Infrared ultraviolet spectrum intelligent auxiliary analysis tool based on ai recognition

Abstract

This invention relates to the field of spectral analysis and processing technology, and discloses an AI-based intelligent auxiliary analysis tool for infrared and ultraviolet spectra, including an optical detection module, a clock synchronization controller, and a data processing device. The data processing device utilizes timestamps to construct a sliding time window, performs weighted aggregation on ultraviolet spectral data to align the infrared time series; extracts feature signals as anchor points for reconstruction and removes the theoretical absorption baseline of the principal components; performs an outer product operation on the infrared and ultraviolet residual matrices to generate a two-dimensional cross-covariance matrix; generates a feature matrix through singular value decomposition and low-rank reconstruction; converts the feature matrix into a one-dimensional eigenvector and inputs it into a partial least squares regression model, outputting impurity concentration data. Through a combination of hardware and software, strict time-domain alignment of heterogeneous spectra is achieved, and cross-domain covariance and orthogonal filtering effectively remove the principal component background and instrument noise, enabling stable and accurate prediction of trace impurity concentrations in dynamic flow systems.

Description

Technical Field

[0001] This invention relates to the field of spectral analysis and processing technology, specifically to an AI-based intelligent auxiliary analysis tool for infrared and ultraviolet spectra. Background Technology

[0002] In online flow systems such as liquid chromatography elution pipelines or continuous flow microchannel reactions, real-time monitoring of trace impurity concentration changes within the fluid plays a crucial role in process control. Conventional online spectroscopic monitoring often employs single infrared or ultraviolet spectroscopy techniques. However, in complex real-world fluids, high concentrations of the main component generate strong absorption baselines, severely masking the weak spectral signals of trace impurities. Single-spectral detection methods often suffer from significant overlap of characteristic bands and low signal-to-noise ratios, making it difficult to accurately identify and quantify structurally similar trace impurities.

[0003] To improve analytical accuracy, the introduction of combined infrared and ultraviolet (UV) spectroscopy has become a solution. However, in practical applications of such heterogeneous detection systems, the hardware operating mechanisms differ significantly. Infrared spectrometers typically rely on interferometers for step-by-step scanning, resulting in long single sampling periods; while UV spectrometers often employ array detectors for continuous snapshot scanning, with a data acquisition frequency far exceeding that of infrared spectrometers. This difference in mechanisms leads to severe misalignment in the time series of the two spectral data streams. Under flow conditions where the concentration of the analyte dynamically changes over time, if the high-frequency UV data and low-frequency infrared data cannot be strictly aligned in the time domain, subsequent algorithms will incorrectly correlate the spectral characteristics at different physical moments, resulting in analytical results that deviate from the true concentration.

[0004] Furthermore, existing spectral data processing methods lack effective decoupling mechanisms for the strong dynamic background of principal components when dealing with heterogeneous spectral streams. Directly stitching infrared and ultraviolet spectral data not only preserves the independent thermal noise of the two instruments but also fails to effectively separate cross-domain correlation signals driven by perturbations of the same impurity concentration, resulting in weak true absorptions still being submerged in instrument noise. Simultaneously, the combination of two instruments leads to a significant increase in the number of spatial detection channels. Simple data stacking introduces severe multicollinearity, making conventional parameter prediction models prone to matrix inversion problems or overfitting when calculating impurity concentrations, thus failing to meet the reliability requirements of industrial online real-time monitoring. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an AI-based intelligent auxiliary analysis tool for infrared and ultraviolet spectra, which solves the problems of data fusion under inconsistent sampling frequencies of heterogeneous spectrometers and the accurate extraction of perturbation concentration signals in low signal-to-noise ratio environments.

[0006] To achieve the above objectives, the present invention provides an intelligent auxiliary analysis tool for infrared and ultraviolet spectra based on AI recognition, comprising an optical detection module, a clock synchronization controller, and a data processing device. The data processing device includes: The time-domain warping module is used to construct a sliding time window using the timestamp of the acquired infrared detector as the reference axis, and to perform weighted aggregation operations on the ultraviolet spectral data falling within the sliding time window to generate an infrared time-series spectral matrix and an ultraviolet time-series spectral matrix.

[0007] By setting a unified time reference and using a weighted algorithm, the sampling time difference caused by different detection mechanisms is eliminated, and the alignment of multi-source heterogeneous spectral data in the time dimension is achieved.

[0008] The residual extraction module is used to extract feature signals from the infrared and ultraviolet time-series spectral matrices as anchor points, calculate the theoretical absorption baseline of the principal component in the fluid under test, and remove the theoretical absorption baseline to generate infrared and ultraviolet residual matrices. Utilizing the independent absorption characteristics of the principal component in specific bands as an internal concentration benchmark, it dynamically reverse-engineers and strips away strong absorption backgrounds, exposing trace impurity perturbation signals masked by the principal component signals.

[0009] The cross-domain mapping module performs matrix multiplication on the infrared and ultraviolet residual matrices in the time dimension to generate a two-dimensional cross-covariance matrix. Utilizing the physical correlation of the synchronous response induced by trace substance concentration perturbations in the dual-spectral channels, and the time-independent nature of the independent thermal noise of heterogeneous detectors, a cross-domain covariance mapping is established to achieve the accumulation of real physical absorption signals and the positive and negative cancellation of random thermal noise.

[0010] The orthogonal filtering module performs singular value decomposition on the two-dimensional cross-covariance matrix, extracts singular values ​​and corresponding eigenvectors, and performs low-rank matrix reconstruction to generate a feature matrix. By utilizing orthogonal transformation methods, the stable physical structure components and discrete residual background noise are separated in the mathematical space, thereby improving the spectral signal-to-noise ratio of the feature matrix.

[0011] The prediction output module is used to convert the feature matrix into a one-dimensional feature vector, input the one-dimensional feature vector into a partial least squares regression model, and output the spectral analysis results, which include the concentration data of impurities in the fluid to be tested.

[0012] Preferably, the time-domain warping module is specifically used for: Extract each infrared reference time node from the timestamp of the infrared detector, and construct the sliding time window at the time axis position where the infrared reference time node is located; Based on the constructed sliding time window, all the ultraviolet spectral data falling within the sliding time window are extracted; A Gaussian weighted average model is introduced to map the extracted ultraviolet spectral data to the infrared reference time node, and the weighted aggregation operation is completed to generate the infrared time series spectral matrix and the ultraviolet time series spectral matrix.

[0013] Preferably, the residual extraction module is specifically used for: In the infrared time series spectral matrix and the ultraviolet time series spectral matrix, the characteristic bands that belong only to the strong absorption of the principal component and do not overlap with other components are locked as the anchor points. Based on the locked anchor point, extract the spectral response intensity value corresponding to the anchor point at the current time node; Based on the extracted spectral response intensity values, the prior benchmark database is queried to reverse-calculate the actual concentration data of the principal component at the current moment; The full-spectrum data corresponding to the standard concentration gradient closest to the calculated actual concentration data is retrieved from the prior benchmark database, and the retrieved full-spectrum data is weighted along the concentration dimension using an interpolation algorithm to reconstruct the theoretical absorption baseline. The reconstructed theoretical absorption baseline is removed from the infrared time series spectral matrix and the ultraviolet time series spectral matrix, respectively, to generate the infrared residual matrix and the ultraviolet residual matrix.

[0014] Preferably, the cross-domain mapping module is specifically used for: Calculate the row mean vectors of the infrared residual matrix and the ultraviolet residual matrix along the time dimension respectively; Based on the calculated row mean vector, subtract the corresponding value in the row mean vector from each element in the infrared residual matrix and the ultraviolet residual matrix to complete the zero-mean processing of the residual data. Perform matrix transpose on the zero-mean UV residual matrix to generate a transposed UV residual matrix; In the time dimension, the matrix multiplication operation is performed on the infrared residual matrix after zero-mean processing and the generated transposed ultraviolet residual matrix to construct the two-dimensional cross covariance matrix.

[0015] This operation eliminates the static residual DC bias present in the time series, satisfying the mathematical premise for cross-domain covariance calculation.

[0016] Preferably, the orthogonal filtering module is specifically used for: Extract all non-zero singular values ​​from the singular value diagonal matrix generated by performing the singular value decomposition operation on the two-dimensional cross covariance matrix; The cumulative contribution rate of each extracted non-zero singular value is calculated in descending order of numerical value. When the calculated cumulative contribution rate is greater than or equal to the preset lower limit and less than or equal to the preset upper limit, the cutoff threshold integer is determined by combining the number of known independently changing impurity types in the fluid to be tested. Based on the established truncation threshold integer, the principal component singular values, the corresponding left singular vector and right singular vector are truncated and retained in a number equal to the truncation threshold integer as the feature vector. The matrix low-rank reconstruction is performed through matrix multiplication to generate the feature matrix.

[0017] Preferably, the partial least squares regression model in the prediction output module is pre-trained in the following manner: A standard sample set with a known trace impurity concentration gradient is injected into a simulated flow pipeline, causing the standard sample set to produce concentration gradient fluctuations. For the standard sample set in a dynamic circulation state, the operations corresponding to the time domain warping module, the residual extraction module, the cross-domain mapping module, and the orthogonal filtering module are executed sequentially to obtain the corresponding standard fusion feature row vector; The obtained standard fusion feature row vectors are fitted and trained with known nominal concentration labels, and the leave-one-out cross-validation method is used to determine the optimal number of latent variables to be retained in the partial least squares regression model. Based on the determined optimal number of latent variables, the prediction model parameters are generated and stored.

[0018] This pre-training algorithm projects high-dimensional and collinear input variables into a low-dimensional latent variable space, overcoming the technical limitation of conventional linear regression models that cannot invert high-dimensional feature variable matrices.

[0019] Preferably, the optical detection module is deployed in an online circulation system, and the optical detection module includes an infrared spectrometer connected to an infrared circulation cell and an ultraviolet spectrometer connected to an ultraviolet circulation cell, respectively. The clock synchronization controller establishes a physical communication connection with the motherboards of the infrared spectrometer and the ultraviolet spectrometer to synchronously send global acquisition start-up commands and record the timestamps of the infrared spectrometer and the ultraviolet spectrometer within the set global monitoring time window.

[0020] Preferably, the detection mechanism of the infrared spectrometer is interferometer step-scan detection, and the detection mechanism of the ultraviolet spectrometer is array detector continuous scan detection, and the single sampling period of the infrared spectrometer is longer than that of the ultraviolet spectrometer. The width of the sliding time window constructed in the time-domain warping module is set based on a preset multiple of the single sampling period of the infrared spectrometer.

[0021] Preferably, the online flow system is a liquid chromatography elution line or a continuous flow microchannel reactor line; During the operation of the online flow system, the fluid to be tested flows sequentially through the infrared flow cell and the ultraviolet flow cell.

[0022] Preferably, the data processing device is further configured to: Extract the concentration data output by the prediction output module and transmit the extracted concentration data to the human-computer interaction interface for real-time numerical display; Determine whether the concentration data is greater than or equal to a preset safety threshold. When the concentration data is determined to be greater than or equal to the preset safety threshold, send an alarm signal to an external programmable logic controller through a digital output interface. The preset safety threshold is a concentration limit constant pre-stored in the data processing device.

[0023] This invention provides an AI-based intelligent auxiliary analysis tool for infrared and ultraviolet spectra. It offers the following advantages: 1. This invention introduces a clock synchronization controller to record absolute timestamps and combines it with a time-domain warping module to construct a sliding time window. It then uses weighted aggregation operations to map high-frequency ultraviolet spectral data to low-frequency infrared reference time nodes. This hardware and software combined data alignment mechanism solves the problem of sampling frequency differences caused by different detection mechanisms, ensures strict synchronization of heterogeneous spectral data in the time series, avoids erroneous analysis results in subsequent correlation operations caused by unaligned data, and improves the reliability of online fluid monitoring data.

[0024] 2. This invention utilizes a residual extraction module to calculate and remove the theoretical absorption baseline of the principal components, and combines it with a cross-domain mapping module to perform an outer product operation on the infrared and ultraviolet residual matrices to generate a two-dimensional cross-covariance matrix. This takes advantage of the fact that the random thermal noise of heterogeneous spectrometers is uncorrelated in the time dimension, while the physical absorption changes of the same trace impurity molecule in the infrared and ultraviolet bands have strict time synchronization. The outer product operation makes the uncorrelated instrument noise cancel out during the summation process, thereby effectively separating and extracting the weak trace impurity cross-response signal from the high-concentration principal component background and instrument noise.

[0025] 3. This invention performs singular value decomposition and low-rank reconstruction on the two-dimensional cross-covariance matrix through an orthogonal filtering module. Then, the flattened one-dimensional eigenvector is input into the partial least squares regression model of the prediction output module for quantitative calculation. Singular value decomposition further truncates the remaining irregular high-frequency background noise at the algebraic level, while the partial least squares regression algorithm overcomes the mathematical defects of a large number of cross-domain spectral feature variables and serious multicollinearity through a dimension reduction projection mechanism, and realizes stable and accurate prediction of fluid trace impurity concentration under low signal-to-noise ratio dynamic conditions. Attached Figure Description

[0026] Figure 1 This is a system architecture diagram according to an embodiment of the present invention; Figure 2 This is a flowchart of a process according to an embodiment of the present invention; Figure 3 This is a bar chart comparing the root mean square error (RMSEP) of prediction under four different data processing modes in this embodiment of the invention. Figure 4 This is a bar chart comparing the system detection limit (LOD) under four different data processing modes in this embodiment of the invention; Figure 5 This is a line graph comparing the real-time measured value of impurity concentration within a local time window with the theoretically given benchmark value in an embodiment of the present invention. Detailed Implementation

[0027] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] See attached document Figure 1 This invention provides an AI-based intelligent auxiliary analysis tool for infrared and ultraviolet spectra, which may include: a time-domain normalization module, a residual extraction module, a cross-domain mapping module, an orthogonal filtering module, and a prediction output module.

[0029] The time-domain warping module is used to construct a sliding time window using the timestamps of the acquired infrared detectors as the reference axis, and to perform weighted aggregation operations on the ultraviolet spectral data falling within the sliding time window to generate infrared time-series spectral matrices and ultraviolet time-series spectral matrices. The residual extraction module is used to extract feature signals from the infrared time series spectral matrix and the ultraviolet time series spectral matrix as anchor points, calculate the theoretical absorption baseline of the principal component in the fluid under test, remove the theoretical absorption baseline, and generate the infrared residual matrix and the ultraviolet residual matrix. The cross-domain mapping module is used to perform matrix multiplication on the infrared residual matrix and the ultraviolet residual matrix in the time dimension to generate a two-dimensional cross covariance matrix. The orthogonal filtering module is used to perform singular value decomposition on the two-dimensional cross covariance matrix, extract singular values ​​and corresponding eigenvectors, perform low-rank matrix reconstruction, and generate a feature matrix. The prediction output module is used to convert the feature matrix into a one-dimensional feature vector, input the one-dimensional feature vector into the partial least squares regression model, and output the spectral analysis results, which include the concentration data of impurities in the fluid to be tested.

[0030] This technology is applied to online flow-through systems where the concentration of the analyte changes gradient over time. The online flow-through system can be a liquid chromatography elution line or a continuous flow microchannel reactor line.

[0031] The hardware topology includes an optical detection module, a clock synchronization controller, and a data processing unit. The optical detection module is deployed in the online flow system. The fluid to be measured flows sequentially through the infrared flow cell and the ultraviolet flow cell in the optical detection module. The infrared flow cell is connected to an infrared spectrometer, and the ultraviolet flow cell is connected to an ultraviolet spectrometer.

[0032] The detection mechanism of an infrared spectrometer is interferometric step-scan detection. The detection mechanism of an ultraviolet spectrometer is array detector continuous scanning detection. The single sampling period of an infrared spectrometer is longer than that of an ultraviolet spectrometer.

[0033] The clock synchronization controller is physically connected to the motherboards of both the infrared and ultraviolet spectrometers via a transistor-to-transistor logic trigger bus. The clock synchronization controller is used to synchronously send global acquisition start-up commands and record the instrument's absolute timestamp signal. The raw infrared and ultraviolet data streams, each with an absolute timestamp, are then transmitted to the data processing unit.

[0034] See attached document Figure 2 This invention provides an AI-based intelligent assisted analysis method for infrared and ultraviolet spectra, comprising the following steps: Within a set global monitoring time window, the data processing device receives infrared and ultraviolet spectrum streams with timestamps. The data processing device constructs a sliding time window using the low-frequency discrete timestamps of the infrared detector as a reference axis.

[0035] The data processing device performs a weighted aggregation operation on the ultraviolet spectral data falling within the sliding time window, and generates an infrared time-series spectral matrix and an ultraviolet time-series spectral matrix containing concentration gradient perturbation information.

[0036] The data processing device extracts characteristic signals from the time-series spectral matrix that do not contain trace impurities, using them as anchor points. Based on the characteristic anchor point data, the device retrieves a priori benchmark database and calculates the theoretical absorption baseline of the principal components over the time series. Simultaneously, matrix subtraction is used to remove the theoretical absorption baseline from the original sequence matrix, generating infrared and ultraviolet residual matrices that have been filtered to remove background signals from the principal components.

[0037] The data processing unit extracts the infrared and ultraviolet residual matrices and performs matrix multiplication operations in the time dimension. The data processing unit then generates a two-dimensional cross-covariance matrix characterizing the temporal co-variance rate of change of the signals from the infrared and ultraviolet detection channels.

[0038] The data processing unit performs singular value decomposition on the two-dimensional cross-covariance matrix. The unit extracts the preceding singular values ​​and their corresponding eigenvectors from the decomposition, performs low-rank matrix reconstruction, and generates the reconstructed feature matrix.

[0039] The data processing device performs a spatial flattening operation on the reconstructed feature matrix, converting the two-dimensional matrix into a one-dimensional feature vector. The data processing device then inputs this one-dimensional feature vector into a parameter discrimination model for computation, and outputs the structural category determination results of impurities in the tested fluid, as well as concentration data over time, through a front-end interface.

[0040] In this embodiment, the configured clock synchronization controller performs clock signal synchronization and data stream time domain alignment operations through a hardware link.

[0041] During the continuous flow of fluid through the detection cell, the concentration of the analyte changes continuously over time. If the sampling times of heterogeneous detectors are not aligned, subsequent algorithms will correlate misaligned infrared and ultraviolet data, resulting in erroneous analysis results. Therefore, this embodiment utilizes a hardware bus to address the time misalignment problem.

[0042] The clock synchronization controller, as the core hardware scheduling unit, is physically implemented using field-programmable gate arrays (FPGAs) or industrial data acquisition cards. It establishes a physical communication connection with the optical detection module motherboard via a transistor-to-transistor logic trigger bus. Within a set global monitoring time window, the clock synchronization controller captures the discrete timestamp sequence of the infrared detector and the high-frequency timestamp sequence of the ultraviolet detector, and transmits the spectral data stream containing absolute timestamps to the data processing device.

[0043] The data processing device reads the aforementioned infrared and ultraviolet timestamp sequences. Because the infrared spectrometer uses interferometer step scanning, while the ultraviolet spectrometer uses array detector continuous snapshot scanning, the acquisition frequency of ultraviolet spectrum data is much higher than that of the infrared spectrum. In this embodiment, the extracted infrared timestamp sequence is set to contain N infrared reference time nodes, and the ultraviolet timestamp sequence contains M high-frequency ultraviolet time nodes. Within the same global monitoring time window, the value of M is much larger than the value of N.

[0044] The data processing device establishes a time coordinate system based on each infrared reference time node in the infrared timestamp sequence. For any infrared reference time node in the infrared timestamp sequence... (K takes the value of an integer from 1 to N), the data processing device at the infrared reference time node A sliding time window is constructed at the location on the time axis. In this embodiment, the width of the sliding time window is set to 0.5 to 1.5 times the single sampling period of the infrared spectrometer to ensure that the sliding time window contains a sufficient amount of ultraviolet sampling data, while avoiding the situation where the time span is too large and thus masks the true concentration change gradient.

[0045] The data processing device extracts all raw ultraviolet (UV) spectral data whose time coordinates fall within a sliding time window. In this embodiment, a weighted aggregation operation is performed on the raw UV spectral data within the sliding time window using a preset algorithm. Specifically, the data processing device introduces a Gaussian weighted average model to map high-frequency UV spectral data to a low-frequency infrared reference time node.

[0046] Set in ultraviolet high frequency timestamp At that location, the ultraviolet detector operates at a specific wavelength. The original response strength of the channel is After weighted aggregation, it is mapped to the infrared reference time node. UV spectral response value The calculation formula is as follows: ; in, Indicates the infrared reference time node Corresponding to the sliding time window; the summation symbol indicates that the summation falls within the sliding time window. Within, corresponding to each ultraviolet high-frequency timestamp The ultraviolet spectral response values ​​are subjected to traversal and accumulation operations; The standard deviation parameter of the Gaussian weighting function is preset in the data processing device.

[0047] In this embodiment, the standard deviation parameter The value is positively correlated with the width of the sliding time window, and is usually set to 1 / 6 of the sliding time window width to ensure that the effective weights for aggregation calculation fall within the window range. The exponential function part serves as the weighting coefficient, used to control the distance from the infrared reference time node. The more recent the original data, the greater its weight in the aggregation result.

[0048] In this embodiment, the data processing device performs dynamic background reconstruction and nonlinear residual extraction operations on the generated infrared time series spectral matrix and ultraviolet time series spectral matrix.

[0049] In online flow systems, high concentrations of principal components generate strong absorption baselines, thus masking the spectral signals of trace impurities. Since the concentration of principal components is much higher than that of trace impurities, the spectral response of the principal components dominates in the detector. This embodiment utilizes the independent absorption characteristics of specific functional groups of the principal components, using the principal components as an internal concentration benchmark to achieve dynamic estimation and subtraction of the global background.

[0050] In practice, before actual measurement, high-purity main component standard solution and known trace impurity standard solution are injected into the flow cell, and standard solution spectra are collected respectively. The data processing device compares the above standard solution spectra and calibrates the characteristic signal of the response band without trace impurities as anchor points in the spectral channel space of the infrared time series spectral matrix and the ultraviolet time series spectral matrix.

[0051] Specifically, the data processing device locks onto characteristic bands that belong only to the strong absorption of the principal component and do not overlap with other components. In a preferred embodiment, the absorbance of the principal component in a specific band is set to account for more than 95% of the total signal response to ensure that the characteristic channel signal can represent the concentration change of the principal component. In this embodiment, the selected set of infrared anchor channels is denoted as... The set of ultraviolet anchor points is denoted as .

[0052] By calibrating the aforementioned characteristic anchor points, the data processing device can accurately reflect the concentration changes of the principal component at different physical moments without interference from trace impurities. This applies to any infrared reference time node in the infrared and ultraviolet time-series spectral matrices. The data processing device extracts the set of infrared anchor point channels at the current time point. With UV anchor point channel set The data processing device retrieves the spectral response intensity values ​​from the prior reference database based on the extracted spectral response intensity values. In this embodiment, the prior reference database is constructed by pre-preparing a series of standard principal component solutions with known concentration gradients and collecting full-channel absorption spectrum data.

[0053] According to Beer-Lambert's law, the absorbance of a substance is directly correlated with its concentration within a certain range. Based on the spectral response intensity of the anchor point at the current moment, the data processing device queries a priori reference database to inversely calculate the actual concentration data of the principal component at the current moment. Furthermore, the data processing device retrieves the full-spectrum data corresponding to the two nearest standard concentration gradients above and below this actual concentration data from the priori reference database.

[0054] The data processing device uses interpolation algorithms (such as linear or polynomial interpolation) to perform weighted calculations on the aforementioned standard full-spectrum data along the concentration dimension, thereby reconstructing the infrared reference time node. Theoretical absorption baseline column vector of the lower principal component across the full spectral channel range .

[0055] The data processing device iterates through all N infrared reference time nodes and calculates the N infrared theoretical absorption baseline column vectors. The infrared theoretical absorption baseline matrix is ​​constructed by splicing the data sequentially in chronological order. Similarly, the data processing device will convert N column vectors of theoretical ultraviolet absorption baselines. The data are spliced ​​together to form a theoretical ultraviolet absorption baseline matrix. Assuming the infrared detector has P spatial detection channels and the ultraviolet detector has q spatial detection channels, then the infrared theoretical absorption baseline matrix is... Dimensions Theoretical UV absorption baseline matrix Dimensions .

[0056] The data processing unit uses matrix subtraction to remove the theoretical absorption baseline from the original time-series spectral matrix, generating a residual matrix that filters out the background signal of the principal components. The specific calculation formula is as follows: ; ; In the formula, This represents the infrared time-series spectral matrix input to this module. This represents the calculated reconstructed infrared theoretical absorption baseline matrix. This represents the infrared residual matrix output after background removal; This represents the ultraviolet time-series spectral matrix input to this module. This represents the calculated reconstructed theoretical ultraviolet absorption baseline matrix. This represents the ultraviolet residual matrix output after background stripping.

[0057] After matrix subtraction, the strong absorption baseline of the high-concentration principal components is stripped away. The resulting infrared residual matrix... With the ultraviolet residual matrix Only the concentration perturbation signals of trace impurities at each infrared reference time point and the random thermal noise of the heterogeneous detector itself are retained. The above operation removes the dominant principal component interference data, providing a data foundation for the subsequent construction of the covariance eigenmap matrix.

[0058] In this embodiment, the data processing device uses the joint spectral change rate generated by the perturbation of trace substance concentration on the time axis to construct a cross-domain physical covariance matrix.

[0059] In online circulation systems, the concentration of trace impurities undergoes dynamic perturbations over time. Based on the principles of molecular spectroscopy, perturbations in specific molecular concentrations induce synchronous changes in both infrared vibrational energy levels and ultraviolet electronic transition energy levels. Since the infrared and ultraviolet spectrometers are independent hardware devices, the random thermal noise generated by their detectors is not correlated over time.

[0060] This embodiment utilizes the aforementioned physical synergistic properties to establish a cross-domain covariance mapping. This mapping operation is used to calculate the degree of linear correlation between each infrared band and each ultraviolet band over time. Band combinations with high correlation can be identified as the true physical absorption produced by the same impurity molecule.

[0061] In the specific implementation process, the data processing device retrieves the infrared residual matrix generated by the residual extraction operation. With the ultraviolet residual matrix As input variables for the operation. In this embodiment, the infrared residual matrix. Dimensions UV residual matrix Dimensions To meet the mathematical prerequisites for covariance calculation, the data processing device calculates the infrared residual matrix along the time dimension. Row mean vector and ultraviolet residual matrix Row mean vector. The data processing device subtracts the corresponding row mean from each element in the original matrix to perform zero-mean processing of the residual data. In this embodiment, this operation eliminates any static residual DC bias that may exist in the time series.

[0062] To perform cross-domain correlation operations, the data processing device performs zero-mean normalization on the ultraviolet residual matrix. Perform matrix transpose processing to generate a matrix with dimension 1. Transposed UV residual matrix .

[0063] The data processing device processes the zero-mean infrared residual matrix over time. With the transposed ultraviolet residual matrix Perform matrix multiplication to construct the two-dimensional cross-covariance matrix C. The specific calculation formula is as follows: ; In the formula, C represents the calculated two-dimensional cross-covariance matrix; N represents the total number of infrared reference time nodes, the specific value of which is determined by the global time window length and the infrared sampling frequency. This represents the infrared residual matrix after zero-mean processing; This represents the UV residual matrix after transpose and zero-mean normalization. After the matrix multiplication operations described above, the resulting two-dimensional cross-covariance matrix C has its dimension mapped from the time series dimension to the spatial band dimension, with a specific dimension size of... .

[0064] The elements within the two-dimensional cross-covariance matrix C represent the physical mapping relationship between the coordinated rate of change of the infrared and ultraviolet detection channels over time. Specifically, the element value corresponding to any row and column index in the two-dimensional cross-covariance matrix C represents the cross-covariance of a specific infrared wavelength channel and a specific ultraviolet wavelength channel within a set time window. Since the independent thermal noise of heterogeneous detectors is randomly distributed over time and has no cross-correlation, the positive and negative thermal noise cancel each other out during matrix multiplication and summation, and the corresponding covariance term value approaches 0. Conversely, trace impurities share the same physical concentration driving source in the absorbance changes within the dual detection domains, exhibiting temporal synchronicity. Therefore, the characteristics of associated impurity signals are accumulated and preserved in the two-dimensional cross-covariance matrix C.

[0065] This embodiment constructs a cross-domain physical covariance matrix to transform the irregular instrument noise floor in a one-dimensional time series into covariance element values ​​approaching zero. This operation establishes a mathematical model for the orthogonal separation of signal and noise, mapping physically corresponding cross-domain spectral features into a two-dimensional matrix space, providing data support for subsequent extraction of trace component features in low signal-to-noise ratio environments.

[0066] In this embodiment, the data processing device utilizes the difference between the thermal noise independence of the two instruments and the physical correlation of trace signals to perform orthogonal decoupling and noise reduction operations through singular value decomposition.

[0067] The absorption signals of trace substances containing true concentration perturbations in a two-dimensional cross-covariance matrix have a fixed spectral structure and temporal synchronization, concentrating on the principal component singular values ​​with larger values ​​after singular value decomposition. Conversely, the small amount of incompletely canceled instrument noise remaining after the pre-processing steps, lacking stable structural characteristics, will be uniformly dispersed in the sequence of smaller singular values. This embodiment utilizes the orthogonal transformation method in linear algebra to separate the components of different physical properties in mathematical space.

[0068] In the specific implementation process, the data processing device performs singular value decomposition on the two-dimensional cross-covariance matrix C. The specific decomposition formula is as follows: ; In the formula, C represents the dimension as follows: The two-dimensional cross-covariance matrix, where P represents the total number of infrared detection channels, q represents the total number of ultraviolet detection channels; U represents the dimensionless cross-covariance matrix. The left singular vector matrix represents the spatial characteristics of the infrared detection channel; The dimension is A diagonal matrix of singular values, with the singular values ​​on the diagonal arranged in descending order; The dimension is The right singular vector matrix transpose represents the spatial characteristics of the ultraviolet detection channel.

[0069] After decomposition, the data processing device establishes a truncation threshold integer K. In this embodiment, the rule for establishing the truncation threshold integer K is based on the combined constraints of the cumulative contribution rate of singular values ​​and the prior physical information of the fluid system. The data processing device extracts the singular value diagonal matrix. For all non-zero singular values, calculate the proportion of the square of each singular value to the sum of the squares of all singular values. The data processing device accumulates the singular values ​​in descending order. When the cumulative contribution rate reaches a preset range of 95% to 99%, a preliminary calculated value is obtained. Simultaneously, the data processing device retrieves the number of known independently varying trace impurities to be measured within the flow system, using this number as a physical constraint. The data processing device compares the preliminary calculated value with the number of independently varying impurities, typically using the number of independently varying impurities as the final cutoff threshold integer K. In conventional fluid trace detection systems, due to the limited number of analyte components, the cutoff threshold integer K is usually between 1 and 5.

[0070] After establishing the truncation threshold integer K, the data processing device performs feature extraction and matrix reconstruction operations. The data processing device truncates the singular value diagonal matrix. The top-ranked principal component singular values ​​are used to construct a new truncated diagonal matrix, retaining the first K singular values. Simultaneously, the data processing device extracts the first K columns of the left singular vector matrix U to form the feature matrix. And extracting the transpose matrix of the right singular vector. The first K rows constitute the feature matrix .

[0071] The data processing unit performs matrix multiplication on the retained singular values ​​and eigenvectors to reconstruct the feature matrix after filtering out high-frequency random noise. The specific calculation formula is as follows: ; In the formula, This represents the reconstructed feature matrix, with dimensions maintained at 1. ; This represents the truncated left singular vector matrix; This represents a truncated diagonal matrix containing the singular values ​​of the first K principal components; This represents the transpose of the truncated right singular vector.

[0072] After singular value truncation and reconstruction operations, the characteristic matrix Only cross-domain correlation signals driven by fluctuations in the actual trace molecule concentrations are retained. This embodiment uses a low-rank approximation operation to filter out residual random noise from the instrument, thereby improving the signal-to-noise ratio of the two-dimensional feature data spectrum and providing reliable input variables for subsequent prediction models.

[0073] In this embodiment, the data processing device calculates and outputs real-time concentration data of trace impurities in the fluid under test based on the feature matrix generated after orthogonal noise reduction and combined with a pre-trained chemometric model.

[0074] In the specific implementation process, the data processing device extracts the feature matrix output from the previous steps. To meet the dimensionality requirements of input data for conventional chemometrics prediction models, the data processing device performs feature matrix... Perform a vectorization unpacking operation. Specifically, the data processing device will unpack the data into vectorized units of dimension 1. The feature matrices are concatenated in row-major order, generating a dimension of . The fused feature row vector X. In this embodiment, the two-dimensional spatial band cross-correlation features are reconstructed into a one-dimensional input sequence through the matrix flattening operation described above.

[0075] In practical spectral analysis systems, the number of infrared and ultraviolet detection channels is typically large, resulting in tens of thousands of feature variables in the expanded fused feature row vector. Due to synchronous spectral changes caused by perturbations of the same physical concentration, severe multicollinearity exists among these feature variables. Therefore, in this embodiment, the data processing device is equipped with a pre-trained partial least squares regression model. When establishing a linear mapping, the partial least squares regression algorithm can project high-dimensional and highly collinear input variables into a low-dimensional latent variable space, thereby effectively extracting feature components associated with concentration changes. This algorithm overcomes the limitation of traditional linear regression models where matrices cannot be inverted under such data.

[0076] Before online monitoring, the data processing device performs pre-training using a standard sample set with known trace impurity concentration gradients. To ensure that the training data is consistent with the physical characteristics of the online monitoring conditions, the aforementioned standard sample set is sequentially injected into a simulated flow pipeline containing fluid pumping and dynamic mixing, so that the standard samples also produce slight concentration gradient fluctuations as they flow through the detection cell.

[0077] The data processing device repeatedly performs time-domain warping, residual extraction, cross-domain mapping, and orthogonal filtering operations on a standard sample set in a dynamic flow state to obtain the corresponding standard fused feature row vectors. The pre-training process uses these standard fused feature row vectors to fit and train the model against known nominal concentration labels. During the training phase, the data processing device typically employs leave-one-out cross-validation to determine the optimal number of latent variables to retain in the partial least squares regression model.

[0078] In this embodiment, the data processing device generates and stores prediction model parameters containing a column vector W of regression coefficients and a bias constant b based on the optimal number of latent variables.

[0079] In the online prediction phase, the data processing device inputs the real-time acquired fusion feature row vector x into the pre-trained partial least squares regression model. The specific concentration calculation formula is as follows: ; In the formula, y represents the predicted trace impurity concentration value; x represents the flattened fused feature row vector with dimension 1. w represents a column vector of regression coefficients generated from training on the standard sample set, with dimension 1. b represents the bias constant generated by training with the standard sample set.

[0080] After acquiring the trace impurity concentration value y, the data processing device transmits the prediction result to the human-machine interface for real-time display and simultaneously stores it in the industrial database for retrospective analysis. When the concentration value y is greater than or equal to a preset safety threshold, the data processing device sends an alarm signal to an external programmable logic controller through a digital output interface.

[0081] In this embodiment, the preset safety threshold is pre-written into the memory of the data processing device by the operator according to the purity specifications of the specific fluid production process. To allow for response time for process adjustments, the preset safety threshold is typically set to 80% to 90% of the upper limit of impurity concentration allowed by the fluid product quality standard. This embodiment achieves automated monitoring of trace impurities in the online flow system through the aforementioned data transfer and threshold determination mechanism.

[0082] Specific application example: Online monitoring of trace drug impurities in a continuous flow microchannel reactor. Implementation scenarios and hardware deployment: In a continuous flow microchannel synthesis process for an antitumor drug intermediate, the concentration of the main product (main component) is typically around 100 mg / mL, while the concentration of mutagenic trace impurities (analytes) generated by accompanying side reactions is only 0.05 mg / mL to 0.5 mg / mL. The online flow system tubing sequentially connects an infrared flow cell (stepped scanning, single sampling period of 1 second, i.e., 1 Hz) and an ultraviolet flow cell (array detection continuous snapshot, sampling frequency of 10 Hz). The system is equipped with a clock synchronization controller based on an industrial data acquisition card, aligning absolute timestamps via a TTL bus.

[0083] Data Acquisition and Time Domain Warping: Within a set global monitoring time window (e.g., 60 seconds), the data processing device receives 60 infrared reference time nodes (N=60) and 600 ultraviolet high-frequency time nodes (M=600). The data processing device uses infrared nodes... Construct a sliding time window with a width of 1 second centered on the target element. Using the preset standard deviation parameter The Gaussian weighted average model maps 10Hz high-frequency ultraviolet data to a 1Hz low-frequency infrared time node. According to the formula of this invention: ; Successfully generated time-aligned infrared and ultraviolet time-series spectral matrices.

[0084] Residual extraction and cross-domain mapping: Because the main component concentration is as high as 100 mg / mL, its strong absorption baseline can mask impurity signals. The system locks the main component at 1750 cm⁻¹ infrared. -1 By identifying characteristic anchor points at 100 nm and 270 nm in the ultraviolet region, the actual concentration of the principal components was deduced, and the theoretical absorption baseline of the full spectrum was reconstructed by referring to a table. and By matrix subtraction and Background removal. Then, the generated residual matrix is ​​zero-mean normalized, and matrix multiplication is performed along the time dimension to construct the cross-domain cross covariance matrix: .

[0085] Orthogonal decoupling noise reduction and prediction output: Perform singular value decomposition on a two-dimensional matrix C containing significant instrument thermal noise. Since there is only one independently varying trace impurity in the system, the system selects forward singularities with a cumulative contribution rate greater than 98% (cutoff threshold K=1), and uses the formula... Perform low-rank reconstruction. Finally, The space is flattened into a one-dimensional fused feature row vector X, which is input into a pre-trained partial least squares regression (PLS) model. It outputs the concentration change curve of the mutagenic impurity in real time. When predicting the concentration... When this happens, the system will automatically trigger an alarm signal.

[0086] Experimental verification and effect comparison: Experimental verification and effect comparison To test the data processing flow of this system, a prepared validation sample was injected into the fluid pipeline. The validation sample contained the main product and known specific impurities, with the impurity concentration variable set to fluctuate within the range of 0.01% to 0.50%, while the main product concentration remained at the process standard level.

[0087] Four different data processing modes were set up for comparative analysis in the test. The first group used only data collected by a single ultraviolet detector to build a partial least squares (PLS) model. The second group used only data collected by a single infrared detector to build a PLS model. The third group extracted data from both ultraviolet and infrared dual-channel detectors, but did not perform timestamp alignment and weighted aggregation operations on the hardware bus; instead, it directly concatenated the misaligned data and input it into the model. The fourth group used the complete data processing program of this invention, which includes time-domain warping, cross-domain cross-covariance calculation, and orthogonal decomposition and reconstruction.

[0088] See attached document Figure 3 The following are the statistical values ​​for the root mean square error (RMSEP) of prediction. The RMSEP for a single ultraviolet channel is 0.122%. The RMSEP for a single infrared channel is 0.085%. The dual-channel stitching method, which does not include a time-domain warming step, introduces incorrect temporal variable correspondences, leading to an RMSEP of 0.151%. After processing using the method of this invention, the measured RMSEP is 0.015%.

[0089] See attached document Figure 4 The system detection limit (LOD) statistics are listed. The system detection limits obtained from the first to third comparison modes are 0.150%, 0.100%, and 0.180%, respectively. After the multi-dimensional noise reduction and feature extraction of this invention, the system detection limit reaches 0.008%, indicating that the data cleaning algorithm reduces the masking effect of instrument random thermal noise on weak signals.

[0090] See attached document Figure 5 The diagram shows a line graph illustrating impurity concentration monitoring within a local time window. This data is derived from a comparison of continuous measurement output values ​​from the fourth experimental model within a set time period with theoretically given values. The concentration curve trajectory calculated by the model closely matches the actual injected fluid concentration gradient fluctuation trend, without significant hysteresis or distortion. Tests confirm that this method can output reliable concentration parameters under online flow conditions.

Claims

1. An AI-based intelligent auxiliary analysis tool for infrared and ultraviolet spectra, comprising an optical detection module, a clock synchronization controller, and a data processing device, characterized in that, The data processing device includes: The time-domain warping module is used to construct a sliding time window using the timestamp of the acquired infrared detector as the reference axis, and to perform weighted aggregation operations on the ultraviolet spectral data falling within the sliding time window to generate an infrared time-series spectral matrix and an ultraviolet time-series spectral matrix. The residual extraction module is used to extract feature signals as anchor points from the infrared time series spectral matrix and the ultraviolet time series spectral matrix, calculate the theoretical absorption baseline of the principal component in the fluid to be tested, and remove the theoretical absorption baseline to generate the infrared residual matrix and the ultraviolet residual matrix. The cross-domain mapping module is used to perform matrix multiplication on the infrared residual matrix and the ultraviolet residual matrix in the time dimension to generate a two-dimensional cross covariance matrix. The orthogonal filtering module is used to perform singular value decomposition on the two-dimensional cross covariance matrix, extract singular values ​​and corresponding eigenvectors, perform low-rank matrix reconstruction, and generate a feature matrix. The prediction output module is used to convert the feature matrix into a one-dimensional feature vector, input the one-dimensional feature vector into a partial least squares regression model, and output the spectral analysis results, which include the concentration data of impurities in the fluid to be tested.

2. The AI-based intelligent auxiliary analysis tool for infrared and ultraviolet spectra according to claim 1, characterized in that, The time-domain warping module is specifically used for: Extract each infrared reference time node from the timestamp of the infrared detector, and construct the sliding time window at the time axis position where the infrared reference time node is located; Based on the constructed sliding time window, all the ultraviolet spectral data falling within the sliding time window are extracted; A Gaussian weighted average model is introduced to map the extracted ultraviolet spectral data to the infrared reference time node, and the weighted aggregation operation is completed to generate the infrared time series spectral matrix and the ultraviolet time series spectral matrix.

3. The AI-based intelligent auxiliary analysis tool for infrared and ultraviolet spectra according to claim 1, characterized in that, The residual extraction module is specifically used for: In the infrared time series spectral matrix and the ultraviolet time series spectral matrix, the characteristic bands that belong only to the strong absorption of the principal component and do not overlap with other components are locked as the anchor points. Based on the locked anchor point, extract the spectral response intensity value corresponding to the anchor point at the current time node; Based on the extracted spectral response intensity values, the prior benchmark database is queried to reverse-calculate the actual concentration data of the principal component at the current moment; The full-spectrum data corresponding to the standard concentration gradient closest to the calculated actual concentration data is retrieved from the prior benchmark database, and the retrieved full-spectrum data is weighted along the concentration dimension using an interpolation algorithm to reconstruct the theoretical absorption baseline. The reconstructed theoretical absorption baseline is removed from the infrared time series spectral matrix and the ultraviolet time series spectral matrix, respectively, to generate the infrared residual matrix and the ultraviolet residual matrix.

4. The AI-based intelligent auxiliary analysis tool for infrared and ultraviolet spectra according to claim 1, characterized in that, The cross-domain mapping module is specifically used for: Calculate the row mean vectors of the infrared residual matrix and the ultraviolet residual matrix along the time dimension respectively; Based on the calculated row mean vector, each element in the infrared residual matrix and the ultraviolet residual matrix is ​​subtracted from the corresponding value in the row mean vector to complete the zero-mean processing of the residual data. Perform matrix transpose on the zero-mean UV residual matrix to generate a transposed UV residual matrix; In the time dimension, the matrix multiplication operation is performed on the infrared residual matrix after zero-mean processing and the generated transposed ultraviolet residual matrix to construct the two-dimensional cross covariance matrix.

5. The AI-based intelligent auxiliary analysis tool for infrared and ultraviolet spectra according to claim 1, characterized in that, The orthogonal filtering module is specifically used for: Extract all non-zero singular values ​​from the singular value diagonal matrix generated by performing the singular value decomposition operation on the two-dimensional cross covariance matrix; The cumulative contribution rate of each extracted non-zero singular value is calculated in descending order of numerical value. When the calculated cumulative contribution rate is greater than or equal to the preset lower limit and less than or equal to the preset upper limit, the cutoff threshold integer is determined by combining the number of known independently changing impurity types in the fluid to be tested. Based on the established truncation threshold integer, the principal component singular values, the corresponding left singular vector and right singular vector are truncated and retained in a number equal to the truncation threshold integer as the feature vector. The matrix low-rank reconstruction is performed through matrix multiplication to generate the feature matrix.

6. The AI-based intelligent auxiliary analysis tool for infrared and ultraviolet spectra according to claim 1, characterized in that, The partial least squares regression model in the prediction output module is pre-trained in the following manner: A standard sample set with a known trace impurity concentration gradient is injected into a simulated flow pipeline, causing the standard sample set to produce concentration gradient fluctuations. For the standard sample set in a dynamic circulation state, the operations corresponding to the time domain warping module, the residual extraction module, the cross-domain mapping module, and the orthogonal filtering module are executed sequentially to obtain the corresponding standard fusion feature row vector; The obtained standard fusion feature row vectors are fitted and trained with known nominal concentration labels, and the leave-one-out cross-validation method is used to determine the optimal number of latent variables to be retained in the partial least squares regression model. Based on the determined optimal number of latent variables, the prediction model parameters are generated and stored.

7. The AI-based intelligent auxiliary analysis tool for infrared and ultraviolet spectra according to claim 1, characterized in that, The optical detection module is deployed in the online circulation system, and the optical detection module includes an infrared spectrometer connected to an infrared circulation cell and an ultraviolet spectrometer connected to an ultraviolet circulation cell, respectively. The clock synchronization controller establishes a physical communication connection with the motherboards of the infrared spectrometer and the ultraviolet spectrometer to synchronously send global acquisition start-up commands and record the timestamps of the infrared spectrometer and the ultraviolet spectrometer within the set global monitoring time window.

8. The AI-based intelligent auxiliary analysis tool for infrared and ultraviolet spectra according to claim 7, characterized in that, The infrared spectrometer uses an interferometer step-scan detection mechanism, while the ultraviolet spectrometer uses an array detector continuous scan detection mechanism. Furthermore, the single sampling period of the infrared spectrometer is longer than that of the ultraviolet spectrometer. The width of the sliding time window constructed in the time-domain warping module is set based on a preset multiple of the single sampling period of the infrared spectrometer.

9. The AI-based intelligent auxiliary analysis tool for infrared and ultraviolet spectra according to claim 7, characterized in that, The online flow system is a liquid chromatography elution pipeline or a continuous flow microchannel reactor pipeline; During the operation of the online flow system, the fluid to be tested flows sequentially through the infrared flow cell and the ultraviolet flow cell.

10. The AI-based intelligent auxiliary analysis tool for infrared and ultraviolet spectra according to claim 1, characterized in that, The data processing device is also used for: Extract the concentration data output by the prediction output module and transmit the extracted concentration data to the human-computer interaction interface for real-time numerical display; Determine whether the concentration data is greater than or equal to a preset safety threshold. When the concentration data is determined to be greater than or equal to the preset safety threshold, send an alarm signal to an external programmable logic controller through a digital output interface. The preset safety threshold is a concentration limit constant pre-stored in the data processing device.

Images