MFC flow compensation method and system based on multi-dimensional features

Abstract

This invention discloses a method and system for MFC traffic compensation based on multidimensional features. The method includes: acquiring multidimensional traffic data; performing smoothing and noise reduction and steady-state feature extraction on the multidimensional traffic data, and constructing a multidimensional steady-state feature vector; constructing a traffic compensation model composed of a multidimensional physical partition gating switching kernel module, a temperature-independent decoupling compensation module, and a multi-objective partition constraint evaluation module; inputting the multidimensional steady-state feature vector into the traffic compensation model, optimizing the traffic compensation model using a search algorithm, and outputting predicted traffic data. The method and system provided by this invention can correct the measurement deviation of MFC output traffic, improve the accuracy and stability of MFC output traffic measurement, thereby achieving high-precision online compensation and health status assessment.

Description

Technical Field

[0001] This invention relates to the fields of fluid measurement and industrial automation technology, and in particular to an MFC flow compensation method and system based on multidimensional features. Background Technology

[0002] Mass flow controllers (MFCs) are widely used in precision fluid measurement and control fields such as semiconductor manufacturing. Accurate measurement and stable control of fluid flow are core elements for ensuring high-yield processes in advanced manufacturing. However, due to the mechanical hysteresis effect and dead-zone nonlinearity of the electromagnetic proportional valve inside the MFC, as well as the thermal drift of the sensor caused by ambient temperature interference, the measurement and control accuracy of MFCs across the entire range often fails to meet the requirements of advanced processes.

[0003] Currently, the industry primarily uses standard table comparison methods for MFC calibration. Internal data processing and error calibration largely rely on traditional analytical algorithms, but these methods have inherent limitations in multidimensional decoupling and nonlinear fitting. Common calibration methods include univariate polynomial fitting and piecewise linear interpolation based on discrete points. Polynomial fitting is a globally static approximation method, which struggles to accurately describe the drastic local nonlinear changes caused by the coupling of solenoid valve dead zones and thermal drift. Piecewise interpolation requires dense shutdown calibration points to maintain basic accuracy; however, the applicability of the interpolation table significantly decreases when ambient temperature shifts. Most of these traditional algorithms are limited to univariate mappings from "single flow reading" to "actual flow." For the multi-valued mapping divergence problem caused by mechanical hysteresis, where "the same reading corresponds to different actual flow rates," polynomial fitting can only perform mathematical averaging, resulting in limited actual compensation accuracy.

[0004] To overcome the aforementioned accuracy bottleneck, existing technologies attempt to improve both hardware structure and control algorithms. For example, CN121300561B discloses a low-pressure differential gas mass flow controller, employing a dual-channel hardware structure with a main laminar flow element and an auxiliary laminar flow element, combined with an adaptive PID optimized by the Sparrow Search Algorithm (SSA) for flow feedback control. However, this approach relies on increasing the number of physical hardware channels, which not only increases the manufacturing cost and size of the controller but also introduces additional mechanical failure points. Furthermore, conventional PID algorithms are essentially deterministic feedback control, making it difficult to completely decouple the nonlinear multi-valued mapping divergence problem caused by the mechanical hysteresis of the solenoid valve from a physical mechanism perspective.

[0005] In recent years, compensation schemes based on complex machine learning have gradually emerged. For example, CN120233678A discloses a linear correction system for control valve opening / flow rate, which uses an LSTM network for flow rate prediction and calculates the compensation amount through Gaussian process regression (using a squared exponential kernel function) combined with digital twin flow field simulation. This method can improve the linearity of the control valve under specific operating conditions, but its algorithm architecture is relatively complex and depends on massive historical data and high computing power. In addition, the squared exponential kernel used in this scheme is a conventional stationary kernel function, which has consistent smoothing characteristics throughout the entire space. It is difficult to adapt to the heterogeneous fluid dynamic characteristics of MFC, which undergoes drastic changes in the "extremely low flow dead zone" and smooth changes in the "linear operating zone". Therefore, it is prone to over-smoothing and accuracy degradation at the moment of valve opening.

[0006] In summary, existing MFC traffic compensation methods still have shortcomings in terms of hardware cost, limitations of univariate compensation, accuracy of nonlinear fitting, adaptability to small sample scenarios, and reliability early warning. There is an urgent need for a technical solution that can overcome the above technical bottlenecks and achieve high-precision and high-reliability online compensation. Summary of the Invention

[0007] In view of the above-mentioned shortcomings of current MFC fluid measurement, this invention provides an MFC flow compensation method based on multi-dimensional features, which can correct the measurement deviation of MFC output flow, improve the accuracy and stability of MFC output flow measurement, and thus achieve high-precision online compensation and health status assessment.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] A multi-dimensional feature-based MFC traffic compensation method, characterized in that the method includes:

[0010] Obtain multi-dimensional traffic data;

[0011] The multidimensional traffic data is smoothed, denoised, and its steady-state features are extracted, and a multidimensional steady-state feature vector is constructed.

[0012] A flow compensation model is constructed, consisting of a multi-dimensional physical partition gating switching kernel module, a temperature-independent decoupling compensation module, and a multi-objective partition constraint evaluation module.

[0013] The multidimensional steady-state feature vector is input into the flow compensation model, and the flow compensation model is optimized by combining a search algorithm to output predicted flow data.

[0014] According to one aspect of the invention, the multidimensional flow data includes device-indicated instantaneous flow rate, solenoid valve drive status parameters, real-time temperature, reference target tag data, and opening voltage threshold, wherein the solenoid valve drive status parameters are drive voltage or opening ratio.

[0015] According to one aspect of the present invention, the flow compensation method further includes: dividing the working range of the solenoid valve into a dead zone and a linear zone based on the driving voltage and the opening voltage threshold; and determining whether the current working range is a blind zone based on the reference target label data.

[0016] According to one aspect of the present invention, the step of smoothing and denoising the multidimensional traffic data, extracting steady-state features, and constructing a multidimensional steady-state feature vector specifically involves:

[0017] A composite first-order sliding data window is constructed and preliminary data filtering is performed on the multidimensional flow data;

[0018] Based on the first-order sliding data window, the multidimensional flow data after data filtering is used to determine the steady-state condition.

[0019] Based on the steady-state discrimination results of the operating conditions, the multi-dimensional flow data is subjected to precise data filtering and steady-state feature optimization extraction.

[0020] A multidimensional steady-state feature vector is constructed, which includes a normal steady-state feature vector and a blind zone feature vector to be measured.

[0021] According to one aspect of the present invention, the multidimensional physical partition gating switching kernel module is configured with a combined kernel function:

[0022]

[0023] In the formula, x and These are two sets of normal steady-state feature vectors, representing the current and historical data; T and These are the current and historical real-time temperatures, respectively. This is the driving voltage; For Matérn3 / 2 kernel functions; For RBF kernel functions; These are independent linear kernel terms; This is a Sigmoid-gated activation function.

[0024] According to one aspect of the present invention, the Sigmoid-gated activation function is:

[0025]

[0026] In the formula, To enable the voltage threshold, For driving voltage, is the scheduling gain coefficient, and exp is the natural constant.

[0027] According to one aspect of the present invention, the multi-objective partitioning constraint evaluation module constructs a multi-objective fitness function:

[0028]

[0029] In the formula, The baseline target label data for the i-th sample; Let N be the mean of the predicted traffic for the i-th sample; N is the total number of training samples. The dead zone weight function, The driving voltage corresponding to the i-th sample; M is the number of feature samples to be tested in the blind zone; For the blind zone features that were not included in the training, The posterior prediction variance for the blind zone of the flow compensation model; This represents the variance constraint coefficient.

[0030] According to one aspect of the present invention, the search algorithm is a sparrow search algorithm, which employs a multi-objective fitness function for optimization.

[0031] According to one aspect of the present invention, the predicted flow data includes a predicted flow mean and a predicted standard deviation, wherein the predicted flow mean is used to compensate for the instantaneous flow indicated by the device, and the predicted standard deviation triggers a device fault warning when it exceeds a preset fault warning threshold.

[0032] A multi-dimensional feature-based MFC traffic compensation system, wherein the system is applied to the aforementioned MFC traffic compensation method, the system comprising:

[0033] The data acquisition module acquires multi-dimensional traffic data;

[0034] The data processing module performs smoothing and noise reduction and steady-state feature extraction on the multidimensional traffic data, and constructs a multidimensional steady-state feature vector.

[0035] The model building module constructs a flow compensation model consisting of a multi-dimensional physical partition gating switching kernel module, a temperature-independent decoupling compensation module, and a multi-objective partition constraint evaluation module.

[0036] The result generation module inputs the multidimensional steady-state feature vector into the traffic compensation model, optimizes the traffic compensation model using a search algorithm, and outputs predicted traffic data.

[0037] The advantages of this invention are as follows: By collecting multi-dimensional flow data and constructing a high signal-to-noise ratio steady-state feature vector through sliding window steady-state discrimination and mean filtering, a flow compensation model is constructed, consisting of a multi-dimensional physical partition gated switching kernel module, a temperature-independent decoupling compensation module, and a multi-objective partition constraint evaluation module. Specifically, a Sigmoid-gated adaptive switching Matérn3 / 2 kernel and an RBF kernel are used to fit the dead-zone flow mutation and linear zone stationary changes, respectively; the temperature-independent kernel term removes thermal drift interference; the multi-objective fitness function introduces dead-zone error weighting penalties and blind-zone posterior variance constraints, combined with a sparrow search algorithm to optimize hyperparameters. Finally, the predicted flow mean is output to compensate the original indicated flow, while the predicted standard deviation is output for equipment fault early warning. Thus, relying only on a small number of calibration samples, dead-zone nonlinearity and thermal drift errors are significantly suppressed, model overfitting and oscillations are avoided, and high-precision, high-reliability online adaptive flow compensation is achieved under all operating conditions. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. 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.

[0039] Figure 1 This is a schematic flowchart of an MFC traffic compensation method based on multi-dimensional features according to the present invention.

[0040] Figure 2 This is a structural diagram of an MFC flow compensation system based on multidimensional features, as described in this invention. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] Example 1

[0043] like Figure 1 As shown, a traffic compensation method based on multi-dimensional features includes the following specific steps:

[0044] Step S1: Obtain multidimensional traffic data;

[0045] The specific steps for obtaining multi-dimensional traffic data in step S1 are as follows:

[0046] The multidimensional flow data includes the instantaneous flow rate indicated by the equipment, solenoid valve drive status parameters, real-time temperature, reference target label data, and opening voltage threshold. The acquisition methods for multidimensional flow data are divided into real-time acquisition and calibration tests.

[0047] The device's instantaneous flow rate, solenoid valve drive status parameters, and real-time temperature are synchronously acquired in real time by the supporting acquisition components: the device's instantaneous flow rate is acquired through the built-in sensor of the mass flow controller (MFC), the underlying solenoid valve drive status parameters are acquired through the solenoid valve drive acquisition module, and the real-time temperature is acquired through the temperature acquisition module. The solenoid valve drive status parameters can be selected using either drive voltage or opening ratio, adapted to the actual device model; in this embodiment, drive voltage is selected. The real-time temperature includes the device's operating ambient temperature and the fluid medium temperature.

[0048] The baseline target label data and the opening voltage threshold were measured during the equipment calibration phase. The calibration test was conducted under standard operating conditions: standard atmospheric pressure, constant temperature, and adjustable driving voltage in stages. The MFC under test was paired with a high-precision standard flowmeter to calibrate the flow rate and solenoid valve threshold. The data acquisition frequency and timing were consistent with the actual operation of the equipment. The physical flow rate was measured using the high-precision standard flowmeter, and this physical flow rate served as the baseline target label data. However, due to limitations in the calibration test duration and operating conditions, the calibration phase could not cover all operating conditions of the equipment. For some conditions, only characteristic data such as instantaneous flow rate, driving status parameters, and real-time temperature could be collected, and the corresponding baseline target label data for the high-precision standard flowmeter could not be collected. These conditions without corresponding baseline target label data were defined as blind zones. During the test, the solenoid valve driving voltage was gradually increased while simultaneously observing the flowmeter readings. The driving voltage corresponding to the first effective non-zero flow rate output by the high-precision standard flowmeter was determined as the opening voltage threshold at which the solenoid valve exited the dead zone and entered the linear operating range. The opening voltage threshold did not require preprocessing and was re-acquired and calibrated every 3-6 months.

[0049] It should be noted that the dead zone of a solenoid valve is the opening range where the driving voltage is lower than the opening voltage threshold. In this range, even a small change in voltage will cause a drastic nonlinear change in flow. The linear zone of a solenoid valve is the normal operating range where the driving voltage is higher than the opening voltage threshold. The flow changes steadily and approximately linearly with the driving voltage.

[0050] Step S2: Perform smoothing and noise reduction and steady-state feature extraction on the multidimensional traffic data, and construct a multidimensional steady-state feature vector;

[0051] Step S2, which involves data filtering and steady-state feature extraction of the multidimensional flow data and the construction of a multidimensional steady-state feature vector, specifically includes:

[0052] Step S21: Construct a composite first-order sliding data window and perform preliminary data filtering on the multidimensional flow data;

[0053] For the multidimensional flow data obtained in step S1, a composite first-order sliding data window is constructed; the composite first-order sliding data window includes a first-order short sliding sub-window and a first-order long sliding sub-window. Multidimensional flow data is extracted frame by frame through the sliding window to initially filter out discrete and abnormal data. Specifically: for rapidly changing physical quantities such as instantaneous flow indicated by the device and solenoid valve drive status parameters, a continuous sampling point system is constructed. A first-order short sliding data window. The initial value is 20, determined jointly by the MFC sampling frequency and the gas path dynamic response time, with a range of 10~50. If If the value is less than 10, the high-frequency interference filtering effect is poor; if A value greater than 50 will introduce a delay of over 0.5 seconds and is prone to misjudging flow status, thus weakening real-time response capabilities. For slowly varying physical quantities such as real-time temperature, which are affected by environmental factors and sensor thermal drift, a continuous sampling point count of [number missing] is constructed. A first-order long sliding data window. The preferred value range is 200~1000 (corresponding to a time scale of several seconds to tens of seconds). A longer sliding data window with a larger time scale is used to truly reflect the slow drift trend of temperature and avoid the failure to filter out low-frequency noise or the forced alignment with fast variables due to a short window.

[0054] Step S22: Perform steady-state judgment on the multi-dimensional flow data after data filtering;

[0055] The difference between the maximum and minimum instantaneous flow rates indicated by MFC within each frame of the sliding data window is calculated in real time to obtain the flow range value. A fixed steady-state discrimination threshold is pre-configured. The real-time calculated flow range value is compared with the steady-state discrimination threshold. When the range value is less than the preset steady-state discrimination threshold, the device operating condition corresponding to the current sliding window is determined to be a steady-state condition, and the data in the window is valid steady-state data. If the range value is greater than or equal to the preset steady-state discrimination threshold, it is determined to be a transient condition, and the data in the window is transient data caused by step response or external vibration, which is directly discarded. The steady-state discrimination threshold ranges from 0.1%FS to 1%FS, where FS represents the upper limit of the rated maximum measurement range of MFC. In this embodiment, the steady-state discrimination threshold is set to 0.4%FS.

[0056] Step S23: Based on the working condition judgment result, perform precise data filtering and steady-state feature optimization extraction on the multi-dimensional flow data;

[0057] Valid window data corresponding to all steady-state operating conditions are filtered out. For rapidly changing physical quantities, mean filtering is performed on the instantaneous flow rate of the equipment and the solenoid valve drive status parameters within the first-order short sliding data window to extract smoothed indicated flow rate and smoothed drive voltage. For slowly changing physical quantities, mean filtering is performed on the real-time temperature within the first-order long sliding data window to eliminate low-frequency interference and environmental errors and extract smoothed temperature.

[0058] Step S24: Construct a multidimensional steady-state feature vector.

[0059] The smoothed indicated flow rate, smoothed drive voltage (or smoothed opening ratio), and smoothed temperature obtained after mean filtering are vectorized and concatenated to unify the data dimension and data time sequence, generating a standardized, high signal-to-noise ratio three-dimensional steady-state feature vector; the three-dimensional steady-state feature vector can be expressed as:

[0060]

[0061] In the formula, To smooth out the flow rate indication, To smooth the drive voltage, To smooth out the temperature.

[0062] Furthermore, based on whether or not they are accompanied by reference target label data, all three-dimensional steady-state feature vectors are distinguished into: normal steady-state feature vectors with reference target label data and blind zone features to be measured without reference target label data. Specifically:

[0063] (1) A structure is constructed by stacking N sets of normal steady-state feature vectors with baseline target labels vertically in time sequence, forming a dimension of Normal steady-state characteristic matrix Simultaneously, the corresponding baseline target label data is combined to form a dimension of... The baseline target label column vector The normal steady-state characteristic matrix X t Compared with the baseline target label column vector Y t A one-to-one correspondence is formed.

[0064] (2) The feature vectors of blind zone conditions without benchmark target labels in M ​​groups are stacked to form a dimension of Blind zone variance constraint feature matrix .

[0065] Step S3: Construct a flow compensation model consisting of a multi-dimensional physical partition gating switching kernel module, a temperature-independent decoupling compensation module, and a multi-objective partition constraint evaluation module;

[0066] In step S3, the flow compensation model consists of a multi-dimensional physical partition gating switching core module, a temperature-independent decoupling compensation module, and a multi-objective partition constraint evaluation module. Data is sequentially fed into each module for computation. Through multi-physical partitioning, dead-zone nonlinearity errors are corrected, temperature thermal drift interference is separated, and small-sample overfitting is constrained. After model computation, predicted flow data is output. The multi-dimensional physical partitioning is an embedded operating condition division and control logic within the model. It relies on the driving voltage to divide dead-zone or linear operating condition partitions, and then combines this with independent temperature compensation partitions to achieve partitioned control across three physical fields: voltage, flow, and temperature. The specific structure, calculation formulas, and functions of each module are as follows:

[0067] The multi-dimensional physical partition gating switching core module is used to realize the partition control of voltage and flow. Based on the main framework of Gaussian process regression (GPR), a non-stationary combined kernel is designed as the core computing unit of the model. It distinguishes between dead zone partitioning and linear operating condition partitioning by relying on the driving voltage threshold. It adaptively switches between two computing kernels, Matérn3 / 2 kernel and RBF kernel, with Sigmoid gating weights to adapt to the differentiated characteristics of the solenoid valve dead zone flow change and the solenoid valve linear zone flow steady change, respectively. At the same time, a dedicated computing position for temperature compensation is reserved in the combined kernel to adapt to the independent decoupling requirements of temperature characteristics. The combined kernel function expression is as follows:

[0068]

[0069] In the formula, x and These are two sets of normal steady-state feature vectors, representing the current and historical data; T and These are current and historical real-time temperature data, respectively. This is the real-time drive voltage of the solenoid valve; Using the Matérn3 / 2 kernel function, fit the sudden change in flow rate in the dead zone of the solenoid valve; Use the RBF kernel function to ensure stable flow in the linear region; An independent linear kernel term is reserved for subsequent temperature drift correction; This is a Sigmoid-gated activation function used to implement logical scheduling switching between different physical partitions. Its expression is:

[0070]

[0071] In the formula, This refers to the threshold voltage at which the solenoid valve transitions from the dead zone to the linear zone. The real-time drive voltage of the solenoid valve serves as the dynamic input variable for the gate function. The scheduling gain coefficient is initially set to 10, with a range of 2 to 30. It is used to adjust the steepness of gating switching and optimize the fitting accuracy of the transition interval between the dead zone and the linear zone. exp is the natural constant e.

[0072] The temperature-independent decoupling compensation module is used for independent temperature compensation, addressing the slow linear thermal drift error caused by ambient temperature and fluid medium temperature, and compensating for the aforementioned combined core. The independent linear kernel term is used for functional limitation. It relies solely on temperature time series data to complete temperature drift fitting calculation. At the mathematical level, temperature thermal drift interference and the nonlinear flow change of the valve body itself are physically separated and do not participate in the gated kernel switching logic. As a result, a temperature partition is added on the basis of the original voltage and flow partition, which improves the full-dimensional and multi-dimensional physical partition system of voltage, flow and temperature, and enhances the flow compensation adaptation capability of the model under full temperature range and full opening conditions.

[0073] The multi-objective partitioning constraint evaluation module is used to supplement and improve the multi-dimensional physical partitioning control rules, adapt to small sample calibration conditions, and construct a multi-objective fitness function that includes dead zone error weighting penalty and blind zone posterior variance constraint, serving as the evaluation standard for subsequent Sparrow Search (SSA) algorithm iteration optimization. By weighting the residuals of samples in the dead zone voltage range, the model's dead zone accuracy fitting ability is enhanced; simultaneously, the posterior prediction variance of the blind zone test points of the flow compensation model is introduced as a regularization term to constrain the model's prediction uncertainty, addressing the shortcomings of blind zone oscillation and overfitting in small sample calibration scenarios, thus completing the overall construction of the dedicated flow compensation model. The formula for the multi-objective fitness function is:

[0074]

[0075] In the formula, the first term is the dead-zone weighted fitting error term. is the true physical flow rate of the i-th sample, and is the high-precision standard flow meter reference target label data collected in step S1; The predicted flow value for the i-th sample is obtained by inputting the three-dimensional steady-state feature vector into the flow compensation model constructed in step S3; N is the total number of training samples. The dead zone weight function, The driving voltage is the i-th sample; the error penalty weight is automatically increased for samples in the dead zone; the second term is the blind zone posterior variance constraint term, and M is the number of blind zone feature samples to be tested. For the blind zone features that were not included in the training, The posterior prediction variance for the blind zone of the flow compensation model; This is the variance constraint coefficient, initially set to 0.5, with a range of 0.05 to 5.0; it is used to constrain the numerical uncertainty of uncalibrated blind zone predictions, avoiding overfitting oscillations in small sample environments.

[0076] To automatically increase the error penalty weight within the dead zone voltage range and overcome the severe nonlinear fitting difficulties caused by the mechanical dead zone, the dead zone weighting function... A segmented weighting strategy based on physical workspace partitioning is adopted, and its specific mathematical expression is as follows:

[0077]

[0078] In the formula, The threshold voltage for the solenoid valve to transition from the dead zone to the linear zone; This is the dead zone error penalty coefficient, and it is limited. >1 is used to explicitly apply a weighted penalty to the prediction residuals within the dead zone interval in the mathematical model. In this embodiment, The initial value is preferably set to 5, and the range is limited to 2 to 10. The threshold voltage for activation... The specific physical meaning is as follows: Due to the inherent mechanical hysteresis and dead-zone nonlinearity of the electromagnetic proportional valve inside the mass flow controller (MFC), when the underlying drive voltage... When the electromagnetic proportional valve fails to fully overcome the mechanical friction or spring tension of the internal valve needle, the valve barely opens or remains in a dead zone. Within this zone, even a small change in the driving voltage can cause a drastic nonlinear jump in the physical flow rate. When the electromagnetic proportional valve opens normally and leaves the dead zone, it enters the normal linear operating range, and the output flow rate changes smoothly and approximately linearly with the increase of the drive voltage.

[0079] Step S4: Input the multidimensional steady-state feature vector into the flow compensation model, optimize the flow compensation model using a search algorithm, and output predicted flow data.

[0080] In step S4, the multidimensional steady-state feature vector is input into the flow compensation model, and the flow compensation model is optimized by combining a search algorithm to output predicted flow data. Specifically, this is as follows:

[0081] Step S41: SSA algorithm initialization;

[0082] Initialize the basic parameters of the SSA algorithm. Set the initial population size to 30 and the initial maximum number of iterations to 100. Simultaneously, initialize the hyperparameters to be optimized for the traffic compensation model, including the scheduling gain coefficient. Variance constraint coefficient The length scales of the Matérn3 / 2 kernel function, the RBF kernel function, and the temperature kernel variance are all included. The initial values ​​and ranges of each hyperparameter can be set conventionally according to model training requirements. In this embodiment, the scheduling gain coefficient... The initial value is set to 10, and the variance constraint coefficient is... The initial value of the kernel is set to 0.5; the initial value of the length scale of the Matérn3 / 2 kernel function is set to 0.5, with a range limited to 0.05~2.0; the initial value of the length scale of the RBF kernel function is set to 2.5, with a range limited to 1.0~8.0; the temperature independent decoupling compensation module... The initial value of the temperature kernel variance is set to 0.05, and the range is limited to 0.001~0.3.

[0083] Step S42: Iterative optimization of the SSA algorithm and output of optimal hyperparameters;

[0084] Using the multi-objective fitness function (Fitness) as the sole iterative evaluation metric and minimizing the fitness value as the optimization objective, the flow compensation model's hyperparameters are iteratively optimized generation by generation through SSA (Simulated Swarm Optimization) simulation of population foraging, early warning, and update mechanisms. Specifically, the normal steady-state feature matrix of the training samples is used... The input flow compensation model is processed sequentially through a multi-dimensional physical partition gating switching kernel module and a temperature-independent decoupling compensation module to calculate the full-sample covariance matrix, thereby obtaining the model's predicted flow mean for the training samples. Simultaneously, the blind zone variance constraint feature matrix is... Input the model and calculate the posterior prediction variance corresponding to the features to be tested in each blind zone. The multi-objective partitioning constraint evaluation module combines the training sample prediction results and the baseline target label column vector. The posterior prediction variance of the blind zone is substituted into the multi-objective fitness function Fitness to calculate the fitness value corresponding to the current hyperparameter combination. Based on the fitness value of the current population individuals, the SSA algorithm updates the sparrow population position, removes inferior parameter combinations with high fitness values, retains high-quality individuals, and completes one generation iteration.

[0085] The update process is repeated until the preset maximum number of iterations is reached or the convergence condition is met, ultimately outputting the globally optimal hyperparameter combination of the model. The maximum number of iterations is limited to 50-200, used to balance the computational cost of the non-stationary combined kernel flow compensation model with the global optimization accuracy of the multi-dimensional hyperparameter space. Since high-order covariance matrix operations are involved in a single iteration, if the maximum number of iterations is less than 50, the SSA algorithm is prone to premature convergence and getting trapped in local optima, resulting in insufficient extrapolation accuracy. If the maximum number of iterations is greater than 200, the algorithm essentially enters a stable convergence phase in the later stages of optimization; continuing to add iterations will not significantly reduce uncertainty or residuals, but will instead waste computational resources.

[0086] Step S43: Calibrate the flow compensation model.

[0087] The optimal hyperparameters obtained from SSA optimization are assigned to the flow compensation model to complete the final model calibration. The normal steady-state eigenvectors are input into the calibrated model, and dual-channel results are output to achieve a closed loop between flow compensation and equipment status monitoring.

[0088] (1) Flow compensation output: The model predicts the average flow rate as the accurate physical flow rate, which replaces the original MFC indicated flow rate (i.e. the instantaneous flow rate indicated by the equipment) to complete the error compensation under all operating conditions;

[0089] (2) Status monitoring output: Model prediction standard deviation, used to characterize the reliability of the prediction results. When the prediction standard deviation exceeds the preset fault warning threshold, it is determined that the MFC has a fault risk and an automatic warning is triggered. The initial value of the fault warning threshold is set to 3%FS, and the value range is 1.5%~5%FS, which is set according to the 3σ confidence criterion. During normal operation, the prediction standard deviation is stable within 1.5%FS; if the prediction standard deviation is greater than 5%FS, it indicates a deviation from the calibration range, which can be determined as sensor drift, component aging or valve body jamming, and a fault alarm will be triggered.

[0090] Example 2

[0091] like Figure 2 As shown, an MFC traffic compensation system based on multi-dimensional features is described. This embodiment applies to Embodiment 1, and the system includes:

[0092] Data acquisition module M1 acquires multidimensional traffic data;

[0093] The data processing module M2 performs smoothing and noise reduction and steady-state feature extraction on the multidimensional traffic data, and constructs a multidimensional steady-state feature vector.

[0094] The model building module M3 constructs a flow compensation model consisting of a multi-dimensional physical partition gating switching kernel module, a temperature-independent decoupling compensation module, and a multi-objective partition constraint evaluation module.

[0095] The result generation module M4 inputs the multidimensional steady-state feature vector into the traffic compensation model, optimizes the traffic compensation model using a search algorithm, and outputs predicted traffic data.

[0096] The advantages of this invention are as follows: By collecting multi-dimensional flow data and constructing a high signal-to-noise ratio steady-state feature vector through sliding window steady-state discrimination and mean filtering, a flow compensation model is constructed, consisting of a multi-dimensional physical partition gated switching kernel module, a temperature-independent decoupling compensation module, and a multi-objective partition constraint evaluation module. Specifically, a Sigmoid-gated adaptive switching Matérn3 / 2 kernel and an RBF kernel are used to fit the dead-zone flow mutation and linear zone stationary changes, respectively; the temperature-independent kernel term removes thermal drift interference; the multi-objective fitness function introduces dead-zone error weighting penalties and blind-zone posterior variance constraints, combined with a sparrow search algorithm to optimize hyperparameters. Finally, the predicted flow mean compensates for the original indicated flow, and the predicted standard deviation is also output for equipment fault early warning. Therefore, relying only on a small number of calibration samples, dead-zone nonlinearity and thermal drift errors are significantly suppressed, model overfitting and oscillations are avoided, and high-precision, high-reliability online adaptive flow compensation is achieved under all operating conditions.

[0097] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A multi-dimensional feature-based MFC traffic compensation method, characterized in that, The method includes: Obtain multi-dimensional traffic data; The multidimensional traffic data is smoothed, denoised, and its steady-state features are extracted, and a multidimensional steady-state feature vector is constructed. A flow compensation model is constructed, consisting of a multi-dimensional physical partition gating switching kernel module, a temperature-independent decoupling compensation module, and a multi-objective partition constraint evaluation module. The multidimensional steady-state feature vector is input into the flow compensation model, and the flow compensation model is optimized by combining a search algorithm to output predicted flow data.

2. The MFC traffic compensation method according to claim 1, characterized in that, The multidimensional flow data includes the device-indicated instantaneous flow rate, solenoid valve drive status parameters, real-time temperature, reference target label data, and opening voltage threshold, wherein the solenoid valve drive status parameters are drive voltage or opening ratio.

3. The MFC flow compensation method according to claim 2, characterized in that, The flow compensation method further includes: dividing the working range of the solenoid valve into a dead zone and a linear zone based on the driving voltage and the opening voltage threshold; and determining whether the current working range is a blind zone based on the reference target label data.

4. The MFC traffic compensation method according to claim 3, characterized in that, The specific steps for smoothing and denoising the multidimensional traffic data, extracting steady-state features, and constructing a multidimensional steady-state feature vector are as follows: A composite first-order sliding data window is constructed and preliminary data filtering is performed on the multidimensional flow data; Based on the first-order sliding data window, the multidimensional flow data after data filtering is used to determine the steady-state condition. Based on the steady-state discrimination results of the operating conditions, the multi-dimensional flow data is subjected to precise data filtering and steady-state feature optimization extraction. A multidimensional steady-state feature vector is constructed, which includes a normal steady-state feature vector and a blind zone feature vector to be measured.

5. The MFC traffic compensation method according to claim 4, characterized in that, The multi-dimensional physical partition gating switching kernel module is configured with combined kernel functions: In the formula, x and These are two sets of normal steady-state feature vectors, representing the current and historical data; T and These are the current and historical real-time temperatures, respectively. This is the driving voltage; For Matérn3 / 2 kernel functions; For RBF kernel functions; These are independent linear kernel terms; This is a Sigmoid-gated activation function.

6. The MFC traffic compensation method according to claim 5, characterized in that, The Sigmoid gated activation function is: In the formula, To enable the voltage threshold, For driving voltage, is the scheduling gain coefficient, and exp is the natural constant.

7. The MFC traffic compensation method according to claim 6, characterized in that, The multi-objective partitioning constraint evaluation module constructs a multi-objective fitness function: In the formula, The baseline target label data for the i-th sample; Let be the mean of the predicted flow rate for the i-th sample; N is the total number of training samples; The dead zone weight function, The driving voltage corresponding to the i-th sample; M is the number of feature samples to be tested in the blind zone; For the blind zone features that were not included in the training, The posterior prediction variance for the blind zone of the flow compensation model; This represents the variance constraint coefficient.

8. The MFC traffic compensation method according to any one of claims 1 to 7, characterized in that, The search algorithm is the Sparrow Search Algorithm, which uses a multi-objective fitness function for optimization.

9. The MFC traffic compensation method according to claim 8, characterized in that, The predicted flow data includes the predicted flow mean and the predicted standard deviation. The predicted flow mean is used to compensate for the instantaneous flow indicated by the device. When the predicted standard deviation exceeds a preset fault warning threshold, a device fault warning is triggered.

10. An MFC flow compensation system based on multi-dimensional features, characterized in that, The system is applied to the MFC traffic compensation method according to any one of claims 1 to 9, and the system comprises: The data acquisition module acquires multi-dimensional traffic data; The data processing module performs smoothing and noise reduction and steady-state feature extraction on the multidimensional traffic data, and constructs a multidimensional steady-state feature vector. The model building module constructs a flow compensation model consisting of a multi-dimensional physical partition gating switching kernel module, a temperature-independent decoupling compensation module, and a multi-objective partition constraint evaluation module. The result generation module inputs the multidimensional steady-state feature vector into the traffic compensation model, optimizes the traffic compensation model using a search algorithm, and outputs predicted traffic data.

Images