Chemical production tail gas emission automatic monitoring system

Abstract

The application provides a chemical production tail gas emission automatic monitoring system, through collecting tail gas spectrum and corresponding concentration data, relying on wavelength response variance and wavelength interval ratio to calculate overlapping weight distribution, differentiating segmentation full waveband spectrum according to the difference, obtaining candidate subinterval set, combining cross correlation coefficient and characteristic redundancy to solve interval quality factor, screening optimal cooperative subinterval combination, building calibration subinterval covariance matrix; extracting de-redundancy potential factor vector therefrom, combining local prediction residual and redundancy to set subinterval initial fusion weight, constructing full waveband fusion regression coefficient matrix, when measuring tail gas spectrum, relying on interval overlapping waveband response consistency deviation to correct initial weight, obtaining fusion weight and updating fusion regression coefficient matrix, measuring tail gas emission concentration, realizing tail gas concentration fine monitoring.

Description

Technical Field

[0001] This application belongs to the field of monitoring, and in particular relates to an automated monitoring system for chemical production exhaust gas emissions. Background Technology

[0002] The exhaust gases generated during chemical production processes have a significant impact on the ecological environment. Stable and accurate monitoring of exhaust gas emission concentrations is a crucial foundation for pollutant emission control. Spectroscopic analysis technology, with its advantages of fast detection speed, non-contact operation, and suitability for online applications, has been widely used in exhaust gas detection scenarios. Partial Least Squares (PLS) is a commonly used method for constructing a quantitative mapping relationship between spectral data and exhaust gas concentrations. However, full-band spectra typically contain significant background noise and redundant variables. Directly modeling the entire spectrum can easily increase model complexity and limit prediction accuracy. In actual chemical monitoring sites, the exhaust gas spectra are also easily affected by environmental factors such as temperature, humidity, and airflow disturbances, leading to shifts in the response consistency of overlapping bands in adjacent sub-intervals. Therefore, how to overcome the limitations of conventional equal-width spectral segmentation and static weight fusion methods, reduce inter-interval feature redundancy, and achieve adaptive adjustment of fusion weights to improve the accuracy of chemical exhaust gas emission concentration monitoring is a technical problem that needs to be solved in this field. Summary of the Invention

[0003] To improve the stability and prediction accuracy of the quantitative detection model for chemical exhaust gas under complex operating conditions, this invention provides an automated monitoring system for chemical production exhaust gas emissions.

[0004] The first aspect of this invention provides a chemical production exhaust gas emission monitoring system, comprising the following modules:

[0005] The generation module is used to collect spectral data and corresponding tail gas concentration data of chemical production tail gas, calculate the overlap weight distribution based on the ratio of the response variance of each wavelength point to the wavelength position interval, and perform differential width segmentation on the full-band spectral data according to the overlap weight distribution to generate a set of candidate sub-intervals. The calculation module is used to calculate the cross-correlation coefficient and feature redundancy of each candidate sub-interval to obtain the interval quality factor, select the optimal combination of cooperative sub-intervals from the candidate sub-interval set based on the interval quality factor, and construct the corrected sub-interval covariance matrix using the interval quality factor. The construction module is used to extract the redundancy-removing latent factor vector based on the corrected sub-interval covariance matrix, determine the initial fusion weight of each sub-interval using the local prediction residual and feature redundancy, and construct the full-band fusion regression coefficient matrix based on the redundancy-removing latent factor vector. The monitoring module is used to calculate the response consistency deviation on the overlapping bands of adjacent sub-intervals when acquiring the spectral data of the chemical exhaust gas to be measured, adjust the initial fusion weight according to the response consistency deviation to obtain the fusion weight, reconstruct the fusion regression coefficient matrix using the fusion weight, and calculate the monitoring value of the chemical exhaust gas emission concentration by combining the spectral data of the chemical exhaust gas to be measured.

[0006] A second aspect of this invention provides a method for monitoring emissions from chemical production tail gas, comprising the following steps: Collect spectral data and corresponding concentration data of chemical production tail gas, calculate the overlap weight distribution based on the ratio of response variance of each wavelength point to wavelength position interval, and perform differential width segmentation on the full-band spectral data according to the overlap weight distribution to generate a set of candidate sub-intervals. The cross-correlation coefficient and feature redundancy of each candidate sub-interval are calculated to obtain the interval quality factor. Based on the interval quality factor, the optimal combination of cooperative sub-intervals is selected from the set of candidate sub-intervals, and the corrected sub-interval covariance matrix is ​​constructed using the interval quality factor. Based on the corrected sub-interval covariance matrix, the redundancy-removing latent factor vector is extracted, the initial fusion weight of each sub-interval is determined by the local prediction residual and feature redundancy, and the full-band fusion regression coefficient matrix is ​​constructed based on the redundancy-removing latent factor vector. When acquiring the spectral data of the chemical exhaust gas to be tested, the response consistency deviation on the overlapping bands of adjacent sub-intervals is calculated. The initial fusion weight is adjusted according to the response consistency deviation to obtain the fusion weight. The fusion regression coefficient matrix is ​​reconstructed using the fusion weight. The monitoring value of the chemical exhaust gas emission concentration is calculated by combining the spectral data of the chemical exhaust gas to be tested.

[0007] This invention obtains the overlap weight distribution by calculating the ratio of the response variance at each wavelength point to the wavelength position interval, and then performs differentiated width segmentation on the full-band spectral data accordingly. This reduces the impact of traditional equal-width segmentation on the integrity of spectral characteristic peaks and preserves effective feature information in the exhaust gas spectrum. Based on this, interval quality factors are constructed using cross-correlation coefficients and feature redundancy to screen the optimal combination of cooperative sub-intervals and correct the covariance matrix, thereby weakening the interference of invalid bands and redundant features, improving the rationality of interval selection and the relevance of model modeling. Simultaneously, by extracting the redundancy-removing potential factor vector and combining it with local prediction residuals to determine the initial fusion weights, the fusion weights are adjusted in real time during actual monitoring based on the response consistency deviation of overlapping bands in adjacent sub-intervals. This reduces the impact of spectral data fluctuations in the detection environment on the monitoring results, improving the accuracy, reliability, and overall stability of the chemical exhaust gas emission concentration monitoring results. Attached Figure Description

[0008] Figure 1This is a schematic diagram of the curve of the overlap weight integral as a function of the spectral wavenumber; Figure 2 This is a diagram showing the comparison between the initial fusion weights and the fusion weights for each interval. Detailed Implementation

[0009] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0010] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0011] In the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0012] References to "one embodiment" or "some embodiments" as described in this application specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in another embodiment," "in yet another embodiment," etc., appearing in different parts of this application specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0013] Example 1 In a first embodiment of the present invention, a chemical production exhaust gas emission monitoring system includes the following modules: The generation module is used to collect spectral data and corresponding concentration data of chemical production tail gas, calculate the overlap weight distribution based on the ratio of the response variance of each wavelength point to the wavelength position interval, and perform differential width segmentation on the full-band spectral data according to the overlap weight distribution to generate a set of candidate sub-intervals.

[0014] A Fourier transform infrared spectrometer was used to continuously collect the spectral absorbance sequence of exhaust gas at the outlet of a chemical production pipeline. Simultaneously, a standard gas analyzer was used to measure the corresponding actual exhaust gas concentration values, forming a sample set. The collected full-band spectral data underwent baseline correction and multivariate scattering correction preprocessing to reduce background noise and light scattering. The wavelength points are spectral variable points arranged according to the spectral acquisition coordinates, which include wavelength coordinates or wavenumber coordinates. In this representation, the wavelength position interval corresponds to the wavenumber position interval between adjacent spectral variable points. The sample variance (response variance) of the absorbance at each wavelength point in the preprocessed full-band spectral data is calculated, and the wavelength position interval between adjacent wavelength points is determined. The response variance is divided by the wavelength position interval to obtain the initial weight vector. For the last wavelength point in the full band, the initial weight of a single point is calculated using the wavelength position interval between it and the preceding adjacent wavelength point to ensure that the number of initial weight sequences at a single point is consistent with the number of wavelength points in the full band. Gaussian smoothing filtering is used to smooth the initial weight vector to obtain an overlapping weight distribution. The overlapping weight distribution is integrated to form a cumulative distribution function, and the benchmark sub-interval division points are determined according to a cumulative weight threshold sequence with a fixed step size, so that each benchmark sub-interval contains the same overlapping weight integral value. Then, the boundaries of adjacent benchmark sub-intervals are extended to both sides by a fixed number of wavelength points, forming a common overlapping band between adjacent candidate sub-intervals, thereby generating a set of candidate sub-intervals containing multiple widths with overlapping boundaries.

[0015] In one embodiment, calculating the overlap weight distribution based on the ratio of the response variance at each wavelength point to the wavelength position interval includes: Obtain the response variance of each wavelength point in the full-band spectral data across all samples; Calculate the physical location interval between two adjacent wavelength points; The initial weight of a single point is obtained by dividing the response variance of each wavelength point by the positional interval of the corresponding adjacent wavelength point. The initial weight of the last wavelength point is calculated by using its positional interval with the adjacent previous wavelength point. A Gaussian smoothing function is used to smooth the initial weight sequence at a single point to generate a continuous overlapping weight distribution curve. The value of each point after smoothing is the sum of the initial weights at each point multiplied by the corresponding Gaussian weight factor. The Gaussian weight factor is calculated based on the square of the difference between the wavelength position of the wavelength point corresponding to the initial weight at a single point and the wavelength position of the current point, as well as the standard deviation parameter of the smoothing function.

[0016] Acquire full-band spectral data, for example, ranging from 4000 to 10000. It contains 2000 spectral variable points and 100 sample records. The response variance of each spectral variable point i is calculated one by one by statistically analyzing the variance between samples. Let the physical location interval between adjacent spectral variable points i and i+1 be . For example, the instrument resolution is set to =3 The corresponding response variance Divide by interval Generate initial weights for a single point When i is the last spectral variable point, the physical location interval between the last spectral variable point and the adjacent previous spectral variable point is used as... The initial weight of the last single point is calculated, thus obtaining a sequence of initial weights for each of 2000 discrete values. Then, a weighted smoothing process is performed on the sequence using a preset smoothing window, for example, with a window size of 11 points and a standard deviation parameter of 2.5.

[0017] For the j-th target spectral variable location point in the smooth sequence Extract the initial weights of several adjacent single points centered on the location of the target spectral variable, for example, the initial weights of single points within the range of j-5 to j+5. Based on the initial weights of each individual point. Corresponding spectral variable positions relative to the current target spectral variable position The distance between points determines the corresponding smoothing weights. Points closer to each other have larger initial smoothing weights, while points farther apart have smaller initial smoothing weights. The initial smoothing weights of each point within the window are multiplied by their corresponding smoothing weights and then summed to obtain the smoothing weight value for the j-th target spectral variable location. This process is repeated across the entire data segment to generate a continuous overlapping weight distribution curve.

[0018] In one embodiment, the step of generating a candidate sub-interval set by differentially segmenting the full-band spectral data according to the overlap weight distribution includes: The cumulative distribution function is calculated by integrating the overlapping weight distribution curve over the variable interval from the initial wavelength of the spectrum to the current wavelength using an integral formula. Set a cumulative weight threshold sequence with a fixed step size; The full-band spectral data is divided into adjacent reference sub-intervals at the positions where the cumulative distribution function reaches the corresponding cumulative weight threshold sequence, so that each reference sub-interval contains the same overlapping weight integral value. The boundaries of adjacent reference sub-intervals are extended to both sides by a predetermined fixed number of wavelength points to construct a set of candidate sub-intervals.

[0019] For the generated continuously overlapping weighted distribution curve data, start with the spectral variable point, for example, the 4000th point with index position 1. At the current spectral variable position, the lower bound for integration is used as the upper bound for integration, and the area definite integral below the curve is calculated by continuously shifting the current spectral variable position to the right. The integration result at each wavelength variable position is recorded as a continuously accumulated value, forming an increasing cumulative distribution function curve. If the total global integral value of the full-band overlapping weight curve is 1000, and the full-band cumulative weight value range needs to be divided into 20 sub-ranges, then the fixed step size of the cumulative weight threshold sequence is set to 50, and cumulative weight threshold sequences with values ​​of 50, 100, 150 up to 1000 are generated sequentially.

[0020] Along the generated cumulative distribution function curve, the node closest to the aforementioned threshold sequence is found from front to back, and this node is established as the interval division point. The entire band is divided into 20 reference sub-intervals, with the total internal integral of each reference sub-interval remaining constant at approximately 50. A fixed expansion threshold is set, for example, 10 wavelength points. The left and right boundaries of these 20 reference sub-intervals are expanded outward by 10 points to each side. If the boundary points exceed the extreme values ​​of the entire band, they are truncated. For example, if the original index boundary of a reference sub-interval is 150 to 190 points, the expanded range becomes 140 to 200 points. All interval groups with overlapping and coincident wavelengths formed by this expansion process constitute the candidate sub-interval set. The curve of the overlap weight integral changing with the spectral wavenumber is shown below. Figure 1 As shown.

[0021] The calculation module is used to calculate the cross-correlation coefficient and feature redundancy of each candidate sub-interval to obtain the interval quality factor, select the optimal combination of cooperative sub-intervals from the candidate sub-interval set based on the interval quality factor, and construct the corrected sub-interval covariance matrix using the interval quality factor.

[0022] For each candidate sub-interval in the candidate sub-interval set, the Pearson correlation coefficient between the spectral data and the exhaust gas concentration data at each wavelength point within the candidate sub-interval is calculated, and the average of the absolute values ​​of the Pearson correlation coefficients is used as the cross-correlation coefficient. Principal component analysis is performed on the spectral data of the candidate sub-interval to extract the variance contribution rate of each principal component, and the sum of the variance contribution rates of the first three principal components is used as the feature redundancy of the interval. The cross-correlation coefficient is divided by the feature redundancy to obtain the interval quality factor of the corresponding candidate sub-interval.

[0023] All candidate sub-intervals are sorted in descending order of interval quality factor, and a predetermined number of the top-ranked candidate sub-intervals are selected to form the optimal cooperative sub-interval combination. For each sub-interval in the optimal cooperative sub-interval combination, the initial covariance matrix of the sub-interval spectral data is calculated, and the corresponding adaptive noise variance is determined based on the benchmark measurement noise variance and the interval quality factor. The product of the adaptive noise variance and the identity matrix of the same dimension is subtracted from the initial covariance matrix to obtain the corrected sub-interval covariance matrix.

[0024] In one embodiment, the step of calculating the cross-correlation coefficient and feature redundancy of each candidate sub-interval to obtain an interval quality factor, and selecting the optimal combination of cooperative sub-intervals from the candidate sub-interval set based on the interval quality factor, includes: Calculate the Pearson correlation coefficient between the spectral data and exhaust gas concentration data at each wavelength point in each candidate sub-interval, and use the average of the absolute values ​​of each Pearson correlation coefficient as the cross-correlation coefficient. Extract the principal components of each candidate sub-interval and calculate the variance contribution rate. Set the sum of the variance contribution rates of the first three principal components as the feature redundancy. The interval quality factor is obtained by dividing the cross-correlation coefficient of each candidate sub-interval by the feature redundancy. Candidate subintervals are sorted in descending order of their interval quality factors, and the combination of candidate subintervals that ranks at the top of a preset number is selected as the optimal collaborative subinterval combination.

[0025] An absorbance observation matrix with 60 wavelength points is extracted from a candidate sub-interval, with dimensions such as 100 samples × 60 variables. Simultaneously, the measured concentration vector of the corresponding chemical exhaust gas is retrieved. The correlation coefficients between the response intensity sequence and the exhaust gas concentration vector at each of the 60 wavelength points are calculated. The absolute values ​​of the 60 correlation coefficients are taken, and the average value is calculated; for example, if the average is 0.82, this average value is set as the cross-correlation coefficient. Simultaneously, a centered principal component analysis (PCA) is performed on the absorbance observation matrix of the candidate sub-interval, orthogonally decomposing the matrix into a series of principal components with decreasing variance explanatory power.

[0026] The variance contribution rate of the principal components is extracted. Assuming the variance contribution rates of the first three principal components PC1, PC2, and PC3 are 82%, 9%, and 4% respectively, totaling 95%, the variance contribution rates of the first three principal components are summed and represented as 0.95. This value is used as the feature redundancy. The cross-correlation coefficient of 0.82 is divided by the feature redundancy of 0.95, and the interval quality factor of the candidate sub-interval is calculated to be approximately 0.863. After calculating the interval quality factors of all candidate sub-intervals, they are sorted in descending order. If the preset number of top columns is set to 6, the 6 candidate sub-intervals with the highest quality factors are selected and rearranged according to the physical wavelength position of each candidate sub-interval in the entire band from smallest to largest. In the subsequent response consistency deviation calculation, only the weight adjustment calculation is performed on adjacent selected sub-intervals with common overlapping bands after rearrangement. For adjacent selected sub-intervals without common overlapping bands, no overlap consistency adjustment is performed, and their initial fusion weights are maintained until the normalization stage. This constructs the optimal combination of cooperative sub-intervals.

[0027] In one embodiment, constructing the corrected sub-interval covariance matrix using the interval quality factor includes: Obtain the reference measurement noise variance measured at the calibration time of the spectrometer, and obtain the reference quality factor corresponding to the full-band spectral data; The ratio of the baseline quality factor to the interval quality factor of each individual sub-interval in the optimal collaborative sub-interval combination is used as the scaling factor. The adaptive noise variance of each corresponding sub-interval is obtained by multiplying the benchmark measurement noise variance by the scaling factor. The initial covariance matrix is ​​calculated using the spectral data of each sub-interval. The product of the adaptive noise variance of the corresponding sub-interval and the identity matrix of the same dimension is then subtracted from the initial covariance matrix. Eigenvalues ​​that are less than the preset lower limit of eigenvalues ​​are truncated to obtain the corrected sub-interval covariance matrix.

[0028] In the offline calibration of the spectrometer, the statistical variance of the dark current response fluctuations of multiple collected background spectral data is calculated and regarded as the reference measurement noise variance, for example, a variance value of 0.008 is obtained. Simultaneously, the reference quality factor is calculated for the entire spectrum as a whole interval, assumed to be 5.5. For any specific sub-interval of the optimal cooperative sub-interval combination, for example, if the calculated interval quality factor of the sub-interval is 11.0, the reference quality factor 5.5 is divided by the interval quality factor 11.0 of the specific sub-interval to obtain a scaling factor of 0.5.

[0029] By multiplying the baseline measurement noise variance of 0.008 by the scaling factor of 0.5, the adaptive noise variance for this specific sub-interval is calculated to be 0.004. The centered spectral matrix for this specific sub-interval, with 100 samples and 60 variables, is then obtained. According to the formula An initial covariance matrix of dimension 60×60 is constructed. An identity matrix I, also of dimension 60×60, with 1s on its main diagonal and 0s elsewhere, is constructed and multiplied by an adaptive noise variance of 0.004. This 0.004·I diagonal noise matrix is ​​subtracted from the initial covariance matrix, and eigenvalues ​​smaller than a preset lower limit are truncated to generate the corrected sub-interval covariance matrix.

[0030] The construction module is used to extract the de-redundancy latent factor vector based on the corrected sub-interval covariance matrix, determine the initial fusion weight of each sub-interval using the local prediction residual and feature redundancy, and construct the full-band fusion regression coefficient matrix based on the de-redundancy latent factor vector.

[0031] Partial least squares operations are performed on the centered spectral sample matrix and exhaust gas concentration data of each sub-interval in the optimal coordinated sub-interval combination. The corrected sub-interval covariance matrix is ​​used as the basis for covariance constraints or weight calculation in the latent factor extraction process to extract a predetermined number of initial latent factor vectors. Orthogonalization is performed on the initial latent factor vectors to obtain redundant latent factor vectors. The spectral data of each sub-interval is projected onto the corresponding redundant latent factor vector to obtain the sub-interval score matrix. Linear regression is performed using the sub-interval score matrix and the exhaust gas concentration data to obtain the predicted concentration. The local prediction residual is obtained by subtracting the predicted concentration from the actual concentration.

[0032] The reciprocal of the sum of squares of the local prediction residuals is extracted as the accuracy index. This accuracy index is multiplied by the reciprocal of the feature redundancy corresponding to the sub-interval. The product of all sub-intervals within the optimal collaborative sub-interval combination is normalized to a sum of one, and the normalized value is used as the initial fusion weight for each sub-interval. The local regression coefficient vector of each sub-interval is calculated separately, and the local regression coefficient vector of each sub-interval is multiplied by the initial fusion weight corresponding to the sub-interval. Since there are overlapping bands among the sub-intervals, for the regression coefficients at the overlapping bands, the weighted local regression coefficients of the adjacent sub-intervals involving that band are numerically accumulated. For non-overlapping bands, the weighted local regression coefficients of the sub-interval to which the non-overlapping band belongs are taken. The calculation results of all wavelength points in the entire band are concatenated and merged in wavelength order to construct the fusion regression coefficient matrix of the entire band.

[0033] In one embodiment, the extraction of the deredundant latent factor vector based on the corrected sub-interval covariance matrix includes: The initial latent factor matrix is ​​formed by extracting initial weight vectors based on the spectral sample matrix of each sub-interval, the exhaust gas concentration data, and the corrected sub-interval covariance matrix using the partial least squares algorithm. The first latent factor vector in the initial latent factor matrix is ​​normalized and then set as the benchmark orthogonal basis. The remaining latent factor vectors are sequentially projected onto the directions of the generated orthogonal basis to remove them. The orthogonalized residual vector is generated by subtracting the projection components of the vectors on the directions of the generated orthogonal basis from the remaining latent factor vectors. The redundancy-removing latent factor vector is generated by column concatenating all the orthogonalized residual vectors with the reference orthogonal basis.

[0034] An iterative partial least squares (PLS) method is employed. A 100×60-dimensional centered sub-interval spectral sample matrix and a corresponding 100×1-dimensional chemical exhaust gas calibration concentration vector are input. A 60×60-dimensional corrected sub-interval covariance matrix is ​​used to constrain the initial weight vector. The latent variable extraction dimension parameter for PLS is set to 4. Through iterative computation, four columns of initial weight vectors are extracted to characterize the covariance relationship between the spectral data and the exhaust gas concentration. These four columns are then combined column-wise to form a 60-row × 4-column initial latent factor matrix.

[0035] The initial latent factor matrix is ​​orthogonalized. First, the first column of latent factor vectors is normalized to form the first orthogonal basis vector. Then, each subsequent column of latent factor vectors is processed sequentially, eliminating its projection components along the directions of the generated orthogonal basis vectors. The residual vectors after projection elimination are then normalized to form new orthogonal basis vectors. This process is repeated until all four columns of latent factor vectors are orthogonalized. Finally, the resulting four columns of orthogonal basis vectors are concatenated column-wise to generate a 60-row × 4-column deredundant latent factor vector matrix.

[0036] In one embodiment, the step of determining the initial fusion weights for each sub-interval using local prediction residuals and feature redundancy, and constructing a full-band fusion regression coefficient matrix based on the deredundancy-removed latent factor vector, includes: Calculate the composite residual metric based on the local prediction residual and the feature redundancy. Extract the composite residual measure for each sub-interval and calculate its reciprocal to obtain the corresponding reciprocal value; Divide the reciprocal of a specific single sub-interval by the sum of the reciprocals of all sub-intervals in the optimal collaborative sub-interval combination, and use the result as the initial fusion weight of the sub-interval. Calculate the local regression coefficient matrix of each subinterval under the transformation of the redundant latent factor vector space; The local regression coefficient matrix of each sub-interval is mapped to a global mapping matrix with the same dimension as the full-band spectral data and all other elements being zero, to generate a full-band local regression matrix; Multiply the full-band local regression matrix of each sub-interval by the corresponding initial fusion weight, and then perform an accumulation operation of the same dimension element matrix to form the fusion regression coefficient matrix.

[0037] For each sub-interval within the optimal collaborative sub-interval combination, the root mean square prediction residual of that sub-interval, obtained using leave-one-out cross-validation, is read. Assume a sub-interval has a residual of 0.05. Multiply 0.05 by the feature redundancy extracted from that sub-interval, assumed to be 0.08, to obtain a composite residual metric of 0.004. Take the reciprocal of the composite residual metric for each sub-interval, resulting in 250. Assume the optimal collaborative sub-interval combination contains three sub-intervals, whose reciprocals are calculated to be 250, 300, and 450, respectively, with a total reciprocal sum of 1000. For the first sub-interval with a reciprocal of 250, divide 250 by the sum of 1000 to obtain an initial fusion weight of 0.25 for that sub-interval; similarly, the weights for the other two sub-intervals are 0.30 and 0.45.

[0038] Based on the solved deredundant latent factor vector matrix, partial least squares dimensionality reduction coefficients are inversely calculated to obtain a columnar local regression coefficient matrix for a specific sub-interval, such as one containing 60 variables. A column vector with dimensions matching the entire band (e.g., 2000 wavelength points) and all elements initialized to zero is generated as a global mapping matrix. Based on the true band indices of these 60 variables, the local regression coefficient values ​​are equivalently assigned to replace the corresponding coordinates within this zero vector, generating a full-band local regression matrix. Non-zero numerical elements in the full-band local regression matrix of each sub-interval are then multiplied by their assigned initial fusion weights, for example, by 0.25. The corresponding elements of the weighted full-band local regression matrices of all sub-intervals are then summed to generate a 2000-dimensional full-band fused regression coefficient matrix.

[0039] The monitoring module is used to calculate the response consistency deviation on the overlapping bands of adjacent sub-intervals when acquiring the spectral data of the chemical exhaust gas to be measured, adjust the initial fusion weight according to the response consistency deviation to obtain the fusion weight, reconstruct the fusion regression coefficient matrix using the fusion weight, and calculate the monitoring value of the chemical exhaust gas emission concentration by combining the spectral data of the chemical exhaust gas to be measured.

[0040] The unknown concentration spectral data of the chemical exhaust gas under test is acquired in real time using on-site online spectral detection equipment, and preprocessed using the same baseline correction and multivariate scattering correction as the training samples. The preprocessed spectral data is then segmented according to the optimal cooperating sub-interval combination rule, and adjacent selected sub-intervals are determined according to the physical wavelength position order of the optimal cooperating sub-interval combination across the entire wavelength band. For any two adjacent selected sub-intervals with a common overlapping wavelength range, the current spectral response value and the overlapping segment elements in the local regression coefficient matrix of each sub-interval are extracted within their common overlapping wavelength range, and the average predicted response of the two sub-spectral segments is calculated. If two adjacent selected sub-intervals do not have a common overlapping wavelength range, the calculation of the response consistency deviation for that pair of sub-intervals is skipped.

[0041] The absolute value of the difference between the two predicted response averages is calculated and divided by the sum of the absolute values ​​of the two predicted response averages and the sum of a preset minimum positive number to obtain the response consistency deviation on the overlapping band. This response consistency deviation is multiplied by a set smoothing damping constant to generate a weight adjustment amount. Within the overlapping segment, this weight adjustment amount is compared with the initial fusion weight corresponding to the sub-interval with the smaller interval quality factor, and the smaller value is taken as the actual weight adjustment amount. The initial fusion weight corresponding to the sub-interval with the smaller interval quality factor is subtracted from the actual weight adjustment amount, and the initial fusion weight corresponding to the adjacent sub-interval with the larger interval quality factor is added to the actual weight adjustment amount. The updated weights within the optimal cooperative sub-interval combination are normalized again to a sum of one to obtain the fusion weights of each sub-interval. Following the rule of accumulating overlapping bands with the same initial fusion regression coefficient matrix and directly taking values ​​from non-overlapping bands, the fusion weights and the local regression coefficient vectors of each sub-interval are recombined and calculated to obtain the updated fusion regression coefficient matrix. The row vector formed by the spectral data of the preprocessed chemical exhaust gas to be tested is multiplied by the column vector formed by the reconstructed fusion regression coefficient matrix. The resulting scalar is the monitoring value of the emission concentration of the chemical exhaust gas to be tested at the current time, and the monitoring value is output to the monitoring and control terminal.

[0042] In one embodiment, calculating the response consistency deviation on overlapping bands in adjacent sub-intervals and adjusting the initial fusion weights based on the response consistency deviations to obtain fusion weights includes: Within the common overlapping wavelength range of any two adjacent sub-intervals, the average predicted response of the first sub-spectral band and the average predicted response of the second sub-spectral band are calculated by combining their respective local regression coefficient matrices. The absolute value of the difference between the two predicted response averages is calculated and divided by the sum of the absolute values ​​of the two predicted response averages and the sum of a preset minimum positive number, which is used as the response consistency deviation. The weight adjustment amount is generated by multiplying the response consistency deviation by the set smoothing damping constant. Within the overlapping section, the weight adjustment amount is compared with the initial fusion weight corresponding to the sub-interval with the smaller interval quality factor, and the minimum value of the two is taken as the actual weight adjustment amount. The initial fusion weight corresponding to the sub-interval with the smaller interval quality factor is subtracted from the actual weight adjustment amount, and the initial fusion weight corresponding to the adjacent sub-interval with the larger interval quality factor is added to the actual weight adjustment amount. The individual fusion weights are output after normalization by iterating through and updating all initial fusion weights and dividing by the sum of all updated weights.

[0043] When acquiring the spectral data of the chemical exhaust gas to be tested, it was detected that adjacent sub-intervals A and B within the optimal synergistic sub-interval combination share a common wavelength range, for example, covering 20 wavelength overlap points. Within these 20 overlap points, 20 corresponding spectral response values ​​are extracted from the spectral data to be tested, and local coefficient elements corresponding to these 20 overlap points are extracted from the local regression coefficient matrices of sub-intervals A and B respectively, to ensure that the spectral response values ​​to be tested and the local regression coefficient matrices are consistent in dimension within the overlapping bands. Within these 20 overlap points, the current spectral response value to be tested is multiplied point-by-point by the local regression coefficient of that segment in sub-interval A and the average is calculated to obtain the average predicted response of the first sub-spectral segment, such as an output result of 20.4 mg / m³. The same spectral data is multiplied point-by-point by the local regression coefficient of the corresponding overlapping segment in sub-interval B and the average is calculated to obtain the average predicted response of the second sub-spectral segment, such as an output result of 21.2 mg / m³. The absolute value of the subtraction is calculated to be 0.8, and the result is divided by the sum of the absolute values ​​of the two and the sum of the preset minimum positive number. When the preset minimum positive number is very small, the denominator is approximately 41.6, and the response consistency deviation is calculated to be approximately 0.0192.

[0044] Multiplying this deviation value of 0.0192 by the smoothing damping constant, assumed to be 0.15, yields a weight adjustment of 0.00288. If the interval quality factor of sub-interval A (10.2) is found to be lower than that of sub-interval B (14.5), the initial fusion weight of sub-interval A (originally 0.25000) is compared with the weight adjustment of 0.00288. The smaller value is taken as the actual weight adjustment. This actual weight adjustment is subtracted from the initial fusion weight of sub-interval A, and this actual weight adjustment is used to compensate for the initial fusion weight of sub-interval B, changing it from 0.30000 to 0.30288. This process is repeated for all adjacent overlapping sub-interval weight parameters. The updated and adjusted weight sequences are summed. If the adjusted sum is 0.9995, each weight is divided by this sum to make the total weight of the entire weight sequence 1.0. The output fusion weights are loaded into the algorithm model. The comparison between the initial fusion weights and the fusion weights of each interval is illustrated below. Figure 2 As shown.

[0045] Example 2 Embodiment 2 of the present invention proposes a method for monitoring emissions from chemical production tail gas, comprising the following steps: Collect spectral data and corresponding concentration data of chemical production tail gas, calculate the overlap weight distribution based on the ratio of response variance of each wavelength point to wavelength position interval, and perform differential width segmentation on the full-band spectral data according to the overlap weight distribution to generate a set of candidate sub-intervals. The cross-correlation coefficient and feature redundancy of each candidate sub-interval are calculated to obtain the interval quality factor. Based on the interval quality factor, the optimal combination of cooperative sub-intervals is selected from the set of candidate sub-intervals, and the corrected sub-interval covariance matrix is ​​constructed using the interval quality factor. Based on the corrected sub-interval covariance matrix, the redundancy-removing latent factor vector is extracted, the initial fusion weight of each sub-interval is determined by the local prediction residual and feature redundancy, and the full-band fusion regression coefficient matrix is ​​constructed based on the redundancy-removing latent factor vector. When acquiring the spectral data of the chemical exhaust gas to be tested, the response consistency deviation on the overlapping bands of adjacent sub-intervals is calculated. The initial fusion weight is adjusted according to the response consistency deviation to obtain the fusion weight. The fusion regression coefficient matrix is ​​reconstructed using the fusion weight. The monitoring value of the chemical exhaust gas emission concentration is calculated by combining the spectral data of the chemical exhaust gas to be tested.

[0046] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A chemical production tail gas emission monitoring system, characterized in that, Includes the following modules: The generation module is used to collect spectral data and corresponding tail gas concentration data of chemical production tail gas, calculate the overlap weight distribution based on the ratio of the response variance of each wavelength point to the wavelength position interval, and perform differential width segmentation on the full-band spectral data according to the overlap weight distribution to generate a set of candidate sub-intervals. The calculation module is used to calculate the cross-correlation coefficient and feature redundancy of each candidate sub-interval to obtain the interval quality factor, select the optimal combination of cooperative sub-intervals from the candidate sub-interval set based on the interval quality factor, and construct the corrected sub-interval covariance matrix using the interval quality factor. The construction module is used to extract the redundancy-removing latent factor vector based on the corrected sub-interval covariance matrix, determine the initial fusion weight of each sub-interval using the local prediction residual and feature redundancy, and construct the full-band fusion regression coefficient matrix based on the redundancy-removing latent factor vector. The monitoring module is used to calculate the response consistency deviation on the overlapping bands of adjacent sub-intervals when acquiring the spectral data of the chemical exhaust gas to be measured, adjust the initial fusion weight according to the response consistency deviation to obtain the fusion weight, reconstruct the fusion regression coefficient matrix using the fusion weight, and calculate the monitoring value of the chemical exhaust gas emission concentration by combining the spectral data of the chemical exhaust gas to be measured.

2. The system according to claim 1, characterized in that, The calculation of the overlap weight distribution based on the ratio of the response variance at each wavelength point to the wavelength position interval includes: Obtain the response variance of each wavelength point in the full-band spectral data across all samples; Calculate the physical location interval between two adjacent wavelength points; The initial weight of a single point is obtained by dividing the response variance of each wavelength point by the positional interval of the corresponding adjacent wavelength point. The initial weight of the last wavelength point is calculated by using its positional interval with the adjacent previous wavelength point. A Gaussian smoothing function is used to smooth the initial weight sequence at a single point to generate a continuous overlapping weight distribution curve. The value of each point after smoothing is the sum of the initial weights at each point multiplied by the corresponding Gaussian weight factor. The Gaussian weight factor is calculated based on the square of the difference between the wavelength position of the wavelength point corresponding to the initial weight at a single point and the wavelength position of the current point, as well as the standard deviation parameter of the smoothing function.

3. The system according to claim 2, characterized in that, The step of generating a candidate sub-interval set by differentially segmenting the full-band spectral data according to the overlap weight distribution includes: The cumulative distribution function is calculated by integrating the overlapping weight distribution curve over the variable interval from the initial wavelength of the spectrum to the current wavelength using an integral formula. Set a cumulative weight threshold sequence with a fixed step size; The full-band spectral data is divided into adjacent reference sub-intervals at the positions where the cumulative distribution function reaches the corresponding cumulative weight threshold sequence, so that each reference sub-interval contains the same overlapping weight integral value. The boundaries of adjacent reference sub-intervals are extended to both sides by a predetermined fixed number of wavelength points to construct a set of candidate sub-intervals.

4. The system according to claim 1, characterized in that, The calculation of the cross-correlation coefficient and feature redundancy of each candidate sub-interval yields the interval quality factor. Based on the interval quality factor, the optimal combination of cooperative sub-intervals is selected from the candidate sub-interval set, including: Calculate the Pearson correlation coefficient between the spectral data and exhaust gas concentration data at each wavelength point in each candidate sub-interval, and use the average of the absolute values ​​of each Pearson correlation coefficient as the cross-correlation coefficient. Extract the principal components of each candidate sub-interval and calculate the variance contribution rate. Set the sum of the variance contribution rates of the first three principal components as the feature redundancy. The interval quality factor is obtained by dividing the cross-correlation coefficient of each candidate sub-interval by the feature redundancy. Candidate subintervals are sorted in descending order of their interval quality factors, and the combination of candidate subintervals that ranks at the top of a preset number is selected as the optimal collaborative subinterval combination.

5. The system according to claim 1, characterized in that, The construction of the corrected sub-interval covariance matrix using the interval quality factors includes: Obtain the reference measurement noise variance measured at the calibration time of the spectrometer, and obtain the reference quality factor corresponding to the full-band spectral data; The ratio of the baseline quality factor to the interval quality factor of each individual sub-interval in the optimal collaborative sub-interval combination is used as the scaling factor. The adaptive noise variance of each corresponding sub-interval is obtained by multiplying the benchmark measurement noise variance by the scaling factor. The initial covariance matrix is ​​calculated using the spectral data of each sub-interval. The product of the adaptive noise variance of the corresponding sub-interval and the identity matrix of the same dimension is then subtracted from the initial covariance matrix. Eigenvalues ​​that are less than the preset lower limit of eigenvalues ​​are truncated to obtain the corrected sub-interval covariance matrix.

6. The system according to claim 1, characterized in that, The extraction of the deredundant latent factor vector based on the corrected sub-interval covariance matrix includes: The initial latent factor matrix is ​​formed by extracting initial weight vectors based on the spectral sample matrix of each sub-interval, the exhaust gas concentration data, and the corrected sub-interval covariance matrix using the partial least squares algorithm. The first latent factor vector in the initial latent factor matrix is ​​normalized and then set as the benchmark orthogonal basis. The remaining latent factor vectors are sequentially projected onto the directions of the generated orthogonal basis to remove them. The orthogonalized residual vector is generated by subtracting the projection components of the vectors on the directions of the generated orthogonal basis from the remaining latent factor vectors. The redundancy-removing latent factor vector is generated by column concatenating all the orthogonalized residual vectors with the reference orthogonal basis.

7. The system according to claim 1, characterized in that, The process of determining the initial fusion weights for each sub-interval using local prediction residuals and feature redundancy, and constructing a full-band fusion regression coefficient matrix based on the deredundancy-removed latent factor vector, includes: Calculate the composite residual metric based on the local prediction residual and the feature redundancy. Extract the composite residual measure for each sub-interval and calculate its reciprocal to obtain the corresponding reciprocal value; Divide the reciprocal of a specific single sub-interval by the sum of the reciprocals of all sub-intervals in the optimal collaborative sub-interval combination, and use the result as the initial fusion weight of the sub-interval. Calculate the local regression coefficient matrix of each subinterval under the transformation of the redundant latent factor vector space; The local regression coefficient matrix of each sub-interval is mapped to a global mapping matrix with the same dimension as the full-band spectral data and all other elements being zero, to generate a full-band local regression matrix; Multiply the full-band local regression matrix of each sub-interval by the corresponding initial fusion weight, and then perform an accumulation operation of the same dimension element matrix to form the fusion regression coefficient matrix.

8. The system according to claim 1, characterized in that, The calculation of response consistency deviation on overlapping bands in adjacent sub-intervals, and the adjustment of the initial fusion weights based on the response consistency deviation to obtain fusion weights, includes: Within the common overlapping wavelength range of any two adjacent sub-intervals, the average predicted response of the first sub-spectral band and the average predicted response of the second sub-spectral band are calculated by combining their respective local regression coefficient matrices. The absolute value of the difference between the two predicted response averages is calculated and divided by the sum of the absolute values ​​of the two predicted response averages and the sum of a preset minimum positive number, which is used as the response consistency deviation. The weight adjustment amount is generated by multiplying the response consistency deviation by the set smoothing damping constant. Within the overlapping section, the weight adjustment amount is compared with the initial fusion weight corresponding to the sub-interval with the smaller interval quality factor, and the minimum value of the two is taken as the actual weight adjustment amount. The initial fusion weight corresponding to the sub-interval with the smaller interval quality factor is subtracted from the actual weight adjustment amount, and the initial fusion weight corresponding to the adjacent sub-interval with the larger interval quality factor is added to the actual weight adjustment amount. The individual fusion weights are output after normalization by iterating through and updating all initial fusion weights and dividing by the sum of all updated weights.

9. A method for monitoring emissions from chemical production tail gas, characterized in that, Includes the following steps: Collect spectral data and corresponding concentration data of chemical production tail gas, calculate the overlap weight distribution based on the ratio of response variance of each wavelength point to wavelength position interval, and perform differential width segmentation on the full-band spectral data according to the overlap weight distribution to generate a set of candidate sub-intervals. The cross-correlation coefficient and feature redundancy of each candidate sub-interval are calculated to obtain the interval quality factor. Based on the interval quality factor, the optimal combination of cooperative sub-intervals is selected from the set of candidate sub-intervals, and the corrected sub-interval covariance matrix is ​​constructed using the interval quality factor. Based on the corrected sub-interval covariance matrix, the redundancy-removing latent factor vector is extracted, the initial fusion weight of each sub-interval is determined by the local prediction residual and feature redundancy, and the full-band fusion regression coefficient matrix is ​​constructed based on the redundancy-removing latent factor vector. When acquiring the spectral data of the chemical exhaust gas to be tested, the response consistency deviation on the overlapping bands of adjacent sub-intervals is calculated. The initial fusion weight is adjusted according to the response consistency deviation to obtain the fusion weight. The fusion regression coefficient matrix is ​​reconstructed using the fusion weight. The monitoring value of the chemical exhaust gas emission concentration is calculated by combining the spectral data of the chemical exhaust gas to be tested.

10. The method according to claim 9, characterized in that, The calculation of the overlap weight distribution based on the ratio of the response variance at each wavelength point to the wavelength position interval includes: Obtain the response variance of each wavelength point in the full-band spectral data across all samples; Calculate the physical location interval between two adjacent wavelength points; The initial weight of a single point is obtained by dividing the response variance of each wavelength point by the positional interval of the corresponding adjacent wavelength point. The initial weight of the last wavelength point is calculated by using its positional interval with the adjacent previous wavelength point. A Gaussian smoothing function is used to smooth the initial weight sequence at a single point to generate a continuous overlapping weight distribution curve. The value of each point after smoothing is the sum of the initial weights at each point multiplied by the corresponding Gaussian weight factor. The Gaussian weight factor is calculated based on the square of the difference between the wavelength position of the wavelength point corresponding to the initial weight at a single point and the wavelength position of the current point, as well as the standard deviation parameter of the smoothing function.

Images