Cereal quality analysis method based on prediction-correction double neural network fusion

Abstract

The invention discloses a cereal quality analysis method based on prediction-correction double neural network fusion. The cereal quality analysis method comprises the steps of constructing a standard sample data set, constructing a spectrum screening method, constructing a correction sample data set, constructing a prediction neural network model, constructing a correction neural network model and evaluating quality components. Through dual-network fusion collaborative design of the prediction neural network and the correction neural network, the problems of overfitting, weak generalization ability and the like of a traditional model in a small sample scene are effectively solved, and the stability and prediction accuracy of the model are improved.

Description

Technical Field

[0001] The present invention relates to the field of grain analysis, and in particular to a grain quality analysis method based on prediction-correction dual neural network fusion. Background Art

[0002] In modern agriculture and food processing, quality testing of grains (such as wheat, rice, corn, and sorghum) is a key component in ensuring agricultural product quality and promoting industrial upgrading. Near-infrared (NIR) spectroscopy, due to its non-destructive, rapid, and online capabilities, has become a mainstream method for grain quality assessment. This technique, based on the absorption characteristics of organic molecules (such as amylopectin, proteins, crude fiber, and carbohydrates) in grains, enables rapid detection by establishing mathematical models linking spectral data with quality parameters.

[0003] However, traditional near-infrared spectroscopy (NIR) analysis relies on large, high-quality calibration datasets to ensure model accuracy. Obtaining this data requires complex chemical analysis (such as high-performance liquid chromatography for sugars and Kjeldahl nitrogen determination for protein), which is time-consuming, labor-intensive, and costly. For example, the cost of chemical analysis for a single sample of corn kernel protein can be 5-10 times that of NIR spectroscopy. Furthermore, large-scale sample testing requires significant manpower and equipment resources, making it difficult to meet the low-cost testing requirements of scenarios such as early screening in grain breeding, dynamic quality monitoring during the grain filling period of fresh corn, and real-time monitoring of processed raw materials.

[0004] In terms of data processing and model building, existing technologies often use partial least squares (PLS) or single artificial neural networks (ANN) to build prediction models. While PLS algorithms offer high computational efficiency, their prediction accuracy is limited when processing complex nonlinear relationships in grain spectral data (such as spectral differences between different grain varieties and multi-component interactions). Traditional ANN models require a large number of training samples to avoid overfitting. When the sample size is insufficient (e.g., less than 200), the model's generalization ability significantly decreases, making it difficult to meet the high-precision detection requirements of low-cost, small-sample scenarios.

[0005] The core bottleneck of existing technologies lies in modeling the mapping relationship between spectral data and quality parameters as a single process, failing to distinguish between the fundamentally different issues of "prediction with large-scale data" and "correction with small sample sizes." Traditional methods either rely on large amounts of real data to build a single model, increasing costs, or sacrifice accuracy with small sample sizes, reducing the method's adaptability and failing to simultaneously meet the dual requirements of "low cost" and "high accuracy" in grain quality testing. Summary of the Invention

[0006] The main purpose of the present invention is to provide a grain quality analysis method based on prediction-correction dual neural network fusion, aiming to solve the above technical problems.

[0007] To achieve the above objectives, the present invention proposes a grain quality analysis method based on prediction-correction dual neural network fusion, which includes: (1) Constructing a standard sample data set: preparing artificially synthesized standard samples and recording the composition data of each sample, using a near-infrared spectrometer to collect near-infrared spectral data of the artificially synthesized standard samples, and constructing a standard sample data set containing spectral data and corresponding composition data; (2) Constructing a spectral screening method: performing a correlation analysis between the spectrum and the composition data of the standard sample data set, and constructing a spectral screening method based on the analysis results; (3) Constructing a calibration sample data set: preparing actual grain samples, collecting near-infrared spectral data of each sample, combining the spectral screening method, and screening out calibration samples using the SELECT algorithm; using chemical analysis methods to determine the quality component composition of the calibration samples, and constructing a calibration sample data set; (4) Constructing a prediction neural network model: using the standard sample data set to train the prediction neural network model, and adjusting the network parameters to make the quality parameter prediction value output by the model converge with the mean square error of the standard sample composition data, thereby completing the parameter optimization of the prediction neural network model; (5) Constructing a correction neural network model: performing deviation calculation on the prediction results of the prediction neural network model for the spectral data of the correction sample and the measured values of the correction sample data set to obtain deviation data, using the deviation data as the correction measured value, and constructing a correction neural network model with the correction measured value as the target output. By adjusting the network parameters, the mean square error between the correction prediction value output by the model and the correction measured value converges, thereby completing the parameter optimization of the correction neural network model; (6) Quality component evaluation: Near-infrared spectral data of the grain sample to be evaluated is collected, and after the near-infrared spectral data is processed by the spectral screening method, it is simultaneously input into the trained prediction neural network model and correction neural network model to obtain the predicted value and correction value of the quality component respectively; by fusing the predicted value and correction value, the final grain quality component evaluation result is output.

[0008] In one embodiment, the steps of constructing the standard sample data set in step (1) are specifically as follows: Set the spectrum acquisition wavenumber range to 4000-10000cm -1 , spectral resolution 4cm -1 , spectrum collection was performed on each artificially synthesized standard sample, the number of scans was 32, and the near-infrared spectrum data of the sample was obtained after taking the average value; The collected near-infrared spectral data are associated with the corresponding component data to construct a standard sample data set containing spectral data and corresponding component data.

[0009] In one embodiment, the steps of constructing the spectral screening method in step (2) are specifically as follows: A preset threshold R of the absolute value of the correlation coefficient was defined, 0.5≤R≤0.9, and the spectral bands with absolute values of the correlation coefficient greater than R were retained as characteristic wavelengths to construct a spectral screening method.

[0010] In one embodiment, the cereal grains in step (3) are powdered or whole grain samples.

[0011] In one embodiment, the step of determining the quality component composition of the calibration sample using a chemical analysis method in step (3) is specifically as follows: The carbohydrate composition of cereal grains was determined by high performance liquid chromatography; The protein content of cereal grains was determined by Kjeldahl method; The crude fiber content of cereal grains was determined by acid-base washing method; The amylopectin content in cereal grains was determined by iodine colorimetry.

[0012] In one embodiment, the step (4) of constructing a prediction neural network model is specifically as follows: Preprocessing the spectral data of the standard sample data set using the spectral screening method to extract spectral data corresponding to characteristic wavelengths; Select a feedforward neural network as the prediction neural network model. Set the network to contain 1-4 hidden layers, at least one neuron, and use the tanh function, ReLU function, or swish function as the neuron activation function. The known component ratio of the artificially synthesized standard sample was used as the target output, and the characteristic spectrum data was used to train the prediction neural network model. During the training process, the Adam optimization algorithm was used, and the learning rate was set to 0.001, the batch size was 32, and the number of training rounds was 200.

[0013] In one embodiment, the step of constructing the correction neural network model in step (5) is specifically as follows: In combination with the spectral screening method, the spectral data of the calibration sample data set is preprocessed to extract the spectral data corresponding to the characteristic wavelength; The trained prediction neural network model is used to predict the spectral data of the screened calibration sample data set obtained in step (5) to obtain prediction results of various quality parameters; Calculate the deviation between the predicted result and the measured value of the calibration sample data set to obtain the deviation data of each calibration sample, and use the deviation data as the calibration measured value; Select a feedforward neural network as the correction neural network model. Set the network to contain 1-4 hidden layers, at least one neuron, and the neuron activation function to be the tanh function, ReLU function, or swish function. The obtained corrected measured value is used as the target output, and the characteristic spectrum data obtained in step (5) is used to train the corrected neural network model. During the training process, the Adam optimization algorithm is used, the learning rate is set to 0.001, the batch size is 16, and the number of training rounds is 150.

[0014] The technical solution of the present invention achieves a multi-dimensional technological breakthrough in grain quality assessment by constructing a near-infrared spectroscopy detection system that integrates a prediction-correction dual neural network. Its beneficial effects are specifically reflected in the following aspects: (1) Reduced testing costs and data dependency: By artificially synthesizing standard samples, the reliance on large-scale real experimental data is reduced. Combined with small sample calibration, the time, manpower, and financial costs of data collection are significantly reduced. Traditional methods rely on hundreds to thousands of real samples and implement high-cost chemical analysis methods to determine quality component data to build prediction models. However, this invention uses artificially synthesized standard samples and combines them with the SELECT algorithm to screen small sample calibration sets. This reduces the amount of real grain kernel samples required for chemical analysis to determine quality components to 10%-40%, effectively reducing system costs.

[0015] (2) Achieve high-precision detection under small sample conditions: The dual-network collaborative mechanism breaks through the performance bottleneck of the traditional single model. The prediction neural network learns the theoretical distribution of quality parameters based on artificially synthesized data, and the correction neural network compensates for prediction deviations through small sample measured data. The fusion of the two makes the detection accuracy of key quality parameters such as amylopectin, protein, crude fiber, and sugars basically the same as that of the traditional ANN model. Experimental data show that when the sample size is only 15% of the traditional method, the present invention has an R&D accuracy of 100% in the detection of quality components such as amylopectin, protein, and crude fiber in fresh corn during the filling period. 2 It can reach 0.92.

[0016] (3) Significantly improved model generalization and stability: The prediction network learns theoretical distribution characteristics, making it resistant to noise, while the correction network adapts to the spectral variation of grains of different varieties and origins through a bias compensation mechanism. The dual-network fusion and collaborative design of the prediction neural network and the correction neural network effectively solves the problems of overfitting and weak generalization ability of traditional models in small sample scenarios, and improves the stability and prediction accuracy of the model.

[0017] (4) Constructing a universal detection framework across multiple grain varieties: Through spectral feature screening and a dual-network parameter adaptive adjustment mechanism, this invention achieves compatible detection of multiple grains, including corn, wheat, and rice. When switching from corn detection to rice detection, there is no need to rebuild the entire dataset; only 20-60 rice calibration samples are needed to complete the model migration, saving costs and significantly improving the efficiency of technology promotion. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying any creative work.

[0019] Figure 1 It is the process of constructing and training a predictive neural network model based on artificially synthesized standard sample data; Figure 2 It is the process of constructing and training the correction neural network model based on the correction sample data; Figure 3 It is a process of evaluating grain quality based on near infrared spectroscopy using a prediction-correction dual neural network fusion. The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0020] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0021] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative position relationship, movement status, etc. between the various components under a certain specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indication will also change accordingly.

[0022] In addition, the terms "first," "second," and so on, used in this disclosure are for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of the technical features being referenced. Thus, a feature specified as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of this disclosure, "plurality" means at least two, such as two or three, unless otherwise specifically defined.

[0023] Moreover, the technical solutions between the various embodiments of the present invention may be combined with each other, but this must be based on the fact that ordinary technicians in this field can implement them. When the combination of technical solutions is mutually contradictory or cannot be implemented, it should be deemed that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0024] The present invention provides a grain quality analysis method based on prediction-correction dual neural network fusion.

[0025] The grain quality analysis method based on the prediction-correction dual neural network fusion provided by the embodiment of the present invention is specifically as follows: Step (1), build a standard sample data set: Step (1a), preparation of artificial synthetic standard samples: Based on the theoretical or experimental distribution patterns of quality parameters such as amylopectin, protein, crude fiber, and carbohydrates in grain kernels, the distribution range and probability density function of each quality parameter are first determined. Monte Carlo simulation is then used to simulate the distribution patterns of each component in the grain kernels, generating diverse and consistent data on the proportions of each quality component. High-purity raw materials are then mixed according to the aforementioned proportions to create a synthetic standard sample. For example, in experiments evaluating the quality of fresh corn during the grain-filling period, the main quality components include amylose, amylopectin, protein, crude fiber, and carbohydrates (including fructose, glucose, and sucrose), which together account for 95%. Specifically, the protein content of fresh corn during the grain-filling period exhibits a normal distribution between 8% and 15%. For sweet, glutinous, and sweet-glutinous corn, the amylopectin content follows a normal distribution with multiple means and standard deviations. The Python SciPy library can be used to generate diverse and consistent data on the proportions of each quality component. Prepare synthetic standard samples using high-purity amylopectin (purity ≥98%), amylose, soy protein, corn fiber, glucose, fructose, and sucrose as raw materials. These are accurately weighed and thoroughly mixed according to simulated proportions. Strictly control the ambient temperature and humidity during sample preparation.

[0026] Step (1b), component data recording: For each artificially synthesized standard sample, the specific values of its quality parameters such as amylopectin, protein, crude fiber, and carbohydrates are recorded in detail.

[0027] Step (1c), near-infrared spectroscopy data acquisition: A near-infrared spectrometer was used with a spectrum acquisition wavenumber range of 4000-10000 cm-1 and a spectral resolution of 4 cm-1. Spectra were acquired for each artificially synthesized standard sample with 32 scans. The average value was taken to reduce noise and obtain the near-infrared spectrum data of the sample.

[0028] Step (1d) constructs a standard sample dataset: The collected near-infrared spectral data are associated with the corresponding component data to construct a standard sample data set containing spectral data and corresponding component data.

[0029] Step (2): Constructing a spectral screening method: Step (2a), correlation analysis: The correlation analysis between the spectrum and component data of the standard sample data set was performed, and the Pearson correlation coefficient between the spectrum data at each wavelength and quality parameters such as amylose, amylopectin, protein, crude fiber, and carbohydrates was calculated.

[0030] Step (2b), characteristic wavelength screening: By setting the preset threshold of the absolute value of the correlation coefficient to R (generally 0.5 ≤ R ≤ 0.9), and retaining spectral bands with absolute values of the correlation coefficient greater than R as characteristic wavelengths, a spectral screening method is constructed. This method can screen out spectral bands with high correlation with quality parameters, reduce data dimensionality, and improve the efficiency and accuracy of subsequent model training.

[0031] Step (3), construct the calibration sample data set: Step (3a), actual sample preparation and spectrum acquisition: Prepare a certain number of actual grain samples, typically using powdered samples, but also whole grain samples after processing. Use the same near-infrared spectrometer as in step (1c) and set the same parameters to collect near-infrared spectral data for each sample.

[0032] Step (3b), calibration sample screening: Combined with the spectral screening method constructed in step (2), the spectral data of the actual samples are preprocessed to extract the spectral data corresponding to the characteristic wavelength. Then, the calibration samples are screened using the SELECT algorithm. The spectral data of the characteristic wavelength is used as input, and the Euclidean distance is used as the sample space distance metric to screen out representative calibration samples that cover the spectral variation range of the actual grain sample dataset. For example, the Euclidean distance between each sample and other samples is calculated, and a number of samples that can maximize the coverage of the entire sample space are selected as calibration samples.

[0033] Step (3c), quality parameter determination: The selected calibration samples were subjected to chemical analysis to determine their quality and composition. High-performance liquid chromatography was used to determine carbohydrate content, the Kjeldahl method was used to determine protein content, the acid-base washing method was used to determine crude fiber content, and the iodine colorimetric method was used to determine amylopectin content.

[0034] Step (3d), construct the correction sample data set: The near-infrared spectral data of the calibration samples were associated with the measured values of the corresponding quality parameters to construct a calibration sample data set containing the measured values of quality components such as amylose, amylopectin, protein, crude fiber, and carbohydrates.

[0035] Step (4), build a prediction neural network (PNN) model: Step (4a), spectral screening: Please refer to Figure 1 , combined with the spectral screening method constructed in step (2), the spectral data of the standard sample data set constructed in step (1) are preprocessed to extract the spectral data corresponding to the characteristic wavelength.

[0036] Step (4b), model selection and architecture design: A feedforward neural network (FNN) was selected as the prediction neural network model. The network was set to contain 1-4 hidden layers, with the number of neurons in each hidden layer determined experimentally. The neuron activation function used was the tanh function, ReLU function, or swish function.

[0037] Step (4c), model training and parameter optimization: The known component ratio of the artificially synthesized standard sample obtained in step (1) was used as the target output, and the characteristic spectrum data obtained in step (4a) were used to train the prediction neural network model. During the training process, the Adam optimization algorithm was used, with a learning rate of 0.001, a batch size of 32, and a number of training rounds of 200. By adjusting the network parameters, the mean square error between the predicted quality parameter output by the model and the standard sample composition data converged. During the training process, the change in the mean square error was monitored in real time. When the mean square error no longer decreased significantly, the training was stopped, the parameter optimization of the prediction neural network model was completed, and the prediction neural network model and its parameters were saved.

[0038] Step (5): Construct the Correction Neural Network (CNN) model: Step (5a), spectral screening: Please refer to Figure 2, combined with the spectral screening method constructed in step (2), the spectral data of the calibration sample data set constructed in step (3) are preprocessed to extract the spectral data corresponding to the characteristic wavelength.

[0039] Step (5b), deviation calculation: The prediction neural network model trained in step (4) is used to predict the spectral data of the screened calibration sample data set obtained in step (5a) to obtain the prediction results of each quality parameter. Then, the deviation between the prediction results and the measured values of the calibration sample data set is calculated to obtain the deviation data of each calibration sample, and the deviation data is used as the calibration measured value.

[0040] Step (5c), model selection and architecture design: A feedforward neural network (FNN) is selected as the correction neural network model. The network is set to contain 1-4 hidden layers, with the number of neurons in each hidden layer determined by experiments. The neuron activation function is the tanh function, the ReLU function, or the swish function.

[0041] Step (5d), model training and parameter optimization: The corrected measured value obtained in step (5b) is used as the target output, and the characteristic spectrum data obtained in step (5a) are used to train the corrected neural network model. During the training process, the Adam optimization algorithm is used, with a learning rate of 0.001, a batch size of 16, and 150 training rounds. By adjusting the network parameters, the mean square error between the corrected predicted value output by the model and the corrected measured value converges. Similarly, the mean square error is monitored during the training process. When the convergence condition is met, training is stopped, completing the parameter optimization of the corrected neural network model. The corrected neural network model and its parameters are saved.

[0042] Step (6), quality component assessment: Step (6a), spectral data acquisition: Please refer to Figure 3 For the grain samples to be evaluated, the sample preparation and near-infrared spectral data collection were completed according to the method in step (3a).

[0043] Step (6b), spectral screening: Combined with the spectral screening method constructed in step (2), the spectral data collected in step (6a) is preprocessed to extract the spectral data corresponding to the characteristic wavelength.

[0044] Step (6c), model prediction and correction: The spectral data filtered in step (6b) is input into the prediction neural network model trained in step (4) to obtain the quality component prediction value; at the same time, the spectral data is input into the correction neural network model trained in step (5) to obtain the quality component correction value.

[0045] Step (6d), result fusion and output: The predicted value and the corrected value are fused through preset fusion rules, for example, by summing the predicted value and the corrected value, and finally the evaluation results of the grain quality components are output, including the specific values of parameters such as amylose, amylopectin, protein, crude fiber, and carbohydrates.

[0046] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. All equivalent structural transformations made by using the contents of the present invention description and drawings under the concept of the present invention, or direct / indirect application in other related technical fields are included in the patent protection scope of the present invention.

Claims

1. A grain quality analysis method based on prediction-correction dual neural network fusion, characterized in that: The grain quality analysis method based on the prediction-correction dual neural network fusion includes: (1) Constructing a standard sample data set: preparing artificially synthesized standard samples and recording the composition data of each sample, using a near-infrared spectrometer to collect near-infrared spectral data of the artificially synthesized standard samples, and constructing a standard sample data set containing spectral data and corresponding composition data; (2) Constructing a spectral screening method: performing a correlation analysis between the spectrum and the composition data of the standard sample data set, and constructing a spectral screening method based on the analysis results; (3) Constructing a calibration sample data set: preparing actual grain samples, collecting near-infrared spectral data of each sample, combining the spectral screening method, and screening out calibration samples using the SELECT algorithm; using chemical analysis methods to determine the quality component composition of the calibration samples, and constructing a calibration sample data set; (4) Constructing a prediction neural network model: using the standard sample data set to train the prediction neural network model, and adjusting the network parameters to make the quality parameter prediction value output by the model converge with the mean square error of the standard sample composition data, thereby completing the parameter optimization of the prediction neural network model; (5) Constructing a correction neural network model: performing deviation calculation on the prediction results of the prediction neural network model for the spectral data of the correction sample and the measured values of the correction sample data set to obtain deviation data, using the deviation data as the correction measured value, and constructing a correction neural network model with the correction measured value as the target output. By adjusting the network parameters, the mean square error between the correction prediction value output by the model and the correction measured value converges, thereby completing the parameter optimization of the correction neural network model; (6) Quality component evaluation: Near-infrared spectral data of the grain sample to be evaluated is collected, and after the near-infrared spectral data is processed by the spectral screening method, it is simultaneously input into the trained prediction neural network model and correction neural network model to obtain the predicted value and correction value of the quality component respectively; by fusing the predicted value and correction value, the final grain quality component evaluation result is output.

2. The grain quality analysis method based on prediction-correction dual neural network fusion according to claim 1, characterized in that: The steps of constructing the standard sample data set in step (1) are specifically as follows: Set the spectrum acquisition wavenumber range to 4000-10000cm -1 , spectral resolution 4cm -1 , spectrum collection was performed on each artificially synthesized standard sample, the number of scans was 32, and the near-infrared spectrum data of the sample was obtained after taking the average value; The collected near-infrared spectral data are associated with the corresponding component data to construct a standard sample data set containing spectral data and corresponding component data.

3. The grain quality analysis method based on prediction-correction dual neural network fusion according to claim 1, characterized in that: The steps of constructing the spectral screening method in step (2) are specifically as follows: A preset threshold R of the absolute value of the correlation coefficient was defined, 0.5≤R≤0.9, and the spectral bands with absolute values of the correlation coefficient greater than R were retained as characteristic wavelengths to construct a spectral screening method.

4. The grain quality analysis method based on prediction-correction dual neural network fusion according to claim 1, characterized in that: The cereal grains in step (3) are powdered or whole grain samples.

5. The grain quality analysis method based on prediction-correction dual neural network fusion according to claim 1, characterized in that: The steps of determining the quality component composition of the calibration sample by chemical analysis method in step (3) are specifically as follows: The carbohydrate composition of cereal grains was determined by high performance liquid chromatography; The protein content of cereal grains was determined by Kjeldahl method; The crude fiber content of cereal grains was determined by acid-base washing method; The amylopectin content in cereal grains was determined by iodine colorimetry.

6. The grain quality analysis method based on prediction-correction dual neural network fusion according to claim 1, characterized in that: The steps of constructing the prediction neural network model in step (4) are specifically as follows: Preprocessing the spectral data of the standard sample data set using the spectral screening method to extract spectral data corresponding to characteristic wavelengths; Select a feedforward neural network as the prediction neural network model. Set the network to contain 1-4 hidden layers, at least one neuron, and use the tanh function, ReLU function, or swish function as the neuron activation function. The known component ratio of the artificially synthesized standard sample was used as the target output, and the characteristic spectrum data was used to train the prediction neural network model. During the training process, the Adam optimization algorithm was used, and the learning rate was set to 0.001, the batch size was 32, and the number of training rounds was 200.

7. The grain quality analysis method based on prediction-correction dual neural network fusion according to claim 1, characterized in that: The steps of constructing the correction neural network model in step (5) are specifically as follows: In combination with the spectral screening method, the spectral data of the calibration sample data set is preprocessed to extract the spectral data corresponding to the characteristic wavelength; The trained prediction neural network model is used to predict the spectral data of the screened calibration sample data set obtained in step (5) to obtain prediction results of various quality parameters; Calculate the deviation between the predicted result and the measured value of the calibration sample data set to obtain the deviation data of each calibration sample, and use the deviation data as the calibration measured value; Select a feedforward neural network as the correction neural network model. Set the network to contain 1-4 hidden layers, at least one neuron, and the neuron activation function to be the tanh function, ReLU function, or swish function. The obtained corrected measured value is used as the target output, and the characteristic spectrum data obtained in step (5) is used to train the corrected neural network model. During the training process, the Adam optimization algorithm is used, the learning rate is set to 0.001, the batch size is 16, and the number of training rounds is 150.

Images