A method for reducing calibration error
By configuring an environmental simulation component and a machine learning model in the eddy flux system, in-situ calibration of the equipment was achieved, solving the problem of large errors in traditional calibration methods and improving calibration accuracy and field monitoring quality.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GANSU ACAD OF FORESTRY SCI
- Filing Date
- 2026-01-29
- Publication Date
- 2026-05-26
AI Technical Summary
Existing calibration methods for eddy flux systems have large errors, and the traditional calibration process differs greatly from the actual measurement environment, which affects the accuracy of field monitoring and measurement.
By configuring the environment simulation section to simulate the target environment, using machine learning models for parameter recommendation, and combining closed-loop adaptive feedback control and multi-parameter collaborative calibration, the equipment can be calibrated in situ, reducing calibration errors.
It significantly improved calibration accuracy, shortened calibration time, reduced errors, and improved the quality of field data monitoring.
Smart Images
Figure CN122084831A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of vortex gas analyzer calibration technology, specifically relating to a method for reducing calibration error. Background Technology
[0002] Eddy covariance is currently the recognized standard method for flux measurement. Open-circuit flux observations, in particular, do not lose high-frequency data due to structural issues, thus ensuring measurement accuracy. However, open-circuit measurement structures are susceptible to interference from environmental factors such as rain and snow. Therefore, equipment needs periodic calibration after a period of field operation to ensure measurement accuracy. Currently, the Integrated Open-Circuit Eddy Covariance System (IRGASON) is a high-precision integrated sensor specifically designed and manufactured for determining eddy covariance. It can simultaneously measure the molar density of CO2 / H2O in air, three-dimensional wind speed, ultrasonic virtual temperature (sound field temperature), atmospheric pressure, and air temperature. It is suitable for various underlying surface environments in the field, including forests, grasslands, farmland, deserts, cities, and water bodies, and is widely used in the study of regional carbon and water cycle processes. As an effective means of measuring the exchange of matter and energy between ecosystems and the atmosphere, it provides an important data foundation for analyzing the interactions between the geosphere, biosphere, and atmosphere, supporting large-scale, long-term, and continuous scientific research.
[0003] Currently, the calibration method for IRGASON flux analyzers involves removing the equipment on-site and performing calibration in a laboratory. This typically involves first fixing a simple calibration tube to both ends of the infrared gas sensing path (covering only the optical path), then injecting standard gases (N2, CO2) at a fixed flow rate into the calibration tube to verify the measured values and correct the instrument's calibration coefficients. The general calibration procedure includes: first, replacing the absorbent inside the flux analyzer; then running the analyzer for at least 24 hours to allow the absorbent to fully absorb the residual H2O and CO2 in the chamber; and then performing zero-point calibration, CO2 span calibration, and H2O span calibration sequentially. However, during this process, because the standard gas only passes through the infrared analyzer sensing path of the IRGASON integrated probe, while other parts of the instrument are exposed to the atmosphere, differences in temperature and pressure inside and outside the calibration tube can occur during calibration. This can lead to phenomena such as standard gas leakage, resulting in a longer calibration time and process to reach sensing equilibrium, and even fluctuations in measured values at the equilibrium point. Furthermore, since the calibration gas introduced during calibration flows in a fixed and uniform manner within the calibration tube, and the temperature and pressure data during calibration are only the indoor environmental parameters during calibration, not the environmental parameters of the calibration gas inside the calibration tube, the actual air flow in the measurement environment is irregular turbulent, and the temperature and pressure of the airflow also vary with the magnitude of turbulence. Therefore, there is a significant difference between the traditional calibration process and the actual measurement process, resulting in a large calibration error in the calibration results. This affects the measurement accuracy of the instrument during field monitoring, and the error may reach more than 10%. Summary of the Invention
[0004] In view of the above-mentioned defects or deficiencies in the prior art, this application aims to provide a calibration error reduction method.
[0005] According to an embodiment of the present invention, a method for reducing calibration error is provided, which includes the following steps:
[0006] The calibration error reduction device collects historical calibration data and current environmental parameters to construct a dataset. The calibration error reduction device includes an integrated probe, an environmental simulation part, and an analysis part. The environmental simulation part has an internal space and can simulate the target environment within the internal space. The integrated probe is entirely housed within the environmental simulation part, and both the integrated probe and the environmental simulation part are connected to the analysis part. The machine learning model is trained using the dataset to obtain a parameter recommendation model that correlates parameters with calibration results; Based on the parameter recommendation model, a calibration gas is input into the environmental simulation section, and the air pressure values inside and outside the environmental simulation section are measured. Based on the internal and external air pressure values of the environmental simulation section, automatic zero-point calibration, CO2 span multi-parameter collaborative calibration, H2O cross-validation span calibration, and operational verification are performed in sequence.
[0007] In some embodiments, the environmental simulation section includes: a sealed body, an environmental regulator, and a plurality of environmental sensors, wherein the environmental regulator and the plurality of environmental sensors are all disposed within the sealed body, wherein the environmental regulator is used to adjust the environmental parameters inside the environmental simulation section, and the environmental sensors are used to measure the environmental parameters inside the environmental simulation section.
[0008] In some embodiments, the environmental conditioner is capable of simulating an asymmetric turbulent distribution with wind speeds of 0-7.
[0009] In some embodiments, the environmental simulation section further includes a calibration gas supply unit for providing calibration gas.
[0010] In some embodiments, the environmental simulation section further includes a cleaning mechanism for cleaning the integrated probe.
[0011] In some embodiments, the machine learning model includes a gradient boosting tree model, the loss function of which is as follows: The mean square error (MSE) optimized for the calibration scenario is:
[0012] in, This represents the actual measurement error (label). This represents the error value predicted by the model.
[0013] In some embodiments, the gradient descent optimization process of the gradient boosting tree model includes: The label mean of all samples is used as the initial prediction value. The first round residual is calculated and multiple rounds of iteration are performed. In each iteration, a CART regression tree is trained based on the current residual, and the optimal splitting feature is selected using a greedy algorithm; The predictions from the new trees are weighted according to the learning rate and added to the overall model, using the following formula:
[0014] in For learning rate, Let be the predicted value for the t-th tree; When the validation set MSE does not decrease for 5 consecutive rounds or increases by ≥0.01ppm², the iteration stops, and the model at this point is the optimal model.
[0015] In some embodiments, the automatic zero-point calibration includes: The environmental simulation section simulates external airflow and measures air pressure data inside the environmental simulation section in real time. When the measured value fluctuates beyond the threshold, the parameters of the environmental simulation section regarding the simulated external airflow are dynamically adjusted. The parameters are automatically iteratively optimized based on variance analysis stability, test response time, and cross-validation calibration coefficients until the full wind force level calibration is qualified.
[0016] In some embodiments, the CO2 span multi-parameter collaborative calibration includes: First, based on the zero-point calibration, standard CO2 gas is introduced, and the temperature inside the simulated environment is actively adjusted by ±5℃ to verify the effect of temperature. The least squares method is used to fit and determine the temperature compensation coefficient to correct the measurement error caused by temperature changes. Then, after the atmospheric pressure in the simulated environment reaches equilibrium and the measurement curve stabilizes, standard CO2 gas span calibration is performed multiple times to ensure reliability. Finally, a secondary calibration verification is introduced. A standard gas is generated by mixing CO2 and N2. A linear regression model of ultrasonic virtual temperature and concentration is established using the corrected parameters, and the correlation index is evaluated. If the error exceeds the limit, an adaptive iterative optimization based on gradient descent is triggered, which gradually approaches the optimal solution through parameter perturbation and learning rate adjustment.
[0017] In some embodiments, the H2O cross-validation span calibration includes: Output a fixed dew point gas and monitor the impact of pressure changes on water vapor measurement, as well as correct the pressure compensation coefficient for water vapor measurement; Then, after the air pressure in the simulated environment is balanced and the measurement curve is stable, the H2O span is calibrated at multiple intervals. Then, by increasing the dew point temperature and outputting dew point gas, the consistency of related parameters, including temperature and pressure, is cross-validated using the corrected H2O parameters. If the error exceeds the limit, parameter iterative adjustment is triggered until the multi-parameter matching meets the standard, and the H2O span calibration value is determined. Finally, input the H2O span calibration value into the secondary setting dew point temperature value to complete the H2O cross-validation span calibration.
[0018] Compared with the prior art, the beneficial effects of this invention are: This invention simulates the target environment through an environmental simulation component, eliminating the need for on-site equipment removal and laboratory calibration. This enables in-situ calibration of the equipment, effectively monitoring and improving the simulation accuracy of the eddy flux system during calibration. Through closed-loop adaptive feedback control and machine learning algorithms, dynamic optimization and personalized adjustment of calibration parameters are achieved. Furthermore, multi-parameter collaborative calibration and cross-validation strategies avoid the limitations of single-parameter correction, significantly improving the calibration accuracy of the flux system. Using the method described in this invention, calibration time is reduced by more than 5 hours, and calibration error is reduced by more than 15%, thereby improving the quality of field data monitoring. Attached Figure Description
[0019] To more clearly illustrate the solutions in the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0020] Figure 1 This is a flowchart of a calibration error reduction method according to an embodiment of the present invention; Figure 2 This is a partial structural diagram of the environment simulation section according to an embodiment of the present invention. Detailed Implementation
[0021] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.
[0022] like Figure 1-2 As shown, in one embodiment, a calibration error reduction method is provided, which includes the following steps: Step S1: Collect historical calibration data and current environmental parameters using the calibration error reduction device to construct a dataset. The calibration error reduction device includes an integrated probe, an environmental simulation section, and an analysis section. The environmental simulation section has an internal space and can simulate the target environment within the internal space. The integrated probe is entirely housed within the environmental simulation section, and both the integrated probe and the environmental simulation section are connected to the analysis section. Before collecting historical calibration data and current environmental parameters using the calibration error reduction device, clean the exterior of the IRGASON sensing "window" (integrated probe) of the vortex gas analyzer at the monitoring site to ensure a signal strength of 0.9 or higher.
[0023] In some embodiments, the environmental simulation section includes: a sealed body, an environmental regulator, and a plurality of environmental sensors, wherein the environmental regulator and the plurality of environmental sensors are disposed within the sealed body, wherein the environmental regulator is used to adjust the environmental parameters inside the environmental simulation section, and the environmental sensors are used to measure the environmental parameters inside the environmental simulation section.
[0024] In some embodiments, the sealed enclosure is box-shaped, having an inlet and an outlet, and an aviation connector; in some embodiments, the inlet and outlet are circular with a diameter of 5 mm. In some embodiments, the inner wall of the sealed enclosure is coated with a nano-coating material to reduce gas adsorption rate. Inlet A of the sealed enclosure can be fitted with a high-precision flow meter and a gas mixer to mix CO2 and N2 to achieve multiple CO2 concentrations; inlet B of the sealed enclosure can be directly connected to a dew point meter for water vapor span calibration. The size of the sealed enclosure should not be too large, ideally allowing it to completely accommodate the IRGASON vortex integrated probe. This facilitates rapid gas balance within and outside the integrated probe device during subsequent calibration, and also ensures that the entire instrument is in a unified measurement environment. This environmental simulation section is powered by a mobile power module and has a built-in intelligent fault diagnosis chip, which can monitor abnormalities during the calibration process in real time (such as pipeline leakage or sensor drift). The air inlet A of the sealed enclosure is connected to the N2 and CO2 cylinders via a gas mixer and a high-precision flow meter. Temperature and pressure sensors within the sealed enclosure are connected to the external analysis unit (e.g., the EC100 analyzer) via aviation connectors. The EC100 analyzer is also connected to a computer or other electronic equipment. After all connections are complete, the N2 cylinder is opened (the CO2 cylinder is closed). The calibration error reduction device first collects historical calibration data (including flow rate, error, etc.) and current environmental parameters (such as temperature, humidity, and air pressure).
[0025] In some embodiments, the plurality of environmental sensors include a temperature sensor, a pressure sensor, and a high-precision flow rate sensor installed within the enclosure to monitor the temperature, pressure, and airflow conditions within the enclosure accordingly.
[0026] In some embodiments, the environmental conditioner includes an air agitation module, which may be a controllable variable frequency fan that remotely controls its wind speed changes. The fan's layout and parameters are optimized based on CFD simulation, enabling it to simulate asymmetric turbulent distributions with wind speeds ranging from 0 to 7.
[0027] In some embodiments, the environmental simulation section further includes a cleaning mechanism for cleaning the integrated probe. In some embodiments, the cleaning mechanism may be a retractable brush self-cleaning module for automatically cleaning the surface of the integrated probe, the retractable brush self-cleaning module being integrated into a sealed body.
[0028] In some embodiments, the environmental simulation section further includes a calibration gas supply unit for providing calibration gas, such as the N2 cylinder and CO2 cylinder described above.
[0029] Step S2: Train the machine learning model using the dataset to obtain a parameter recommendation model that correlates parameters with calibration results. The parameter recommendation model can output recommended initial N2 flow rate and fan speed based on the input data, and can also dynamically provide feedback optimization to achieve intelligent parameter recommendation, adjust the flow meter and gas mixer to ensure that the introduced standard high-purity N2 with a purity of 99.999%.
[0030] Specifically, (1) Input feature data Based on the core features of "historical calibration data + environmental parameters", it specifically includes: Historical calibration data: IRGASON calibration records for the past 3 years under different scenarios, including zero-point calibration coefficient, CO2 / H2O span coefficient, standard gas concentration (e.g., 300ppm / 400ppm CO2), calibration equilibrium time, measurement error value (deviation from standard value), etc. Environmental parameters: atmospheric temperature (-10-40℃), air pressure (80-110kPa), wind speed (0-7), humidity (10%-90% RH) at the calibration time, as well as real-time monitoring data such as cavity temperature and airflow disturbance intensity inside the calibration equipment; Equipment status parameters: sensor signal strength (≥0.9), cleanliness level (self-cleaning module feedback value), equipment running time, etc.
[0031] (2) Labeling effect Labeled as "Measurement error value within 24 hours after calibration," specifically, it represents the deviation (unit: ppm) between the IRGASON measured value and the standard gas concentration in a natural environment. For example, the absolute error of the CO2 measurement value compared to the 400ppm standard value: |Measured -400| (3) Machine learning model selection XGBoost is adopted as the gradient boosting tree implementation framework, which supports custom loss functions and has excellent efficiency in processing high-dimensional features, making it suitable for the needs of multi-parameter collaborative optimization in calibration scenarios.
[0032] Tree structure parameters: The tree depth is set to 5 levels, the number of leaf nodes is 32, and the minimum number of sample splits per tree is 20. Ensemble parameters: The number of iterations is initially set to 100, the learning rate (step size) is set to 0.1, and "leaf node regularization" (L2 regularization coefficient = 1) is used to suppress overfitting.
[0033] In some embodiments, the machine learning model includes a gradient boosting tree model, the loss function of which is as follows: The mean square error (MSE) optimized for the calibration scenario is:
[0034] in, This represents the actual measurement error (label). This represents the error value predicted by the model.
[0035] (4) Training process and parameter optimization The training and validation sets were divided in a 7:3 ratio. "Time-stratified sampling" was used to ensure that the two sets of data included samples from different seasons and weather conditions, thus avoiding model bias caused by uneven data distribution. Continuous features such as wind speed and temperature are divided into boxes (wind speed is divided into 3 boxes: 0-2, 3-5, and 6-7) to improve the model's ability to capture nonlinear relationships.
[0036] The gradient descent optimization process of the gradient boosting tree model includes: The label mean of all samples is used as the initial prediction value. The first round residual is calculated and multiple rounds of iteration are performed. In each iteration, a CART regression tree is trained based on the current residual, and the optimal splitting feature (such as "temperature difference between the sealed body and the standard temperature" and "wind speed level") is selected through a greedy algorithm. The predictions from the new trees are weighted according to the learning rate (0.1) and added to the overall model, as follows:
[0037] in For learning rate, Let be the predicted value for the t-th tree; When the validation set MSE does not decrease for 5 consecutive rounds or increases by ≥0.01ppm², the iteration stops, and the model at this point is the optimal model.
[0038] (5) Supplementary measures for over-compatibility control By introducing column sampling (feature sampling rate = 0.8), each tree is trained using only 80% of its features, reducing the impact of feature collinearity. Leave-one-out cross-validation was used to validate samples from different underlying surface scenarios such as grassland, sea, forest, and city, ensuring the model's generalization ability in specific environments (validation error for each scenario must be ≤ ±0.5ppm).
[0039] (6) Correlation between model output and calibration The trained model can output: 1. Recommendations for optimal calibration parameters based on real-time environmental parameters (such as current temperature 25℃, wind speed level 3), including N2 flow rate (1.0-2.0L / min), fan speed (600-1200rpm), balancing waiting time, etc. 2. The predicted calibration error value will be automatically triggered when the predicted error is greater than 1 ppm, triggering a secondary calibration process (such as adding a CO2 concentration point for verification).
[0040] This model can specifically demonstrate the correlation between parameters and calibration results: for every 1 level increase in wind speed, the model recommends an increase of 200 rpm in fan speed, and at the same time, the N2 flow rate increases by 0.2 L / min to offset the uneven concentration distribution caused by enhanced turbulence, ultimately reducing the calibration error to within ±0.3 ppm (40% improvement over traditional methods).
[0041] Step S3: Based on the parameter recommendation model, input the calibration gas into the environmental simulation part and measure the air pressure values inside and outside the environmental simulation part. Step S4: Based on the internal and external air pressure values in the environmental simulation section, perform automatic zero-point calibration, CO2 span multi-parameter collaborative calibration, H2O cross-validation span calibration, and operational verification in sequence.
[0042] In some embodiments, automatic zero-point calibration includes: The system uses ECMon software to measure the air pressure in the simulated environment in real time and activates an air agitation module (small-power fan) to simulate external airflow. The calibration error reduction device collects measurement data and sensor data from the simulated environment in real time. When the measured air pressure fluctuates beyond a threshold, it automatically adjusts the fan speed or N2 flow rate for dynamic compensation. Based on a closed-loop feedback control mechanism, the calibration error reduction device uses the standard deviation of the measured value fluctuation (e.g., ±0.1ppm) and the rate of change (less than 0.05% / min) as thresholds, and dynamically adjusts parameters through a PID controller. The compensation amount is calculated based on the benchmark values corresponding to the wind speed level (e.g., level 2 wind corresponds to 600rpm speed and 1.0 L / min flow rate), combined with the real-time error (deviation between measured and target values). At low wind speeds, the N2 flow rate is increased to dilute interference, and at high wind speeds, the speed is increased to enhance mixing uniformity. Adaptive PID parameters (increasing the integral coefficient at low wind speeds) and gradient descent optimization (learning rate 0.05) are used to avoid overcompensation. After each adjustment, stability, response time (less than 90 seconds), and cross-validation calibration coefficient (deviation less than ±0.05%) were tested using Allan variance analysis. If the standards were not met, the process was automatically iterated until the calibration of all wind force levels was qualified. Based on wind speed observations from the Yellow River Delta meteorological station over the past two years, air turbulence at eight wind force levels—0 (no wind), 1 (gentle wind), 2 (light wind), 3 (light breeze), 4 (moderate wind), 5 (fresh wind), 6 (strong wind), and 7 (gale)—was simulated. The air agitation module (low-power fan) was activated eight times, according to the wind force level from lowest to highest. Each time, the system automatically monitored and waited for the air pressure, concentration, and flow velocity in the sealed environment to reach a stable equilibrium. After observing the stable measurement curve of the integrated probe using ECMon software, zero-point calibration was performed, and the zero-point coefficient was corrected. The correction was complete when the calibration value showed 0. This calibration process was repeated eight times. After the final eight simulations were completed, the N2 gas cylinder was shut off.
[0043] In some embodiments, CO2 span multi-parameter collaborative calibration includes: After zero-point calibration, close the N2 gas cylinder at inlet A of the sealed enclosure, then open the CO2 gas cylinder at the inlet, controlling the flow rate using a high-precision flow meter. At this point, CO2 gas with a standard concentration of 400 ppm is introduced into the sealed enclosure. To reduce calibration error, the equipment synchronously adjusts the temperature within the sealed enclosure by ±5℃, verifying the effect of temperature on the measurement results and correcting the temperature compensation coefficient. This is typically a functional relationship between temperature and the deviation of the measured value, used to correct measurement errors caused by temperature changes. It takes the form of linear compensation, and the simplified expression is: Compensation coefficient = k × (T - T0) + b T is the current intracavity temperature (°C), T0 is the standard temperature (25°C), k is the temperature sensitivity coefficient (reflecting the effect of unit temperature change on the measured value), and b is a constant term (obtained by fitting calibration data).
[0044] By adjusting the temperature by ±5℃ and collecting multiple sets of data, the calibration error reduction equipment uses the least squares fitting method to determine k and b, so that the deviation between the compensated measured value and the true concentration is minimized.
[0045] Continue observing the internal and external pressure values of the calibration equipment. Once the pressure inside the sealed chamber and the ambient pressure reach equilibrium, observe the integrated probe measurement curve using ECMon software (as part of the calibration error reduction device). After the curve stabilizes, enter the CO2 standard gas concentration of 400 ppm in the CO2 span calibration value field on the ECMon software page, and then click CO2 span calibration to correct the CO2 span coefficient. To ensure the accuracy and reliability of the instrument and span calibration, repeat the CO2 span calibration steps three times, with a 5-minute interval between each calibration. After all three calibrations meet the requirements, complete the calibration and finally shut off the 400 ppm CO2 cylinder. To ensure the accuracy and stability of CO2 span calibration, the CO2 and N2 cylinders were reopened at a 3:1 flow rate ratio, controlled by a high-precision flow meter and gas mixer to achieve a gas concentration output of 300 ppm for a second CO2 span calibration. This process was then repeated from the first CO2 span calibration. The corrected CO2 parameters were used to back-verify the correlation between ultrasonic virtual temperature and concentration. If the error exceeded the limit, iterative calibration was triggered. (In CO2 span calibration, verifying the correlation between ultrasonic virtual temperature and concentration and iterative calibration are crucial steps to ensure measurement accuracy. To reduce calibration error, the equipment first establishes a linear regression model (T0) between ultrasonic virtual temperature and concentration based on the corrected CO2 parameters.) v =a×C+b), Tv is the acoustic field temperature (unit: ℃) measured by the IRGASON ultrasonic virtual temperature analyzer, reflecting the thermodynamic state of the airflow. Where: A is the proportion of the influence of the regression coefficient concentration on the ultrasonic virtual temperature (e.g., the number of ℃ change in Tv for every 10ppm increase in CO2 concentration); C is the actual concentration of CO2 in the gas concentration environment simulation section (unit: ppm, such as 300ppm, 400ppm standard gas); b is the benchmark offset value of the linear regression of the constant term, ensuring the fitting accuracy in the low concentration range.
[0046] The correlation is assessed by fitting coefficients using the least squares method and calculating the coefficient of determination R² and root mean square error (RMSE). If R² < 0.95 or RMSE exceeds a threshold (e.g., ±0.5℃), iterative calibration is triggered. During iteration, the calibration error reduction device applies a ±1% perturbation to parameters such as the temperature compensation coefficient. Calibration is repeated at concentration points of 300ppm and 400ppm, and parameters are adjusted according to the direction of correlation change (new parameter = old parameter + η × 10 ... J), using an adaptive learning rate (initially 0.1 and gradually decreasing to 0.01) to progressively approximate the optimal solution, where η is the learning rate. J represents the gradient of the error with respect to the temperature compensation coefficient k. Leave-one-out cross-validation is used to ensure parameter generalization ability; if divergence occurs, rollback is performed and regularization constraints are added. This process ensures correlation R² > 0.98, improves measurement accuracy to ±2 ppm, and extends the calibration cycle. Finally, in the ECMon software page, input the CO2 standard gas concentration value of 300 ppm for the secondary calibration at the CO2 span calibration value field, until the secondary calibration is complete.
[0047] In some embodiments, H2O cross-validation span calibration includes: First, connect the air inlet B of the sealed enclosure to the dew point meter. Power on the dew point meter and set the dew point temperature (set the dew point value 3-5℃ lower than the ambient temperature). Then, output the gas with a fixed dew point. The system synchronously monitors the impact of pressure changes on water vapor measurement and corrects the pressure compensation coefficient for water vapor measurement. Observe the internal and external air pressure values of the calibration equipment. After the air pressure inside the enclosure and the ambient air pressure reach equilibrium, observe the measurement curve of the integrated probe through the ECMon software until it stabilizes. Then, input the dew point temperature value into the H2O span calibration value on the ECMon software page and click H2O span calibration to complete the H2O span coefficient correction. To reduce equipment error and improve calibration accuracy, repeat the H2O span calibration steps three times, with a 10-minute interval between each calibration. After all three calibrations meet the requirements, complete the H2O span calibration. To ensure the accuracy and stability of the H2O span calibration, perform a second H2O calibration. During the second calibration, the dew point temperature is reset (set 2°C higher than the first calibration), and the gas with the second dew point is output. The process of the first H2O span calibration is then repeated. The corrected H2O parameters are used to cross-validate other related parameters. During the second H2O calibration, the system first uses the corrected water vapor parameters (pressure compensation coefficient) to measure the gas at the new dew point temperature (2°C higher than the first). By comparing the measured dew point with the theoretical value, the error is calculated to ensure it is within the threshold. Simultaneously, the correction results for related parameters such as temperature and pressure are verified to ensure logical consistency among parameters. If the error exceeds the limit, iterative parameter adjustments are triggered until multiple parameters match the target. Finally, the second dew point temperature value is entered into the H2O span calibration value on the ECMon software page, and then the H2O span calibration is clicked until the second calibration is complete.
[0048] This invention replaces the calibration tube with modular, self-cleaning calibration equipment and incorporates an environmental simulation component to mimic the target environment, achieving in-situ calibration without requiring equipment removal and laboratory visits. This effectively monitors and improves the simulation accuracy of the eddy flux system during calibration. Through closed-loop adaptive feedback control and machine learning algorithms, dynamic optimization and personalized adjustment of calibration parameters are achieved. Furthermore, multi-parameter collaborative calibration and cross-validation strategies avoid the limitations of single-parameter correction, significantly improving calibration accuracy. Using this method, calibration time is reduced by more than 5 hours, and calibration error is reduced by more than 15%, thereby improving the quality of field data monitoring.
[0049] The above description is merely a specific implementation of the present invention and does not limit the scope of protection of the present invention. Any modifications and alterations within the technical scope of the present invention will readily be conceived by those skilled in the art, and all of these should be included within the scope of protection of the present invention. Therefore, the scope of protection of the claims should be taken as the scope of protection of the present invention.
Claims
1. A method for reducing calibration error, characterized in that, include: The calibration error reduction device collects historical calibration data and current environmental parameters to construct a dataset. The calibration error reduction device includes an integrated probe, an environmental simulation part, and an analysis part. The environmental simulation part has an internal space and can simulate the target environment within the internal space. The integrated probe is entirely housed within the environmental simulation part, and both the integrated probe and the environmental simulation part are connected to the analysis part. The machine learning model is trained using the dataset to obtain a parameter recommendation model that correlates parameters with calibration results; Based on the parameter recommendation model, a calibration gas is input into the environmental simulation section, and the air pressure values inside and outside the environmental simulation section are measured. Based on the internal and external air pressure values of the environmental simulation section, automatic zero-point calibration, CO2 span multi-parameter collaborative calibration, H2O cross-validation span calibration, and operational verification are performed in sequence.
2. The calibration error reduction method according to claim 1, characterized in that: The environmental simulation section includes: a sealed body, an environmental regulator, and multiple environmental sensors. The environmental regulator and the multiple environmental sensors are all disposed within the sealed body. The environmental regulator is used to adjust the environmental parameters inside the environmental simulation section, while the environmental sensors are used to measure the environmental parameters inside the environmental simulation section.
3. The calibration error reduction method according to claim 1, characterized in that: The environmental regulator is capable of simulating asymmetric turbulent distributions with wind speeds ranging from 0 to 7.
4. The calibration error reduction method according to claim 1, characterized in that, The environmental simulation section also includes a calibration gas supply unit for providing calibration gas.
5. The calibration error reduction method according to claim 1, characterized in that, The environmental simulation section also includes a cleaning mechanism for cleaning the integrated probe.
6. The calibration error reduction method according to claim 1, characterized in that, The machine learning model includes a gradient boosting tree model, and the loss function of the gradient boosting tree model is as follows: The mean square error (MSE) optimized for the calibration scenario is: in, This represents the actual measurement error (label). This represents the error value predicted by the model.
7. The calibration error reduction method according to claim 6, characterized in that, The gradient descent optimization process of the gradient boosting tree model includes: The label mean of all samples is used as the initial prediction value. The first round residual is calculated and multiple rounds of iteration are performed. In each iteration, a CART regression tree is trained based on the current residual, and the optimal splitting feature is selected using a greedy algorithm; The predictions from the new trees are weighted according to the learning rate and added to the overall model, using the following formula: in For learning rate, Let be the predicted value for the t-th tree; When the validation set MSE does not decrease for 5 consecutive rounds or increases by ≥0.01ppm², the iteration stops, and the model at this point is the optimal model.
8. The calibration error reduction method according to claim 1, characterized in that, The automatic zero-point calibration includes: The environmental simulation section simulates external airflow and measures air pressure data inside the environmental simulation section in real time. When the measured value fluctuates beyond the threshold, the parameters of the environmental simulation section regarding the simulated external airflow are dynamically adjusted. The parameters are automatically iteratively optimized based on variance analysis stability, test response time, and cross-validation calibration coefficients until the full wind force level calibration is qualified.
9. The calibration error reduction method according to claim 1, characterized in that, The CO2 span multi-parameter collaborative calibration includes: First, based on the zero-point calibration, standard CO2 gas is introduced, and the temperature inside the simulated environment is actively adjusted by ±5℃ to verify the effect of temperature. The least squares method is used to fit and determine the temperature compensation coefficient to correct the measurement error caused by temperature changes. Then, after the atmospheric pressure in the simulated environment reaches equilibrium and the measurement curve stabilizes, standard CO2 gas span calibration is performed multiple times to ensure reliability. Finally, a secondary calibration verification is introduced. A standard gas is generated by mixing CO2 and N2. A linear regression model of ultrasonic virtual temperature and concentration is established using the corrected parameters, and the correlation index is evaluated. If the error exceeds the limit, an adaptive iterative optimization based on gradient descent is triggered, which gradually approaches the optimal solution through parameter perturbation and learning rate adjustment.
10. The calibration error reduction method according to claim 1, characterized in that, The H2O cross-validation span calibration includes: Output a fixed dew point gas and monitor the impact of pressure changes on water vapor measurement, as well as correct the pressure compensation coefficient for water vapor measurement; Then, after the air pressure in the simulated environment is balanced and the measurement curve is stable, the H2O span is calibrated at multiple intervals. Then, by increasing the dew point temperature and outputting dew point gas, the consistency of related parameters, including temperature and pressure, is cross-validated using the corrected H2O parameters. If the error exceeds the limit, parameter iterative adjustment is triggered until the multi-parameter matching meets the standard, and the H2O span calibration value is determined. Finally, input the H2O span calibration value into the secondary setting dew point temperature value to complete the H2O cross-validation span calibration.