Infrared measurement system nonlinear correction method based on convolutional neural network

Abstract

The invention provides an infrared measurement system nonlinear correction method based on a convolutional neural network, and the method comprises the steps: S1, experiment data preparation: obtaining a series of input signal standard values and measurement system response values through an experiment, and taking the values as an experiment training data set; s2, model training: training the experimental training data set by using a convolutional neural network algorithm to obtain a convolutional neural network model between the response value of the measurement system and the standard value of the input signal, and training one convolutional neural network model for each wavelength value or each waveband range of the infrared measurement system; and S3, model verification: evaluating each trained convolutional neural network model by using an experimental verification data set to verify the accuracy of the model to predict the input signal standard value. According to the method, the nonlinear correction of the infrared measurement system is realized, the limitation of N-order polynomial or piecewise linear function fitting can be broken through, and the method is suitable for nonlinear correction of various infrared measurement systems.

Description

Technical Field

[0001] This invention relates to the field of infrared measurement, and more specifically, to a nonlinear correction method for infrared measurement systems based on convolutional neural networks. Background Technology

[0002] The nonlinear effect of the measurement system means that the relationship between the response signal of the measuring instrument and the incident signal is nonlinear. In infrared measurement systems, photodetectors such as mercury cadmium telluride (MCT), indium gallium arsenide (InGaAs) and indium antimonide (InSb) are widely used. These measuring instruments generally have nonlinear effects. The reported nonlinear measurement and correction methods include the double aperture method and the polynomial fitting method. For example, Theocharous et al. of the National Physical Laboratory (NPL) of the United Kingdom used the double aperture method to study the nonlinear characteristics of photoconductive and photovoltaic MCT detectors in the 10.3 µm and 3.8 µm bands. In the experiment, they used a chopper to modulate the radiation source signal to be measured into an AC signal, while filtering out the unmodulated background radiation signal as a DC component, thus realizing the separation of the radiation signal to be measured from the background noise [1]. Hamadani et al. of the National Institute of Standards and Technology (NIST) evaluated the nonlinear relationship between the measurement signal and the incident flux using the double-aperture method for 627 nm and 890 nm light-emitting diode (LED) systems. They also tried to use the linear least squares method to solve an nth-order polynomial to obtain the best fit between the measurement signal and the incident flux [2]. He et al. of the National Institute of Metrology of China used the polynomial fitting method and piecewise linear interpolation method to correct the nonlinearity of the infrared measurement system [3]. However, according to the working principle of the double-aperture method, this method is generally applicable to the visible and near-infrared bands, or to the nonlinear evaluation of infrared band measurement systems that have chopper modulation to separate the radiation source signal and the background radiation signal. In infrared measurement systems that are not cooled by vacuum liquid nitrogen and do not have choppers, the method has a large deviation due to the presence of background noise. Although the polynomial fitting method and piecewise linear interpolation method can be used in the infrared band, it is difficult to find a suitable polynomial or piecewise linear function to achieve a good fitting effect. There is currently no nonlinear correction method for infrared measurement systems based on convolutional neural networks. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a nonlinear correction method for infrared measurement systems based on convolutional neural networks, so as to solve the problem that the fitting effect of polynomial fitting method and piecewise linear interpolation method in existing infrared measurement systems is not good when used in the infrared band.

[0004] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0005] The present invention provides a nonlinear correction method for an infrared measurement system based on a convolutional neural network, comprising the following steps: S1. Experimental data preparation: Obtaining a series of standard values ​​of input signals and response values ​​of the measurement system through experiments as an experimental training dataset; S2. Model training: Training the experimental training dataset using a convolutional neural network algorithm to obtain a convolutional neural network model between the response values ​​of the measurement system and the standard values ​​of the input signals, and training a convolutional neural network model for each wavelength value or each band range of the infrared measurement system; S3. Model validation: Evaluating each trained convolutional neural network model using an experimental validation dataset to verify the accuracy of the model in predicting the standard values ​​of the input signals.

[0006] Optionally, in the above-mentioned nonlinear correction method for infrared measurement system based on convolutional neural network, in step S1, a standard blackbody is used as the target to be measured, and the temperature of the standard blackbody is set to a series of different values ​​in sequence. After it stabilizes, the temperature of the standard blackbody is measured and calibrated using a thermometer.

[0007] Optionally, in the above-mentioned nonlinear correction method for infrared measurement systems based on convolutional neural networks, in step S2, the temperature of the standard blackbody and the response value of the measurement system in the experimental training dataset are defined as input features, and the standard value of the input signal is defined as the training target value; during the training process of the convolutional neural network model, a convolutional neural network model is trained for each wavelength value or each band range of the infrared measurement system to fit the relationship between the input features and the training target value.

[0008] Optionally, in the above-mentioned nonlinear correction method for infrared measurement systems based on convolutional neural networks, in step S3, during the model verification stage, within the range covered by the standard values ​​of the input signals in the experimental training dataset, another series of different standard values ​​of input signals and measurement system response values ​​are obtained experimentally as experimental verification datasets; the input features are defined using the same method as in the experimental training dataset, and substituted into the aforementioned trained convolutional neural network model to predict the standard values ​​of the input signals; the predicted standard values ​​of the input signals are compared with the standard values ​​of the input signals in the experimental verification dataset obtained through experiments to obtain the relative deviation of the nonlinear correction of the prediction model.

[0009] Optionally, in the above-mentioned nonlinear correction method for an infrared measurement system based on a convolutional neural network, the infrared measurement system is an infrared spectral radiance measurement system, and in step S1, a standard blackbody is used as the target to be measured, and the temperature of the standard blackbody is sequentially set to a series of different values. After it stabilizes, a thermometer is used to measure and calibrate the temperature of the standard blackbody. ; Calculate the spectral radiance of a standard blackbody at this temperature according to Planck's formula (1). The standard value of the input signal is used as the reference value; the response value of the measurement system is obtained by measuring the standard blackbody using an infrared spectral radiance measurement system. The temperature T of a standard blackbody t Wavelength λ i Input signal standard value and measurement system response value The experimental training dataset is formed, where Planck's formula is:

[0010] (1)

[0011] Where c1 and c2 represent the first radiation constant and the second radiation constant, respectively. This refers to a specific wavelength or wave value corresponding to the infrared spectral radiance measurement system.

[0012] Optionally, in the above-described nonlinear correction method for infrared measurement systems based on convolutional neural networks, in step S2, the temperature T of a standard blackbody is included in the experimental training dataset. t Measurement system response value As an input feature, the standard value of the input signal As the training target value, for each wavelength value λ i A convolutional neural network M(λ) is obtained through training. i ).

[0013] Optionally, in the above-described nonlinear correction method for an infrared measurement system based on a convolutional neural network, in step S3, the temperature of the standard blackbody is set to any value within the temperature range used in the experimental training dataset. After it stabilizes, the temperature of the standard blackbody is measured and calibrated using a thermometer. Calculate the spectral radiance of a standard blackbody at this temperature using Planck's formula. The standard value of the input signal is used as the reference value; the response value of the measurement system is obtained by measuring the standard blackbody using an infrared spectral radiance measurement system. ; at each wavelength λ i Under the condition that the temperature T of the standard blackbody is... v and measurement system response value The input features of the experimental validation dataset are substituted into the model trained using the experimental training dataset to obtain the standard values ​​of the predicted input signal. Input signal standard value As a standard value, it is compared with the standard value of the input signal predicted by the model. By comparing the results, the deviation of the prediction results can be analyzed using equation (2). ,

[0014] (2).

[0015] Optionally, in the above-described nonlinear correction method for an infrared measurement system based on a convolutional neural network, the infrared measurement system is an infrared thermal imager or an infrared thermometer, and in step S1, the temperature of the calibrated standard blackbody is... As the input signal standard value; after setting the emissivity value of the infrared thermal imager or infrared thermometer to the emissivity value of the standard blackbody, the temperature of the standard blackbody is measured using the infrared thermal imager or infrared thermometer to obtain the measurement system response value. ,in, The wavelength or band range of an infrared thermal imager or infrared thermometer, and the temperature T of a standard blackbody. t and measurement system response value A training dataset for the experiment is formed; in step S2, the system response values ​​are measured in the training dataset. As an input feature, the temperature T of a standard blackbody t As the training target value, it is used for a specific wavelength value or band range λ. i A convolutional neural network M(λ) is obtained through training. i ).

[0016] Optionally, in the above-mentioned nonlinear correction method for an infrared measurement system based on a convolutional neural network, the infrared measurement system is an infrared radiometer, and in step S1, a standard blackbody is used as the target to be measured. The temperature of the standard blackbody is sequentially set to a series of different values, and after it stabilizes, the temperature of the standard blackbody is measured and calibrated using a thermometer. The radiance of a standard blackbody at this temperature is calculated using Planck's integral formula (3). The standard value of the input signal is used; the response value of the measurement system is obtained by measuring the standard blackbody using an infrared radiometer. ,in, The wavelength range of the infrared radiometer, and the temperature T of the standard blackbody. t Input signal standard value and measurement system response value The experimental training dataset is formed, where Planck's formula for the integral is:

[0017] (3),

[0018] In step S2, the temperature T of the standard blackbody in the experimental training dataset is... t Measurement system response value As an input feature, the standard value of the input signal As the training target value, for each band range λ i A convolutional neural network M(λ) is obtained through training. i ).

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0020] This invention presents a nonlinear correction method for infrared measurement systems based on convolutional neural networks. It employs a one-dimensional convolutional neural network to construct a model relationship between the response value of the infrared measurement system and the standard value of the input signal. Through this model, when the infrared measurement system obtains any response value, the standard value of the input signal can be predicted, thus achieving nonlinear correction of the infrared measurement system. Compared with traditional methods using polynomial fitting or piecewise linear interpolation, this invention's method, using a one-dimensional convolutional neural network, overcomes the limitations of fitting N-order polynomials or piecewise linear functions, significantly improving the accuracy of nonlinear correction in infrared measurement systems. Furthermore, this invention is applicable to the nonlinear correction of various infrared measurement systems. Attached Figure Description

[0021] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below.

[0022] Figure 1 This is a flowchart illustrating the nonlinear correction method for an infrared measurement system based on a convolutional neural network according to the present invention.

[0023] Figure 2 This is a flowchart of the training process for the convolutional neural network (1D-CNN) model of this invention;

[0024] Figure 3 This is a graph comparing the nonlinear correction of three methods of the present invention: the one-dimensional convolutional neural network method, the piecewise linear interpolation method, and the polynomial fitting method. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0026] like Figure 1 The nonlinear correction method for infrared measurement systems based on convolutional neural networks of the present invention includes the following steps:

[0027] S1. Experimental Data Preparation: A series of standard values ​​of input signals and response values ​​of the measurement system are obtained through experiments, which serve as the experimental training dataset;

[0028] S2. Model Training: The experimental training dataset is trained using a convolutional neural network (1D-CNN) algorithm to obtain a convolutional neural network model between the response value of the measurement system and the standard value of the input signal. A convolutional neural network model is trained for each wavelength value or each band range of the infrared measurement system.

[0029] During the model training phase, the temperature T of the standard blackbody in the experimental training dataset from step S1 is used. t Measurement system response value Defined as input features, the standard value of the input signal Defined as the training target value; during the training process of the convolutional neural network model, a convolutional neural network model is trained for each wavelength value or each band range of the infrared measurement system to fit the relationship between the input features and the training target value.

[0030] S3. Model Validation: After training, use experimental validation datasets for each trained convolutional neural network (1D-CNN) model. An evaluation was conducted to verify the accuracy of the model's prediction of the standard values ​​of the input signal.

[0031] During the model validation phase, within the range covered by the standard values ​​of the input signals in the experimental training dataset, another series of different standard values ​​of input signals and measurement system response values ​​are obtained through experiments to serve as the experimental validation dataset. Input features are defined using the same method as in the aforementioned experimental training dataset and substituted into the relevant model obtained from the previous training to predict the standard values ​​of the input signals. The predicted standard values ​​of the input signals are compared with the standard values ​​of the input signals in the experimental validation dataset obtained through experiments to obtain the relative deviation of the nonlinear correction of the prediction model.

[0032] Taking the nonlinear calibration of an infrared spectral radiance measurement system as an example, the method of the present invention will be specifically described. The nonlinear calibration method of the infrared spectral radiance measurement system includes the following steps:

[0033] S1. Experimental Data Preparation: A series of standard values ​​of input signals and response values ​​of the measurement system are obtained through experiments, which serve as the experimental training dataset;

[0034] In step S1, a standard blackbody is used as the target to be measured. The temperature of the standard blackbody is set to a series of different values ​​in sequence. After it stabilizes, the temperature of the standard blackbody is measured and calibrated using a thermometer. ; Calculate the spectral radiance of a standard blackbody at this temperature according to Planck's formula (1). The standard value of the input signal is used as the standard value; the infrared spectral radiance measurement system is used to measure the standard blackbody to obtain the infrared spectral radiance response value of the measurement system (i.e., the measurement system response value). The temperature T of a standard blackbody t Wavelength λ i Input signal standard value and measurement system response value A training dataset for the experiment is formed. Planck's formula is:

[0035] (1)

[0036] Where c1 and c2 represent the first radiation constant and the second radiation constant, respectively. This refers to a specific wavelength or wave value corresponding to the infrared spectral radiance measurement system.

[0037] S2. Model Training: The experimental training dataset is trained using a convolutional neural network (1D-CNN) algorithm to obtain a convolutional neural network model between the response value of the measurement system and the standard value of the input signal. A convolutional neural network model is trained for each wavelength value of the infrared measurement system.

[0038] In the experimental training dataset, the temperature T of the standard blackbody t and measurement system response value Defined as input characteristics, standard value of the input signal Defined as the training target value, for each wavelength value λ i A convolutional neural network M(λ) is obtained through training. i This allows us to obtain the mapping relationship between the measurement system response value and the standard value of the input signal.

[0039] Specifically, based on the experimental conditions, for each wavelength value λ in the infrared spectrum i The input features (such as measurement system response values, temperature and other related parameters) and training target values ​​(standard values ​​of input signals) are set, and the feature values ​​are standardized. In the model training phase, hyperparameter optimization and training parameter settings are first performed, and then the relevant model is obtained through training.

[0040] First, for each wavelength value λ in the infrared spectrum i The temperature T of the standard blackbody in the experimental training dataset in step S1 t and measurement system response value Defined as input feature, the standard value of the input signal Defined as the training target value. Before model training, all feature values ​​are standardized.

[0041] Secondly, during the training of the 1D-CNN model, a 1D-CNN model is trained for each wavelength value of the infrared spectrum. This is done to fit the relationship between input features and training target values. The process involves two key steps: hyperparameter optimization (…). Figure 2 The "first step" and model training (in the context of model training) Figure 2 (Step Two) First, for each wavelength condition, the optimal hyperparameters are configured using a hyperband hyperparameter search algorithm based on the KerasTuner hyperparameter tuning framework. These hyperparameters include input and output hyperparameters. In this step, the algorithm's input hyperparameters include the maximum number of training epochs per trial, the reduction factor, and the number of iterations; the output hyperparameters include the number of filters and activation functions in the convolutional layers, the number of units in the fully connected layers, the dropout rate in the dropout layers, and the learner's learning rate. Subsequently, these output hyperparameters are used as inputs to the convolutional layers, batch normalization layers, flattening layers, fully connected layers, and dropout layers to train the convolutional neural network (1D-CNN) model. To prevent overfitting and ensure optimal model convergence, methods such as early stopping and dynamic learning rate adjustment (ReduceLROnPlateau) were employed during training.

[0042] S3. Model Validation: After training, use experimental validation datasets for each trained convolutional neural network (1D-CNN) model. An evaluation was conducted to verify the accuracy of the model's prediction of the standard values ​​of the input signal.

[0043] During the model validation phase, within the range covered by the standard values ​​of the input signals in the experimental training dataset, another series of different standard values ​​of input signals and measurement system response values ​​are obtained experimentally as the experimental validation dataset; for each wavelength value λ of the infrared spectrum... i The system response value, temperature, and other parameters from the experimental verification dataset are used as input features and substituted into the previously trained model to predict the standard value of the input signal. The predicted standard value of the input signal is then compared with the standard value of the input signal from the experimental verification dataset to obtain the relative deviation of the nonlinear correction of the prediction model.

[0044] Specifically, the temperature of the standard blackbody is set to any value within the temperature range used in the experimental training dataset. After it stabilizes, the temperature of the standard blackbody is measured and calibrated using a thermometer. Similarly, the spectral radiance of a standard blackbody at this temperature is calculated using Planck's formula. The standard value of the input signal is used as the standard value; the infrared spectral radiance measurement system is used to measure the standard blackbody to obtain the infrared spectral radiance response value of the measurement system (i.e., the measurement system response value). At each wavelength λ i Under the condition that the temperature T of the standard blackbody is...v and measurement system response value The input features of the experimental validation dataset are substituted into the model trained using the experimental training dataset to obtain the predicted spectral radiance values, i.e., the predicted standard values ​​of the input signal. Input signal standard value As a standard value, it is compared with the standard value of the input signal predicted by substituting the input features into the model. By comparing the results, the deviation of the prediction results can be analyzed using equation (2). ,

[0045] (2).

[0046] In the experiment, the model can be repeatedly trained by modifying its parameters to improve the accuracy of its predictions.

[0047] This invention's method, based on the mapping relationship between the input signal value and the response value of the infrared measurement system, trains a model using a convolutional neural network (1D-CNN) algorithm. When the measurement system obtains any response value, it can predict the corresponding input signal value, thus achieving nonlinear correction of the infrared measurement system. Besides the nonlinear correction of infrared spectral radiance measurement systems mentioned above, this invention's method is also applicable to the nonlinear correction of various optical radiation measurement systems such as Fourier transform infrared spectrometers, infrared thermal imagers, infrared thermometers, and infrared radiometers. Specifically, the nonlinear calibration of Fourier transform infrared spectrometers is similar to that of infrared spectral radiance measurement systems; these systems use the standard value of spectral radiance as the standard value of the input signal. Systems like infrared thermal imagers and infrared thermometers use temperature values ​​as the standard value of the input signal. Systems like infrared radiometers use the integrated radiance value as the standard value of the input signal.

[0048] In the nonlinear calibration of infrared thermal imagers and infrared thermometers, the temperature of the calibrated standard blackbody is... As the input signal standard value; after setting the emissivity value of the infrared thermal imager or infrared thermometer to the emissivity value of the standard blackbody, the infrared thermal imager or infrared thermometer is used to measure the temperature of the standard blackbody to obtain the response value of the infrared thermal imager or infrared thermometer (i.e., the response value of the measurement system). ,in, This refers to the wavelength or band range of an infrared thermal imager or infrared thermometer, such as 1.25 micrometers, 8–14 micrometers, or 3–5 micrometers. The temperature T of a standard blackbody. t and the response value of the measurement system Create an experimental training dataset. In the experimental training dataset, measure the system response values. As an input feature, the temperature T of a standard blackbody tAs the training target value, it is used for a specific wavelength value or band range λ. i A convolutional neural network M(λ) is obtained through training. i ).

[0049] Taking an infrared thermometer with a detector response wavelength range of 8~14 micrometers as an example, the experimental training dataset and experimental verification dataset are shown in Table 1 and Table 2 below.

[0050] Table 1. Experimental training dataset for infrared thermometers with detector response wavelengths ranging from 8 to 14 micrometers.

[0051] Measurement system response value (Unit: K) <![CDATA[Temperature T of the standard black body t (unit: K)]]> 373 373 472.7 473 572.8 573 672.7 673 772.4 773 872.3 873

[0052] Table 2. Experimental verification dataset for infrared thermometers with detector response wavelength range of 8–14 micrometers.

[0053] Measurement system response value (Unit: K) <![CDATA[Temperature T of the standard boldface v (unit: K)]]> 423 423 522.9 523

[0054] In the nonlinear calibration of an infrared radiometer, a standard blackbody is used as the target object. The temperature of the standard blackbody is sequentially set to a series of different values. After it stabilizes, a thermometer is used to measure and calibrate the temperature of the standard blackbody. ; Calculate the radiance of a standard blackbody at this temperature using Planck's integral formula (3). The standard value of the input signal is used; the response value of the measurement system is obtained by measuring the standard blackbody using an infrared radiometer. .in, The wavelength range of the infrared radiometer, such as 8–14 micrometers or 3–5 micrometers, etc. The temperature T of the standard blackbody. t Input signal standard value and the response value of the measurement system The experimental training dataset is formed, wherein Planck's formula for the integral is:

[0055] (3)

[0056] In the experimental training dataset, the temperature T of the standard blackbody t Measurement system response value As an input feature, the standard value of the input signal As the training target value, for each band range λ i A convolutional neural network M(λ) is obtained through training. i ).

[0057] Effect verification

[0058] A specific embodiment of a nonlinear calibration method for an infrared spectral radiance measurement system is illustrated. This method includes:

[0059] Data collection for the experimental training dataset: (i) The temperature of the standard blackbody was set to a value in the range of 473 K to 823 K (interval of 50 K); (ii) The spectral radiance value of the standard blackbody was calculated according to Planck's formula (1) and used as the standard value of the input signal; the standard blackbody was measured using an infrared spectral radiance measurement system to obtain the corresponding measurement system response value under each temperature condition; (iii) The standard blackbody temperature, wavelength, and measurement system response value were used as the input features of the experimental training dataset, and the standard value of the input signal was used as the training target value. The model was trained using these data.

[0060] Data collection for experimental verification dataset: (i) The temperature of the standard blackbody was set to 698 K and 798 K, and the data was collected at these two temperature points respectively; (ii) The spectral radiance value of the standard blackbody was calculated according to Planck's formula (1), and used as the standard value of the input signal. The standard blackbody was measured using an infrared spectral radiance measurement system to obtain the corresponding measurement system response value under each temperature point condition; (iii) For each wavelength value λ of the infrared spectrum... i (iv) The predicted spectral radiance value is obtained by substituting the standard blackbody temperature and the measurement system response value as input features of the experimental verification dataset into the model trained using the experimental training dataset; and (iv) The predicted spectral radiance value is compared with the standard value of the input signal in the experimental verification dataset to analyze the accuracy of the model prediction.

[0061] The training and validation of the prediction model can be implemented in Python. In the optimal hyperparameter configuration, the maximum number of training epochs per trial is set to 200, the reduction factor is 3, and the number of iterations per wavelength is 3. Table 3 lists two examples of output-optimized hyperparameters. Through model training and validation, at 5 micrometers, the model's prediction biases at 698 K and 798 K are 0.97% and 0.42%, respectively; at 10 micrometers, the model's prediction biases at 698 K and 798 K are 0.06% and 0.02%, respectively.

[0062] Table 3. Two examples of output-optimized hyperparameters

[0063] Wavelength parameters Number of filters Activation function of convolutional layer Number of units in a fully connected layer The proportion of discarded layers Optimizer learning rate 5.0 micrometers 32 ReLU 192 0.4 0.0032 10.0 micrometers 32 ReLU 192 0.4 0.0100

[0064] To analyze the effectiveness of the method, the relative deviation value calculated by the one-dimensional convolutional neural network (1D-CNN) method proposed in this invention is used. Compared with previous methods based on the Piecewise linear interpolation method and the Polynomial fitting method, in 698 K ( Figure 3 a) and 798 K ( Figure 3 The results obtained from predictions in (b) were compared. Figure 3 As shown in a and b in the figure, the prediction results obtained by the method of the present invention have a smaller relative deviation from the theoretical values, which verifies the effectiveness of the method of the present invention.

[0065] The above embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and are not intended to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or improve the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in the present invention; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

[0066] References

[0067] [1]Theocharous, E., Ishii, J., and Nigel, PF, “Absolute linearity measurements on HgCdTe detectors in the infrared region,” Applied Optics 43(21), 4182-4188 (2004).

[0068] [2]Hamadani, BH, Shore, A., Roller, J., Yoon, HW, andCampanelli, M., “Non-linearity measurements of solar cells with an led-basedcombinatorial flux addition method,” Metrologia 53(1), 76-85 (2016).

[0069] [3]He, S., Sun, R., Dai, C., Liu, J., Feng, G., and Wang, J.,“Research of non-linearity correction in infrared spectral radiancemeasurement,” Proc. SPIE 13154, Sixth Conference on Frontiers in OpticalImaging and Technology: Novel Detector Technologies, 131540G (30 April 2024).

Claims

1. A nonlinear correction method for an infrared measurement system based on a convolutional neural network, characterized in that, Includes the following steps: S1. Experimental Data Preparation: A series of standard values ​​of input signals and response values ​​of the measurement system are obtained through experiments as experimental training datasets; S2. Model Training: The experimental training dataset is trained using a convolutional neural network algorithm to obtain a convolutional neural network model between the response value of the measurement system and the standard value of the input signal, and a convolutional neural network model is trained for each wavelength value or each band range of the infrared measurement system. S3. Model Validation: Each trained convolutional neural network model is evaluated using an experimental validation dataset to verify the accuracy of the model in predicting standard values ​​of the input signal.

2. The nonlinear correction method for an infrared measurement system based on a convolutional neural network according to claim 1, characterized in that, In step S1, a standard blackbody is used as the target to be measured. The temperature of the standard blackbody is set to a series of different values ​​in sequence. After it stabilizes, the temperature of the standard blackbody is measured and calibrated using a thermometer.

3. The nonlinear correction method for an infrared measurement system based on a convolutional neural network according to claim 1, characterized in that, In step S2, the temperature of the standard blackbody and the response value of the measurement system in the experimental training dataset are defined as input features, and the standard value of the input signal is defined as the training target value. During the training process of the convolutional neural network model, a convolutional neural network model is trained for each wavelength value or each band range of the infrared measurement system to fit the relationship between the input features and the training target value.

4. The nonlinear correction method for an infrared measurement system based on a convolutional neural network according to claim 1, characterized in that, In step S3, during the model validation phase, within the range covered by the standard values ​​of the input signals in the experimental training dataset, another series of different standard values ​​of input signals and measurement system response values ​​are obtained experimentally as the experimental validation dataset. Input features are defined using the same method as in the experimental training dataset and substituted into the convolutional neural network model obtained from the previous training to predict the standard values ​​of the input signals. The predicted standard values ​​of the input signals are compared with the standard values ​​of the input signals in the experimental validation dataset obtained experimentally to obtain the relative deviation of the nonlinear correction of the prediction model.

5. The nonlinear correction method for an infrared measurement system based on a convolutional neural network according to claim 1, characterized in that, The infrared measurement system is an infrared spectral radiance measurement system. In step S1, a standard blackbody is used as the target to be measured. The temperature of the standard blackbody is sequentially set to a series of different values. After it stabilizes, a thermometer is used to measure and calibrate the temperature of the standard blackbody. ; Calculate the spectral radiance of a standard blackbody at this temperature according to Planck's formula (1). , as the standard value of the input signal; The standard blackbody was measured using the infrared spectral radiance measurement system to obtain the system response value. The temperature T of the standard blackbody t Wavelength λ i Input signal standard value and the response value of the measurement system The experimental training dataset is formed, wherein Planck's formula is: （1） Where c1 and c2 represent the first radiation constant and the second radiation constant, respectively. This refers to a specific wavelength or wave value corresponding to the infrared spectral radiance measurement system.

6. The nonlinear correction method for an infrared measurement system based on a convolutional neural network according to claim 5, characterized in that, In step S2, the temperature T of the standard blackbody is included in the experimental training dataset. t The measurement system response value As an input feature, the standard value of the input signal As the training target value, for each wavelength value λ i A convolutional neural network M(λ) is obtained through training. i ).

7. The nonlinear correction method for an infrared measurement system based on a convolutional neural network according to claim 5, characterized in that, In step S3, the temperature of the standard blackbody is set to any value within the temperature range used in the experimental training dataset. After it stabilizes, the temperature of the standard blackbody is measured and calibrated using a thermometer. Calculate the spectral radiance of the standard blackbody at this temperature using Planck's formula. , as the standard value of the input signal; The standard blackbody was measured using the infrared spectral radiance measurement system to obtain the system response value. ; At each wavelength λ i Under the condition that the temperature T of the standard blackbody is... v and the response value of the measurement system The input features of the experimental verification dataset are substituted into the model trained using the experimental training dataset to obtain the predicted standard values ​​of the input signal. The standard value of the input signal As a standard value, it is compared with the standard value of the input signal predicted by the model. By comparing the results, the deviation of the prediction results can be analyzed using equation (2). , (2)。 8. The nonlinear correction method for an infrared measurement system based on a convolutional neural network according to claim 1, characterized in that, The infrared measurement system is an infrared thermal imager or an infrared thermometer, and in step S1, the temperature of the calibrated standard blackbody is... As the input signal standard value; after setting the emissivity value of the infrared thermal imager or the infrared thermometer to the emissivity value of the standard blackbody, the infrared thermal imager or the infrared thermometer is used to measure the temperature of the standard blackbody to obtain the measurement system response value. ,in, The wavelength value or band range of the infrared thermal imager or the infrared thermometer, and the temperature T of the standard blackbody. t and the response value of the measurement system The experimental training dataset is formed; in step S2, the measurement system response values ​​are added to the experimental training dataset. As an input feature, the temperature T of the standard blackbody t As the training target value, it is used for a specific wavelength value or band range λ. i A convolutional neural network M(λ) is obtained through training. i ).

9. The nonlinear correction method for an infrared measurement system based on a convolutional neural network according to claim 1, characterized in that, The infrared measurement system is an infrared radiometer. In step S1, the standard blackbody is used as the target to be measured. The temperature of the standard blackbody is sequentially set to a series of different values. After it stabilizes, the temperature of the standard blackbody is measured and calibrated using a thermometer. The radiance of the standard blackbody at this temperature is calculated using Planck's integral formula (3). The standard value of the input signal is used; the infrared radiometer is used to measure the standard blackbody to obtain the response value of the measurement system. ,in, The wavelength range of the infrared radiometer, and the temperature T of the standard blackbody. t The input signal standard value and the response value of the measurement system The experimental training dataset is formed, wherein Planck's formula for the integral is: （3）， In step S2, the temperature T of the standard blackbody in the experimental training dataset is... t The measurement system response value As an input feature, the standard value of the input signal As the training target value, for each band range λ i A convolutional neural network M(λ) is obtained through training. i ).

Images