to obtain the prediction result corresponding to the current control cycle; Based on the prediction results, multiple sets of compensation candidate sequences are generated. Combined with stage control pressure constraints, disturbance type, disturbance intensity, disturbance confidence and pressure propagation time delay fingerprint vector, each compensation candidate sequence is screened to determine the target compensation candidate sequence. Based on the target compensation candidate sequence, target pressure curve, allowable fluctuation band, and stage switching window, safety constraint correction is performed to obtain the compensation control instruction set.

7. The vacuum furnace AI pressure control method based on multimodal data fusion as described in claim 6, characterized in that, The prediction results include at least the next time-domain pressure trajectory; The compensation control command set includes pump start / stop adjustment amount, pump speed adjustment amount, throttle valve opening adjustment amount, vacuum valve switching amount, bypass valve switching amount, air charging valve opening adjustment amount, and heating rate correction amount.

8. The vacuum furnace AI pressure control method based on multimodal data fusion as described in claim 1 or 7, characterized in that, The steps for writing back the correction result are as follows: The system executes the compensation control instruction set and collects the actual pressure response. It then compares the actual pressure response with the predicted results to determine the deviation. Based on the deviation results, it performs online correction on the predicted results for the next control cycle to generate the correction results. The compensation control instruction set, actual pressure response, prediction results, deviation results, and correction results are written and written back.

Images