Infrared spectrum atmospheric absorption interference correction method and system

Abstract

This invention discloses a method and system for correcting atmospheric absorption interference in infrared spectroscopy. The method includes: establishing a historical spectral database; selecting nearest-neighbor reference base spectra from the historical spectral database based on the sample to be corrected; setting correction branches: a reference spectrum generation branch generates a reference spectrum for the sample to be corrected based on the sample to be corrected, the nearest-neighbor reference base spectra, and environmental and instrument parameters; a physical correction branch constructs differential interference components for the sample to be corrected based on a time-series spectral window, and performs physical constraint correction using multispectral robust estimation and adaptive piecewise scaling factors to obtain the physically corrected spectrum; a deep learning branch performs deep residual learning on the sample to be corrected, the reference spectrum, and the physically corrected spectrum to obtain a deep corrected spectrum; and adaptively fusion of the corrected spectra based on the uncertainty or consistency of the correction branches to obtain the final corrected spectrum. This invention achieves more refined and robust correction of atmospheric absorption interference in infrared spectroscopy.

Description

Technical Field

[0001] This invention relates to the field of infrared spectral data processing technology, specifically to a method and system for correcting atmospheric absorption interference in infrared spectroscopy. Background Technology

[0002] In infrared spectroscopy measurements, background spectra are typically acquired first, followed by sample spectra, with background subtraction used to reduce the influence of environmental absorption. However, in practical applications, background and sample measurements are not always perfectly synchronized. Especially under conditions such as in the measurement chamber, open optical paths, or in-situ online environments, water vapor and carbon dioxide concentrations change over time, resulting in residual fine atmospheric absorption structures remaining in the sample spectrum. These residual peaks are usually distributed within several characteristic absorption regions, exhibiting sharp peak shapes and rapid changes, directly impacting subsequent qualitative identification, quantitative modeling, characteristic peak analysis, and effective band selection.

[0003] Existing methods fall into two categories: one relies on direct subtraction using background or standard gas spectra, while the other constructs a difference spectrum from two adjacent sample spectra and then calculates the scaling factor by minimizing the point-to-point spectral length to correct for water vapor or carbon dioxide interference. These methods are simple in concept and have some effectiveness, but they still have some significant shortcomings:

[0004] 1. It relies only on a small number of adjacent spectra, and historical sample information is not utilized; 2. Most methods use a single scaling factor, which makes it difficult to adapt to changes in the interference spectrum in different wavenumber bands; 3. When the sample itself contains a true strong peak within the calibration region, overcalibration or undercalibration is likely to occur; 4. Insufficient adaptability across instruments, resolutions, and scenarios; 5. Limited ability to handle nonlinear residuals, complex drifts, and unknown samples; 6. The calibration results lack multi-source verification and reliability closed-loop.

[0005] In long-term applications, a large amount of historical sample spectra, instrument parameters, and environmental information has often been accumulated. If this historical data can be effectively utilized, it is possible to construct a reference spectrum that more closely approximates the actual sample state for the current sample to be tested. Meanwhile, deep learning also has advantages in spectral feature extraction, similar sample retrieval, conditional generation, and residual correction. If historical sample reference, physically interpretable correction, and deep learning compensation can be unified, the correction effect can be further improved while maintaining interpretability.

[0006] Therefore, it is necessary to propose a new technical solution that, based on existing multispectral robust estimation and adaptive piecewise optimization, further introduces historical data to generate reference spectra and deep learning residual compensation, thereby forming a more complete and robust infrared spectral atmospheric absorption correction method. Summary of the Invention

[0007] The technical problem to be solved by this invention is that existing methods usually only rely on the current sample and local time-series spectra for correction, which does not make sufficient use of historical knowledge and does not fully consider the differences in conditions such as instrument number, resolution, number of scans, temperature, humidity and pressure, resulting in insufficient robustness of the method when applied across instruments and scenarios.

[0008] To solve the above-mentioned technical problems, the present invention provides the following technical solution: An infrared spectral atmospheric absorption interference correction method includes: Establish a historical spectral database; based on the sample to be calibrated, filter from the historical spectral database to obtain the nearest neighbor reference spectrum; The calibration branches are set up as follows: a reference spectrum generation branch, a physical calibration branch, and a deep learning branch. The reference spectrum generation branch generates a reference spectrum for the sample to be calibrated based on the sample to be calibrated, the nearest neighbor reference spectrum, and environmental and instrument parameters. The physical calibration branch constructs differential interference components for the sample to be calibrated based on a time-series spectral window and performs physical constraint calibration using multispectral robust estimation and adaptive piecewise scaling factors to obtain the physically calibrated spectrum. The deep learning branch performs deep residual learning on the sample to be calibrated, the reference spectrum, and the physically calibrated spectrum to obtain the deep-calibrated spectrum. Based on the uncertainty or consistency of the correction branches, the corrected spectra are adaptively fused to obtain the final corrected spectrum.

[0009] Technical benefits: Historical sample spectra are not just used for classification or comparison, but are directly used to generate the reference spectrum of the current sample. The reference spectrum does not directly replace the measured spectrum, but serves as a correction constraint and fusion reference. The physical correction branch is responsible for removing most of the interpretable atmospheric interference, while the deep learning branch is responsible for compensating for nonlinear residuals that are difficult to model. The final corrected spectrum is not a fixed weighted spectrum, but is adaptively fused based on local uncertainty, self-consistency, and reference consistency.

[0010] Furthermore, obtaining the nearest neighbor reference spectrum includes: Process the historical sample spectra in the historical spectral database to match the dimensions of the sample to be corrected; The spectra of the sample to be corrected and the historical samples of the same dimension are used to perform feature encoding based on a three-level one-dimensional convolutional residual coding network to obtain the current sample code and the historical sample code. Cosine similarity is calculated for the current sample code and historical sample codes. All historical sample spectra are then sorted from highest to lowest similarity, and the top-ranked spectra are selected. The spectrum is used as a set of nearest-neighbor historical samples; The nearest neighbor reference spectrum is obtained by weighting the set of nearby historical samples.

[0011] Furthermore, the step of generating the reference spectrum of the sample to be calibrated employs a retrieval-enhanced conditional coding-decoding network; The retrieval enhancement conditional encoding includes the current spectrum to be calibrated branch, the historical nearest neighbor weighted base spectrum branch, and the environmental and instrument parameters branch. These branches encode the sample to be calibrated, the nearest neighbor reference base spectrum, and the environmental and instrument parameters, respectively, to obtain feature vectors. The three feature vectors are then concatenated into a joint latent variable. The decoder uses a three-layer one-dimensional deconvolution structure to restore the latent variables to the reference spectrum.

[0012] Furthermore, the adaptive piecewise scaling factor is: The interference windows for water vapor and carbon dioxide are divided into several continuous sub-intervals, and a constant scaling factor is set for each sub-interval. The constant scaling factor is obtained by minimizing the following objective function: ; ; ; ; ; ; In the formula, For physical optimization loss, , , , , These are, respectively, point-to-point spectral length loss, residual atmospheric absorption loss, sample peak fidelity loss, boundary continuity loss, and reference consistency loss. , , , , These are the weights corresponding to point-to-point spectral length loss, residual atmospheric absorption loss, sample peak fidelity loss, boundary continuity loss, and reference consistency loss, respectively. , The first , Wavenumber points, , The first , The physically corrected spectral values ​​corresponding to each wavenumber point The total number of wavenumber points within the band. As an atmospheric interference mask, The Hadamard product is a point-by-point multiplication. wave number The corresponding physically corrected spectral values, It is the square of the L2 norm. As a mask for sample peaks, For the sample to be corrected, It is an L1 norm. To correct the set of interval boundaries, , These are the physically corrected spectral values ​​for the left and right sides of the boundary, respectively. This is the reference spectrum.

[0013] Furthermore, the physically corrected spectrum is obtained by applying the following formula: ; In the formula, wave number The corresponding physically corrected spectral values, For the sample to be corrected, , These are the piecewise scaling factor functions for water vapor and carbon dioxide, respectively. , These are the differential interference components corresponding to water vapor and carbon dioxide, respectively.

[0014] Furthermore, the sample peak mask is obtained through the following steps: For the current spectrum to be corrected Calculate the second derivative spectrum ; In the second derivative spectrum Find the peak center by locating the negative extreme value; For the peak center, candidate sample peak regions are formed by expanding based on the half-width at half-maximum threshold; then compared with the reference spectrum. The peak regions are intersected; The intersection region is defined as the true peak region of the sample. .

[0015] The final sample peak mask is defined as follows: .

[0016] Furthermore, obtaining depth-corrected spectra includes: The deep learning branch uses a one-dimensional U-Net residual compensation network. The encoding end uses three downsampling convolutional blocks, and the decoding end uses a three-layer upsampling structure corresponding to the encoding end. The upsampling result of each layer is spliced ​​or fused with the feature map of the same scale as the encoding end through skip connections, so as to preserve the sample peak shape, local details and boundary features while restoring the spectral resolution. Finally, two one-dimensional sequences are output: residual correction amount and depth branch uncertainty. Based on the residual correction, the local complex residuals that the physical correction branch failed to eliminate are compensated to obtain the deep correction spectrum.

[0017] Furthermore, the final corrected spectrum is obtained and expressed using the following formula: ; ; In the formula, b is the branch number, and r is the summation index. For the final calibrated spectrum, , , These are the weights for the physically corrected spectrum, the depth-corrected spectrum, and the reference spectrum, respectively. , , Wavenumber Reference spectrum, physically corrected spectrum, depth corrected spectrum, For the first The weights of the correction branches, It is an exponential function. For the first The correction branch at wavenumber Uncertainty at the location, For the first The correction branch at wavenumber The uncertainty at, where, In the formula, it is a constant. In the formula, it is a variable.

[0018] Furthermore, the calibration method also includes: generating a calibration confidence result based on a comprehensive index for the final calibration spectrum; and, based on the calibration confidence result, after the calibration spectrum result has passed quality control, reverting the final calibration spectrum, environmental and instrument parameters of the current calibration sample back to the historical spectral database for subsequent reference spectrum generation.

[0019] The present invention also provides a system for applying the above-described infrared spectral atmospheric absorption interference correction method, comprising: Database and similarity filtering module: used to establish a historical spectral database; based on the sample to be calibrated, it filters from the historical spectral database to obtain the nearest neighbor reference spectrum; The calibration branch module is used to set up calibration branches: a reference spectrum generation branch, a physical calibration branch, and a deep learning branch. The reference spectrum generation branch generates a reference spectrum for the sample to be calibrated based on the sample to be calibrated, the nearest neighbor reference spectrum, and environmental and instrument parameters. The physical calibration branch constructs differential interference components for the sample to be calibrated based on a time-series spectral window and performs physical constraint calibration using multispectral robust estimation and adaptive piecewise scaling factors to obtain the physically calibrated spectrum. The deep learning branch performs deep residual learning on the sample to be calibrated, the reference spectrum, and the physically calibrated spectrum to obtain the deep-calibrated spectrum. The adaptive fusion final correction module is used to adaptively fuse the corrected spectrum based on the uncertainty or consistency of the correction branches to obtain the final corrected spectrum.

[0020] Compared with the prior art, the beneficial effects of the present invention are: To address the problem that a single scaling factor is insufficient to adapt to changes in interference spectral shape across different wavenumber bands: Existing methods mostly employ a single scaling factor to uniformly correct the entire water vapor or carbon dioxide interference window. This is insufficient to adapt to the differences in interference spectral shapes across different wavenumber bands caused by varying environmental conditions, easily leading to local undercorrection or overcorrection. To address this issue, this invention employs a piecewise scaling factor joint correction scheme based on robust time-series window differentials in the physical correction branch. This invention enables different wavenumber bands to use correction parameters that better reflect local spectral shape variations, thereby improving adaptability to complex atmospheric interference spectral shape changes, reducing local residual peaks and local overcorrection, and enhancing the accuracy and stability of the physical correction branch correction results.

[0021] To address the issue of overcorrection or undercorrection when the true strong peak of a sample overlaps with atmospheric interference peaks: Existing methods, if a true strong peak exists within the correction window, it is easily misidentified as part of atmospheric interference, leading to weakening, distortion, or even abnormal negative peaks. This invention employs a combined scheme of sample peak masking, automatic window selection, and reference spectral constraints. During automatic window selection, candidate correction windows with excessive overlap with the true peak region are eliminated. The generated reference spectrum is directly incorporated into the sample peak fidelity and reference consistency terms, simultaneously constraining both physical and depth correction results. Through this scheme, the present invention effectively avoids the true peak region during atmospheric interference correction and uses the reference spectrum to constrain the sample structure, thereby reducing the risk of miscorrection when the true peak overlaps with atmospheric interference peaks, improving peak shape fidelity, and reducing abnormal negative peaks and local peak shape distortion.

[0022] To address the issue of insufficient adaptability across instruments, resolutions, and scenarios: existing methods typically rely solely on the current sample and local time-series spectra for correction, failing to adequately utilize historical knowledge and not fully considering differences in instrument number, resolution, scan count, temperature, humidity, and pressure, resulting in insufficient robustness when applied across instruments and scenarios. To solve this problem, this invention proposes a scheme for historical sample retrieval and conditional reference spectrum generation. The current sample spectrum, nearest-neighbor reference spectra, environmental parameters, and instrument parameters are all input into a conditional generation network to generate a reference spectrum matching the current state. This allows the invention to move beyond relying solely on current single-measurement information, incorporating historical sample knowledge, environmental conditions, and instrument conditions into the reference spectrum generation process. This results in a reference spectrum that more closely approximates the true state of the current sample under current operating conditions, thereby improving the adaptability and scalability of this invention across instruments, resolutions, and scenarios.

[0023] To address the limited processing capabilities for nonlinear residuals, complex drifts, and unknown samples: Existing physical correction methods are mainly suitable for interpretable atmospheric disturbance subtraction, but their processing capabilities are limited for complex background disturbances, local drifts, unknown samples, and nonlinear residuals that are difficult for physical models to fully describe, often leaving significant errors in local regions. To solve this problem, this invention proposes a scheme where a physical correction branch and a deep residual compensation branch work in tandem. This allows the invention to retain the interpretability of the physical branch while introducing a deep branch to handle nonlinear residuals and complex drifts, making the invention more adaptable to unknown samples, complex operating conditions, and local anomalies, thereby improving the completeness and robustness of the final correction results.

[0024] To address the lack of multi-source verification and reliability closed-loop in calibration results: Existing technologies often output calibration results via a single path, lacking quantitative evaluation of result reliability and cross-validation mechanisms between physical branches, deep branches, and reference branches. Therefore, it is difficult to objectively assess the credibility of the output results in complex applications. To solve this problem, this invention proposes an uncertainty-driven point-by-point fusion and confidence evaluation scheme. Based on indicators such as the intensity of residual water vapor / carbon dioxide characteristic peaks, noise changes in featureless regions, the number and amplitude of anomalous negative peaks, the consistency of multi-window scaling factors, the consistency between the final output spectrum and the reference spectrum, the degree of anomalousness of deep branch residuals, and the difference between the physical branch and deep branch results, the confidence of the final result is evaluated. Through this scheme, this invention no longer uses fixed-weight fusion but dynamically allocates weights based on the credibility of each branch in local bands, thereby improving the fusion robustness in complex regions. Simultaneously, by constructing a reliability closed loop through confidence evaluation, the output results not only have "values" but also "credibility explanations," making them more suitable for online monitoring, long-term operation, and engineering deployment.

[0025] To address the issue of insufficient continuous evolution capability after utilizing historical samples: existing technologies, even when incorporating historical samples, often only use online update mechanisms during offline phases. As application scenarios change, sample types increase, or instrument status changes, historical knowledge cannot continuously evolve, leading to a gradual decline in model adaptability. To solve this problem, this invention further proposes an online backflow update scheme after quality control, enabling the historical sample library to continuously expand and optimize as the system operates. This enhances the system's continuous adaptability to new scenarios, new samples, and new operating conditions, giving the entire system self-growing characteristics and making it more suitable for long-term engineering deployment.

[0026] In summary, this invention addresses the problems in existing technologies, such as the difficulty of adapting to complex spectral variations with a single scaling factor, the susceptibility to miscorrection when sample peaks overlap with atmospheric peaks, insufficient adaptability across instruments and scenarios, limited ability to handle nonlinear residuals and complex drift, and the lack of a reliable closed-loop correction result. It proposes several technical solutions, including joint correction with segmented scaling factors, sample peak masking and automatic window selection, historical sample retrieval and conditional reference spectrum generation, synergy between physical correction and deep residual compensation, uncertainty-driven fusion, and online updates. These solutions enable more refined, robust, and generalizable correction of atmospheric absorption interference in infrared spectra. Attached Figure Description

[0027] Figure 1 This is a flowchart of an infrared spectral atmospheric absorption interference correction method according to an embodiment of the present invention.

[0028] Figure 2 This is a schematic diagram showing the comparison of spectra before and after correction in an embodiment of the present invention. Detailed Implementation

[0029] To facilitate understanding of the technical solution of the present invention by those skilled in the art, the technical solution of the present invention will now be further described in conjunction with the accompanying drawings.

[0030] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0031] Please see Figure 1 As shown, the present invention provides a method for correcting atmospheric absorption interference in infrared spectroscopy, comprising: S10, Establish a historical spectral database. Based on the sample to be calibrated, filter from the historical spectral database to obtain the nearest neighbor reference spectrum.

[0032] In one embodiment of the present invention, S11, historical spectral database, is denoted as: ; In the formula, Characterized as a historical spectral database, For the first Historical sample spectra, For sample category and operating condition category, These are environmental parameters, including temperature, humidity, pressure, and data collection time. This includes instrument parameters such as resolution, number of scans, instrument serial number, and optical path configuration. Total historical sample size For wavenumber variables.

[0033] In one embodiment of the present invention, in order to organize historical data into a standard library that can be retrieved, generated and updated later, preprocessing such as baseline correction, wavenumber drift correction and abnormal spectrum removal is performed uniformly before historical samples are put into the library.

[0034] In this embodiment, S12, the current sample time series window is constructed: Continuous collection of current samples Sample spectrum: In the formula, For the first Sample spectrum.

[0035] Select the first from them The following are samples to be corrected: In the formula, For the sample to be corrected, For the first Sample spectrum.

[0036] Constructing a timing window: ; In the formula, For the current sample's time window, For the first Sample spectrum, For the first The sample spectrum, of which, All are constants. Generally, it is acceptable That is, a fixed seven-time window consisting of three spectra at the beginning and three at the end. or When this is the case, timing windows can be constructed using methods such as shortening the window, using a single-sided window, or mirroring. This represents the total number of spectra in the sample.

[0037] In this embodiment, S13, feature encoding and similarity retrieval of the spectra of the sample to be corrected and historical samples: First, the preprocessed samples to be corrected are resampled to a uniform wavenumber dimension, forming a one-dimensional vector of length L: ; In the formula, To characterize the sample to be calibrated, its length is A one-dimensional vector, Indicates the spectrum to be corrected in the th... L Spectral intensity values ​​at wavenumber points.

[0038] Historical sample spectra are uniformly processed into vectors of the same length as follows: ; In the formula, To characterize the spectrum of historical samples, its length is A one-dimensional vector, For the first The historical sample spectrum in the first L Spectral values ​​at wavenumber points.

[0039] The feature encoding network is fixed to a three-level one-dimensional convolutional residual encoding network, with the following specific structure: 1. First convolutional block, Conv1D, kernel size 7, output channels 32; 2. Second convolutional block: Conv1D, kernel size 5, output channels 64; 3. Third convolutional block: Conv1D, kernel size 3, output channels 128; 4. Each convolutional block adopts a structure of "convolution + batch normalization + GELU activation + residual connection"; 5. Add a channel attention weighting module after the third convolutional block; 6. Finally, after global average pooling and a fully connected layer, a 128-dimensional encoded vector is output.

[0040] Therefore, the current sample to be corrected is encoded as follows: ; In the formula, Encode the sample to be corrected. For the feature encoding network, the original spectrum is compressed into a low-dimensional vector to facilitate similar sample retrieval, where For encoding mapping functions, These are the model parameters for the feature encoding network. This represents the spectrum of the sample to be corrected input into the feature encoding network. It represents a 128-dimensional real vector space.

[0041] The spectral encoding of historical samples is as follows: ; In the formula, For the first Spectral coding of historical samples, For the first Historical sample spectra, For the first Historical sample spectra through The extracted low-dimensional vector.

[0042] Cosine similarity is calculated for the current sample code and historical sample codes. All historical sample spectra are then sorted from highest to lowest similarity, and the top-ranked spectra are selected. Each spectrum is a set of nearby historical samples.

[0043] To avoid accidental distortion of a single nearest neighbor spectrum, instead of directly using the most similar spectrum, the preceding spectrum is used. The nearest neighbor reference spectrum is obtained by weighting the similarity of the nearest neighbor historical samples: ; In the formula, As the nearest reference spectrum, The weights of the nearest historical samples, Sort by similarity. The original historical spectra corresponding to each nearest-neighbor historical sample. Among them, the weights... It can be obtained through the following formula: ; In the formula, For temperature coefficient, For the first Feature vectors of the nearest neighbor historical samples The feature vector of the sample to be corrected. For similarity function, It is an exponential function. For the first The feature vector corresponding to each nearest historical sample.

[0044] S20, set up calibration branches: a reference spectrum generation branch, a physical calibration branch, and a deep learning branch. The reference spectrum generation branch generates a reference spectrum for the sample to be calibrated based on the sample to be calibrated, the nearest neighbor reference spectrum, and environmental and instrument parameters. The physical calibration branch constructs differential interference components for the sample to be calibrated based on a time-series spectral window and performs physical constraint calibration using multispectral robust estimation and an adaptive piecewise scaling factor to obtain the physically calibrated spectrum. The deep learning branch performs deep residual learning on the sample to be calibrated, the reference spectrum, and the physically calibrated spectrum to obtain the deep-calibrated spectrum.

[0045] In one embodiment of the present invention, S21, in the reference spectrum generation branch: The step of generating the reference spectrum of the sample to be calibrated employs a retrieval-enhanced conditional coding-decoding network. The retrieval-enhanced conditional coding includes a branch for the current spectrum to be calibrated, a branch for historical nearest-neighbor weighted base spectra, and a branch for environmental and instrumental parameters. These branches encode the sample to be calibrated, the nearest-neighbor reference base spectra, and the environmental and instrumental parameters, respectively, to obtain feature vectors. Specifically, the environmental and instrumental parameters are uniformly represented as parameter vectors: ; In the formula, A parameter vector representing the environment and instruments. For temperature, Relative humidity, For pressure, For resolution, For the number of scans, This is the instrument number encoding value. The same convolutional encoder structure is used for the branch to be calibrated and the nearest neighbor reference spectrum branch, and the outputs are as follows: ; In the formula, For the current sample spectrum After encoder The extracted feature vectors, For the current sample spectrum The encoder's function is to convert the input spectrum... Extract it into a low-dimensional feature representation. Nearest neighbor reference spectrum After encoder The extracted feature vectors, The encoder, which is a nearest-neighbor reference spectrum branch, is used to... Extract it into a low-dimensional feature representation.

[0046] The parameter vectors of the environment and instruments are mapped to 32-dimensional conditional vectors through a two-layer fully connected network: ; In the formula, For parameter vectors After encoder The resulting conditional eigenvectors The encoder for the environmental and instrument parameters branch is specifically a fully connected network that maps conditional parameters to feature vectors. This is a parameter vector for the environment and instruments.

[0047] The three feature vectors are concatenated into a joint latent variable: ; In the formula, These are joint latent variables.

[0048] The decoder employs a three-layer one-dimensional deconvolution structure to restore the latent variables to the reference spectrum: ; In the formula, The resulting reference spectrum is generated by the model under the combined constraints of the current sample, historical neighbor information, and environmental instrument conditions. For the reference spectrum generation network, where This represents the parameters of the generator network.

[0049] The reference spectrum generation network is trained using the following loss function: ; in: ; ; ; ; In the formula, The overall loss function of the reference spectrum generation network is... , , , These are reconstruction loss, smoothing loss, peak fidelity loss, and conditional consistency loss, respectively. , , , These are the weights corresponding to reconstruction loss, smoothing loss, peak fidelity loss, and conditional consistency loss, respectively. For reference spectrum, As a reference to the true spectrum, it serves as the standard reference spectrum for the target during training. The total number of wavenumber points after uniform resampling of the spectrum; , The first , The reference spectrum corresponding to each wavenumber point As a mask for sample peaks, The Hadamard product is a point-by-point multiplication. It is an L1 norm. It is the square of the L2 norm. Encoders for the environmental and instrument parameters branch, For the parameter vectors of the environment and instruments, for Feature representation after parametric encoder This is the vector of estimated parameters, reconstructed parameters, and matching parameters corresponding to the generated results.

[0050] The generated reference spectrum does not directly replace the current spectrum, but is used for sample peak mask generation, reference consistency constraints, and final result fusion.

[0051] In one embodiment of the present invention, S22, physical correction branch, multispectral robustness difference and interference component extraction: Constructing a difference spectrum based on a time-series window: ; In the formula, For the first Bar difference spectrum, representing the current spectrum to be corrected relative to the time window. The Middle Between sample spectra, in wavenumber The difference at the point, For the first Sample spectrum, This is the current spectrum to be corrected. When constructing the difference spectrum, the current sample to be corrected is removed. It itself only adopts The differential spectra of adjacent samples are constructed.

[0052] Taking the median of multiple difference spectra yields a robust difference spectrum: ; In the formula, For robust difference spectrum, The median operator is used for all difference spectra. Find the median of the values ​​at a given wavenumber point.

[0053] Then, the corresponding interference components were extracted from the preset absorption windows for water vapor and carbon dioxide, respectively. ; ; In the formula, , These are the differential interference components corresponding to water vapor and carbon dioxide, respectively. , These are water vapor and carbon dioxide interference masks, respectively. This is the Hadamard product of point-by-point multiplication.

[0054] In the physical correction branch optimization, a piecewise scaling factor joint correction model is used. First, the interference windows for water vapor and carbon dioxide are divided into several continuous sub-intervals, and a constant scaling factor is set for each sub-interval. ; ; In the formula, , These are the piecewise scaling factor functions for water vapor and carbon dioxide, respectively. , They are the first two components of water vapor and carbon dioxide, respectively. The constant scaling factor for each sub-interval, , They are water vapor and carbon dioxide, respectively. Sub-intervals.

[0055] The physically corrected spectrum was obtained using the following physical correction model: ; In the formula, wave number The corresponding physically corrected spectral values, For the sample to be corrected, , These are the piecewise scaling factor functions for water vapor and carbon dioxide, respectively. , These are the differential interference components corresponding to water vapor and carbon dioxide, respectively.

[0056] The constant scaling factor is obtained by minimizing the following objective function: ; ; ; ; ; ; In the formula, For physical optimization loss, , , , , These are, respectively, point-to-point spectral length loss, residual atmospheric absorption loss, sample peak fidelity loss, boundary continuity loss, and reference consistency loss. , , , , These are the weights corresponding to point-to-point spectral length loss, residual atmospheric absorption loss, sample peak fidelity loss, boundary continuity loss, and reference consistency loss, respectively. , The first , Wavenumber points, , The first , The physically corrected spectral values ​​corresponding to each wavenumber point The total number of wavenumber points within the band. The atmospheric interference mask is formed by combining water vapor and carbon dioxide interference masks to obtain the overall atmospheric interference mask. The Hadamard product is a point-by-point multiplication. wave number The corresponding physically corrected spectral values, It is the square of the L2 norm. As a mask for sample peaks, For the sample to be corrected, It is an L1 norm. To correct the set of interval boundaries, , These are the physically corrected spectral values ​​for the left and right sides of the boundary, respectively. This is the reference spectrum.

[0057] This optimization problem is solved using the coordinate descent method: fixed renew ,fixed renew Alternating iterations guide the convergence of the objective function.

[0058] In one embodiment of the present invention, the sample peak mask is further obtained through the following steps: For the current spectrum to be corrected Calculate the second derivative spectrum ; In the second derivative spectrum Find the peak center by locating the negative extreme value; For the peak center, candidate sample peak regions are formed by expanding based on the half-width threshold; Then compare with the reference spectrum The peak regions are intersected; The intersection region is defined as the true peak region of the sample. .

[0059] The final sample peak mask is defined as follows: ; In one embodiment of the present invention, a sub-window for optimizing the scaling factor is obtained through automatic window selection. Obtaining the sub-window for optimizing the scaling factor through automatic window selection specifically includes: Candidate sub-intervals are generated within preset windows for water vapor and carbon dioxide; Delete the region that is the actual peak of the sample. Candidate intervals with an overlap ratio exceeding a threshold; The remaining interval is reserved for scaling factor optimization.

[0060] In this embodiment, automatic window selection refers to the system automatically selecting suitable sub-windows for correction calculation from the atmospheric interference area after completing the sample peak masking. Knowing which regions represent the true sample peaks and should be avoided for atmospheric correction, the system then automatically filters out regions corresponding to the true sample peaks from the preset interference windows for water vapor and carbon dioxide. With less overlap, it is more suitable for estimating scaling factors and for entering wavenumber sub-intervals in physical optimization solutions.

[0061] In one embodiment of the present invention, S23, the deep learning branch and the physical correction branch can remove most of the interpretable atmospheric interference, but local undercorrection, overcorrection, and complex nonlinear residuals may still exist. Therefore, a deep residual branch is introduced to learn the mapping from the current spectrum to be corrected, the reference spectrum, the physically corrected spectrum to the residual correction amount. The deep learning branch always uses a one-dimensional U-Net residual compensation network. The network input is a three-channel sequence: In the formula, This is the input sequence for the deep learning branch.

[0062] The encoding layer uses three downsampling convolutional blocks. Specifically, the structure of each downsampling convolutional block is as follows: the fourth convolutional block... +Batch Normalization + Activation + Fifth Convolutional Block +Residual connection +Downsampling, the number of channels in the three layers are set to 64, 128, and 256 respectively.

[0063] The decoding end employs a three-layer upsampling structure corresponding to the encoding end's layer. The upsampling results of each layer are concatenated or fused with feature maps of the same scale as the encoding end via skip connections. This aims to preserve sample peak shapes, local details, and boundary features while restoring spectral resolution. The final output consists of two one-dimensional sequences: residual correction amount... and deep branch uncertainty .

[0064] ; In the formula, This refers to the neural network model corresponding to the deep learning branch. This is the input sequence for the deep learning branch.

[0065] The correction result output by the deep learning branch is as follows: ; In the formula, For depth-corrected spectra, This is the spectrum after physical correction.

[0066] The training loss function for the deep learning branch is: ; in: ; ; ; In the formula, To train the total loss function for the deep learning branch, , , These are residual reconstruction loss, uncertainty constraint loss, and peak fidelity loss, respectively. , , These are the weights of the residual reconstruction loss, uncertainty constraint loss, and peak fidelity loss, respectively. For depth-corrected spectra, For the ground truth spectrum during the training phase, The pointwise uncertainty of the output for the deep learning branch. For positive numbers that are infinitely close to 0, For logarithms, As a mask for sample peaks, The Hadamard product is a point-by-point multiplication. This serves as the reference spectrum. Through the correction of this branch, the deep network can correct the remaining local complex residuals of the physical branches.

[0067] In an embodiment of the present invention, a multi-objective loss function and joint constraints are constructed: based on the multi-objective optimization approach, a reference consistency term and a deep learning branch constraint term are added, and the final optimization adopts a unified objective function. ; The multi-window consistency item is as follows: ;

[0068] In the formula, To optimize the total loss for multiple objectives, For the consistency loss term of multi-window scaling factor, , , These represent the weights of the physics optimization loss, the total training loss function for the deep learning branch, and the consistency loss term for the multi-window scaling factor, respectively. For the first The set of scaling factors obtained from candidate windows. for The average value, This represents the number of valid candidate windows.

[0069] S30, based on the uncertainty or consistency of the correction branch, adaptively fuse the corrected spectrum to obtain the final corrected spectrum.

[0070] In one embodiment of the present invention, the final corrected spectrum is not a fixed-ratio weighted average, but rather an adaptive fusion of the uncertainties of the reference spectrum generation branch, the physical correction branch, and the deep learning branch.

[0071] Among them, the uncertainty of the physical correction branch It mainly consists of three parts: residual atmospheric peak intensity, multi-window consistency, and boundary jump degree.

[0072] ; In the formula, For the uncertainty of the physical correction branch, , , These are the normalized residual atmospheric term, the multi-window consistency term, and the boundary jump term, respectively. , , These are the weights of the normalized residual atmospheric term, the multi-window consistency term, and the boundary jump term, respectively, and they satisfy... + + In other words, at which locations are the residual atmospheric peaks stronger, the multi-window results more inconsistent, and the boundary jumps more pronounced, in the physical correction branch? It will be even bigger.

[0073] First, the residual atmospheric term of the physical correction branch is obtained using the following formula: ; In the formula, For the residual atmospheric term, For masking atmospheric interference areas, These are the spectral values ​​after physical correction.

[0074] Secondly, if the physical correction results are obtained separately under multiple candidate windows... Multi-window consistency is defined as follows: ; ;

[0075] In the formula, For point-to-point multi-window consistency items, Scalar indicators used in confidence assessment Indicates the first Candidate windows, For the first The physically corrected spectrum obtained from candidate windows To represent the average of the physical correction spectra obtained from multiple candidate windows, This represents the number of valid candidate windows.

[0076] Finally, define the boundary jump terms: ; In the formula, For boundary jump terms, , These are the physically corrected spectral values ​​for the left and right sides of the boundary, respectively. To correct the set of interval boundaries.

[0077] In this embodiment, the uncertainty of the deep learning branch As already given above, it will not be repeated here.

[0078] In this embodiment, the uncertainty of the reference spectrum generation branch The weighted variance of the nearest-neighbor reference spectrum at that wavenumber position is given. The reference bifurcation uncertainty is defined as: ; In the formula, Uncertainty in generating branches for the reference spectrum The weights of the nearest historical samples, Sort by similarity. The original historical spectrum corresponding to each nearest neighbor historical sample The nearest neighbor reference spectrum. That is, if... If the nearest neighbor reference base spectra differ significantly from each other at a certain wavenumber position, it indicates that the reference spectrum branching at that position is not stable enough. It will be even bigger.

[0079] Correction branch fusion method: Input the uncertainties of the three correction branches into the softmax weight allocation function: ; Corrected spectrum

[0080]

[0081] In the formula, For the final calibrated spectrum, , , These are the weights for the physically corrected spectrum, the depth-corrected spectrum, and the reference spectrum, respectively. , , Wavenumber Reference spectrum, physically corrected spectrum, depth corrected spectrum, It is an exponential function. For the first The correction branch at wavenumber Uncertainty at the location, For the first The correction branch at wavenumber The uncertainty at, where, In the formula, it is a constant. In the formula, it is a variable. For the first The weights of the correction branches. And, These are the uncertainties of the physical correction branch, the deep learning branch, and the reference spectrum generation branch, respectively.

[0082] In one embodiment of the present invention, the calibration method further includes: S40, generating a calibration confidence result based on a comprehensive index for the final calibration spectrum. And, based on the calibration confidence result, after the calibration spectrum result passes quality control, the final calibration spectrum, environmental and instrument parameters of the current calibration sample are fed back into the historical spectral database for subsequent reference spectrum generation.

[0083] In this embodiment, after correction, a confidence score can be generated by combining the following indicators: 1. Residual water vapor / carbon dioxide characteristic peak intensity: Under atmospheric masking, the residual energy of the final corrected spectrum can be expressed as: ; In the formula, The characteristic peak intensity of residual water vapor / carbon dioxide, This is the final calibrated spectrum.

[0084] 2. Noise Change in Feature Region: This indicates whether the correction has made the originally relatively flat, featureless region noisier. Select a featureless region mask and calculate the standard deviation of noise before and after correction. ; ; ; In the formula, To correct the standard deviation of noise in the featureless region before correction, It is a function of standard deviation; For featureless regions, For the sample to be corrected, The standard deviation of noise in the featureless region after correction; This is an indicator of noise variation in the featureless region.

[0085] 3. Number and amplitude of abnormal negative peaks: This indicates that if a significant downward false peak appears after correction, it signifies overcorrection. Let the negative peak threshold be: ; in, The negative peak threshold, This is for noise bias in featureless regions. 1 is generally taken as 2~5. Count the number of negative peaks that meet the conditions: ; In the formula, The number of negative peaks that meet the conditions is determined. The amplitude of the largest negative peak is also calculated. ; In the formula, The maximum negative peak amplitude, The function is used to maximize the value. Further synthesis yields the abnormal negative peak index: ; In the formula, This is an abnormally negative peak indicator. , These are the weighting coefficients for the negative peak threshold and the maximum negative peak amplitude, respectively.

[0086] 4. Consistency of multi-window scaling factors; This factor indicates whether the scaling factors calculated by different candidate windows are stable and consistent, that is, the multi-window consistency item is used as the consistency index of multi-window scaling factors.

[0087] 5. Consistency between the final output spectrum and the reference spectrum: This indicates the consistency between the final output spectrum and the reference spectrum, checking whether the final output spectrum maintains a reasonable consistency with the reference spectrum in the true peak region of the sample. This is achieved using a sample peak mask. calculate have:

[0088] In the formula, This serves as a consistency index between the final output spectrum and the reference spectrum.

[0089] 6. Are the residual magnitudes of deep learning branches abnormal? Deep learning branch output residual correction amount:

[0090] Its amplitude energy is:

[0091] in, This represents the residual magnitude of the deep learning branch. If this value deviates significantly, it indicates that the deep network may be distorting the true information of the sample.

[0092] 7. Is the difference between the results of the physical correction branch and the deep learning branch too large? (This refers to the physical correction branch.) With deep learning branches If the results differ significantly between the two branches, it indicates that at least one branch is unstable. Define the branch difference measure:

[0093] In the formula, This represents the difference between the physics and deep learning branches.

[0094] The above seven indicators are basic diagnostic indicators. For easier comprehensive evaluation, they are grouped into five categories of upper-level scoring items: residual atmospheric peak score. Noise stability score Negative peak anomaly score Reference consistency score and branch stability score Among them, branch stability score The evaluation is composed of seven original indicators, including the consistency of multi-window scaling factors, the degree of abnormality in deep branch residuals, and the difference between physical branches and deep branches. The comprehensive formula uses five aggregate terms for weighting. That is: ; ; ; ; ; in, , , , , This is the attenuation coefficient, which is generally equal to 3. , These represent the maximum and minimum reference values ​​for residual atmospheric peak indicators from historical data. , These are the maximum and minimum reference values ​​for the noise variation index, respectively. , These are the maximum and minimum reference values ​​for the abnormal negative peak index, respectively. , These are the maximum and minimum reference values ​​for the reference consistency index, respectively. , These represent the maximum and minimum reference values ​​for the multi-window consistency index, respectively. , These are the maximum and minimum reference values ​​for the residual magnitude index of deep learning branches, respectively. , These are the maximum and minimum reference values ​​for the difference between the results of the physical correction branch and the deep learning branch, respectively.

[0095] The overall confidence level can be expressed as:

[0096] in, To assess overall confidence, a higher confidence level indicates a more reliable correction result. , , , , These are the residual atmospheric peak score, noise stability score, negative peak anomaly score, reference consistency score, and branch stability score. , , , , These are the coefficients for the residual atmospheric peak score, noise stability score, negative peak anomaly score, reference consistency score, and branch stability score, respectively.

[0097] In one embodiment of the present invention, the online update mechanism is as follows: After the sample calibration results pass quality control, the final calibrated spectrum, environmental parameters, instrument parameters, and necessary labels of the current sample can be fed back into the historical sample library for subsequent similar sample retrieval and reference spectrum generation. In this way, the historical sample library will continuously expand with use, and the subsequently generated reference spectra will be closer to the actual application scenario.

[0098] Please see Figure 2 The comparison of spectra before and after correction, as shown in the examples, demonstrates that after processing the original infrared spectrum using the method of this invention, the sharp residual peaks and high-frequency oscillations in the atmospheric interference absorption region of the corrected spectrum are significantly suppressed compared to the uncorrected spectrum. The overall spectral lines are smoother and more continuous, and the residual atmospheric absorption characteristics are significantly reduced. Simultaneously, the true absorption peaks of the sample are not significantly weakened, and key feature regions still maintain good peak shape and structural information. This indicates that while weakening atmospheric absorption interference from water vapor, carbon dioxide, etc., this invention can effectively prevent the true peaks of the sample from being incorrectly corrected. These results demonstrate that the robust difference and piecewise scaling factor physical correction, sample peak masking and automatic window selection, reference spectrum constraints, depth residual compensation, and uncertainty adaptive fusion mechanisms proposed in this invention can work synergistically to improve the accuracy, robustness, and peak fidelity of infrared spectral atmospheric interference correction in complex scenarios.

[0099] The present invention also provides a system for applying the above-described infrared spectral atmospheric absorption interference correction method, comprising: The database and similarity filtering module is used to establish a historical spectral database; based on the sample to be calibrated, it filters from the historical spectral database to obtain the nearest neighbor reference spectrum.

[0100] The calibration branch module is used to set up calibration branches: a reference spectrum generation branch, a physical calibration branch, and a deep learning branch. The reference spectrum generation branch generates a reference spectrum for the sample to be calibrated based on the sample to be calibrated, the nearest neighbor reference spectrum, and environmental and instrument parameters. The physical calibration branch constructs differential interference components for the sample to be calibrated based on a time-series spectral window and performs physical constraint calibration using multispectral robust estimation and an adaptive piecewise scaling factor to obtain the physically calibrated spectrum. The deep learning branch performs deep residual learning on the sample to be calibrated, the reference spectrum, and the physically calibrated spectrum to obtain the deep-calibrated spectrum. The adaptive fusion final correction module is used to adaptively fuse the corrected spectrum based on the uncertainty or consistency of the correction branches to obtain the final corrected spectrum.

[0101] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.

[0102] The above embodiments are merely examples of implementation methods of the invention. The scope of protection of the present invention is not limited to the above embodiments. For those skilled in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. A method for correcting atmospheric absorption interference in infrared spectroscopy, characterized in that, include: Establish a historical spectral database; Based on the sample to be calibrated, the nearest neighbor reference spectrum is obtained by screening from the historical spectral database; The calibration branches are set up as follows: reference spectrum generation branch, physical calibration branch, and deep learning branch. Among them, the reference spectrum generation branch generates the reference spectrum of the sample to be calibrated based on the sample to be calibrated, the nearest neighbor reference spectrum, and environmental and instrument parameters. The physical correction branch constructs differential interference components based on the time-series spectral window for the sample to be corrected, and performs physical constraint correction using multispectral robust estimation and adaptive piecewise scaling factor to obtain the physically corrected spectrum; the deep learning branch performs deep residual learning on the sample to be corrected, the reference spectrum, and the physically corrected spectrum to obtain the deep corrected spectrum. Based on the uncertainty or consistency of the correction branches, the corrected spectra are adaptively fused to obtain the final corrected spectrum.

2. The infrared spectral atmospheric absorption interference correction method according to claim 1, characterized in that, Obtaining the nearest neighbor reference spectrum includes: Process the historical sample spectra in the historical spectral database to match the dimensions of the sample to be corrected; The spectra of the sample to be corrected and the historical samples of the same dimension are used to perform feature encoding based on a three-level one-dimensional convolutional residual coding network to obtain the current sample code and the historical sample code. Cosine similarity is calculated for the current sample code and historical sample codes. All historical sample spectra are then sorted from highest to lowest similarity, and the top-ranked spectra are selected. The spectrum is used as a set of nearest-neighbor historical samples; The nearest neighbor reference spectrum is obtained by weighting the set of nearby historical samples.

3. The infrared spectral atmospheric absorption interference correction method according to claim 1, characterized in that, The step of generating the reference spectrum of the sample to be calibrated employs a search-enhanced conditional coding-decoding network; The retrieval enhancement conditional encoding includes the current spectrum to be calibrated branch, the historical nearest neighbor weighted base spectrum branch, and the environmental and instrument parameters branch. These branches encode the sample to be calibrated, the nearest neighbor reference base spectrum, and the environmental and instrument parameters, respectively, to obtain feature vectors. The three feature vectors are then concatenated into a joint latent variable. The decoder uses a three-layer one-dimensional deconvolution structure to restore the latent variables to the reference spectrum.

4. The infrared spectral atmospheric absorption interference correction method according to claim 1, characterized in that, The adaptive piecewise scaling factor is: The interference windows for water vapor and carbon dioxide are divided into several continuous sub-intervals, and a constant scaling factor is set for each sub-interval. The constant scaling factor is obtained by minimizing the following objective function: ； ； ； ； ； ； In the formula, For physical optimization loss, , , , , These are, respectively, point-to-point spectral length loss, residual atmospheric absorption loss, sample peak fidelity loss, boundary continuity loss, and reference consistency loss. , , , , These are the weights corresponding to point-to-point spectral length loss, residual atmospheric absorption loss, sample peak fidelity loss, boundary continuity loss, and reference consistency loss, respectively. , The first , Wavenumber points, , The first , Physically corrected spectral values ​​corresponding to each wavenumber point The total number of wavenumber points within the band. As an atmospheric interference mask, The Hadamard product is a point-by-point multiplication. wave number The corresponding physically corrected spectral values, The square of the L2 norm, As a mask for sample peaks, For the sample to be calibrated, It is an L1 norm. To correct the set of interval boundaries, , These are the physically corrected spectral values ​​for the left and right sides of the boundary, respectively. This is the reference spectrum.

5. The infrared spectral atmospheric absorption interference correction method according to claim 4, characterized in that, To obtain the physically corrected spectrum, apply the following formula: ； In the formula, wave number The corresponding physically corrected spectral values, For the sample to be calibrated, , These are the piecewise scaling factor functions for water vapor and carbon dioxide, respectively. , These are the differential interference components corresponding to water vapor and carbon dioxide, respectively.

6. The infrared spectral atmospheric absorption interference correction method according to claim 4, characterized in that, The sample peak mask is obtained through the following steps: For the current spectrum to be corrected Calculate the second derivative spectrum ; In the second derivative spectrum Find the peak center by locating the negative extreme value; For the peak center, candidate sample peak regions are formed by expanding based on the half-width at half-maximum threshold; then compared with the reference spectrum. The peak regions are intersected; The intersection region is defined as the true peak region of the sample. ; The final sample peak mask is defined as follows: .

7. The infrared spectral atmospheric absorption interference correction method according to claim 1, characterized in that, Obtaining depth-corrected spectra includes: The deep learning branch uses a one-dimensional U-Net residual compensation network. The encoding end uses three downsampling convolutional blocks, and the decoding end uses a three-layer upsampling structure corresponding to the encoding end. The upsampling result of each layer is spliced ​​or fused with the feature map of the same scale as the encoding end through skip connections, so as to preserve the sample peak shape, local details and boundary features while restoring the spectral resolution. Finally, two one-dimensional sequences are output: residual correction amount and depth branch uncertainty. Based on the residual correction, the local complex residuals that the physical correction branch failed to eliminate are compensated to obtain the deep correction spectrum.

8. The infrared spectral atmospheric absorption interference correction method according to claim 1, characterized in that, The final corrected spectrum is obtained using the following formula: ； ； In the formula, For the final calibrated spectrum, , , These are the weights for the physically corrected spectrum, the depth-corrected spectrum, and the reference spectrum, respectively. , , Wavenumber Reference spectrum, physically corrected spectrum, depth corrected spectrum, It is an exponential function. For the first The correction branch at wavenumber Uncertainty at the location, For the first The correction branch at wavenumber The uncertainty at, where, In the formula, it is a constant. In the formula, it is a variable. For the first The weights of the correction branches.

9. The infrared spectral atmospheric absorption interference correction method according to claim 1, characterized in that, The calibration method also includes: generating a calibration confidence result based on a comprehensive index for the final calibration spectrum; and, based on the calibration confidence result, after the calibration spectrum result has passed quality control, re-entering the final calibration spectrum, environmental and instrument parameters of the current calibration sample into the historical spectral database for subsequent reference spectrum generation.

10. A system for applying the infrared spectral atmospheric absorption interference correction method according to any one of claims 1-9, characterized in that, include: Database and similarity filtering module: used to establish a historical spectral database; based on the sample to be calibrated, filter from the historical spectral database to obtain the nearest neighbor reference spectrum; The calibration branch module is used to set calibration branches: reference spectrum generation branch, physical calibration branch, and deep learning branch; among them, the reference spectrum generation branch generates the reference spectrum of the sample to be calibrated based on the sample to be calibrated, the nearest neighbor reference spectrum, and environmental and instrument parameters. The physical correction branch constructs differential interference components based on the time-series spectral window for the sample to be corrected, and performs physical constraint correction using multispectral robust estimation and adaptive piecewise scaling factor to obtain the physically corrected spectrum; the deep learning branch performs deep residual learning on the sample to be corrected, the reference spectrum, and the physically corrected spectrum to obtain the deep corrected spectrum. The adaptive fusion final correction module is used to adaptively fuse the corrected spectrum based on the uncertainty or consistency of the correction branches to obtain the final corrected spectrum.

Images