Quantitative prediction model construction method and device for spectrum reconstruction, and storage medium
By constructing a spectral reconstruction and prediction model using a convolutional neural network, the problem of accuracy and prediction precision in mapping high-dimensional spectra to low-dimensional spectra is solved. This achieves synergistic optimization of the accuracy of spectral reconstruction and prediction precision, and is applicable to the detection of material properties in the chemical, agricultural, and pharmaceutical fields.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINESE ACAD OF AGRI MECHANIZATION SCI GRP CO LTD
- Filing Date
- 2026-01-16
- Publication Date
- 2026-05-19
AI Technical Summary
In existing technologies, high-precision spectrometers are expensive to acquire high-dimensional continuous spectra, while low-cost sensors have limited spectral resolution. This leads to feature space breaks and information loss in spectral reconstruction models under high-dimensional differences. Furthermore, there are inter-instrument differences and sample reproducibility problems between different instruments, making it difficult to achieve synergistic optimization of spectral reconstruction accuracy and prediction precision.
A convolutional neural network-based approach is adopted to construct a spectral reconstruction model using pure standard sample data. A mapping matrix network structure is designed, and the network parameters are optimized by combining channel loss and reconstruction loss to achieve accurate reconstruction of high-dimensional continuous spectra into low-dimensional discrete data. At the same time, a content prediction model is constructed based on measured sample data, and a multi-objective constrained loss function is used to synergistically optimize the accuracy of spectral reconstruction and the accuracy of model prediction.
While ensuring the accuracy of spectral reconstruction, it improves the prediction accuracy of low-dimensional discrete data, solves the problem of feature space fragmentation and information loss in spectral reconstruction models under high-dimensional differences, and enhances the model's transferability and generalization ability.
Smart Images
Figure CN122064948A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of spectral analysis technology, and in particular to a method, apparatus, and storage medium for constructing a quantitative prediction model for spectral reconstruction, especially a method, apparatus, and storage medium for constructing a quantitative prediction model for spectral reconstruction based on a convolutional neural network. Background Technology
[0002] Near-infrared spectroscopy, with its advantages of simultaneous multi-parameter measurement, non-destructive nature, and online real-time detection, is widely used in the detection of material properties in fields such as chemical engineering, agricultural products, and pharmaceuticals. However, in existing technologies, high-precision spectrometers can acquire high-dimensional continuous spectra at n wavelengths, providing rich data information but at a high cost; low-cost sensors output only m discrete channel values, offering low cost but limited spectral resolution. The extreme dimensionality difference between the two (n >> m) easily leads to feature space breaks and loss of key information during model transfer, making it difficult for traditional dimensionality reduction methods to achieve accurate mapping.
[0003] Furthermore, single-objective optimization can lead to contradictions such as "high reconstruction accuracy but weak prediction generalization" or "acceptable prediction accuracy but distorted spectral physical meaning." In addition, there are inter-instrument differences between different instruments and the difficulty in reproducing the measured position of the sample, which further reduces the model's transferability.
[0004] Therefore, there is an urgent need for a collaborative optimization method that balances the accuracy of spectral reconstruction with the precision of the prediction model, in order to solve the problems of accurate mapping of high-dimensional spectra to low-dimensional channels and high prediction accuracy of low-dimensional channel spectral data. Summary of the Invention
[0005] To address the shortcomings of the existing technologies, this invention provides a method for constructing a quantitative prediction model based on spectral reconstruction using convolutional neural networks, comprising:
[0006] Based on pure standard sample data, a spectral reconstruction model is constructed while ensuring the accuracy of spectral reconstruction, so as to realize the spectral reconstruction of high-dimensional continuous spectrum to low-dimensional discrete data.
[0007] Based on measured sample data, a content prediction model is constructed while ensuring the accuracy of model prediction.
[0008] By combining the accuracy of spectral reconstruction with the model prediction precision, a multi-objective constrained loss function is co-optimized to obtain a complete set of calibrated spectral-channel response weights and a prediction model.
[0009] In some embodiments, the construction of the spectral reconstruction model includes:
[0010] The standard dataset is constructed by selecting pure substances with wide-band absorption or reflection characteristics as standards;
[0011] The design incorporates a mapping matrix network structure, including a 1D-CNN feature extraction module, a mapping matrix module, and a deconvolution reconstruction module. The network parameters are optimized using channel loss and reconstruction loss to obtain a spectral reconstruction model.
[0012] In some embodiments, the pure material standard includes an absorption standard, a reflection standard, a scattering standard, and a blank / baseline standard.
[0013] In some embodiments, the mapping matrix network structure is as follows:
[0014] Conv1 - Batch Norm1 - Conv2 - MaxPool - Conv3 - Flatten - Dense1 - Dropout - Dense2 - Reshape - Deconv1 - Deconv2 - Deconv3.
[0015] In some embodiments, the training loss function of the mapping matrix network structure is:
[0016]
[0017]
[0018]
[0019] in, This is due to channel loss; For reconstruction losses; This is a dynamic weighting coefficient used to balance the magnitude of channel loss and reconstruction loss; This represents the k-th discrete channel value predicted by the spectral reconstruction model. This represents the theoretical value of the k-th discrete channel of the target instrument. The value of the reconstructed high-dimensional spectrum at the p-th wavelength point; This represents the true value of the original high-dimensional continuous spectrum acquired by a high-precision spectrometer at the p-th wavelength point.
[0020] In some embodiments, the construction of the content prediction model includes:
[0021] High-dimensional continuous spectra of samples are acquired using high-precision spectrometers, and reconstructed values of discrete channels are output through a spectral reconstruction model.
[0022] The reconstructed spectral data is preprocessed and then divided into a calibration set, a validation set, and a test set.
[0023] The content prediction model network design includes a channel attention module, a multi-scale CNN feature extraction module, a global feature compression module, and a fully connected prediction module;
[0024] Using the reconstructed near-infrared spectral data, a quantitative concentration prediction model is obtained by training the prediction model network structure while minimizing the RMSE loss function.
[0025] In some embodiments, the network structure of the quantitative concentration prediction model is as follows:
[0026] GlobalAvgPool - Dense1 - Dense2 - Multiply - Conv5 - BN5 - AvgPool5 (Conv3 - BN3-AvgPool3) - Concat - GELU - GlobalMaxPool - GlobalAvgPool2 - Flatten - Dense3 - Dropout - Output.
[0027] In some embodiments, the multi-objective constraint loss function is:
[0028]
[0029] in: The equilibrium coefficients are of the order of magnitude, due to the spectral reconstruction loss function. With the loss function of the prediction model The magnitudes of the two may differ, therefore a coefficient is used. To achieve magnitude balance;
[0030] , For the k-th sample, This represents the predicted value of the physical quantity for the k-th sample output by the prediction model.
[0031] In another aspect, the present invention provides an apparatus for constructing a quantitative prediction model for spectral reconstruction, which adopts the above-described method for constructing a quantitative prediction model for spectral reconstruction, the apparatus comprising:
[0032] The spectral reconstruction modeling module is used to build a spectral reconstruction model based on pure standard sample data, while ensuring the accuracy of spectral reconstruction, so as to realize the spectral reconstruction of high-dimensional continuous spectrum to low-dimensional discrete data.
[0033] The content prediction modeling module is used to build a content prediction model based on measured sample data while ensuring the accuracy of the model prediction.
[0034] The joint optimization and deployment module is used to combine the accuracy of spectral reconstruction with the accuracy of model prediction, and to perform multi-objective constrained loss function collaborative optimization to obtain a complete set of calibrated spectral-channel response weights and prediction models.
[0035] Furthermore, the present invention provides a computer-readable storage medium storing at least one piece of program code, which is loaded and executed by a processor to implement the quantitative prediction model construction method for spectral reconstruction as described above.
[0036] Compared with the prior art, the present invention has the following advantages:
[0037] The quantitative prediction model construction method for spectral reconstruction provided by this invention addresses the weighted mapping from continuous spectrum to discrete channels and the mapping relationship of spectral feature space in near-infrared spectra for fruit and vegetable product testing, pharmaceuticals, and chemical industries. It overcomes the problems of feature space breakage and information loss that may occur during model transfer under extreme dimensional differences in the spectrum.
[0038] This invention provides a spectral reconstruction model for high-dimensional continuous spectra to low-discrete channels based on convolutional neural networks and a method for constructing near-infrared spectral prediction models. By learning response weights driven by pure standard sample data and optimizing multi-objective collaborative training with measured sample data, the contradiction between spectral reconstruction accuracy and prediction generalization is resolved. Attached Figure Description
[0039] Figure 1 A schematic diagram of the overall process of constructing a quantitative prediction model for spectral reconstruction provided in an embodiment of the present invention is shown.
[0040] Figure 2 This invention provides an overall flowchart of the method for constructing a quantitative prediction model for spectral reconstruction.
[0041] Figure 3 This diagram shows the overall structure of the quantitative prediction model construction device for spectral reconstruction provided in an embodiment of the present invention.
[0042] In the attached figures, the following labels are used:
[0043] 30: A device for constructing quantitative prediction models for spectral reconstruction;
[0044] 31: Spectral Reconstruction Modeling Module;
[0045] 32: Content prediction modeling module;
[0046] 33: Joint Optimization and Deployment Module;
[0047] S1-S3: Steps. Detailed Implementation
[0048] See Figure 1-3 An embodiment of the present invention provides a method for constructing a quantitative prediction model for spectral reconstruction, comprising:
[0049] S1. Based on pure standard sample data, a spectral reconstruction model is constructed to realize the spectral reconstruction of high-dimensional continuous spectrum into low-dimensional discrete data while ensuring the accuracy of spectral reconstruction.
[0050] S2. Based on measured sample data, construct a content prediction model while ensuring the accuracy of model prediction;
[0051] S3. By combining the accuracy of spectral reconstruction with the model prediction accuracy, multi-objective constrained loss function co-optimization is performed to obtain a complete set of calibrated spectral-channel response weights and prediction models.
[0052] The construction of the spectral reconstruction model includes:
[0053] The construction of the standard dataset involves selecting pure substances with broadband absorption or reflection characteristics as standards. Pure substance standards include absorption standards, reflection standards, scattering standards, and blank / baseline standards.
[0054] The design incorporates a mapping matrix network structure, including a 1D-CNN feature extraction module, a mapping matrix module, and a deconvolution reconstruction module. The network parameters are optimized using channel loss and reconstruction loss to obtain a spectral reconstruction model.
[0055] The mapping matrix network structure is as follows:
[0056] Conv1 - Batch Norm1 - Conv2 - MaxPool - Conv3 - Flatten - Dense1 - Dropout - Dense2 - Reshape - Deconv1 - Deconv2 - Deconv3.
[0057] The training loss function of the mapping matrix network structure is:
[0058]
[0059]
[0060]
[0061] in, This is due to channel loss; For reconstruction losses; This is a dynamic weighting coefficient used to balance the magnitude of channel loss and reconstruction loss; This represents the k-th discrete channel value predicted by the spectral reconstruction model. This represents the theoretical value of the k-th discrete channel of the target instrument. The value of the reconstructed high-dimensional spectrum at the p-th wavelength point; This represents the true value of the original high-dimensional continuous spectrum acquired by a high-precision spectrometer at the p-th wavelength point.
[0062] The construction of the content prediction model includes:
[0063] High-dimensional continuous spectra of samples are acquired using high-precision spectrometers, and reconstructed values of discrete channels are output through a spectral reconstruction model.
[0064] The reconstructed spectral data is preprocessed and then divided into a calibration set, a validation set, and a test set.
[0065] The content prediction model network design includes a channel attention module, a multi-scale CNN feature extraction module, a global feature compression module, and a fully connected prediction module;
[0066] Using the reconstructed near-infrared spectral data, a quantitative concentration prediction model is obtained by training the prediction model network structure while minimizing the RMSE loss function.
[0067] The network structure of the quantitative concentration prediction model is as follows:
[0068] GlobalAvgPool - Dense1 - Dense2 - Multiply - Conv5 - BN5 - AvgPool5 (Conv3 - BN3-AvgPool3) - Concat - GELU - GlobalMaxPool - GlobalAvgPool2 - Flatten - Dense3 - Dropout - Output.
[0069] The multi-objective constraint loss function is as follows:
[0070]
[0071] in: The equilibrium coefficients are of the order of magnitude, due to the spectral reconstruction loss function. With the loss function of the prediction model The magnitudes of the two may differ, therefore a coefficient is used. To achieve magnitude balance;
[0072] , For the k-th sample, This represents the predicted value of the physical quantity for the k-th sample output by the prediction model.
[0073] In addition, corresponding to the aforementioned embodiments of the quantitative prediction model construction method for spectral reconstruction, the present invention also provides embodiments of the apparatus for constructing a quantitative prediction model for spectral reconstruction.
[0074] The quantitative prediction model construction device 30 for spectral reconstruction provided in this embodiment includes:
[0075] The spectral reconstruction modeling module 31 is used to build a spectral reconstruction model based on pure standard sample data, while ensuring the accuracy of spectral reconstruction, so as to realize the spectral reconstruction of high-dimensional continuous spectrum to low-dimensional discrete data.
[0076] Content prediction modeling module 32 is used to build a content prediction model based on measured sample data while ensuring the accuracy of model prediction.
[0077] The joint optimization and deployment module 33 is used to combine the accuracy of spectral reconstruction with the accuracy of model prediction, and to perform multi-objective constrained loss function collaborative optimization to obtain a complete set of calibrated spectral-channel response weights and prediction models.
[0078] Corresponding to the aforementioned embodiments of the quantitative prediction model construction method for spectral reconstruction, this embodiment of the invention also provides a computer-readable storage medium storing at least one piece of program code, which is loaded and executed by a processor to implement the quantitative prediction model construction method for spectral reconstruction as described in the above embodiments.
[0079] The computer-readable storage medium can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be any data processing device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units of any data processing device and external storage devices. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.
[0080] More specifically, the present invention provides more detailed embodiments, as follows: A method for collaborative optimization of spectral reconstruction and quantitative prediction based on convolutional neural networks:
[0081] (1) Spectral reconstruction model based on pure standard sample data
[0082] ① Construction of standard dataset
[0083] Considering the inter-instrumental differences between different instruments and the difficulty in reproducing the measured position of the sample, a dataset of pure substances with wide-band absorption or reflection characteristics was selected to establish the spectral reconstruction model. High-dimensional continuous spectra were acquired using high-precision spectroscopic sensors. (Where n is the number of wavelength points and K is the number of standard samples), "physical tags" are generated through a mathematical model. The response function of each discrete channel of the sensor conforms to a Gaussian distribution. Then, the theoretical discrete channel values of the target spectral reconstruction instrument are obtained according to the Gaussian formula. (where m is the number of channels), the formula is:
[0084]
[0085] in: For the standard sample spectrum at the i-th wavelength point ( The absorbance value of ). For the j-th channel at wavelength The response weight at the point is given by the formula: ,in Let j be the center wavelength of the j-th channel. The half-width and height control parameters; To add random Gaussian noise
[0086] ② Mapping matrix network design
[0087] Design a mapping network structure as follows: Conv1 - Batch Norm1 - Conv2 – MaxPool - Conv3 – Flatten - Dense1 – Dropout - Dense2 – Reshape - Deconv1 - Deconv2 - Deconv3.
[0088] a) The 1D-CNN feature extraction module extracts local and global features from an n-dimensional continuous spectrum to provide a high-dimensional feature representation for mapping matrix learning, and outputs multi-scale feature maps. This includes Conv1 - Batch Norm1 - Conv2 - MaxPool - Conv3 layers. The type, parameter settings, input dimensions, and output dimensions of each layer are shown in the table below:
[0089]
[0090] b. Mapping matrix module, which maps multi-scale feature maps The data is compressed to j-channel values, and a mapping matrix from the n-dimensional spectrum to the j-channel is learned, which contains Flatten - Dense1 - Dropout - Dense2 layers. The type, parameter settings, input dimension, and output dimension of each layer are shown in the table below:
[0091]
[0092] c. Deconvolution reconstruction module, which converts the j-channel predicted values... Reconstructing an n-dimensional spectrum using constrained deconvolution with the MSE loss function. This includes Reshape, Deconv1, Deconv2, and Deconv3 layers. The type, parameter settings, input dimensions, and output dimensions of each layer are shown in the table below:
[0093]
[0094] Loss through the channel With reconstruction losses Joint optimization balances the prediction accuracy of the j-channel with the spectral physical properties, and backpropagation updates the network parameters to obtain the mapping matrix. The loss function formula is as follows:
[0095]
[0096]
[0097]
[0098] in, This is due to channel loss; For reconstruction losses; This is a dynamic weighting coefficient used to balance the magnitude of channel loss and reconstruction loss; This represents the k-th discrete channel value predicted by the spectral reconstruction model. This represents the theoretical value of the k-th discrete channel of the target instrument. The value of the reconstructed high-dimensional spectrum at the p-th wavelength point; This represents the true value of the original high-dimensional continuous spectrum acquired by a high-precision spectrometer at the p-th wavelength point.
[0099] (2) Method for constructing prediction models based on measured sample data
[0100] ① Use high-precision spectroscopic sensors to acquire high-dimensional continuous spectra of samples.
[0101] ② The input is fed into the pre-trained spectral reconstruction model based on pure standard sample data (model 2), and the reconstructed value of channel j is output. Determining real physical quantities using laboratory methods As a supervisory label, the data of channel j is physically enhanced through channel noise simulation.
[0102] ③ Perform spectral preprocessing on the reconstructed spectral data.
[0103] ④ Divide the processed spectral data into a calibration set, a validation set, and a test set.
[0104] ⑤ Predictive model network design
[0105] The network structure of the prediction model is designed as follows: GlobalAvgPool - Dense1 - Dense2 –Multiply - Conv5 - BN5 - AvgPool5 (Conv3 - BN3-AvgPool3) - Concat – GELU –GlobalMaxPool – GlobalAvgPool2 – Flatten - Dense3 – Dropout – Output.
[0106] In principle, this prediction model is constructed while minimizing the RMSE loss function. The formula for the RMSE loss function is as follows:
[0107]
[0108] in: For the k-th sample, This represents the predicted value of the physical quantity for the k-th sample output by the prediction model.
[0109] a) Channel Attention Module: Adaptively learns the importance weights of j channels to enhance the contribution of key channels to prediction. This module includes the GlobalAvgPool-Dense1-Dense2-Multiply layer. The type, parameter settings, input dimensions, and output dimensions of each layer are shown in the table below.
[0110]
[0111] b. Multi-scale CNN feature extraction module, which captures local correlation features between channels through convolutions at different scales, includes Conv5 - BN5 - AvgPool5 (Conv3 - BN3 - AvgPool3) layers. The type, parameter settings, input dimension, and output dimension of each layer are shown in the table below:
[0112] Large convolution kernel:
[0113]
[0114] small convolution kernel
[0115]
[0116] c. The global feature compression module converts the two-dimensional feature map into a one-dimensional vector, which includes the Concat - GELU - GlobalMaxPool - GlobalAvgPool2 - Flatten layers. The type, parameter settings, input dimension, and output dimension of each layer are shown in the table below:
[0117]
[0118] d. The fully connected prediction module maps high-dimensional feature vectors to predicted physical quantity values. It includes a Dense3-Dropout-Output layer that outputs the predicted physical quantity values of the actual samples through the prediction model. The type, parameter settings, input dimensions, and output dimensions of each layer are shown in the table below:
[0119]
[0120] (3) Collaborative optimization mechanism of multi-objective loss function
[0121] Minimizing only the loss function for spectral reconstruction might result in poor model generalization ability; conversely, minimizing only the loss function for prediction accuracy might lead to low prediction accuracy. Therefore, a balance between spectral reconstruction accuracy and prediction accuracy is needed. This requires constructing a complete loss function for co-optimization, ultimately outputting a high-accuracy spectral reconstruction matrix network and a high-precision prediction model. Considering the potential difference in magnitude between the two, a single coefficient is used... To achieve magnitude balance, the new loss function formula is as follows:
[0122]
[0123] The present invention further provides the following practical examples: achieving accurate mapping of spectral data from the AIOX2000 high-precision spectrometer to the AS7265X low-cost sensor.
[0124] 1. Spectral reconstruction model based on pure standard sample data
[0125] (1) Construction of standard dataset
[0126] ① Standard selection: Absorption standard: pure water and glucose solutions with concentrations of 5%, 10%, 20%, 30%, 40%, and 50%; Reflection standard: barium sulfate solution and polytetrafluoroethylene.
[0127] ② Spectral Acquisition: High-dimensional continuous full-spectrum data of the standard samples were acquired using an AIOX2000 spectrometer. Each standard sample was acquired 16 times, and the average was taken as the raw spectrum. Based on the channel response function of AS7265X, calculate the theoretical values for the corresponding 18 channels (where the wavelengths corresponding to the 18 channels are 350, 374, 398, 422, 446, 470, 494, 518, 542, 566, 590, 614, 638, 662, 686, 710, 734, 758, 782, 806, 830, 854, 878, 902, 926, 950, 974, and 998).
[0128] (2) Training and validation of the spectral reconstruction model
[0129] The network structure is Conv1(64,k=7) → BatchNorm → Conv2(128,k=5) → MaxPool → Conv3(36,k=3) → Flatten → Dense1(72,GELU) → Dropout(0.3) → Dense2(18,linear) → Reshape → Deconv1(36,k=3) → Deconv2(128,k=5) → Deconv3(64,k=7). The network is trained using joint channel loss and reconstruction loss, with dynamic weights α=0.6.
[0130] 2. Construction of Potato Starch Prediction Model
[0131] (1) Preparation of experimental data
[0132] ① Use an AIOX2000 spectrometer to collect spectral data and starch content of 100 potatoes. The spectra were measured at the equatorial position of the potato, and the average of ten measurements was taken as the spectral data of the potato sample.
[0133] ② Input the collected spectral data into the spectral reconstruction model and output 18-channel reconstruction values.
[0134] (2) Prediction model design
[0135] The network structure is GlobalAvgPool → Dense1(9,ReLU) → Dense2(18,Sigmoid) →Multiply (channel attention) → [Conv5(k=5)+Conv3(k=3)] → Concat → GELU →GlobalMaxPool → Flatten → Dense3(4,GELU) → Dropout(0.3) → Output(1,linear). The network is trained in conjunction with the loss function of spectral reconstruction and model prediction. Finally, a set of weight matrices that map high-dimensional continuous spectrum to low-dimensional discrete channels and a high-precision prediction model of potato starch content by low-cost spectral sensors are obtained.
[0136] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only.
[0137] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for constructing a quantitative prediction model for spectral reconstruction, characterized in that: include: Based on pure standard sample data, a spectral reconstruction model is constructed while ensuring the accuracy of spectral reconstruction, so as to realize the spectral reconstruction of high-dimensional continuous spectrum to low-dimensional discrete data. Based on measured sample data, a content prediction model is constructed while ensuring the accuracy of model prediction. By combining the accuracy of spectral reconstruction with the model prediction precision, a multi-objective constrained loss function is co-optimized to obtain a complete set of calibrated spectral-channel response weights and a prediction model.
2. The method for constructing a quantitative prediction model for spectral reconstruction according to claim 1, characterized in that: The construction of the spectral reconstruction model includes: The standard dataset is constructed by selecting pure substances with wide-band absorption or reflection characteristics as standards; The design incorporates a mapping matrix network structure, including a 1D-CNN feature extraction module, a mapping matrix module, and a deconvolution reconstruction module. The network parameters are optimized using channel loss and reconstruction loss to obtain a spectral reconstruction model.
3. The method for constructing a quantitative prediction model for spectral reconstruction according to claim 2, characterized in that: Pure material standards include absorption standards, reflection standards, scattering standards, and blank / baseline standards.
4. The method for constructing a quantitative prediction model for spectral reconstruction according to claim 2, characterized in that: The mapping matrix network structure is as follows: Conv1 - Batch Norm1 - Conv2 - MaxPool - Conv3 - Flatten - Dense1 - Dropout - Dense2 - Reshape - Deconv1 - Deconv2 - Deconv3.
5. The method for constructing a quantitative prediction model for spectral reconstruction according to claim 2, characterized in that: The training loss function for the mapping matrix network structure is: in, This is due to channel loss; For reconstruction losses; This is a dynamic weighting coefficient used to balance the magnitude of channel loss and reconstruction loss; This represents the k-th discrete channel value predicted by the spectral reconstruction model. This represents the theoretical value of the k-th discrete channel of the target instrument. The value of the reconstructed high-dimensional spectrum at the p-th wavelength point; This represents the true value of the original high-dimensional continuous spectrum acquired by a high-precision spectrometer at the p-th wavelength point.
6. The method for constructing a quantitative prediction model for spectral reconstruction according to claim 1, characterized in that: The construction of the content prediction model includes: High-dimensional continuous spectra of samples are acquired using high-precision spectrometers, and reconstructed values of discrete channels are output through a spectral reconstruction model. The reconstructed spectral data is preprocessed and then divided into a calibration set, a validation set, and a test set. The content prediction model network design includes a channel attention module, a multi-scale CNN feature extraction module, a global feature compression module, and a fully connected prediction module; Using the reconstructed near-infrared spectral data, a quantitative concentration prediction model is obtained by training the prediction model network structure while minimizing the RMSE loss function.
7. The method for constructing a quantitative prediction model for spectral reconstruction according to claim 6, characterized in that: The network structure of the quantitative concentration prediction model is as follows: GlobalAvgPool - Dense1 - Dense2 - Multiply - Conv5 - BN5 - AvgPool5 (Conv3- BN3-AvgPool3) - Concat - GELU - GlobalMaxPool - GlobalAvgPool2 - Flatten -Dense3 - Dropout - Output.
8. The method for constructing a quantitative prediction model for spectral reconstruction according to claim 1, characterized in that: The multi-objective constraint loss function is: in: The equilibrium coefficients are of the order of magnitude, due to the spectral reconstruction loss function. With the loss function of the prediction model The magnitudes of the two may differ, therefore a coefficient is used. To achieve magnitude balance; , For the k-th sample, This represents the predicted value of the physical quantity for the k-th sample output by the prediction model.
9. A device for constructing a quantitative prediction model for spectral reconstruction, characterized in that: The apparatus for constructing a quantitative prediction model for spectral reconstruction according to any one of claims 1-8 comprises: The spectral reconstruction modeling module is used to build a spectral reconstruction model based on pure standard sample data, while ensuring the accuracy of spectral reconstruction, so as to realize the spectral reconstruction of high-dimensional continuous spectrum to low-dimensional discrete data. The content prediction modeling module is used to build a content prediction model based on measured sample data while ensuring the accuracy of the model prediction. The joint optimization and deployment module is used to combine the accuracy of spectral reconstruction with the accuracy of model prediction, and to perform multi-objective constrained loss function collaborative optimization to obtain a complete set of calibrated spectral-channel response weights and prediction models.
10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores at least one piece of program code, which is loaded and executed by a processor to implement the quantitative prediction model construction method for spectral reconstruction as described in any one of claims 1 to 8.