Trace gas multi-filling repeatability error analysis and compensation method

Abstract

The invention relates to the technical field of error analysis, in particular to a trace gas multi-filling repeatability error analysis and compensation method which comprises the following steps: collecting gas pressure pulsation data and container inner wall adsorption layer thickness change data in a filling stage; inputting the pressure pulsation data into a dynamic coupling analysis module to calculate correlation degree parameters of an adsorption hysteresis effect and a pressure fluctuation frequency spectrum; an adsorption-pressure dynamic coupling correction model is established through a phase-space reconstruction algorithm, and composite correction parameters including the valve response lead and the temperature compensation slope are generated; and the composite correction parameters are input into a filling execution mechanism, and reconstruction of an electromagnetic valve control time sequence and dynamic correction of a temperature control curve are completed before the next filling cycle is started. According to the method, the deep coupling characteristics causing the filling error can be identified, and the sensitivity and pertinence of error analysis are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of error analysis, in particular to a trace gas multiple filling repeatability error analysis and compensation method. BACKGROUND

[0002] In the process of filling high-purity or trace gas, due to the adsorption effect of the inner wall of the container, the gas will be unevenly adsorbed to the inner surface for a period of time after entering the container, resulting in a deviation between the actual filling amount and the target set amount. Especially in the multiple repeated filling task, due to the superimposed disturbance of the adsorption layer state, gas pressure fluctuation and temperature control response, it often causes unpredictable filling deviation and difficult to converge repeatability error, which challenges the high-precision process control.

[0003] In the prior art, the filling control system usually only relies on a single channel pressure sensor to obtain the gas flow state, and controls and corrects through simple threshold judgment or fixed compensation coefficient, lacking real-time perception and response ability to the adsorption dynamic behavior. In addition, the formation rate of the adsorption film is jointly affected by temperature, pressure and container material, which has strong nonlinearity and hysteresis. The traditional method is difficult to effectively quantify the coupling relationship between the adsorption lag process and the fluid disturbance, resulting in lack of adaptability of the control strategy, and the error in the filling process cannot be fundamentally compensated.

[0004] Especially in the gas metering scene pursuing high repeatability precision, such as trace standard gas configuration, micro-reaction gas filling or experimental ultra-clean gas control system, only through static or empirical compensation strategy, it is often difficult to realize stable control across the cycle, and there is also a lack of complete system for dynamic coupling modeling. SUMMARY

[0005] The present application provides a trace gas multiple filling repeatability error analysis and compensation method, which integrates multi-source real-time perception, dynamic correlation analysis and control instruction reconstruction of the whole process compensation method to improve the consistency and precision stability of the multiple filling process.

[0006] The trace gas multiple filling repeatability error analysis and compensation method comprises the following steps: S1, deploying a pressure sensor and an adsorption film thickness detection device (ultrasonic thickness measurement probe) in the filling pipeline, and synchronously collecting gas pressure pulsation data and container inner wall adsorption layer thickness change data during the filling stage; S2, inputting the pressure pulsation data obtained in S1 into a dynamic coupling analysis module, combining the adsorption layer thickness change data, and calculating the correlation degree parameters of the adsorption lag effect and the pressure fluctuation spectrum; S3, based on the correlation degree parameters output by S2, establishing an adsorption-pressure dynamic coupling correction model through phase space reconstruction algorithm to generate composite correction parameters including valve response advance and temperature compensation slope; S4, input the composite correction parameters of S3 into the filling actuator, complete the reconstruction of electromagnetic valve control timing and dynamic correction of temperature control curve before the next filling cycle starts.

[0007] Optionally, S1 specifically comprises: S11, three groups of pressure sensor arrays are arranged axially on the filling pipeline, respectively at 5 cm downstream of the inlet control valve, the middle section of the pipeline, and 3 cm close to the container interface; S12, the adsorption film thickness detection device adopts an ultrasonic thickness measurement probe, which is embedded in the inner wall of the container at an oblique incidence angle of 45° and is distributed equidistantly along the circumference at multiple detection points.

[0008] Optionally, the pressure sensor records the gas pressure pulsation data at a sampling rate of 1 kHz The ultrasonic probe scans the adsorption layer thickness at a frequency of 100 Hz to form adsorption layer thickness change data , wherein: represents the instantaneous value of the gas pressure collected by the pressure sensor at time t and the axial position x of the pipeline, represents the change amount of the adsorption film layer thickness corresponding to the angular position x of the container inner wall at time t relative to the reference initial thickness .

[0009] Optionally, S2 specifically comprises: S21, time-frequency joint analysis is performed on the pressure pulsation data , and a energy distribution matrix of a preset frequency band is extracted by wavelet packet decomposition ; S22, differential processing is performed on the adsorption layer thickness change data to obtain adsorption rate timing data ; S23, construct a pressure-adsorption combined phase space, and calculate the time-varying correlation coefficient of the energy distribution matrix and the adsorption rate ; S24, frequency spectrum analysis is performed on the time-varying correlation coefficient to extract the correlation degree parameter K in the preset characteristic frequency band.

[0010] Optionally, the time-varying correlation coefficient is represented as: ; wherein, is the adsorption-pressure coupling correlation coefficient at time t, is the standard deviation of the local energy matrix , standard deviation of the adsorption rate represents the sum of the sliding window in the specified frequency and position range.

[0011] Optionally, the S24 performs spectrum analysis on the time-varying correlation coefficient , calculates the power spectral density thereof, and extracts the correlation degree parameter K in the feature frequency band [0.5, 5] Hz, which satisfies: ; wherein represents the Fourier transform of the correlation coefficient, K is an energy proportion index of the adsorption-pressure coupling feature frequency band, which is an input correlation degree parameter of the subsequent dynamic correction model, and df is a frequency integral variable.

[0012] Optionally, the S3 specifically comprises: S31, constructing an adsorption-pressure joint phase space based on the Takens theorem, and constructing a multi-dimensional joint vector sequence from the correlation degree parameter, the pressure fluctuation data, and the adsorption layer thickness change rate : S32, determining an optimal embedding dimension m, and reconstructing the joint vector sequence into a phase space trajectory matrix M according to a time sliding window; S33, training a radial basis function neural network, and establishing a mapping relationship from the phase space trajectory to the compensation parameter: wherein is the electromagnetic valve response advance amount, is a temperature compensation curve slope, M is the phase space trajectory matrix, and RBFNN(·) is a radial basis function neural network function mapper, the input of which is a principal component vector, and the output of which is a two-dimensional compensation parameter.

[0013] Optionally, the radial basis function neural network function mapper comprises an input layer, a hidden layer, and an output layer, wherein: In the input layer structure design: the phase space trajectory matrix collected in the historical filling task is used as input data, the principal components covering the main characteristic energy are extracted through singular value decomposition method, and the principal components are normalized to form a standardized feature vector used for network modeling; In the hidden layer structure design, a radial basis function with a Gaussian kernel is selected, the cluster analysis is performed on the feature vectors of the training data to determine the basis function center, and the expansion scale parameter is adaptively set according to the statistical distribution of the samples in the cluster, the number of nodes in the hidden layer is not fixed, and the node size is optimized through the Akaike information criterion; ​In the output layer part, a double output structure is constructed, corresponding to the valve response advance and the temperature compensation slope target parameters respectively, the output node receives the weighted input from all the hidden layer nodes, and a linear mapping mode is adopted to realize the generation of the continuously adjustable compensation amount.

[0014] Optionally, the S4 specifically comprises: S41, electromagnetic valve timing reconstruction: according to the valve response advance generated in the last stage , reconstruct the electromagnetic valve control pulse sequence, including adjusting the original opening time forward; superimpose a high-frequency PWM modulation signal in the filling acceleration stage, and finally generate a discrete pulse sequence as the electromagnetic valve control instruction set; S42, dynamic correction of temperature control curve: construct a temperature compensation slope driven segmented temperature control function, in the pre-correction stage before filling, send the slope instruction to the temperature control unit in advance, and adopt a feedforward-feedback composite control; S43, control instruction fusion: through the time-triggered scheduler of the real-time operating system, complete the following before the next filling cycle starts: valve control instruction set is burned into the waveform generator of the FPGA; temperature control function parameters are written into the temperature control module of the PLC.

[0015] Optionally, the feedforward-feedback composite control comprises: feedforward term: based on predicted temperature compensation amount; feedback term: adjust the PID parameters according to the real-time temperature difference ΔT.

[0016] The beneficial effects of the present application are: In the present application, by deploying pressure sensors and ultrasonic thickness probes in the filling pipeline, a dual-channel data stream of adsorption film thickness and pressure pulsation is constructed, and a dynamic coupling analysis model is constructed based on wavelet packet decomposition, time-varying mutual information and frequency spectrum energy ratio. A quantifiable correlation degree parameter K is proposed, which effectively captures the micro nonlinear relationship between adsorption lag and fluid disturbance. Compared with the traditional single variable model which only relies on pressure fluctuation, the deep coupling characteristics causing filling error can be accurately identified, and the sensitivity and pertinence of error analysis are improved.

[0017] In the present application, Takens theorem is used to construct a joint embedded adsorption-pressure phase space, and radial basis function neural network is used to realize the mapping reasoning from historical trajectory to compensation parameters. Two-dimensional control parameters including electromagnetic valve response advance and temperature compensation slope are output. Through multi-level mechanisms such as dimension reduction and dynamic node optimization screening, the model generalization ability is improved, which has the ability to adaptively identify the source of control error in different filling environments and generate accurate correction parameters in real time. It breaks through the limitation that the traditional static correction mechanism in the cycle cannot cope with fluctuating disturbances.

[0018] This invention designs a hard real-time compensation execution architecture, based on FPGA waveform generator and PLC temperature control module respectively loading compensation instructions, integrating Δt-driven pre-adjustment pulse optimization and α-driven piecewise integral temperature control curve, and adopting a feedforward-feedback composite control strategy to improve the system's synergistic compensation effect for hysteresis adsorption, boundary disturbance and thermal inertia response, effectively avoid the accumulation of repetitive errors, and improve the long-term consistency and control accuracy of the filling system. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Fig. 1 This is a schematic diagram of the method flow according to an embodiment of the present invention; Fig. 2 This is a schematic diagram illustrating the filling process according to an embodiment of the present invention. Detailed Implementation

[0021] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should also be noted that, to make the embodiments more comprehensive, the following embodiments are the best and preferred embodiments, and those skilled in the art can use other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0022] like Figs. 1-2 As shown, the method for analyzing and compensating for repeatability errors in multiple fillings of trace gases includes the following steps: S1, deploy pressure sensors and adsorption film thickness detection devices (ultrasonic thickness probes) in the filling pipeline to simultaneously collect gas pressure pulsation data and adsorption layer thickness change data on the inner wall of the container during the filling stage. S2, input the pressure pulsation data obtained in S1 into the dynamic coupling analysis module, and calculate the correlation parameter between the adsorption hysteresis effect and the pressure fluctuation spectrum by combining the adsorption layer thickness change data. S3, based on the correlation parameters output by S2, establishes an adsorption-pressure dynamic coupling correction model through a phase space reconstruction algorithm, generating composite correction parameters including valve response advance and temperature compensation slope; S4 inputs the composite correction parameters of S3 into the filling actuator to complete the reconstruction of the solenoid valve control timing and the dynamic correction of the temperature control curve before the start of the next filling cycle.

[0023] S1 specifically includes: S11, three groups of pressure sensor arrays are arranged axially along the filling pipeline, respectively at 5 cm downstream of the inlet control valve, in the middle of the pipeline, and 3 cm close to the container interface; S12, the adsorption film thickness detection device uses an ultrasonic thickness measurement probe, which is embedded in the inner wall of the container at an angle of 45°, and multiple detection points are distributed equidistantly along the circumference.

[0024] The pressure sensor records the gas pressure fluctuation data at a sampling rate of 1 kHz The ultrasonic probe scans the adsorption layer thickness at a frequency of 100 Hz to form adsorption layer thickness change data , wherein: represents the instantaneous value of the gas pressure collected by the pressure sensor at time t and the axial position x of the pipeline, represents the change amount of the adsorption film layer thickness corresponding to the angular position x of the inner wall of the container at time t relative to the reference initial thickness .

[0025] S2 specifically includes: S21, time-frequency joint analysis is performed on the pressure fluctuation data , and the energy distribution matrix of the 0.1-100 Hz frequency band is extracted by using wavelet packet decomposition method, obtaining: ; wherein, represents the local pressure energy density at frequency f and position x, represents the wavelet packet decomposition operator based on db6 wavelet basis, is the original pressure signal at time t and position x.

[0026] S22, time differentiation is performed on the adsorption layer thickness change data to obtain adsorption rate time series data: ; wherein, represents the instantaneous rate of adsorption layer thickness change, is the adsorption layer thickness change value.

[0027] S23, construct a pressure-adsorption joint phase space, and use a window sliding mutual information algorithm to calculate the time-varying correlation coefficient of the energy distribution matrix and the adsorption rate in the sliding time window, satisfying: ; wherein, is the adsorption-pressure coupling correlation coefficient at time t, is the standard deviation of the local energy matrix , is the standard deviation of the adsorption layer thickness rate , represents the sum over a sliding window in the specified frequency and position range, with a window width set to 1 / 5 of the adsorption process characteristic time constant, ranging from -200 ms.

[0028] S24, perform spectral analysis on the correlation coefficient , calculate its power spectral density and extract the correlation degree parameter K in the feature frequency band [0.5, 5] Hz, satisfying: ; where, represents the Fourier transform of the correlation coefficient, K is the energy proportion index of the adsorption-pressure coupling characteristic frequency band, as the input correlation degree parameter of the subsequent dynamic correction model, and df is the frequency integral variable.

[0029] The above time-varying correlation coefficient expression is used to calculate the normalized correlation coefficient between the frequency energy distribution of the pressure signal and the adsorption layer thickness rate at a certain time t. Specifically, it performs the following operations: 1. At a given time t, perform a weighted integral calculation on the signals at all spatial positions x and frequency components f, where, represents the pressure fluctuation energy at that position and frequency, represents the adsorption layer thickness rate at position x.

[0030] 2. Divide the above weighted result by the two standard deviations and , represents the overall fluctuation of , represents the overall fluctuation of the adsorption layer thickness rate.

[0031] 3. Obtain a normalized coefficient , whose value ranges between positive and negative, and the larger the value, the more synchronized and coupled the pressure spectral features and adsorption change trends at that time.

[0032] The spectral features of pressure fluctuations directly affect the adsorption / desorption process (such as boundary layer disturbance, molecular collision frequency change), and the two have a correlation, but this correlation is not a simple linear response, so a time-varying correlation coefficient is used.

[0033] The molecular part reflects the "same direction coupling" degree through weighted inner product; The denominator eliminates the problem of inconsistent dimensions and data amplitude inequality by standard deviation normalization; It can be regarded as a standardized cross-correlation index in essence, which is convenient for cross-time sliding window comparison.

[0034] The coefficient is the basis for subsequent spectral integral analysis and dynamic modeling, which not only retains the time domain trend, but also introduces the frequency domain energy factor, and embodies the advantages of time-frequency coupling analysis.

[0035] S3 specifically includes: S31, based on Takens embedding theorem, the adsorption-pressure joint phase space is constructed, and the correlation degree parameter, pressure fluctuation data and adsorption layer thickness change rate are combined into a joint vector sequence according to the delay time Embedding constitutes a joint vector sequence : ; Among them, is the delay time, which is determined by the first minimum value of mutual information function (the value is =8-15ms), n is the time series index, is the adsorption-pressure coupling correlation degree parameter at time t, is the pressure fluctuation data at time t, is the adsorption layer thickness change rate.

[0036] S32, according to Takens theorem, if the real dimension of a dynamic system is d, its state can be reconstructed by a one-dimensional time series to construct at least m>2d Embedding dimension to reconstruct the phase space of the system, so that the dynamic trajectory does not overlap (topological equivalence is maintained), therefore the embedding dimension m must be large enough to expand the full degree of freedom of the system state, but not too large to introduce noise dimension, determine the optimal embedding dimension m=6, the joint vector is reconstructed into a phase space trajectory matrix according to the time sliding window: ; Among them, M is the reconstructed phase space trajectory matrix, and N is the total sampling number of time series.

[0037] S33, d. Train radial basis function neural network (RBFNN), establish the nonlinear mapping relationship from phase space trajectory to compensation parameter: , wherein, is the electromagnetic valve response advance, is the temperature compensation curve slope, and RBFNN(·) represents the radial basis function neural network function mapper, the input is the principal component vector, and the output is a two-dimensional compensation parameter. The specific scheme of the radial basis function neural network function mapper is as follows: 1. Input layer design: Input data preprocessing: The phase space trajectory matrix M is processed as follows: Singular value decomposition (SVD) dimension reduction: SVD decomposition is performed on M: , and the first three principal components (covering more than 95% of the energy) are retained to generate the feature vector .

[0038] Normalization: Max-min normalization is performed on the feature vector to unify the value range of each dimension to [0, 1].

[0039] 2. Hidden layer architecture: Radial basis function configuration: Base function form: Gaussian kernel function is used: , where is the jth cluster center, is the corresponding width parameter; Center point selection: a. K-means clustering (cluster number dynamically optimized) is performed on the training set feature vector, and the cluster center is used as ; b. The base function width , is the number of samples in the cluster.

[0040] Node dynamic optimization: The number of hidden layer nodes J is dynamically adjusted based on the AIC criterion: , where p is the number of samples, and the optimal J (typical value 12-18) is determined by minimizing the AIC value; 3. Output layer mapping: Two-parameter output mechanism: Weight connection: Full connection is used from the hidden layer to the output layer, and the output node is 2, corresponding to: ; ; , where is the weight coefficient, is the bias term, and the output layer uses a linear activation function.

[0041] 4. Training and optimization: Supervised learning strategy is used: Training data: The input is the phase space trajectory matrix M of historical filling tasks, and the output is the measured optimal compensation parameter , and the loss function is: , where is the weight coefficient of the temperature compensation term; is the loss function value, which considers both time and temperature targets, and the parameters are updated based on the optimizer , and the regularization coefficient prevents overfitting, and the training is terminated when the validation set loss decreases continuously for 5 consecutive iterations.

[0042] S4 specifically includes: S41, Solenoid valve timing reconfiguration: Based on the valve response advance amount generated in the previous stage. The solenoid valve control pulse sequence is reconstructed, including adjusting the original opening time forward; a high-frequency PWM modulation signal is superimposed during the filling acceleration stage, and the final solenoid valve control instruction set is a discrete pulse sequence. S42, Dynamic Correction of Temperature Control Curve: Constructing Temperature Compensation Slope The segmented temperature control function driven sends a ramp command to the temperature control unit in advance during the pre-correction stage before filling, and adopts feedforward-feedback composite control. S43, Control Command Fusion: Completed before the start of the next filling cycle via the time-triggered scheduler of the real-time operating system. A waveform generator that programs the valve control instruction set into the FPGA; The temperature control function parameters are written into the temperature control module of the PLC.

[0043] Feedforward-feedback composite control includes: Feedforward term: based on The predicted temperature compensation amount; Feedback item: Adjust PID parameters based on real-time temperature difference ΔT.

[0044] S4 specifically includes: S41, Solenoid valve timing reconfiguration: Based on the valve response advance amount generated in the previous stage. The original solenoid valve control timing is reconstructed, specifically including: 1. Set the original solenoid valve opening time. Adjust forward to: ; 2. During the filling acceleration phase, i.e. from Within a continuous 50ms period, a high-frequency PWM (Pulse Width Modulation) signal is superimposed, and its duty cycle increases according to the following pattern: ,in, The duty cycle of the PWM signal. The characteristic response time for material adsorption; 3. The final generated solenoid valve control instruction set is a discrete pulse sequence, in the form of: .

[0045] S42, Dynamic correction of temperature control curve: Construct temperature-compensated slope Driven dynamic temperature control function : ,when ; wherein, Ttarget is the target temperature of the temperature control unit, Tinitial is the initial set temperature, P is the real-time pressure signal, Pthreshold is the pressure threshold for activating the integral compensation (value 0.3 MPa).

[0046] In the pre-correction phase before filling, a temperature ramp instruction is sent to the temperature control unit 10-15 ms in advance, and the control slope is set as: wherein, Tcalibration is the temperature calibration coefficient, which is obtained through a thermodynamic calibration experiment, and the default value is 0.6-1.0.

[0047] A feedforward-feedback composite control mechanism is adopted: The feedforward term uses to predict the temperature offset trend; The feedback term is based on the real-time temperature difference to adaptively adjust the PID parameters.

[0048] S43, 5 ms before the start of the next filling cycle, the following operations are completed through the time-triggered scheduler of the real-time operating system (RTOS): Burn the valve control instruction set to the waveform generator in the FPGA controller; Write the temperature control function parameters to the temperature regulation module in the PLC controller.

[0049] The valve control instruction set is a set of control commands generated in advance, including the opening degree that the electromagnetic valve should be in at each time, the waveform generator points to the PWM / pulse control module inside the FPGA, which is responsible for generating the required electrical signal waveform according to the precise time axis, for driving the electromagnetic valve to act, and writing the preset instructions or parameters into the registers or storage units of the control hardware, so that they can be called when running.

[0050] The present application encompasses any substitutions, modifications, equivalent methods and schemes made on the essence and scope of the present application. In order for the public to have a thorough understanding of the present application, specific details are described in the following preferred embodiments of the present application, and the present application can also be fully understood without these details by those skilled in the art. In addition, in order to avoid unnecessary confusion to the essence of the present application, well-known methods, processes, procedures, elements and circuits, etc. are not described in detail.

[0051] The above is only the preferred embodiment of the present application, and it should be pointed out that for ordinary skilled in the art, without departing from the principle of the present application, a number of improvements and refinements can be made, and these improvements and refinements should be considered as the protection scope of the present application.

Claims

1. A method for analyzing and compensating for repeatability errors in multiple fillings of trace gases, characterized in that, The method comprises the following steps: S1, deploying a pressure sensor and an adsorption film thickness detection device in the filling pipeline, and synchronously collecting gas pressure pulsation data and container inner wall adsorption layer thickness change data in the filling stage; S2, inputting the pressure pulsation data obtained in S1 into a dynamic coupling analysis module, combining the adsorption layer thickness change data, and calculating an association degree parameter of the adsorption hysteresis effect and the pressure fluctuation spectrum; S3, based on the association degree parameter output in S2, establishing an adsorption-pressure dynamic coupling correction model through a phase space reconstruction algorithm, and generating composite correction parameters including a valve response advance amount and a temperature compensation slope; S4, inputting the composite correction parameters in S3 into a filling execution mechanism, and completing the reconstruction of electromagnetic valve control timing and the dynamic correction of temperature control curve before the start of the next filling cycle.

2. The method of claim 1, wherein, The S1 specifically comprises: S11, arranging three groups of pressure sensor arrays in the axial direction of the filling pipeline, respectively at the downstream of the air inlet control valve, the middle section of the pipeline, and near the container interface; S12, the adsorption film thickness detection device adopts an ultrasonic thickness measurement probe, which is embedded in the container inner wall at an oblique incidence angle of 45°, and multiple detection points are distributed equidistantly along the circumference.

3. The method of claim 2, wherein, The pressure sensor records the gas pressure pulsation data at a sampling rate of 1 kHz The ultrasonic probe scans the adsorption layer thickness at a frequency of 100 Hz to form adsorption layer thickness change data Wherein: represents the gas pressure instantaneous value collected by the pressure sensor at time t and at the pipe axial position x, represents the variation of the adsorbent membrane layer thickness corresponding to the container inner wall angular position x at time t with respect to the reference initial thickness .

4. The method of claim 3, wherein, The S2 specifically comprises: S21, to pressure pulsation data Time-frequency joint analysis is performed, and an energy distribution matrix of a preset frequency band is extracted through wavelet packet decomposition ; S22, the adsorption layer thickness change data The differential processing is performed to obtain the adsorption rate time series data ; S23, constructing pressure-adsorption combined phase space, using window sliding mutual information algorithm to calculate energy distribution matrix time-varying correlation coefficient of adsorption rate time-varying correlation coefficient of adsorption rate time-varying correlation coefficient of adsorption rate S24, performing frequency spectrum analysis on the time-varying correlation coefficient, and extracting the association degree parameter K in the preset characteristic frequency band.

5. The method of claim 4, wherein, The time-varying correlation coefficient is represented as: ； wherein, is the adsorption-pressure coupling correlation coefficient at time t, is the standard deviation of the local energy matrix , is the standard deviation of the adsorption rate , denotes the sum over a sliding window in the specified frequency and position range.

6. The method of claim 5, wherein, The S24 is a time-varying correlation coefficient The spectrum is analyzed, the power spectral density is calculated, and the correlation degree parameter K in the characteristic frequency band [0.5, 5] Hz is extracted, which satisfies: ； wherein, K is the adsorption-pressure coupling characteristic frequency band energy proportion index, as the input correlation degree parameter of the subsequent dynamic correction model, df is the frequency integral variable.

7. The method of claim 1, wherein, The S3 specifically comprises: S31, based on Takens theorem, constructing adsorption-pressure combined phase space, and constructing multi-dimensional combined vector sequence with correlation degree parameter, pressure fluctuation data and adsorption layer thickness change rate : S32, determining the optimal embedding dimension m, and reconstructing the joint vector sequence into a phase space trajectory matrix M according to the time sliding window; S33, training a radial basis function neural network, and establishing a mapping relationship from the phase space trajectory to the compensation parameter: wherein, is an electromagnetic valve response advance, is a temperature compensation curve slope, M is a phase space trajectory matrix, RBFNN(·) is a radial basis function neural network function mapper, the input is a principal component vector, and the output is a two-dimensional compensation parameter.

8. The method of claim 7, wherein, The radial basis function neural network function mapper comprises an input layer, a hidden layer, and an output layer, wherein: In the input layer structure design: the phase space trajectory matrix obtained by collecting in the historical filling task is taken as the input data, the principal components covering the main characteristic energy are extracted through singular value decomposition method, and normalization processing is performed on them to form standardized feature vectors for network modeling; In the hidden layer structure design, the radial basis function with Gaussian kernel is selected, the cluster analysis on the feature vectors of the training data is performed to determine the basis function center, and the expansion scale parameter is adaptively set according to the statistical distribution of the samples in the cluster, the number of hidden layer nodes is not fixed, and the node size is optimized through the Akaike information criterion; In the output layer part, a double-output structure is constructed, which corresponds to the valve response advance amount and the temperature compensation slope target parameter respectively, the output nodes receive weighted input from all hidden layer nodes, and a linear mapping mode is adopted to realize the generation of continuously adjustable compensation amount.

9. The method of claim 7, wherein the method further comprises: determining a plurality of fill levels of the plurality of trace gas cylinders; and determining a plurality of fill levels of the plurality of reference cylinders. The S4 specifically comprises: S41, solenoid timing reconstruction: according to the valve response advance generated in the last stage , reconstruct the solenoid control pulse sequence, including adjusting the original opening time forward; superimpose a high-frequency PWM modulation signal during the filling acceleration stage, and the finally generated solenoid control instruction set is a discrete pulse sequence; S42, dynamic correction of temperature control curve: construct temperature compensation slope The segmented temperature control function is driven, the slope instruction is sent to the temperature control unit in advance in the pre-correction stage before filling, and feed-forward-feedback compound control is adopted. S43, control instruction fusion: through the time-triggered scheduler of the real-time operating system, the following is completed before the start of the next filling cycle: Valve control instruction set is burned into the waveform generator of FPGA; The temperature control function parameters are written into the temperature control module of PLC.

10. The method of claim 9, wherein, The feedforward-feedback composite control comprises: Feedforward term: based on a predicted temperature compensation amount; Feedback term: adjusting the PID parameter according to the real-time temperature difference ΔT.

Images