Artificial intelligence-based total temperature sensor sparse calibration method and system
By employing an AI-based sparse calibration method and utilizing active learning and machine learning models to dynamically identify key calibration points, the problem of excessive number of tests and insufficient accuracy in total temperature sensor calibration is solved, achieving efficient and low-cost high-precision calibration.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-24
- Publication Date
- 2026-03-31
AI Technical Summary
Existing total temperature sensor calibration techniques suffer from problems such as excessive number of tests, waste of resources, and insufficient calibration accuracy. In particular, in multi-parameter coupled scenarios, existing methods cannot achieve high-precision calibration efficiently and at low cost.
An artificial intelligence-based sparse calibration method is adopted. By actively learning loops and machine learning models, key calibration points are dynamically identified, iterative sampling and model updates are performed, and high-precision calibration with the fewest number of trials is achieved.
It achieves the same or even higher calibration accuracy with far fewer tests than traditional methods, significantly reducing calibration costs and time, and improving calibration efficiency and accuracy.
Smart Images
Figure CN121762068A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of aerospace instrumentation and aircraft atmospheric data systems, and in particular to a sparse calibration method and system for total temperature sensors based on artificial intelligence. Background Technology
[0002] Atmospheric Measurement Systems (ADS) are widely used in the aerospace field and are an important component of aircraft control systems. As a component of an ADS, the total temperature sensor measures the stagnation temperature of incoming flow to provide a basis for calculating parameters such as static temperature and true airspeed. However, the total temperature sensor measurement results are affected by multiple physical factors such as temperature, flow velocity, and pressure. To achieve high-precision measurement, multi-parameter coupling calibration is necessary. Current calibration techniques mainly employ the following two inefficient methods:
[0003] 1. Global Grid Calibration Method: Taking a two-parameter (temperature and flow rate) calibration as an example, this method performs calibration experiments on an m×n full grid consisting of m temperature points and n flow rate points. Although this method provides complete data, the number of experiments increases exponentially with the number of parameters. For example, if 10 points are taken for each parameter, 100 experiments are required for a two-parameter calibration, and 1000 experiments are required for a three-parameter calibration. This results in a long calibration cycle and consumes expensive standard testing equipment (such as high-precision wind tunnels and temperature control chambers), leading to extremely high costs.
[0004] 2. Simplified orthogonal test method: Although the number of tests can be reduced by arranging orthogonal arrays, this method is difficult to accurately capture the complex nonlinear characteristics of the sensor response surface. In scenarios where the model is unknown and the nonlinearity is strong, the calibration accuracy cannot be guaranteed, which has inherent defects.
[0005] Therefore, the drawback of existing calibration techniques is that a large number of redundant or suboptimal test points waste resources, while the characterization of key nonlinear regions may be insufficient. A new calibration technique is needed that can meet the calibration accuracy with the fewest calibration test points, so as to achieve a coordinated balance between cost, efficiency and accuracy. Summary of the Invention
[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a sparse calibration method and system for total temperature sensors based on artificial intelligence, which can solve the problems in the prior art.
[0007] The technical solution of this invention: a sparse calibration method for a total temperature sensor based on artificial intelligence, wherein the method includes:
[0008] S1. Determine the working space of the sensor according to the calibration requirements of the total temperature sensor, with the number of parameters as the spatial dimension and the parameter range as the spatial range;
[0009] S2, using the space-filling design method to determine the initial calibration point set within a defined spatial range;
[0010] S3, Perform a physical calibration test on the initial point set to collect the original output signal of the total temperature sensor and the corresponding standard reference value;
[0011] S4. Using the initial point set as the training set, the initial training machine learning proxy model is used as the initial proxy model. This initial proxy model is used to establish the mapping relationship from the original output signal of the total temperature sensor to the measured temperature.
[0012] S5, Active Learning-Driven Iterative Sampling and Initial Agent Model Evolution, includes:
[0013] S51, perform uncertainty quantification and candidate point generation;
[0014] S52, determine the optimal calibration candidate points;
[0015] S53, conduct physical experiments and expand the dataset;
[0016] S54, perform the initial agent model update and evolution to obtain the evolved agent model;
[0017] S6, determine whether the evolved model meets the accuracy requirement. If it does, execute model deployment; otherwise, proceed to S7.
[0018] S7: Determine if the number of calibrations exceeds the preset limit. If yes, return to S4; otherwise, return to S5.
[0019] Preferably, performing uncertainty quantification and candidate point generation includes:
[0020] Generate a set of virtual candidate points within the entire untested workspace;
[0021] Using the initial proxy model, the prediction uncertainty of each candidate point in the virtual candidate point set is quantified.
[0022] Preferably, using the initial surrogate model, an ensemble learning method or a Gaussian process method is employed to quantify the prediction uncertainty of each point in the virtual candidate point set, wherein,
[0023] Ensemble learning methods include: training an ensemble model consisting of K isomorphic neural networks, and calculating the variance of the prediction results of the K ensemble models at candidate points as a measure of uncertainty. The larger the variance, the greater the divergence of the ensemble models at that point and the higher the uncertainty.
[0024] The Gaussian process method includes: if the initial surrogate model is a Gaussian process regression (GPR), then the prediction variance built into the initial surrogate model is directly used as a measure of uncertainty.
[0025] Preferably, determining the optimal point includes:
[0026] Based on the calculated uncertainty metric, the M candidate points with the highest uncertainty are selected from the set of virtual candidate points as the optimal calibration candidate points for the current iteration.
[0027] Preferably, conducting physical experiments and expanding the dataset includes:
[0028] Physical calibration experiments are conducted at the identified M optimal calibration candidate points to obtain new data, which is then used to expand the dataset.
[0029] Preferably, the initial agent model is updated and evolved to obtain the evolved agent model, which includes:
[0030] The newly acquired data is added to the training set to obtain an expanded dataset. The initial agent model is then retrained using the expanded dataset to obtain an evolved agent model.
[0031] Preferably, the space-filling design method is Latin hypercube sampling (LHS) or Sobol sequence.
[0032] Preferably, the machine learning proxy model is a multilayer perceptron (MLP), a Gaussian process regression (GPR), or a random forest.
[0033] This invention also provides a sparse calibration system for a total temperature sensor based on artificial intelligence, wherein the system includes experimental equipment, data acquisition equipment, and information processing equipment, wherein:
[0034] The testing equipment is used for physical calibration tests;
[0035] The data acquisition device is used to acquire the raw output signal of the total temperature sensor;
[0036] The information processing equipment includes an information transmission module, an experimental design initialization module, an active learning engine, a surrogate model trainer, and a convergence judgment and decision-making unit, wherein:
[0037] The information transmission module is used to send test conditions and instructions to the test equipment and to read and record test data from the data acquisition equipment;
[0038] The experimental design initialization module is used to execute S1-S3:
[0039] S1. Determine the working space of the sensor according to the sensor calibration requirements, with the number of parameters as the spatial dimension and the parameter range as the spatial range;
[0040] S2, using the space-filling design method to determine the initial calibration point set within a defined spatial range;
[0041] S3, Perform a physical calibration test on the initial point set to collect the original output signal of the total temperature sensor and the corresponding standard reference value;
[0042] The active learning engine, surrogate model trainer, and convergent decision and judgment unit are used to execute S4-S7:
[0043] S4. Using the initial point set as the training set, the initial training machine learning proxy model is used as the initial proxy model. This initial proxy model is used to establish the mapping relationship from the original output signal of the total temperature sensor to the measured temperature.
[0044] S5, Active Learning-Driven Iterative Sampling and Initial Agent Model Evolution, includes:
[0045] S51, perform uncertainty quantification and candidate point generation;
[0046] S52, determine the optimal calibration candidate points;
[0047] S53, Conducting physical experiments and expanding the dataset.
[0048] S54, perform the initial agent model update and evolution to obtain the evolved agent model;
[0049] S6, determine whether the evolved model meets the accuracy requirement. If it does, execute model deployment; otherwise, proceed to S7.
[0050] S7: Determine if the number of calibrations exceeds the preset limit. If yes, return to S4; otherwise, return to S5.
[0051] The above technical solution utilizes artificial intelligence methods and introduces an active learning loop to dynamically identify the "information enrichment points" that are most effective in improving model accuracy. This allows for the achievement of equal or even better calibration accuracy with far fewer trials than the traditional grid method, ultimately resulting in a leap in calibration efficiency and a significant reduction in cost. Attached Figure Description
[0052] The accompanying drawings, which form part of this specification, are provided to further illustrate embodiments of the invention and, together with the textual description, explain the principles of the invention. It is obvious that the drawings described below are merely some embodiments of the invention, and those skilled in the art can obtain other drawings based on these drawings without any creative effort.
[0053] Figure 1 This is a flowchart of a sparse calibration method for a total temperature sensor based on artificial intelligence, provided as an embodiment of the present invention. Detailed Implementation
[0054] Specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings. In the following description, specific details are set forth for purposes of explanation and not limitation, in order to aid in a thorough understanding of the invention. However, it will be apparent to those skilled in the art that the invention may be practiced in other embodiments departing from these specific details.
[0055] It should be noted that, in order to avoid obscuring the invention with unnecessary details, only the device structure and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.
[0056] Figure 1 This is a flowchart of a sparse calibration method for a total temperature sensor based on artificial intelligence, provided as an embodiment of the present invention.
[0057] exist Figure 1 The document demonstrates a closed-loop process from conditional input and initial sampling to active learning iterations and final deployment.
[0058] This method integrates active learning, machine learning, and optimal experimental design theory to achieve high-precision calibration of multi-parameter total temperature sensors with minimal physical experiments. This invention is particularly suitable for addressing the challenge of high-cost high-precision calibration of total temperature sensors affected by multiple physical factors such as temperature, flow rate, and pressure.
[0059] like Figure 1 As shown, this embodiment of the invention provides a sparse calibration method for a total temperature sensor based on artificial intelligence, wherein the method includes:
[0060] S1. Determine the working space of the sensor according to the calibration requirements of the total temperature sensor, with the number of parameters as the spatial dimension and the parameter range as the spatial range;
[0061] S2, using the space-filling design method to determine the initial calibration point set within a defined spatial range;
[0062] The purpose of determining this initial calibration point set is to provide an unbiased skeleton support for the original model that covers the entire workspace with the fewest possible points.
[0063] S3, Perform a physical calibration test on the initial point set to collect the original output signal (original potential signal) of the total temperature sensor and the corresponding standard reference value;
[0064] S4. Using the initial point set as the training set, the initial training machine learning proxy model is used as the initial proxy model. This initial proxy model is used to establish the mapping relationship from the original output signal of the total temperature sensor to the measured temperature.
[0065] S5, Active Learning-Driven Iterative Sampling and Initial Agent Model Evolution, includes:
[0066] S51, perform uncertainty quantification and candidate point generation;
[0067] S52, determine the optimal calibration candidate point (i.e., the optimal point);
[0068] S53, conduct physical experiments and expand the dataset;
[0069] S54, perform the initial agent model update and evolution to obtain the evolved agent model;
[0070] S6, determine whether the evolved model meets the accuracy requirement. If it does, execute model deployment; otherwise, proceed to S7.
[0071] S7: Determine if the number of calibrations exceeds the preset limit. If yes, return to S4; otherwise, return to S5.
[0072] In other words, S5 is executed repeatedly until the accuracy condition is met: the model's accuracy (e.g., mean absolute error) on an independent validation set is greater than or equal to a preset threshold. The model parameters and structure that meet the condition can then be fixed and deployed to the terminal application system. If the accuracy is still not met when the number of calibration trials is greater than or equal to the preset upper limit, the completed calibration trial data is analyzed and evaluated to determine how to adjust the algorithm model.
[0073] The above technical solution utilizes artificial intelligence methods and introduces an active learning loop to dynamically identify the "information enrichment points" that are most effective in improving model accuracy. This allows for the achievement of equal or even better calibration accuracy with far fewer trials than the traditional grid method, ultimately resulting in a leap in calibration efficiency and a significant reduction in cost.
[0074] According to one embodiment of the present invention, performing uncertainty quantification and candidate point generation includes:
[0075] Generate a set of virtual candidate points (e.g., a dense exploration set) within the entire untested workspace;
[0076] Using the initial proxy model, the prediction uncertainty of each candidate point in the virtual candidate point set is quantified.
[0077] According to one embodiment of the present invention, using an initial surrogate model, an ensemble learning method or a Gaussian process method is employed to quantify the prediction uncertainty of each point in the virtual candidate point set, wherein...
[0078] Ensemble learning methods include: training an ensemble model consisting of K isomorphic neural networks, and calculating the variance of the prediction results of the K ensemble models at candidate points as a measure of uncertainty. The larger the variance, the greater the divergence of the ensemble models at that point and the higher the uncertainty.
[0079] The Gaussian process method includes: if the initial surrogate model is a Gaussian process regression (GPR), then the prediction variance built into the initial surrogate model is directly used as a measure of uncertainty.
[0080] According to one embodiment of the present invention, determining the optimal point includes:
[0081] Based on the calculated uncertainty metric, the M candidate points with the highest uncertainty are selected from the set of virtual candidate points as the optimal calibration candidate points for the current iteration.
[0082] According to one embodiment of the present invention, conducting physical experiments and expanding the dataset includes:
[0083] Physical calibration experiments are conducted at the identified M optimal calibration candidate points to obtain new data, which is then used to expand the dataset.
[0084] According to one embodiment of the present invention, performing an initial agent model update evolution to obtain an evolved agent model includes:
[0085] The newly acquired data is added to the training set to obtain an expanded dataset. The initial agent model is then retrained (or fine-tuned) using the expanded dataset to obtain an evolved agent model (i.e., a new model).
[0086] According to one embodiment of the present invention, the space-filling design method is Latin hypercube sampling (LHS) or Sobol sequence.
[0087] According to one embodiment of the present invention, the machine learning agent model is a multilayer perceptron (MLP), a Gaussian process regression (GPR), or a random forest.
[0088] This invention also provides an artificial intelligence-based sparse calibration system for a total temperature sensor, wherein the system includes experimental equipment, data acquisition equipment, and information processing equipment, wherein:
[0089] The testing equipment is used for physical calibration tests;
[0090] For example, testing equipment may include temperature chambers, high-precision wind tunnels, and their control systems.
[0091] The data acquisition device is used to acquire the raw output signal of the total temperature sensor;
[0092] The information processing equipment includes an information transmission module, an experimental design initialization module, an active learning engine, a surrogate model trainer, and a convergence judgment and decision-making unit, wherein:
[0093] The information transmission module is used to send test conditions and instructions to the test equipment and to read and record test data from the data acquisition equipment;
[0094] The experimental design initialization module is used to execute S1-S3 (i.e., for calibration parameter input and initial space filling sampling):
[0095] S1. Determine the working space of the sensor according to the sensor calibration requirements, with the number of parameters as the spatial dimension and the parameter range as the spatial range;
[0096] S2, using the space-filling design method to determine the initial calibration point set within a defined spatial range;
[0097] S3, Perform a physical calibration test on the initial point set to collect the original output signal of the total temperature sensor and the corresponding standard reference value;
[0098] The active learning engine, surrogate model trainer, and convergent decision and judgment unit are used to execute S4-S7:
[0099] S4. Using the initial point set as the training set, the initial training machine learning proxy model is used as the initial proxy model. This initial proxy model is used to establish the mapping relationship from the original output signal of the total temperature sensor to the measured temperature.
[0100] S5, Active Learning-Driven Iterative Sampling and Initial Agent Model Evolution, includes:
[0101] S51, perform uncertainty quantification and candidate point generation;
[0102] S52, determine the optimal calibration candidate points;
[0103] S53, Conducting physical experiments and expanding the dataset.
[0104] S54, perform the initial agent model update and evolution to obtain the evolved agent model;
[0105] S6, determine whether the evolved model meets the accuracy requirement. If it does, execute model deployment; otherwise, proceed to S7.
[0106] S7: Determine if the number of calibrations exceeds the preset limit. If yes, return to S4; otherwise, return to S5.
[0107] The above technical solution utilizes artificial intelligence methods and introduces an active learning loop to dynamically identify the "information enrichment points" that are most effective in improving model accuracy. This allows for the achievement of equal or even better calibration accuracy with far fewer trials than the traditional grid method, ultimately resulting in a leap in calibration efficiency and a significant reduction in cost.
[0108] In summary, the active learning engine integrates uncertainty quantification and optimal sampling point selection algorithms; a surrogate model trainer for building and updating machine learning models; and a convergence judge and decision maker for evaluating the iterative process and determining termination.
[0109] The above system and the aforementioned Figure 1 For specific examples corresponding to the description, please refer to the aforementioned section. Figure 1 The methods described will not be elaborated upon here.
[0110] The following describes, with examples, a sparse calibration method for a total temperature sensor based on artificial intelligence, as described in this invention.
[0111] In this example, the calibration of a dual-parameter total temperature sensor (temperature and flow rate) is used as an example to illustrate the present invention in detail.
[0112] 1. System Setup and Parameter Definition
[0113] Sensor to be calibrated: Total temperature sensor with dual parameters of temperature and flow rate.
[0114] Calibration parameter range: Temperature: -100℃~50℃, Flow rate: 100m / s~270m / s.
[0115] Test equipment: High-precision temperature-controlled wind tunnel.
[0116] Comparison benchmark: The traditional method uses a 10×10 full grid and requires a total of 100 trials.
[0117] 2. Implementation process of the invention
[0118] a. Latin hypercube sampling was used to select 15 initial calibration points within the workspace.
[0119] b. Conduct experiments at these 15 points and collect data. Use this data to train an ensemble model consisting of 5 neural networks as the initial proxy model.
[0120] c. Proactive learning iteration:
[0121] Iteration 1: The ensemble model predicts for 10,000 virtual candidate points and calculates the variance. The two points with the largest variances (located in a high-temperature, low-flow-rate region) are selected for testing, and the model is updated. Total number of trials: 17.
[0122] Iteration 2: The updated model selects two new high-uncertainty points (located in the low-temperature, high-flow-rate region) from the new candidate set for testing and updates the model. Total number of trials: 19.
[0123] Iterations 3 & 4: Repeat the above process, adding 4 more points. Total number of trials: 23.
[0124] d. After 4 rounds of iteration, with a total of 23 experiments, the model's average temperature prediction error across the entire working domain is below 0.3℃, meeting the accuracy requirements. The iteration is then terminated, and the model parameters and structure are solidified and deployed to the terminal application system.
[0125] 3. Effect Comparison
[0126] Efficiency improvement: This invention achieves the required accuracy in only 23 tests, compared to 100 tests for traditional methods, reducing the number of tests by 77%, and significantly reducing the calibration cycle and cost.
[0127] Accuracy Guarantee: On 30 independent validation points, the accuracy of the model trained by this method is comparable to that of the model built on 100 points. Furthermore, in some regions with strong nonlinearity, the model exhibits better robustness due to the more representative data points.
[0128] Therefore, applying artificial intelligence (algorithms such as active learning loops and iterative sampling) to the physical calibration process of total temperature sensors can achieve a sparse, efficient, and high-precision calibration method.
[0129] As can be seen from the above embodiments, the total temperature sensor sparse calibration method and system based on artificial intelligence described in this invention has at least the following advantages compared to the prior art:
[0130] 1. A leap in cost and efficiency: By changing the number of calibration tests from exponential growth to near linear growth, 50% to 90% of testing resources can be saved, greatly reducing calibration costs and shortening product development and production cycles.
[0131] 2. Adaptability and Intelligence: The method and system described in this invention are adaptive, capable of automatically identifying the most complex and difficult-to-model areas of sensor response, and concentrating resources for precise calibration, thus realizing the intelligence of the calibration process.
[0132] 3. Data-driven high-precision guarantee: Since each new data point maximizes the knowledge boundary of the model, a more robust model can be trained in the entire working domain, especially in nonlinear regions, with less data.
[0133] 4. High versatility: The method and system described in this invention do not depend on a specific sensor physical model. It is a general framework that can be widely applied to the calibration of total temperature sensors with various numbers of parameters.
[0134] The features described and / or illustrated above with respect to one embodiment may be used in the same or similar manner in one or more other embodiments, and / or in combination with or in lieu of features in other embodiments.
[0135] It should be emphasized that the term "including / comprises" as used herein refers to the presence of a feature, whole, step, or component, but does not exclude the presence or addition of one or more other features, wholes, steps, components, or combinations thereof.
[0136] The apparatus and methods described above can be implemented in hardware or in combination with software. This invention relates to computer-readable programs that, when executed by a logic component, enable that logic component to implement the apparatus or constituent parts described above, or to implement the various methods or steps described above. This invention also relates to storage media for storing the above programs, such as hard disks, magnetic disks, optical disks, DVDs, flash memory, etc.
[0137] Many features and advantages of these embodiments are apparent from this detailed description, and therefore the appended claims are intended to cover all such features and advantages of these embodiments that fall within their true spirit and scope. Furthermore, since many modifications and alterations will readily occur to those skilled in the art, the embodiments of the invention are not intended to be limited to the precise structures and operations illustrated and described, but rather to encompass all suitable modifications and equivalents falling within their scope.
[0138] The parts of this invention not described in detail are techniques known to those skilled in the art.
Claims
1. A sparse calibration method for a total temperature sensor based on artificial intelligence, characterized in that, The method includes: S1. Determine the working space of the sensor according to the calibration requirements of the total temperature sensor, with the number of parameters as the spatial dimension and the parameter range as the spatial range; S2, using the space-filling design method to determine the initial calibration point set within a defined spatial range; S3, Perform a physical calibration test on the initial point set to collect the original output signal of the total temperature sensor and the corresponding standard reference value; S4. Using the initial point set as the training set, the initial training machine learning proxy model is used as the initial proxy model. This initial proxy model is used to establish the mapping relationship from the original output signal of the total temperature sensor to the measured temperature. S5, Active Learning-Driven Iterative Sampling and Initial Agent Model Evolution, includes: S51, perform uncertainty quantification and candidate point generation; S52, determine the optimal calibration candidate points; S53, Conducting physical experiments and expanding the dataset. S54, perform the initial agent model update and evolution to obtain the evolved agent model; S6, determine whether the evolved model meets the accuracy requirement. If it does, execute model deployment; otherwise, proceed to S7. S7: Determine if the number of calibrations exceeds the preset limit. If yes, return to S4; otherwise, return to S5.
2. The method according to claim 1, characterized in that, Performing uncertainty quantification and candidate point generation includes: Generate a set of virtual candidate points within the entire untested workspace; Using the initial proxy model, the prediction uncertainty of each candidate point in the virtual candidate point set is quantified.
3. The method according to claim 2, characterized in that, Using the initial surrogate model, ensemble learning or Gaussian process methods are employed to quantify the prediction uncertainty for each point in the virtual candidate point set. Ensemble learning methods include: training an ensemble model consisting of K isomorphic neural networks, and calculating the variance of the prediction results of the K ensemble models at candidate points as a measure of uncertainty. The larger the variance, the greater the divergence of the ensemble models at that point and the higher the uncertainty. The Gaussian process method includes: if the initial surrogate model is a Gaussian process regression (GPR), then the prediction variance built into the initial surrogate model is directly used as a measure of uncertainty.
4. The method according to claim 3, characterized in that, Determining the optimal point includes: Based on the calculated uncertainty metric, the M candidate points with the highest uncertainty are selected from the set of virtual candidate points as the optimal calibration candidate points for the current iteration.
5. The method according to claim 4, characterized in that, Conducting physical experiments and expanding the dataset includes: Physical calibration experiments are conducted at the identified M optimal calibration candidate points to obtain new data, which is then used to expand the dataset.
6. The method according to claim 5, characterized in that, The initial agent model is updated and evolved, resulting in the following evolved agent models: The newly acquired data is added to the training set to obtain an expanded dataset. The initial agent model is then retrained using the expanded dataset to obtain an evolved agent model.
7. The method according to any one of claims 1-6, characterized in that, The space-filling design method is Latin hypercube sampling (LHS) or Sobol sequences.
8. The method according to any one of claims 1-7, characterized in that, The machine learning agent model is a multilayer perceptron (MLP), a Gaussian process regression (GPR), or a random forest.
9. A sparse calibration system for a total temperature sensor based on artificial intelligence, characterized in that, The system includes testing equipment, data acquisition equipment, and information processing equipment, among which: The testing equipment is used for physical calibration tests; The data acquisition device is used to acquire the raw output signal of the total temperature sensor; The information processing equipment includes an information transmission module, an experimental design initialization module, an active learning engine, a surrogate model trainer, and a convergence judgment and decision-making unit, wherein: The information transmission module is used to send test conditions and instructions to the test equipment and to read and record test data from the data acquisition equipment; The experimental design initialization module is used to execute S1-S3: S1. Determine the working space of the sensor according to the sensor calibration requirements, with the number of parameters as the spatial dimension and the parameter range as the spatial range; S2, using the space-filling design method to determine the initial calibration point set within a defined spatial range; S3, Perform a physical calibration test on the initial point set to collect the original output signal of the total temperature sensor and the corresponding standard reference value; The active learning engine, surrogate model trainer, and convergent decision and judgment unit are used to execute S4-S7: S4. Using the initial point set as the training set, the initial training machine learning proxy model is used as the initial proxy model. This initial proxy model is used to establish the mapping relationship from the original output signal of the total temperature sensor to the measured temperature. S5, Active Learning-Driven Iterative Sampling and Initial Agent Model Evolution, includes: S51, perform uncertainty quantification and candidate point generation; S52, determine the optimal calibration candidate points; S53, Conducting physical experiments and expanding the dataset. S54, perform the initial agent model update and evolution to obtain the evolved agent model; S6, determine whether the evolved model meets the accuracy requirement. If it does, execute model deployment; otherwise, proceed to S7. S7: Determine if the number of calibrations exceeds the preset limit. If yes, return to S4; otherwise, return to S5.