A multi-dimensional feature hierarchical sampling based global calibration method and system for a mass flow controller

Abstract

The application discloses a quality flow controller global calibration method and system based on multi-dimensional feature hierarchical sampling. The method extracts four-dimensional feature vectors of the equipment, divides the same batch of equipment into homogeneous sample clusters through multi-level hierarchical clustering, and adaptively and dynamically samples according to the number and dispersion within the cluster. The sample equipment is calibrated in the full flow domain under multiple working conditions, the working condition related features are extracted, the attention mechanism is introduced to weight and fuse the feature parameters, the double-branch neural network mapping model is constructed and the training and verification are completed. For the remaining equipment in the batch, the key flow point data are collected and corrected, and the model is input to predict the global calibration parameters. At the same time, the risk priority sampling is reviewed and verified, and the model is iteratively updated by combining the deviation judgment and transfer learning. The method effectively solves the problem of time and resource consumption of traditional global calibration, greatly shortens the calibration period, reduces the energy consumption, and significantly improves the calibration accuracy and working condition adaptability.

Description

Technical Field

[0001] This application relates to fluid measurement and control technology, and in particular to a method and system for global calibration of a mass flow controller based on multidimensional feature hierarchical sampling. Background Technology

[0002] Mass flow controllers are core devices for precise control of fluid mass flow rates in high-precision industrial processes such as semiconductor manufacturing and biopharmaceuticals. Their calibration accuracy directly determines the stability of the process and the quality of the final product. In current calibration practices, to ensure the reliability of the controller across its entire operating range, high-precision standard flow meters must be used to measure dozens or even hundreds of discrete flow points across the entire flow domain. Each flow point requires waiting for the fluid flow state, temperature field distribution, and sensor output to reach steady-state equilibrium, a process that typically lasts from tens of seconds to several minutes. Subsequently, multiple sets of repeating data must be collected to eliminate random fluctuations, and statistical analysis is used to establish a precise mapping relationship between the setpoint and the actual flow rate, calculating the error correction coefficient. The complete calibration cycle for a single device often takes hours or even days, resulting in the long-term occupation of expensive standard testing equipment and professional technical personnel. In mass production scenarios, this unit-by-unit, full-range calibration mode significantly increases unit manufacturing costs, extends product delivery cycles, and becomes a major constraint on large-scale production.

[0003] To improve calibration efficiency, the industry has attempted to simplify calibration strategies, testing only a few key flow points and then extrapolating parameters for other points using linear interpolation or polynomial fitting. However, due to the strong nonlinear characteristics of mass flow controllers across the entire flow domain, especially the significant inflection point effect and sensitivity abrupt changes in the low flow region, such simplified methods struggle to accurately characterize actual behavior, leading to a significant increase in calibration errors in the low flow region and failing to meet high-precision process requirements. Another type of calibration transfer scheme involves performing complete calibration on a few representative devices in the same batch and then transferring the calibration parameters to other devices through scaling. However, this method is highly dependent on product consistency. In actual production, factors such as batch differences in flow detection elements, fluctuations in the machining tolerances of control valve cores, changes in temperature and humidity in the assembly environment, and the diversity of gas media types can all cause subtle but critical deviations in the response characteristics between devices. When manufacturing tolerances accumulate beyond a threshold, the calibration transfer results quickly become invalid, resulting in unstable performance of batch products. In addition, existing calibration methods lack the ability to adapt to dynamic changes in production and cannot effectively cope with the impact of process parameter drift, long-term equipment aging, or environmental condition fluctuations. They often require frequent interruptions of the production process for recalibration, which further exacerbates resource consumption and efficiency loss. Summary of the Invention

[0004] The purpose of this application is to provide a global calibration method and system for a mass flow controller based on multidimensional feature hierarchical sampling, which can significantly shorten calibration time, reduce resource consumption, improve calibration accuracy and adaptability.

[0005] The first aspect of this invention provides a method for global calibration of a mass flow controller based on multidimensional feature hierarchical sampling, comprising the following steps: Obtain the four-dimensional device feature vector of each mass flow controller in the production batch. The four-dimensional device feature vector includes the batch identifier of the flow detection element, the machining tolerance level of the control valve core, the temperature and humidity range code of the assembly environment, and the code of the gas medium adaptation type. A multi-level hierarchical clustering model is constructed based on four-dimensional device feature vectors to divide the same batch of quality flow controllers into multiple homogeneous sample clusters, so that the difference in device features within each sample cluster is less than a preset threshold and the difference between clusters is greater than a preset threshold. For each sample cluster, sample devices are drawn according to a dynamic sampling ratio, which is adaptively adjusted according to the number of devices and the feature dispersion within the sample cluster. The sample equipment was calibrated across the entire flow domain under multiple operating conditions, covering normal temperature, high temperature, normal pressure, and low pressure conditions. Sample response characteristic data, operating condition coupling parameters, and corresponding sample calibration parameter data were collected. The operating condition correlation feature set is extracted from the sample response characteristic data and operating condition coupling parameters. The operating condition correlation feature set includes the flow-voltage nonlinear coefficient, valve opening-flow hysteresis, temperature drift compensation coefficient, and pressure fluctuation sensitivity. An attention-based weighted fusion algorithm is introduced to assign differentiated weights to feature parameters under different working conditions; A dual-branch neural network mapping model is constructed. The first branch inputs flow detection features and the second branch inputs valve control features. A nonlinear dynamic mapping relationship between the working condition related feature set and the calibration parameter data is established through a feature cross-fusion layer. A regularization term is introduced to suppress overfitting and complete the model training and validation. For the remaining mass flow controllers in the batch, three non-equal interval key test flow points were selected, including the low flow inflection point, the medium flow linear region, and the high flow saturation region, and the remaining response characteristic data were collected under standard operating conditions. The remaining response characteristic data are feature-completed and operating condition-corrected. Non-test flow point data are completed using interpolation algorithms, and feature deviations caused by temperature drift and pressure interference are corrected based on real-time environmental parameters. The corrected feature set to be mapped is input into the dual-branch neural network mapping model, which outputs the prediction calibration parameter matrix under the full flow domain and multiple operating conditions to complete the calibration of the remaining mass flow controller. Validation samples are drawn from the calibrated mass flow controllers in each sample cluster according to the risk-priority sampling rule. Perform full-condition verification calibration on the validation samples, obtain validation calibration parameter data, and calculate the multidimensional deviation matrix between the predicted calibration parameters and the validation calibration parameters; When the deviation of any operating condition or flow point in the multidimensional deviation matrix exceeds the preset dynamic threshold, the verification sample data will be included in the incremental sample library. When the incremental sample library reaches a threshold, a transfer learning fine-tuning strategy is adopted to reuse the original model's basic parameters, update the weights of the feature fusion layer and the output layer, and complete the iterative update of the mapping model.

[0006] Furthermore, a multi-level hierarchical clustering model is constructed based on the four-dimensional device feature vectors, including: The four-dimensional device feature vectors are normalized, the Euclidean distance of each feature dimension is calculated, and the K-means++ clustering algorithm is used to complete the initial cluster division. The silhouette coefficient is introduced to evaluate the clustering effect, and the number of clusters is iteratively optimized until the intra-cluster compactness and inter-cluster separation both meet the preset indicators. The dynamic sampling ratio is: the sampling ratio equals the sum of the base ratio multiplied by 1 and the feature dispersion coefficient. The base ratio is 5%-10%, and the feature dispersion coefficient is 0-0.5.

[0007] Furthermore, the sample equipment undergoes full-flow-domain multi-condition calibration, including: The flow domain covers 0-100% of the full range, and 15-20 equally spaced flow points are set; Operating conditions cover a temperature range of 25-50℃ and a pressure range of 0.1-0.5MPa; The stabilization time of each operating condition-flow point is greater than or equal to 60 seconds. Simultaneously, the output voltage of the flow detection element, the opening degree of the regulating valve, the ambient temperature and humidity, and the pipeline pressure data are collected to form sample response characteristic data. The sample calibration parameter data includes the flow deviation correction coefficient, temperature drift compensation value, and pressure correction value for each operating condition and flow point.

[0008] Furthermore, an attention-based weighted fusion algorithm is introduced to assign differentiated weights to feature parameters under different operating conditions, including: Calculate the Pearson correlation coefficient between each feature parameter and the calibration parameter, and generate initial attention weights based on the correlation coefficients. Iteratively optimize the weights through backpropagation of the model, so that the weight of strongly correlated features accounts for more than or equal to 60%, and the weight of weakly correlated features accounts for less than or equal to 10%, thereby achieving differentiated weighting of feature importance.

[0009] Furthermore, the two-branch neural network mapping model includes: The first branch is a one-dimensional convolutional neural network, used to extract local correlation information of traffic detection features; The second branch is a long short-term memory network, used to capture the temporal variation patterns of valve control features; The feature cross-fusion layer adopts a splicing and fully connected structure to fuse the output features of the two branches; The output layer outputs the flow deviation correction coefficient matrix for the entire flow domain and multiple operating conditions. The model loss function is the sum of mean square error and L2 regularization term.

[0010] Furthermore, feature completion and operating condition correction are performed on the remaining response characteristic data, including: The response data of non-test flow points are completed using a cubic spline interpolation algorithm, with the error controlled to be less than or equal to 0.2% of the full scale. The operating condition correction is based on real-time collected ambient temperature and humidity and pipeline pressure data, and calls the preset temperature drift-pressure compensation formula to correct the deviation of characteristic parameters.

[0011] Furthermore, the three non-equal interval key test flow points are as follows: The low flow rate inflection point is 5%-15% of full scale, the medium flow rate linear region is 45%-55% of full scale, and the high flow rate saturation region is 85%-95% of full scale. The stabilization time for each test flow point is greater than or equal to 30 seconds, and two types of data are collected: flow detection voltage and valve opening.

[0012] Furthermore, validation samples are drawn according to the risk-priority sampling rule, including: Assign a risk score to each calibrated mass flow controller. The risk score is equal to the feature dispersion score multiplied by the operating condition fluctuation sensitivity score. Select the top 5%-8% of devices by risk score as the validation sample.

[0013] Furthermore, a transfer learning fine-tuning strategy is adopted to reuse the original model's basic parameters and update the weights of the feature fusion layer and output layer, including: Freeze the parameters of the bottom feature extraction layer of the dual-branch neural network, and train the feature cross-fusion layer and output layer only with incremental sample library data; The number of training rounds is less than or equal to 20, and the learning rate is 1 / 10 of the original model.

[0014] Secondly, this application also provides a mass flow controller global calibration system based on multidimensional feature stratified sampling, used for the aforementioned mass flow controller global calibration method based on multidimensional feature stratified sampling, comprising: The stratified sampling and differential calibration unit is used to perform batch multidimensional stratified sampling and sample differential full-condition calibration, and output sample response characteristic data, condition coupling parameters and sample calibration parameter data. The working condition coupling mapping model building unit is used to extract the working condition associated feature set, perform weighted fusion through an attention mechanism, and build and train a two-branch neural network mapping model. The simplified testing and intelligent prediction unit is used to perform key flow point testing, feature completion and operating condition correction on the remaining mass flow controller, and input the model to output the prediction calibration parameter matrix. The dynamic verification and closed-loop update unit is used for risk-priority sampling verification, multidimensional deviation calculation, incremental sample library management, and iterative updates of the mapping model transfer learning.

[0015] Compared to existing technologies, this application provides a mass flow controller global calibration method and system based on multi-dimensional feature hierarchical sampling. It achieves homogeneous sample clustering by acquiring four-dimensional device feature vectors to construct a hierarchical clustering model, reduces the total calibration quantity by combining dynamic sampling ratio, establishes a nonlinear dynamic relationship between the operating condition-related feature set and calibration parameters using a dual-branch neural network mapping model, and completes the calibration of the remaining equipment through key flow point testing and feature completion. At the same time, it introduces dynamic verification and model iteration mechanisms to ensure accuracy. It effectively solves the problems of long cycle, high resource consumption and insufficient accuracy of traditional calibration methods, and has the advantages of significantly shortening calibration time, reducing resource consumption, improving calibration accuracy and adaptability. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a method for a global calibration system for a mass flow controller based on multidimensional feature hierarchical sampling, as described in an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of a mass flow controller global calibration system based on multidimensional feature hierarchical sampling in an embodiment of the present invention. Detailed Implementation

[0017] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0018] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0019] Traditional mass flow controllers, when calibrated to high precision across the entire process before leaving the factory, suffer from problems such as long calibration time per unit, the need for a large amount of expensive standard equipment and professional personnel, severe constraints on production efficiency, high manufacturing costs, and long delivery cycles. Simplified calibration schemes cannot accurately reflect the nonlinear characteristics across the entire operating range, with significant errors in the low flow rate region. Calibration transfer schemes heavily rely on product consistency; when batch raw materials, process parameters, or core components change, manufacturing tolerance-induced response differences cause calibration transfer failures. Furthermore, existing methods lack adaptability to dynamic changes in production and cannot cope with factors such as process drift and equipment aging, requiring frequent recalibration of the calibration baseline.

[0020] In this regard, such as Figure 1 As shown, this embodiment of the application provides a global calibration method for a mass flow controller based on multidimensional feature hierarchical sampling.

[0021] One implementation method involves acquiring the four-dimensional device feature vectors of each mass flow controller within a production batch. These feature vectors can include the batch identifier of the flow sensing element, the machining tolerance level of the control valve core, the temperature and humidity range code of the assembly environment, and the gas medium compatibility type code. This feature vector can be obtained through manual input, scanning equipment labels, or exporting data from the production management system. For example, the batch identifier of the flow sensing element can be obtained by reading the QR code on the element; the machining tolerance level of the control valve core can be queried from production records; the temperature and humidity range code of the assembly environment can be determined based on environmental monitoring data from the assembly workshop; and the gas medium compatibility type code can be determined based on the equipment model or configuration list.

[0022] As one implementation approach, a multi-level hierarchical clustering model is constructed based on the four-dimensional device feature vector. This model divides the same batch of mass flow controllers into multiple homogeneous sample clusters, ensuring that the feature differences within each cluster are less than a preset threshold, and the differences between clusters are greater than a preset threshold. This model can be constructed using various clustering algorithms; for example, it can group devices based on feature vector similarity calculations using hierarchical clustering or density clustering methods. During the partitioning process, a fixed threshold can be set to evaluate the similarity of devices within clusters and the differences between devices between clusters.

[0023] As one implementation approach, for each sample cluster, sample devices are selected according to a dynamic sampling ratio, which adaptively adjusts based on the number of devices and the feature dispersion within the sample cluster. The sampling ratio can be set based on empirical values ​​or preset rules. For example, a base sampling ratio can be set and then fine-tuned using a simple linear function based on the size of the sample cluster or the distribution of its internal features.

[0024] As one implementation method, full-flow-domain multi-condition calibration is performed on the sample equipment, covering normal temperature, high temperature, normal pressure, and low pressure conditions. Sample response characteristic data, operating condition coupling parameters, and corresponding sample calibration parameter data are collected. The calibration process can be completed by manually adjusting the flow rate and operating condition parameters and using high-precision standard equipment to measure point by point. For example, multiple flow points can be set, and the temperature and pressure can be manually adjusted at each flow point. After the equipment stabilizes, its output signal and actual flow value are recorded.

[0025] From the sample response characteristic data and the coupling parameters of this operating condition, a set of operating condition-related features is extracted. This set of features includes the flow-voltage nonlinearity coefficient, valve opening-flow hysteresis, temperature drift compensation coefficient, and pressure fluctuation sensitivity. Feature extraction can be performed using data analysis software. For example, the nonlinear relationship between flow and voltage can be calculated through curve fitting, the hysteresis can be obtained by comparing the time difference between valve opening and flow rate changes, and the influence coefficients of temperature and pressure changes on flow rate can be calculated using experimental data.

[0026] An attention-based weighted fusion algorithm is introduced to assign differentiated weights to feature parameters under different operating conditions. The weights of different feature parameters can be determined through expert experience or simple statistical analysis methods. For example, based on the empirical correlation between feature parameters and calibration results, weight values ​​can be manually set, or the importance of features can be initially assessed through methods such as principal component analysis.

[0027] A two-branch neural network mapping model is constructed, where the first branch receives flow detection features and the second branch receives valve control features. A nonlinear dynamic mapping relationship between the operating condition-related feature set and calibration parameter data is established through a feature cross-fusion layer. A regularization term is introduced to suppress overfitting, completing model training and validation. A standard multilayer perceptron can be used as the two-branch neural network structure, with each branch receiving its own feature input. The outputs of the two branches are then simply concatenated and input into a fully connected layer as a fusion layer. Regularization can employ L1 regularization or simple weight decay. Model training can be performed using a gradient descent optimizer, and validation is conducted using a subset of data.

[0028] For the remaining mass flow controllers within the batch, three non-equally spaced key test flow points were selected, including the low flow inflection point, the medium flow linear region, and the high flow saturation region. Remaining response characteristic data were collected under standard operating conditions. Key test flow points can be selected based on equipment type or experience. For example, 10%, 50%, and 90% of full scale can be selected as test points. Under standard operating conditions, the flow rate was manually set and the equipment response data was recorded.

[0029] The remaining response characteristic data is then subjected to feature completion and operating condition correction. Non-test flow point data is completed using an interpolation algorithm, and characteristic deviations caused by temperature drift and pressure disturbances are corrected based on real-time environmental parameters. Feature completion can employ linear or polynomial interpolation methods. Operating condition correction can be performed using pre-established temperature drift and pressure compensation lookup tables, finding the corresponding correction values ​​based on real-time environmental parameters.

[0030] The corrected feature set to be mapped is input into the dual-branch neural network mapping model, which outputs a prediction calibration parameter matrix for the entire flow domain and multiple operating conditions, thus completing the calibration of the remaining mass flow controller. The completed and corrected feature data is input into the pre-trained neural network model, which directly calculates and outputs the corresponding calibration parameters.

[0031] Validation samples are drawn from the calibrated mass flow controllers in each sample cluster according to a risk-priority sampling rule. Risk-priority sampling can be based on simple rules. For example, a subset of calibrated devices can be randomly selected for validation, or selection can be based on certain anomalies recorded during the production process.

[0032] Perform a full-condition verification calibration on the validation samples, obtain the validation calibration parameter data, and calculate the multidimensional deviation matrix between the predicted calibration parameters and the validation calibration parameters. The verification calibration can employ a full-condition calibration method similar to that used for the initial sample calibration. The deviation matrix can be calculated by comparing the differences between the predicted and actual measured values ​​point by point.

[0033] When the deviation of any operating condition or flow point in the multidimensional deviation matrix exceeds a preset dynamic threshold, the validation sample data is added to the incremental sample library. The preset dynamic threshold can be a fixed upper limit of error. When the deviation of any test point exceeds this fixed threshold, the data of that validation sample is added to the incremental sample library.

[0034] When the incremental sample database reaches a threshold, a transfer learning fine-tuning strategy is employed. This reuses the original model's basic parameters, updates the weights of the feature fusion layer and output layer, and completes the iterative update of the mapping model. Transfer learning fine-tuning can be achieved by retraining the entire model on the incremental sample database data, or by training only the last few layers of the model.

[0035] In one embodiment, this application proposes to normalize the four-dimensional device feature vector, calculate the Euclidean distance of each feature dimension, and use the K-means++ clustering algorithm to complete the initial cluster division; introduce the silhouette coefficient to evaluate the clustering effect, iteratively optimize the number of clusters until the intra-cluster compactness and inter-cluster separation both meet the preset indicators; the dynamic sampling ratio is: the sampling ratio is equal to the sum of the base ratio multiplied by 1 and the feature dispersion coefficient, the base ratio is 5%-10%, and the feature dispersion coefficient is 0-0.5.

[0036] As an implementation approach, the four-dimensional device feature vectors are first normalized when constructing a multi-level hierarchical clustering model. Normalization aims to eliminate differences in units and numerical ranges between different feature dimensions, ensuring all features have the same weight during clustering and preventing certain features with larger values ​​from dominating distance calculations. Common normalization methods include Min-Max standardization (scaling the data to the [0,1] interval) or Z-score standardization (converting the data to a distribution with a mean of 0 and a standard deviation of 1). After feature vector normalization, the Euclidean distance between each feature dimension is calculated. Euclidean distance is a commonly used distance metric to quantify the similarity or difference between feature vectors of different devices; it is calculated as the square root of the sum of the squares of the differences between corresponding dimensions. Based on these distances, the K-means++ clustering algorithm is used for initial cluster partitioning. The K-means++ algorithm optimizes the selection of initial cluster centers, avoiding the local optimum problem that might result from the random selection of initial centers by the K-means algorithm, thus improving the stability and accuracy of the clustering results.

[0037] To ensure the quality of clustering results, this application introduces the silhouette coefficient to evaluate the clustering effect and uses it as a basis for iterative optimization of the number of clusters. The silhouette coefficient is an indicator of clustering quality, comprehensively considering both the compactness (cohesion) of samples within a cluster and the separation of samples between clusters. The silhouette coefficient ranges from [-1, 1], with a larger value indicating better clustering. By trying different cluster number values ​​K, the silhouette coefficient is calculated for each K value, and the K value that maximizes the silhouette coefficient or meets the preset criteria is selected to determine the optimal number of clusters. Intra-cluster compactness refers to the minimum distance between samples within the same cluster, indicating high similarity among samples within the cluster; inter-cluster separation refers to the maximum distance between samples in different clusters, indicating significant differences between different clusters. When both indicators meet the preset requirements, it indicates that the clustering results have good structure and discriminative power.

[0038] Furthermore, the dynamic sampling ratio used in this application is defined as the sum of the base ratio multiplied by 1 and the feature dispersion coefficient. The base ratio is a preset fixed sampling ratio, ranging from 5% to 10%, used to ensure that a certain number of devices in each sample cluster are sampled for full-flow-domain multi-condition calibration. The feature dispersion coefficient is a dynamically adjusted factor, ranging from 0 to 0.5, used to reflect the degree of variation in device characteristics within the sample cluster. For example, the feature dispersion can be obtained by calculating the standard deviation, variance, or mean absolute deviation of the device feature vectors within the cluster and mapping it to the range of 0-0.5. When the feature dispersion of devices within a sample cluster is high, the feature dispersion coefficient will increase accordingly, thereby increasing the sampling ratio of that cluster to ensure sufficient coverage of the feature diversity of devices within the cluster; conversely, when the feature dispersion of devices within a sample cluster is low, the feature dispersion coefficient will decrease, reducing the sampling ratio and avoiding unnecessary duplicate calibration.

[0039] In one embodiment, a full-flow-domain multi-condition calibration of the sample device is proposed, specifically including: the flow domain covering 0-100% full scale, setting 15-20 equally spaced flow points; the operating conditions covering a temperature range of 25-50℃ and a pressure range of 0.1-0.5MPa; the stabilization time of each operating condition-flow point being greater than or equal to 60s, and synchronously collecting data on the output voltage of the flow detection element, the opening degree of the regulating valve, the ambient temperature and humidity, and the pipeline pressure to form sample response characteristic data; the sample calibration parameter data including the flow deviation correction coefficient, temperature drift compensation value, and pressure correction value for each operating condition-flow point.

[0040] As an implementation approach, the flow domain covers 0-100% of the full scale, aiming to ensure that the performance of the mass flow controller is fully evaluated across the entire measurable flow range, thereby comprehensively understanding its response characteristics in the low-flow, medium-flow linear range, and high-flow saturation range. Setting 15-20 equally spaced flow points provides sufficient sampling data density to accurately capture the nonlinear characteristics of the flow response curve, avoiding the loss of crucial information or excessive subsequent interpolation errors due to sparse sampling points. These flow points are typically precisely set and controlled using a high-precision flow generator or standard flow meter in conjunction with a precision regulating valve opening.

[0041] The operating conditions cover a temperature range of 25-50℃ and a pressure range of 0.1-0.5MPa, designed to simulate typical environmental temperature and pressure conditions that mass flow controllers may encounter in real-world industrial applications. Temperature and pressure are key environmental parameters affecting the physical properties of gases and the accuracy of flow measurement. Calibration within these ranges can fully reveal the performance drift, cross-sensitivity, and coupling effects of the equipment under different temperature and pressure conditions. Temperature control can be achieved through a high-precision thermostatic chamber or environmental simulation equipment, while pressure control can be accomplished through a combination of a precision pressure regulating valve and a high-precision pressure sensor.

[0042] The stabilization time for each operating condition-flow point is greater than or equal to 60 seconds. This ensures that the mass flow controller, its surrounding gas flow field, and environmental conditions have reached a steady state at each specific test point, thus avoiding interference from transient responses or system fluctuations and guaranteeing the accuracy and representativeness of the collected data. Simultaneous acquisition of data includes the output voltage of the flow sensing element, the opening of the regulating valve, ambient temperature and humidity, and pipeline pressure. This aims to comprehensively record the equipment's input (regulating valve opening), output (flow sensing element output voltage), and all relevant environmental influencing factors (ambient temperature and humidity, pipeline pressure) under specific operating conditions. These multi-dimensional data collectively constitute the sample response characteristic data, providing a comprehensive raw basis for subsequent in-depth analysis of equipment behavior and the establishment of accurate mapping models. Data acquisition is typically achieved through high-precision sensors and automated data acquisition systems (such as industrial control systems or dedicated testing platforms).

[0043] The sample calibration parameter data includes flow deviation correction coefficients, temperature drift compensation values, and pressure correction values ​​for each operating condition and flow point. These parameters are key data for quantifying and compensating for potential performance deviations of the mass flow controller under different operating conditions and flow points. The flow deviation correction coefficient is used to correct measurement errors at specific flow rates, ensuring the accuracy of the measured values; the temperature drift compensation value is used to offset the impact of ambient temperature changes on flow measurement results, improving temperature adaptability; and the pressure correction value is used to correct measurement deviations caused by pipeline pressure fluctuations, enhancing pressure robustness. These parameters are typically calculated and derived by comparing the device to be calibrated with a high-precision standard device and combining real-time environmental parameters. They are the core objective of building an accurate mapping model and achieving global calibration prediction.

[0044] In one embodiment, this application proposes a weighted fusion algorithm that introduces an attention mechanism to assign differentiated weights to feature parameters under different operating conditions. Specifically, it includes: calculating the Pearson correlation coefficient between each feature parameter and the calibration parameter, generating initial attention weights based on the correlation coefficients, and iteratively optimizing the weights through backpropagation of the model so that the weight of strongly correlated features accounts for more than or equal to 60%, and the weight of weakly correlated features accounts for less than or equal to 10%, thereby achieving differentiated weighting of feature importance.

[0045] As one implementation method, when calculating the Pearson correlation coefficient between each feature parameter and the calibration parameter, and generating initial attention weights based on the correlation coefficient, the Pearson correlation coefficient is a statistic that measures the degree of linear correlation between two variables. Its value ranges from -1 to 1, with a larger absolute value indicating a stronger correlation. In this step, for each feature parameter in the set of condition-related features extracted from the sample response characteristic data and the condition coupling parameters (e.g., flow-voltage nonlinear coefficient, valve opening-flow hysteresis, temperature drift compensation coefficient, pressure fluctuation sensitivity), and the corresponding sample calibration parameter data (e.g., flow deviation correction coefficient, temperature drift compensation value, pressure correction value), the Pearson correlation coefficient is calculated. This quantifies the strength of the linear correlation between each feature parameter and the final calibration parameter. Subsequently, based on these correlation coefficients, the initial weights of the attention mechanism can be generated. For example, the absolute value of the correlation coefficient can be normalized and used as the basis for the initial weights; features with stronger correlations are assigned higher initial weights, thus guiding the model to focus on more important features in the early stages of model training.

[0046] When iteratively optimizing weights using model backpropagation, a widely used optimization algorithm in neural network training, the model adjusts its internal parameters, including the weights of the attention mechanism, based on the error between the predicted output and the true calibration parameters (measured by a loss function). In each training iteration, the error signal is backpropagated from the output layer to the input layer, and the attention weights are gradually adjusted according to the gradient descent principle to reduce the prediction error. This iterative optimization process allows the attention weights to adaptively learn and reflect the actual contribution of feature parameters to the calibration parameters under different conditions, thereby continuously improving the model's ability to identify the importance of features.

[0047] To achieve differentiated weighting of feature importance, with strongly correlated features accounting for at least 60% of the weight and weakly correlated features accounting for at least 10%, constraints on weight allocation are introduced during the model's backpropagation iterative optimization process. Specifically, for "strongly correlated features" with high Pearson correlation coefficients to the calibration parameters, their weight in the total weight is mandated to be at least 60%; while for "weakly correlated features" with low correlation to the calibration parameters, their weight is limited to at least 10%. This differentiated weighting strategy ensures that the model can significantly amplify the influence of key features when fusing features, while effectively suppressing the interference of secondary or noisy features. In this way, the model can focus more on the features that have the greatest impact on the calibration results, thereby achieving refined and differentiated weighting of feature importance and improving the model's interpretability and prediction accuracy.

[0048] In one embodiment, a dual-branch neural network mapping model is proposed, specifically including: a first branch is a one-dimensional convolutional neural network, used to extract local correlation information of flow detection features; a second branch is a long short-term memory network, used to capture the temporal variation pattern of valve control features; the feature cross-fusion layer adopts a splicing and fully connected structure to fuse the output features of the two branches; the output layer outputs the flow deviation correction coefficient matrix under the full flow domain and multiple operating conditions, and the model loss function adopts the sum of mean square error and L2 regularization term.

[0049] As an implementation approach, the first branch employs a one-dimensional convolutional neural network (1DCNN), designed to efficiently process flow detection features. Flow detection features, such as flow-voltage nonlinearity coefficients and temperature drift compensation coefficients, typically exhibit specific patterns or trends within a local range. By sliding convolutional kernels across these one-dimensional sequence data, the 1DCNN can automatically learn and extract these local feature patterns, such as the nonlinear trend of voltage output within a specific flow range, or the local correlation of temperature drift coefficients at different temperature points. The input to this branch can be a pre-processed sequence of flow detection features, which, through multiple convolutions, activation functions (such as ReLU), and pooling operations, generates a high-level feature representation with local correlations.

[0050] The second branch employs a Long Short-Term Memory (LSTM) network specifically for processing valve control features. Valve control features, such as valve opening-flow hysteresis and pressure fluctuation sensitivity, often exhibit significant temporal dependencies and dynamic changes. As a special type of recurrent neural network, LSTM, through its internal gating mechanisms (input gate, forget gate, output gate), can effectively capture the long-term dependencies and dynamic evolution of these features over time; for example, the dynamic hysteresis behavior of the flow response when a valve is adjusted from one opening to another. The input to this branch can be a time series of valve control features; after processing by the LSTM network units, the output is a feature representation containing rich temporal information.

[0051] To fully utilize the heterogeneous features extracted from the two branches, this application designs a feature cross-fusion layer. This fusion layer employs a concatenation and fully connected structure to effectively integrate the local correlation features for flow detection output from the first branch (a one-dimensional convolutional neural network) with the temporal dynamic features for valve control output from the second branch (a long short-term memory network). Specifically, a "concatenation" operation connects the feature vectors of the two branches dimensionally, forming a longer comprehensive feature vector that retains all information from different modalities. Subsequently, this comprehensive feature vector is input into one or more fully connected layers. These fully connected layers learn and establish complex interactions and higher-order correlations between the two types of features through nonlinear transformations, thereby generating a unified and more representative fusion feature.

[0052] Finally, the output layer is responsible for mapping the fused features to the final prediction target, namely the flow deviation correction coefficient matrix under full flow domain and multiple operating conditions. This output layer typically consists of one or more fully connected layers, with the number of neurons matching the dimension of the calibration parameter matrix to be predicted. To ensure the model's training effectiveness and generalization ability, the model loss function uses the sum of mean squared error (MSE) and L2 regularization. MSE is used to accurately measure the difference between the flow deviation correction coefficients predicted by the model and the true calibration values, ensuring prediction accuracy. The L2 regularization term effectively suppresses overfitting by penalizing the model weights, enhancing the model's generalization ability and enabling it to maintain stable prediction performance when facing unseen mass flow controllers or new operating conditions.

[0053] In one embodiment, this application proposes feature completion and operating condition correction for the remaining response characteristic data, specifically including: using a cubic spline interpolation algorithm to complete the response data of non-test flow points, with the error controlled to be less than or equal to 0.2% of the full scale; the operating condition correction is based on real-time collected ambient temperature and humidity and pipeline pressure data, and calls a preset temperature drift-pressure compensation formula to correct the deviation of characteristic parameters.

[0054] As one implementation method, cubic spline interpolation is used to complete the response data of non-test flow points, with an error control of less than or equal to 0.2% of full scale. This means that using the cubic spline interpolation algorithm, a smooth and continuous curve is constructed based on the response data collected at three non-equally spaced key test flow points, thereby calculating the response data of all non-test flow points in the entire flow domain. The cubic spline interpolation algorithm approximates the data through piecewise cubic polynomials, ensuring the continuity of function values, first derivatives, and second derivatives at the connection points. This results in a smooth and continuous data curve after completion, which can more accurately reflect the true response characteristics of the mass flow controller across the entire flow range. By strictly controlling the error of the completed data to less than or equal to 0.2% of full scale, high accuracy of data completion is ensured, providing a reliable foundation for subsequent calibration parameter prediction. In addition to cubic spline interpolation, other higher-order interpolation methods, such as Lagrange interpolation, Newton interpolation, or radial basis function interpolation, can be used, but cubic spline interpolation achieves a better balance between smoothness and computational efficiency.

[0055] Operating condition calibration, based on real-time acquired ambient temperature, humidity, and pipeline pressure data, uses preset temperature drift-pressure compensation formulas to correct characteristic parameter deviations. This involves simultaneously collecting real-time data on ambient temperature, humidity, and internal pipeline pressure while testing key flow points on the remaining mass flow controller. This real-time operating condition data is a key external factor affecting the measurement accuracy of the mass flow controller. The preset temperature drift-pressure compensation formulas are established beforehand using extensive experimental data or physical models. They quantify the deviations of characteristic parameters such as the output voltage of the flow sensing element and the opening of the regulating valve under different temperature and pressure conditions relative to standard operating conditions. By inputting the real-time acquired temperature, humidity, and pressure data into these preset compensation formulas, the corresponding compensation values ​​can be accurately calculated and applied to the originally acquired characteristic parameters, effectively correcting the characteristic parameter deviations caused by fluctuations in ambient temperature, humidity, and pipeline pressure. This allows the corrected characteristic parameters to more accurately represent the performance of the mass flow controller under standard operating conditions, providing a more stable and accurate input for subsequent mapping model predictions.

[0056] In one embodiment, this application proposes three non-equal interval key test flow points: the low flow inflection point is 5%-15% of full scale, the medium flow linear region is 45%-55% of full scale, and the high flow saturation region is 85%-95% of full scale; the stabilization time of each test flow point is greater than or equal to 30s, and two types of data are collected: flow detection voltage and valve opening.

[0057] As an implementation approach, the low-flow inflection point refers to the region where the mass flow controller's output response may exhibit nonlinearity or unstable startup characteristics in the low-flow range. Setting this test point within 5%-15% of full scale aims to capture the device's sensitivity and nonlinear behavior during initial startup or at low flow rates, which is crucial for accurately modeling the device's low-flow performance. The medium-flow linear region is typically the area where the mass flow controller operates most frequently and exhibits the most stable performance, with a good linear relationship between its output response and flow rate. Setting this test point within 45%-55% of full scale effectively captures the device's linear characteristics within its main operating range, providing reliable benchmark data for subsequent interpolation and calibration. The high-flow saturation region refers to the region where the mass flow controller's output response may saturate, exhibit increased nonlinearity, or decrease in response speed as it approaches its maximum flow capacity. Setting this test point within 85%-95% of full scale aims to capture the device's performance under high-flow-limit conditions, which is significant for ensuring prediction accuracy across the entire flow domain. Stabilization time refers to the time the device needs to maintain a given flow state before data acquisition at each key test flow point. Setting a settling time of 30 seconds or more ensures that the mass flow controller reaches steady-state operating conditions at each test point, eliminating the influence of transient responses or fluctuations on the measurement results, thereby obtaining more accurate and representative response characteristic data. The flow detection voltage is the direct output signal of the flow detection element, reflecting the actual flow rate; the valve opening is the control input of the regulating valve, reflecting the equipment's ability to regulate flow. Simultaneously collecting these two types of data comprehensively characterizes the input-output relationship and control characteristics of the mass flow controller at specific flow points, providing necessary and sufficient data support for subsequent feature completion, operating condition correction, and the input of the dual-branch neural network mapping model.

[0058] In one embodiment, this application proposes to extract verification samples according to a risk-priority sampling rule, specifically including: assigning a risk score to each calibrated mass flow controller, wherein the risk score is equal to the feature dispersion score multiplied by the operating condition fluctuation sensitivity score; and then, extracting the top 5%-8% of devices with the highest risk scores as verification samples.

[0059] As an implementation method, risk scoring aims to quantify the potential errors or uncertainties that may exist in the prediction calibration process for each calibrated mass flow controller. This scoring mechanism helps identify devices that are more likely to cause inaccurate model predictions, thus prioritizing their validation. Risk scores can be calculated based on multiple dimensions, including the device's own characteristics, historical data performance, and the confidence level of the model predictions. Feature dispersion scoring measures the degree of deviation of the mass flow controller's own characteristics (such as batch identification of flow sensing elements, valve core machining tolerance level, assembly environment temperature and humidity range coding, gas medium adaptation type coding, etc.) within its sample cluster. Devices with high dispersion may represent marginal or anomalies within the sample cluster, and their prediction results may be more unstable. This score can be quantified by calculating the distance between the device's feature vector and the central feature vector of the sample cluster, or by using statistical methods (such as variance, standard deviation). For example, a threshold can be set, assigning a higher dispersion score when the device's feature dispersion exceeds this threshold. Operating condition fluctuation sensitivity scoring assesses the sensitivity of the mass flow controller's response characteristics or calibration parameters to changes in operating conditions (such as temperature and pressure variations). Highly sensitive equipment may exhibit significant performance fluctuations even with minor changes in operating conditions, making it difficult for predictive models to accurately capture its behavior. This rating can be determined by analyzing the trends and magnitudes of calibration parameters (such as flow deviation correction coefficients, temperature drift compensation values, and pressure correction values) as the equipment undergoes multi-condition calibration across the entire flow domain. For example, partial derivatives of calibration parameters with respect to temperature or pressure can be calculated, or historical data can be analyzed to assess performance stability under different operating conditions, thereby assigning a corresponding sensitivity score. After calculating the risk scores for all calibrated mass flow controllers, the system sorts them according to these scores. The equipment with the highest risk scores (5%-8%) is selected as the validation sample, ensuring that the validation process focuses on those devices most likely to have predictive errors or pose the greatest challenge to the model's generalization ability.

[0060] By assigning a risk score to each calibrated mass flow controller and calculating this risk score based on feature dispersion and operating condition fluctuation sensitivity scores, this application can accurately identify devices within a batch that have high feature deviations or are highly sensitive to changes in operating conditions. These devices are often the "difficult points" or "marginal" samples where the prediction model is most prone to errors. By prioritizing the sampling of devices with the top 5%-8% of risk scores as validation samples, limited validation resources can be concentrated on devices that best reflect potential model defects and insufficient generalization ability. This targeted sampling strategy significantly improves the efficiency and effectiveness of the validation process, enabling the model to more quickly identify and correct its prediction biases under complex or abnormal conditions, thereby accelerating the iterative update of the mapping model and ultimately improving the accuracy and reliability of the overall calibration of the mass flow controllers in the entire batch.

[0061] In one embodiment, this application proposes a transfer learning fine-tuning strategy that reuses the basic parameters of the original model and updates the weights of the feature fusion layer and the output layer. Specifically, this strategy includes freezing the parameters of the bottom feature extraction layer of the dual-branch neural network, training the feature cross-fusion layer and the output layer only with incremental sample library data; at the same time, limiting the number of training rounds to less than or equal to 20 rounds, and setting the learning rate to 1 / 10 of the original model.

[0062] As an implementation approach, transfer learning fine-tuning is a machine learning method that aims to transfer knowledge learned from one task to another related task, thereby improving learning efficiency and model performance. Here, by leveraging the foundational knowledge of a pre-trained bi-branch neural network mapping model, local adjustments are made to the newly added incremental sample database, thus avoiding training from scratch, accelerating model convergence, and improving update efficiency. Reusing the original model's basic parameters means retaining and using the weights and biases obtained during the initial training phase of the bi-branch neural network mapping model during iterative updates. These basic parameters contain general feature representations and mapping rules learned from a large amount of sample device data. Reusing them ensures that the model can still utilize its existing powerful feature extraction capabilities and understanding of complex operating conditions when processing new data. Updating the weights of the feature fusion layer and output layer is crucial. The feature fusion layer integrates and crosses different types of features extracted by the bi-branch neural network (first and second branches) to form a more discriminative comprehensive feature representation. The output layer then predicts the calibration parameter matrix for the entire flow domain and multiple operating conditions based on these comprehensive features. Updating the weights of these two layers means the model will focus on learning how to map the specific changes reflected by new incremental sample data to the final calibration parameters, thereby achieving localized and refined adjustments to the model. Freezing the parameters of the bottom feature extraction layer of the bi-branch neural network—which is typically responsible for learning general, basic feature representations from the original input data, such as local correlation information for flow detection features or temporal variation patterns for valve control features—effectively preserves the general feature extraction capabilities learned by the model in the initial training phase, preventing interference or "catastrophic forgetting" of these basic features when the incremental data volume is small. Training the feature cross-fusion layer and output layer only with incremental sample data means that during model iteration and updates, only newly collected incremental sample data with specific biases or representing new operating conditions is used to adjust the parameters of the feature cross-fusion layer and output layer. This targeted training method enables the model to quickly adapt to the local changes reflected by incremental data, while avoiding the computational resources required for large-scale training of the entire model, thus improving update efficiency. The number of training epochs should be less than or equal to 20. An epoch refers to the number of times the model traverses the entire dataset during training. Limiting the number of training epochs to a small number (e.g., less than 20) is to prevent overfitting when the model is trained using only incremental sample data. Since the incremental sample data is typically much smaller than the initial training dataset, too many training epochs can cause the model to overlearn the noise and idiosyncratic patterns of the incremental data, thus reducing its generalization ability on unseen data.The learning rate is set to 1 / 10 of the original model's value. The learning rate is a hyperparameter in neural network training that determines the step size at which the model updates weights in each iteration. Setting the learning rate to 1 / 10 of the original model's training value means that the model will adjust the weights more cautiously and precisely during the fine-tuning phase. A smaller learning rate helps the model fine-tune based on its existing good parameters, avoiding drastic fluctuations in model parameters due to an excessively large learning rate, which could destroy the original learning results and help the model converge stably to a better solution.

[0063] It should be noted that the above description describes some embodiments of the present invention. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps described in the claims may be performed in a different order than that shown in the above embodiments and still achieve the desired results. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0064] Based on the same inventive concept, corresponding to any of the above embodiments, such as Figure 2 As shown, this application also provides a mass flow controller global calibration system based on multidimensional feature stratified sampling, used for the above-mentioned mass flow controller global calibration method based on multidimensional feature stratified sampling, including: The stratified sampling and differential calibration unit is used to perform batch multidimensional stratified sampling and sample differential full-condition calibration, and output sample response characteristic data, condition coupling parameters and sample calibration parameter data. The working condition coupling mapping model building unit is used to extract the working condition associated feature set, perform weighted fusion through an attention mechanism, and build and train a two-branch neural network mapping model. The simplified testing and intelligent prediction unit is used to perform key flow point testing, feature completion and operating condition correction on the remaining mass flow controller, and input the model to output the prediction calibration parameter matrix. The dynamic verification and closed-loop update unit is used for risk-priority sampling verification, multidimensional deviation calculation, incremental sample library management, and iterative updates of the mapping model transfer learning.

[0065] The system described above is used to implement the corresponding methods in the foregoing embodiments and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0066] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for global calibration of a mass flow controller based on multidimensional feature stratified sampling, characterized in that, Includes the following steps: Obtain the four-dimensional device feature vector of each mass flow controller in the production batch. The four-dimensional device feature vector includes the batch identifier of the flow detection element, the processing tolerance level of the regulating valve core, the temperature and humidity range code of the assembly environment, and the code of the gas medium adaptation type. Based on the four-dimensional device feature vector, a multi-level hierarchical clustering model is constructed to divide the same batch of mass flow controllers into multiple homogeneous sample clusters, so that the difference in device features within each sample cluster is less than a preset threshold and the difference between clusters is greater than a preset threshold. For each sample cluster, sample devices are drawn according to a dynamic sampling ratio, which is adaptively adjusted according to the number of devices and the feature dispersion within the sample cluster. Perform full-flow-domain multi-condition calibration on the sample equipment, and collect sample response characteristic data, condition coupling parameters and corresponding sample calibration parameter data; The working condition correlation feature set is extracted from the sample response characteristic data and the working condition coupling parameters. The working condition correlation feature set includes the flow-voltage nonlinear coefficient, valve opening-flow hysteresis, temperature drift compensation coefficient, and pressure fluctuation sensitivity. An attention-based weighted fusion algorithm is introduced to assign differentiated weights to feature parameters under different working conditions; A dual-branch neural network mapping model is constructed. The first branch inputs flow detection features and the second branch inputs valve control features. A nonlinear dynamic mapping relationship between the working condition related feature set and the calibration parameter data is established through a feature cross-fusion layer. A regularization term is introduced to suppress overfitting and complete the model training and validation. For the remaining mass flow controllers in the batch, select several non-equal interval key test flow points, including low flow inflection point, medium flow linear region, and high flow saturation region, and collect the remaining response characteristic data under standard operating conditions; The remaining response characteristic data are feature-completed and operating condition-corrected. Non-test flow point data are completed by interpolation algorithm, and feature deviations caused by temperature drift and pressure interference are corrected based on real-time environmental parameters. The corrected feature set to be mapped is input into the dual-branch neural network mapping model, which outputs the prediction calibration parameter matrix under the full flow domain and multiple operating conditions, thus completing the calibration of the remaining mass flow controller. From the calibrated mass flow controllers in each sample cluster, validation samples are drawn according to the risk-priority sampling rule, and full-condition verification calibration is performed on the validation samples to obtain validation calibration parameter data in order to calculate the multidimensional deviation matrix between the predicted calibration parameters and the validation calibration parameters. When the deviation of any operating condition or flow point in the multidimensional deviation matrix exceeds the preset dynamic threshold, the verification sample data will be included in the incremental sample library. When the incremental sample library reaches a threshold, a transfer learning fine-tuning strategy is adopted to reuse the original model's basic parameters, update the weights of the feature fusion layer and the output layer, and complete the iterative update of the mapping model.

2. The method according to claim 1, characterized in that, The construction of a multi-level hierarchical clustering model based on the four-dimensional device feature vector includes: The four-dimensional device feature vector is normalized, the Euclidean distance of each feature dimension is calculated, and the K-means++ clustering algorithm is used to complete the initial cluster division. The silhouette coefficient is introduced to evaluate the clustering effect, and the number of clusters is iteratively optimized until the intra-cluster compactness and inter-cluster separation both meet the preset indicators. The dynamic sampling ratio is: the sampling ratio is equal to the sum of the base ratio multiplied by 1 and the feature dispersion coefficient, the base ratio is 5%-10%, and the feature dispersion coefficient is 0-0.

5.

3. The method according to claim 1, characterized in that, The calibration of the sample equipment across the entire flow domain under multiple operating conditions includes: The flow domain covers 0-100% of the full range, and 15-20 equally spaced flow points are set; Operating conditions cover a temperature range of 25-50℃ and a pressure range of 0.1-0.5MPa; The stabilization time of each operating condition-flow point is greater than or equal to 60 seconds. Simultaneously, the output voltage of the flow detection element, the opening degree of the regulating valve, the ambient temperature and humidity, and the pipeline pressure data are collected to form the sample response characteristic data. The sample calibration parameter data includes the flow deviation correction coefficient, temperature drift compensation value, and pressure correction value for each operating condition-flow point.

4. The method according to claim 1, characterized in that, The weighted fusion algorithm incorporating an attention mechanism assigns differentiated weights to feature parameters under different operating conditions, including: Calculate the Pearson correlation coefficient between each feature parameter and the calibration parameter, and generate initial attention weights based on the correlation coefficients. Iteratively optimize the weights through backpropagation of the model, so that the weight of strongly correlated features accounts for greater than or equal to 60%, and the weight of weakly correlated features accounts for less than or equal to 10%, thereby achieving differentiated weighting of feature importance.

5. The method according to claim 1, characterized in that, The dual-branch neural network mapping model includes: The first branch is a one-dimensional convolutional neural network, used to extract local correlation information of traffic detection features; The second branch is a long short-term memory network, used to capture the temporal variation patterns of valve control features; The feature cross-fusion layer adopts a splicing and fully connected structure to fuse the output features of the two branches; The output layer outputs the flow deviation correction coefficient matrix for the entire flow domain and multiple operating conditions. The model loss function is the sum of mean square error and L2 regularization term.

6. The method according to claim 1, characterized in that, The feature completion and operating condition correction of the remaining response characteristic data includes: The response data of non-test flow points are completed using a cubic spline interpolation algorithm, with the error controlled to be less than or equal to 0.2% of the full scale. The operating condition correction is based on real-time collected ambient temperature and humidity and pipeline pressure data, and calls the preset temperature drift-pressure compensation formula to correct the deviation of characteristic parameters.

7. The method according to claim 1, characterized in that, The three non-equal interval key test flow points are as follows: The low flow rate inflection point is 5%-15% of full scale, the medium flow rate linear region is 45%-55% of full scale, and the high flow rate saturation region is 85%-95% of full scale. The stabilization time for each test flow point is greater than or equal to 30 seconds, and two types of data are collected: flow detection voltage and valve opening.

8. The method according to claim 1, characterized in that, The sampling of verification samples according to the risk-priority sampling rule includes: Assign a risk score to each calibrated mass flow controller. The risk score is equal to the feature dispersion score multiplied by the operating condition fluctuation sensitivity score. Select the top 5%-8% of devices by risk score as the validation sample.

9. The method according to claim 1, characterized in that, The aforementioned transfer learning fine-tuning strategy reuses the original model's basic parameters and updates the weights of the feature fusion layer and output layer, including: Freeze the parameters of the bottom feature extraction layer of the dual-branch neural network, and train the feature cross-fusion layer and output layer only with incremental sample library data; The number of training rounds is less than or equal to 20, and the learning rate is 1 / 10 of the original model.

10. A global calibration system for a mass flow controller based on multidimensional feature stratified sampling, characterized in that, To implement the method according to any one of claims 1-9, comprising: The stratified sampling and differential calibration unit is used to perform batch multidimensional stratified sampling and sample differential full-condition calibration, and output sample response characteristic data, condition coupling parameters and sample calibration parameter data. The working condition coupling mapping model building unit is used to extract the working condition associated feature set, perform weighted fusion through an attention mechanism, and build and train a two-branch neural network mapping model. The simplified testing and intelligent prediction unit is used to perform key flow point testing, feature completion and operating condition correction on the remaining mass flow controller, and input the model to output the prediction calibration parameter matrix. The dynamic verification and closed-loop update unit is used for risk-priority sampling verification, multidimensional deviation calculation, incremental sample library management, and iterative updates of the mapping model transfer learning.

Images