A vacuum swing valve intelligent control method and system based on a hybrid model

Abstract

The application belongs to the field of valve control, and specifically discloses a vacuum swing valve intelligent control method and system based on a hybrid model. The hybrid prediction model is used to realize early prediction and comprehensive compensation of the cavity pressure change, so that good dynamic response and steady-state accuracy can be maintained when the working condition changes, and response lag and control fluctuation can be reduced. Through equivalent flow conductance / equivalent pumping speed modeling and residual compensation mechanism across the working condition, the prediction accuracy and control robustness are improved, the dependence on artificial experience parameter adjustment is reduced, and the consistency and automation degree of process control are improved. Through the credibility gating, amplitude limiting and physical constraint mechanism, the risk of non-physical output of data-driven compensation is inhibited under complex or abnormal working conditions, and the degenerative protection strategy is automatically switched when the solution fails or the credibility is insufficient, so that the system stability and safety are improved. By introducing the opening change suppression term into the objective function and applying the opening constraint, the valve plate frequent action and mechanical impact are reduced, which helps to improve the pressure stability.

Description

Technical Field

[0001] This application belongs to the field of valve control, and more specifically, relates to a smart control method and system for vacuum swing valves based on a hybrid model. Background Technology

[0002] Vacuum swing valves, as a key component of semiconductor coating equipment, control the pressure in the process chamber by precisely adjusting the swing angle of the valve core. Currently, the commonly used PID control method has significant shortcomings in practical applications: due to the nonlinear changes in gas flow characteristics during the coating process, traditional PID control algorithms struggle to achieve fast and accurate responses, often exhibiting adjustment lag. When the system detects pressure deviations, fixed PID parameters cannot adapt to the dynamic changes in different process stages, leading to frequent fluctuations in valve core movement. This not only affects pressure stability but also shortens the valve's lifespan. Furthermore, this control method is highly dependent on the operator's professional experience; manual intervention increases the uncertainty of process control and reduces production efficiency. Existing control schemes are no longer sufficient to meet the high requirements of modern semiconductor manufacturing for process stability and automation.

[0003] To address these issues, existing technologies have proposed several improvements, such as feedforward control, adaptive gain adjustment, or pressure prediction methods based on empirical models. While these methods can improve the dynamic response performance of the system to some extent, most still rely on the current or historical pressure change rate for judgment, or require waiting for the pressure change trend to gradually emerge before effective correction can be made. This makes it difficult to accurately determine the valve opening before the pressure has stabilized. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the purpose of this application is to provide a smart control method and system for vacuum swing valves based on a hybrid model, aiming to solve the problems of slow response, insufficient model accuracy, poor control stability, and strong reliance on parameter tuning and human experience in existing vacuum swing valve pressure control under complex working conditions.

[0005] To achieve the above objectives, in a first aspect, this application provides a smart control method for a vacuum swing valve based on a hybrid model. The vacuum swing valve is installed between a vacuum chamber and a vacuum pump. The smart control method includes: real-time acquisition of chamber-side pressure, chamber temperature, equivalent inlet load, pump-side pumping speed, and swing valve opening; determining flow pattern equivalent variables based on the current chamber-side pressure, chamber temperature, and swing valve opening; constructing a continuously differentiable flow pattern weighting function based on the flow pattern equivalent variables; smoothly fusing the molecular flow conductance function and viscous flow conductance function of the swing valve according to the flow pattern weighting function to obtain a continuous equivalent conductance across flow patterns; calculating the equivalent pumping speed on the chamber side based on the pump-side pumping speed and the continuous equivalent conductance across flow patterns; and establishing a discrete prediction model for the chamber pressure at the next moment based on mass conservation, combining the equivalent pumping speed on the chamber side, the equivalent inlet load on the chamber side, and the equivalent volume of the chamber. The dynamic correlation feature vector is input into a pre-trained lightweight neural network to obtain the equivalent pumping speed residual compensation amount. Based on the difference statistics between the measured cavity pressure on the cavity side and the one-step prediction value of the current moment output by the hybrid model at the previous moment, a confidence gating coefficient is generated. According to the confidence gating coefficient, the equivalent pumping speed residual compensation amount is gated and fused to obtain the compensated equivalent pumping speed. The compensated equivalent pumping speed is used to replace the cavity side equivalent pumping speed in the discrete prediction model of cavity pressure to obtain the hybrid prediction model. Using the hybrid prediction model, a hybrid prediction pressure sequence is generated by rolling over the prediction step size. Based on the hybrid prediction pressure sequence, the objective function in the current moment is constructed. Under the constraints of opening amplitude and opening change rate, constraint optimization is performed to obtain the valve opening command at the next moment. When the confidence is abnormal or the solution fails, the degradation protection strategy is switched to output a safe opening command.

[0006] Preferably, the real-time acquired cavity-side pressure measurement value is subjected to low-pass filtering to obtain the filtered cavity-side pressure. :

[0007] in, This is the original cavity pressure measurement value. Initialize to the pressure measurement value at power-on; Based on the filtered cavity-side cavity pressure, the trend of cavity-side cavity pressure variation is calculated. :

[0008] The sampling period.

[0009] Preferably, based on the Knudsen number criterion or the equivalent pressure interval criterion, the following is generated: S-shaped continuous divertible weighting function:

[0010] in, For a moment Molecular flux weight, For a moment Viscous flow weight, The critical flow pattern is an equivalent variable, determined based on offline experimental calibration. For a moment The manifold equivalent variable, given by time... The equivalent flow channel dimensions and gas physical property parameters corresponding to the cavity side pressure, cavity temperature, and valve plate opening of the swing valve are calculated. This is the smoothing coefficient.

[0011] Preferably, the equivalent pumping speed on the cavity side is calculated based on the pump-side pumping speed and the continuous equivalent conductance of the crossflow pattern, as follows:

[0012] in, For a moment The equivalent pumping speed on the cavity side. For a moment The transflow continuous equivalent conductance, For a moment Pump side pumping speed, For a moment The opening degree of the swing valve plate. For a moment The manifold equivalent variable.

[0013] Preferably, by combining the equivalent pumping speed on the cavity side, the equivalent intake load on the cavity side, and the equivalent volume of the cavity, a discrete prediction model for the cavity pressure at the next moment is established based on mass conservation, as follows:

[0014] in, This is a one-step prediction of the cavity pressure at the next moment. This is the equivalent volume of the cavity; for Equivalent intake load; for The comprehensive perturbation term is used to characterize low-frequency, difficult-to-model equivalent gas load changes. for Equivalent pumping speed on the cavity side.

[0015] Preferably, the pressure control dynamic correlation feature vector includes at least one of the following variables: cavity-side cavity pressure, cavity-side cavity pressure change trend, swing valve opening, equivalent intake load, pump-side pumping speed, cavity temperature, flow pattern equivalent variable, and the one-step prediction residual from the previous moment.

[0016] Preferably, the lightweight neural network is a feedforward fully connected network, deployed on the embedded controller via fixed-point quantization, with no more than two hidden layers and no more than 5000 network parameters, as detailed below:

[0017] in, for The equivalent pumping rate residual compensation amount of the lightweight neural network output. This is a nonlinear mapping function implemented by a lightweight neural network, used to establish the correspondence between the input feature vector and the equivalent pumping rate residual compensation amount. For a moment The input feature vector, These are parameters for a lightweight neural network.

[0018] Preferably, the compensation amount of the equivalent pumping speed residual is limited, and a non-negative constraint, an upper bound constraint limited by the pump-side pumping speed, and a rate of change constraint are applied to the compensated equivalent pumping speed.

[0019] Preferably, the one-step prediction error The gating coefficient is generated by comparing it with a preset threshold. :

[0020]

[0021]

[0022] in, , Separately control the lower and upper threshold values. for Error statistics The length of the error statistics window. For the first sliding window The prediction error at each time point, Indicates time The actual cavity pressure state quantity. This represents the predicted cavity pressure value output by the hybrid model at the previous time step. Indicates time The actual pressure is the prediction error relative to the previous cycle's forecast results.

[0023] To achieve the above objectives, in a second aspect, this application provides a vacuum swing valve intelligent control system based on a hybrid model, including a memory and one or more processors; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code including computer instructions; the one or more processors call the computer instructions to cause the system to execute the intelligent control method as described in the first aspect.

[0024] It is understandable that the beneficial effects of the second aspect mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here.

[0025] Overall, the technical solutions conceived in this application have the following beneficial effects compared with the prior art: (1) Compared with traditional fixed parameter PID control, this application proposes a vacuum swing valve intelligent control method based on a hybrid model. By using a hybrid predictive model, it can achieve advance prediction and comprehensive compensation for pressure changes, maintain good dynamic response and steady-state accuracy when operating conditions change, nonlinearity increases or disturbances occur, and reduce response lag and control fluctuations.

[0026] (2) This application proposes a smart control method for vacuum swing valve based on a hybrid model. By continuously modeling equivalent flow conductance / equivalent pumping speed across operating conditions and using a residual compensation mechanism, the prediction accuracy and control robustness are improved, the dependence on manual experience in parameter tuning is reduced, and the consistency and automation of process control are enhanced.

[0027] (3) This application proposes a smart control method for vacuum swing valve based on a hybrid model. Through confidence gating, amplitude limiting and physical constraint mechanisms, it suppresses the non-physical output risk of data-driven compensation under complex or abnormal working conditions, and automatically switches the degradation protection strategy when the solution fails or the confidence is insufficient, thereby improving the stability and security of the system.

[0028] (4) This application proposes a smart control method for vacuum swing valve based on a hybrid model. By introducing an opening change suppression term into the objective function and applying an opening constraint, the frequent action of the valve plate and mechanical shock are reduced, which helps to improve pressure stability.

[0029] In summary, this application proposes a smart control method for vacuum swing valves based on a hybrid model. By introducing a pressure prediction mechanism that integrates a mechanistic model and a data-driven model, and combining it with a credibility-gated and constraint-controlled strategy, it achieves rapid, stable, and high-precision regulation of vacuum chamber pressure. This method outperforms existing technologies in terms of modeling accuracy, control speed, stability, and adaptability, effectively improving the pressure regulation performance of vacuum systems and providing more intelligent and precise control solutions for high-end vacuum process equipment (such as single crystal furnaces, ion beam coating systems, and semiconductor etching equipment). Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the structure of a vacuum swing valve intelligent control system based on a hybrid model provided in an embodiment of this application.

[0031] Figure 2 This is a flowchart of a smart control method for a vacuum swing valve based on a hybrid model, provided in an embodiment of this application. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0033] In this application, the term "and / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A existing alone, A and B existing simultaneously, and B existing alone. In this application, the symbol " / " indicates that the related objects are in an "or" relationship, for example, A / B means A or B.

[0034] In this application, the terms "first" and "second," etc., are used to distinguish different objects, not to describe a specific order of objects. For example, "first response message" and "second response message," etc., are used to distinguish different response messages, not to describe a specific order of response messages.

[0035] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0036] In the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more, for example, multiple processing units means two or more processing units, multiple elements means two or more elements, etc.

[0037] The embodiments of this application are described below with reference to the accompanying drawings.

[0038] like Figure 1As shown, this application provides a vacuum swing valve intelligent control system based on a hybrid model, including: a vacuum chamber, a vacuum swing valve, a pressure sensor, a stepper motor drive module, a controller, a signal processing and computing unit, a mass flow controller, and a vacuum pump; wherein, the vacuum swing valve is installed on the exhaust pipe of the vacuum chamber for regulating the gas pressure in the chamber, and the vacuum pump is connected to the exhaust pipe of the vacuum swing valve for evacuating the vacuum chamber; the pressure sensor is set in the chamber for real-time detection of the chamber pressure and outputting a pressure signal to the controller; the stepper motor drive module is connected to the arc-shaped valve plate of the vacuum swing valve for driving the valve plate in the vacuum swing valve to rotate according to the control signal output by the controller; the controller is signal-connected to the pressure sensor, the drive module, and the mass flow controller for receiving signals from the pressure sensor and outputting control commands to the drive module; the signal processing and computing unit is the computing module of the controller.

[0039] like Figure 2 As shown, this application provides a smart control method for a vacuum swing valve based on a hybrid model, specifically including: S1. Establishment of a mechanism-based discrete prediction model for cavity pressure.

[0040] To adapt to cross-condition operation (including cross-flow pattern intervals, pump speed variations, and intake load variations) during the pressure control process of the swing valve, a mechanism-based discrete prediction model for chamber pressure is constructed. This model is used to predict the pressure at the next moment in each control cycle, serving as the basis for subsequent residual compensation, reliability determination, and constraint optimization control steps. The discrete prediction model includes: acquisition and preprocessing of operating state variables, calculation of continuous equivalent conductance and equivalent pumping speed across flow patterns, and discrete pressure update equations.

[0041] S11. Acquisition and preprocessing of operating status data.

[0042] At the beginning of each control cycle, the controller acquires the pressure measurement value output by the cavity pressure sensor. and the current opening degree of the swing valve plate Equivalent intake load information obtained by converting the flow signal from the mass flow controller (MFC). And pump-side pumping speed related to pump speed and cavity temperature .

[0043] To improve noise immunity, the controller performs low-pass filtering on the pressure measurement values ​​to obtain filtered pressure. The filtering process uses a first-order exponential filter:

[0044] in, This is the original cavity pressure measurement value. The filtered cavity pressure can be initialized to equal the pressure measurement value at power-on, with the subscript... Indicates time Corresponding variables.

[0045] The controller calculates the pressure change trend based on the filtered pressure. :

[0046] The sampling period.

[0047] Define the predictor and with , , , , As input for subsequent modeling and solving.

[0048] S12, Calculation of continuous equivalent conductance and equivalent pumping speed for transflow patterns.

[0049] To avoid discontinuous changes in the equivalent conductance and pumping velocity models caused by changes in flow pattern determination, which could lead to instability in predictions or difficulties in optimization, this application introduces equivalent flow pattern variables. And based on this, a continuously changing manifold weight function is constructed.

[0050] (1) Construction of manifold weight function Molecular current weights are generated based on the Knudsen number criterion or the equivalent pressure interval criterion, where, Furthermore, it is continuously variable and continuously differentiable to avoid model abrupt changes caused by manifold switching.

[0051] Employing a continuously differentiable sigmoid function:

[0052] in, As an equivalent variable to the critical flow pattern, it can be predetermined based on offline experimental calibration results. The equivalent variable for the flow pattern can be obtained from time... The equivalent flow channel dimensions and gas physical property parameters corresponding to the cavity pressure, temperature, valve opening, and other parameters are calculated. This is the smoothing coefficient.

[0053] (2) Parameterization of molecular flow and viscous flow conductance The controller establishes the molecular flow conductance function respectively. .

[0054] Molecular flow conductance depends primarily on the valve geometry and valve opening, and is independent of pressure; it is modeled as a function of the valve opening. Continuous functions:

[0055] The conductance of viscous flow depends not only on the opening degree but also on the current cavity pressure. To correct the applicability of the mechanistic model over a wide range, it is modeled as the product of a geometric factor and pressure:

[0056] in, The parameters to be calibrated are obtained in advance through polynomial fitting or table lookup interpolation. After the parameters are calibrated, the valve opening at the current moment can be used as the reference. and pressure The calculation yields the corresponding working conditions. and .

[0057] (3) Construction and smooth fusion of manifold weights The controller smoothly integrates the molecular flow and viscous flow conductance based on the weights to obtain a continuous equivalent conductance across the flow pattern:

[0058] (4) Calculation of equivalent pumping speed on the cavity side The controller determines the pumping speed on the pump side. This can be obtained by converting the pump frequency / pump speed using the pump characteristic curve or by looking up a table. Based on this, the controller calculates the equivalent pumping speed on the cavity side according to the pump-side pumping speed and the continuous equivalent conductance.

[0059] S13, Discrete prediction equation for cavity pressure.

[0060] The controller establishes a discrete prediction equation for cavity pressure based on mass conservation, which is used to predict the pressure in the next control cycle.

[0061] Preferably, with As the prediction state, construct a discrete prediction model:

[0062] in, This is a one-step prediction of the cavity pressure in the next control cycle, calculated based on the mechanistic model. This is the equivalent volume of the cavity; The equivalent intake load term is preferably calculated by combining the MFC flow signal with the gas state equation, so that its dimensions are consistent with the extraction term. As a comprehensive disturbance / unmodeled term, it is used to characterize the equivalent gas load changes that are low-frequency and difficult to explicitly model. The initial estimate can be obtained offline from historical operating data, and during the control operation, it is slowly corrected according to the deviation between the actual pressure change and the prediction of the mechanism model. The residual structural deviation of the model is corrected by the subsequent neural network residual compensation module.

[0063] Calculate the one-step prediction error based on the mechanism model Used to characterize the model's prediction accuracy:

[0064] This information is intended for use in subsequent reliability assessments, residual compensation gating, or degradation protection strategies.

[0065] S2. Construction of a lightweight neural network residual compensation model.

[0066] After obtaining the output of the discrete prediction master model across operating conditions, in order to further compensate for the prediction deviations caused by non-ideal valve mechanisms (such as friction, hysteresis, dead zone), pump end characteristic drift, process venting and model parameter uncertainties, this application constructs a lightweight neural network residual compensation model to generate an equivalent pumping speed residual compensation amount, and merges it with the master model to form a hybrid prediction model.

[0067] The lightweight neural network has a lightweight structure, preferably a feedforward fully connected network, and its input is a state / exogenous feature vector that is dynamically related to the pressure control:

[0068] in, This is the equivalent pumping rate residual compensation amount output by the lightweight neural network. This is a nonlinear mapping function implemented by the lightweight neural network, used to establish the correspondence between the input feature vector and the equivalent pumping rate residual compensation amount. For a moment The input feature vector, These are the parameters of the neural network.

[0069] During the model training phase, the one-step prediction error of the above-mentioned mechanism model Training labels are used to construct the residual compensation network so that the lightweight neural network learns the equivalent pump rate compensation amount caused by unmodeled dynamics, parameter mismatch, or external perturbations.

[0070] Input feature vector Includes one or more of the following variables:

[0071] To ensure the feasibility and real-time performance of the project, the number of layers, nodes, and parameters of the lightweight neural network must meet preset constraints; for example, the number of network layers must not exceed two hidden layers, and the number of parameters must be lower than a preset threshold, so that the single-cycle inference latency meets the control cycle requirements.

[0072] S21. Physical constraints and amplitude limiting of residual compensation.

[0073] To prevent the neural network output from introducing non-physical compensation under abnormal operating conditions and affecting system stability, the controller compensates for the equivalent pumping speed residual. Apply physical constraints and amplitude limiting.

[0074] The original output of the neural network is clipped to obtain the clipped compensation amount. :

[0075] in, For the amplitude limiting function, This is a preset compensation limit.

[0076] Based on this, the controller constructs the compensated equivalent pumping speed. :

[0077] in, This is the confidence gating coefficient, used to control the degree of participation in neural network compensation.

[0078] To satisfy physical rationality, the controller imposes at least one or more constraints, including: nonnegativity constraints: Upper bound constraint (limited by pump-side pumping speed): Rate of change constraint (suppressing compensatory mutations): The above constraints can be achieved through projection / truncation, that is, projecting the constraint onto a feasible region that satisfies the constraint.

[0079] S22, Generation of hybrid prediction models.

[0080] The controller uses the compensated equivalent pumping velocity instead of that in the main model. The output of a hybrid discrete prediction model is used to predict the pressure at the next time step.

[0081] in, For mixed predicted pressure; when hour, The hybrid prediction model degenerates into a pure mechanism prediction model.

[0082] The controller predicts the step size. Above, the comprehensive disturbance term It can be approximated as a constant or a slow variable, taking Rolling generation of mixed prediction sequences This is used to solve for the valve opening command in subsequent constraint optimization.

[0083] S23. Neural network parameter acquisition and update strategy.

[0084] The controller acquires neural network parameters and uses them for online inference. These parameters can be obtained through offline training, with training data including at least one or more historical operating data, pressure gradient experimental data, step response experimental data, or equipment normal production process log data.

[0085] Offline training aims to minimize prediction error, making the neural network output approximate the residual or equivalent pumping bias of the mechanistic model; for example, the following residual objective is used to construct training labels. :

[0086] Parameter training is achieved by minimizing the loss function:

[0087] in, For regularization terms, , where is the weighting coefficient. If the neural network output is an equivalent pumping rate residual compensation amount, then the training objective can be equivalently set as the pumping rate compensation amount that minimizes the mixed prediction error, and is obtained by minimizing the prediction error in the same way.

[0088] To adapt to pump aging, seal changes, and process drift, the controller performs online fine-tuning and updates of network parameters when the reliability conditions are met. To ensure stability, the online update is subject to upper limits on the learning rate and constraints on the parameter change rate. When the reliability is insufficient or the prediction error is abnormal, the update is stopped and the network parameters are frozen.

[0089] S3, Reliability Gating and Constraint Solving: Swing Valve Opening Command.

[0090] Based on a cross-condition hybrid prediction model, the controller performs a confidence assessment in each control cycle to determine the degree of participation of the neural network residual compensation. Under the condition of satisfying constraints, it solves for the valve opening command of the swing valve, so that the cavity pressure tracks the target pressure and ensures the stable and safe operation of the system. When prediction anomalies, compensation anomalies, or solution failures occur, the system switches to a degradation protection strategy to output safe control commands.

[0091] S31, Credibility determination and gating coefficient generation.

[0092] The controller calculates a confidence gating coefficient based on prediction error, residual compensation output, and consistency determination based on sensor redundancy verification or actuator feedback deviation, which is used to adjust the participation of neural network residual compensation in the hybrid model.

[0093] The controller calculates the prediction error:

[0094] in, Indicates time The actual cavity pressure state quantity. This indicates that the time interval obtained from the hybrid prediction model of the previous control cycle is... Predicted pressure, This indicates the prediction error of the actual pressure at the current moment relative to the prediction result of the previous period.

[0095] And construct error statistics (e.g., sliding window mean square or absolute mean):

[0096] in, The length of the error statistics window. For the first in the sliding window The prediction error at each time point.

[0097] The controller will use the mechanistic model to predict the error in one step. The gating coefficient is generated by comparing it with a preset threshold. The following piecewise function is used to represent it:

[0098] in, , To preset the gate threshold, .

[0099] Optionally, the gating coefficients can be further adjusted based on the consistency of the neural network output or constraint violations. For example, a mandatory command can be applied when any of the following conditions occur. or reduce (1) And it occurs more than the preset number of times; (2) Pumping speed after compensation (3) Trigger the upper bound / non-negative / rate of change constraint projection number more than the preset number; (4) Control the solver to fail to solve or return an infeasible solution in this cycle.

[0100] S32. Constraints and Objective Function Construction.

[0101] The controller is based on a hybrid prediction model at prediction step size Generate predicted pressure sequences And construct the objective function and constraints for solving the opening of the swing valve.

[0102] 1) Prediction Model Using the above hybrid model, rolling generation Step prediction.

[0103]

[0104] 2) Objective function Preferably, the objective function includes at least a pressure tracking term and an opening change suppression term:

[0105] in, To set target pressure, , To control the step size , To predict the step size, These are the weighting coefficients. To predict the first in the time domain The increment of the swing valve opening at each control moment relative to the previous moment. To predict the first in the time domain The control amount of the swing valve opening at each control moment is the target opening of the swing valve plate at that moment.

[0106] Optionally, the objective function may also include additional terms such as valve centering / energy consumption, for example, .

[0107] 3) Constraints Constraints are imposed on the opening degree and its rate of change of the swing valve, including: 1) Opening degree amplitude constraint: ;2) Opening change rate constraint (speed limit): Optionally, it also includes 3) safety constraints for predicting pressure, such as: , .

[0108] S33. The constraint solution yields the swing valve opening command.

[0109] Within each control cycle, the controller solves for the valve plate opening sequence of the swing valve based on the objective function and constraints. It outputs the first step control quantity as the valve opening command for this cycle. :

[0110] in, This represents the first step of the optimal swing valve opening increment obtained through constraint optimization in the current control cycle. This indicates the valve opening command for the next control cycle.

[0111] Preferably, to reduce the computational complexity of online calculations, the equivalent pumping rate is locally linearized with respect to the aperture.

[0112] For example, the controller at the current opening degree Sensitivity calculation:

[0113] And construct a linear approximation:

[0114]

[0115] in, This is the equivalent pumping speed after neural network compensation and physical constraint processing.

[0116] Under the aforementioned linear approximation, the controller transforms the prediction model into a form suitable for rapid solution and obtains the desired result using rolling optimization, quadratic programming, or other constraint-based solution methods that meet real-time requirements. .

[0117] S34. Anomaly detection and degradation protection strategy.

[0118] When the credibility of the hybrid prediction model is insufficient, the constraint solution fails, or the prediction error is abnormal, the controller triggers the degradation protection strategy to ensure the safe operation of the system.

[0119] Triggering conditions include, but are not limited to: (1) And it continues for more than the preset period; (2) Prediction error statistics (3) The constraint solver returns infeasibility or timeout; (4) Pumping speed after compensation. Multiple boundary projections or abnormal jumps may occur.

[0120] When the degradation protection strategy is triggered, the controller uses a safety control law to output an opening command.

[0121] Preferably, PI control with amplitude and speed limiting is used:

[0122] And on Apply rate of change constraint .

[0123] In degradation protection mode, the controller can pause online updates of the neural network and parameter self-calibration updates, or reduce the update frequency, to avoid abnormal data from polluting the model.

[0124] S35, Instruction Output and Closed-Loop Update.

[0125] The controller outputs the final valve plate opening command to the valve drive module, driving the valve to perform opening adjustment. The controller records operational data for the current cycle, including pressure measurements, opening commands, exogenous operating information, prediction errors, and solution status. This data is used for the following in the next control cycle: prediction error calculation and reliability gating update; neural network residual compensation input construction; optional online network fine-tuning and parameter updates; and optional updates to the comprehensive disturbance term and mechanism model parameters, thus forming a cross-condition adaptive closed-loop control update process.

[0126] It should be understood that the above-described device is used to execute the methods in the above embodiments. The implementation principle and technical effect of the corresponding program modules in the device are similar to those described in the above methods. The working process of the device can be referred to the corresponding process in the above methods, and will not be repeated here.

[0127] Based on the methods in the above embodiments, this application provides an electronic device that may include a processor, a communications interface, a memory, and a communication bus, wherein the processor, communications interface, and memory communicate with each other via the communication bus. The processor may invoke logical instructions stored in the memory to execute the methods in the above embodiments.

[0128] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application.

[0129] Based on the methods in the above embodiments, this application provides a computer-readable storage medium storing a computer program that, when run on a processor, causes the processor to execute the methods in the above embodiments.

[0130] Based on the methods in the above embodiments, this application provides a computer program product that, when run on a processor, causes the processor to execute the methods in the above embodiments.

[0131] It is understood that the processor in the embodiments of this application can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. A general-purpose processor can be a microprocessor or any conventional processor.

[0132] The method steps in this application embodiment can be implemented in hardware or by a processor executing software instructions. The software instructions can consist of corresponding software modules, which can be stored in random access memory (RAM), flash memory, read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, hard disks, portable hard disks, CD-ROMs, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor, enabling the processor to read information from and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can reside in an ASIC.

[0133] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted through the computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0134] It is understood that the various numerical designations used in the embodiments of this application are merely for the convenience of description and are not intended to limit the scope of the embodiments of this application.

[0135] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A smart control method for a vacuum swing valve based on a hybrid model, characterized in that, The vacuum swing valve is installed between the vacuum chamber and the vacuum pump, and the intelligent control method includes: Real-time acquisition of cavity pressure, cavity temperature, equivalent intake load, pump speed, and swing valve opening; Based on the current cavity pressure, cavity temperature, and swing valve opening, determine the equivalent variables of the flow pattern; Based on manifold equivalent variables, a continuously differentiable manifold weight function is constructed; Based on the flow pattern weighting function, the molecular flow conductance function and the viscous flow conductance function of the swing valve are smoothly fused to obtain the continuous equivalent conductance across flow patterns; The equivalent pumping speed on the cavity side is calculated based on the pump-side pumping speed and the crossflow continuous equivalent conductance. Based on mass conservation, a discrete prediction model for the cavity pressure at the next moment is established by combining the equivalent pumping speed on the cavity side, the equivalent intake load on the cavity side, and the equivalent volume of the cavity. The pressure control dynamic correlation feature vector is input into a pre-trained lightweight neural network to obtain the equivalent pumping speed residual compensation amount. Based on the difference statistics between the measured value of the cavity pressure on the cavity side and the one-step prediction value of the current time output by the hybrid model at the previous time, a confidence gating coefficient is generated. Based on the confidence gating coefficient, the equivalent pumping rate residual compensation amount is gated and fused to obtain the compensated equivalent pumping rate; By replacing the cavity-side equivalent pumping velocity in the discrete prediction model of cavity pressure with the compensated equivalent pumping velocity, a hybrid prediction model is obtained. A hybrid forecasting model is used to generate a hybrid forecasting stress sequence by rolling over the forecasting step size; Based on the hybrid predicted pressure sequence, an objective function is constructed for the current moment. Under the constraints of opening amplitude and opening change rate, constraint optimization is performed to obtain the valve opening command for the next moment. When the credibility is abnormal or the solution fails, switch to the degradation protection strategy and output the safe opening command.

2. The intelligent control method as described in claim 1, characterized in that, The real-time measured cavity pressure values ​​are low-pass filtered to obtain the filtered cavity pressure: in, for Original cavity pressure measurement value, After filtering, the pressure in the cavity side chamber. Initialize to the pressure measurement value at power-on; Based on the filtered cavity-side cavity pressure, the trend of cavity-side cavity pressure variation is calculated. : The sampling period.

3. The intelligent control method as described in claim 1, characterized in that, Based on the Knudsen number criterion or the equivalent pressure interval criterion, the following is generated: S-shaped continuous divertible weight function: in, For a moment Molecular flux weight, For a moment Viscous flow weight, The critical flow pattern is an equivalent variable, determined based on offline experimental calibration. For a moment The manifold equivalent variable, given by time... The equivalent flow channel dimensions and gas physical property parameters corresponding to the cavity side pressure, cavity temperature, and valve plate opening of the swing valve are calculated. This is the smoothing coefficient.

4. The intelligent control method as described in claim 1, characterized in that, Based on the pump-side pumping velocity and the equivalent flow conductance across the flow pattern, the equivalent pumping velocity on the cavity side is calculated as follows: in, For a moment The equivalent pumping speed on the cavity side. For a moment The transflow-mode continuous equivalent conductance, For a moment Pump side pumping speed, For a moment The opening degree of the swing valve plate. For a moment The manifold equivalent variable.

5. The intelligent control method as described in claim 1, characterized in that, Combining the equivalent pumping speed on the cavity side, the equivalent intake load on the cavity side, and the equivalent volume of the cavity, a discrete prediction model for the cavity pressure at the next moment is established based on mass conservation, as follows: in, This is a one-step prediction of the cavity pressure at the next moment. This is the equivalent volume of the cavity; for Equivalent intake load; for The comprehensive perturbation term is used to characterize low-frequency, difficult-to-model equivalent gas load changes. for Equivalent pumping speed on the cavity side.

6. The intelligent control method as described in claim 1, characterized in that, The pressure control dynamic correlation feature vector includes at least one of the following variables: cavity side pressure, cavity side pressure change trend, swing valve opening, equivalent intake load, pump side pumping speed, cavity temperature, flow pattern equivalent variable, and the one-step prediction residual of the previous moment.

7. The intelligent control method as described in claim 1, characterized in that, The lightweight neural network is a feedforward fully connected network, deployed on an embedded controller via fixed-point quantization. It has no more than two hidden layers and no more than 5000 network parameters, as detailed below: in, for The equivalent pumping rate residual compensation amount of the lightweight neural network output. This is a nonlinear mapping function implemented by a lightweight neural network, used to establish the correspondence between the input feature vector and the equivalent pumping rate residual compensation amount. For a moment The input feature vector, These are parameters for a lightweight neural network.

8. The intelligent control method as described in claim 7, characterized in that, The compensation amount of the equivalent pumping speed residual is limited, and a non-negative constraint, an upper bound constraint limited by the pump-side pumping speed, and a rate of change constraint are applied to the compensated equivalent pumping speed.

9. The intelligent control method as described in claim 1, characterized in that, One-step prediction error The gating coefficient is generated by comparing it with a preset threshold. : in, , Separately control the lower and upper threshold values. for Error statistics The length of the error statistics window. For the first sliding window The prediction error at each time point, Indicates time The actual cavity pressure state quantity. This represents the predicted cavity pressure value output by the hybrid model at the previous time step. Indicates time The actual pressure is the prediction error relative to the previous cycle's forecast results.

10. A smart control system for a vacuum swing valve based on a hybrid model, characterized in that, Includes memory and one or more processors; The memory is coupled to the one or more processors, and the memory is used to store computer program code, the computer program code including computer instructions; The one or more processors invoke the computer instructions to cause the system to perform the intelligent control method as described in any one of claims 1 to 9.

Images