Machine learning based multi-module sequence data recognition and analysis method

Abstract

The application discloses a multi-module sequence data identification and analysis method based on machine learning for chemical in-situ and working condition characterization, and comprises the following steps: reading chemical characterization sequence data files to form an ordered measurement sequence, completing verification, background deduction and smoothing denoising; setting a background suppression channel and a peak shape continuation channel, identifying background changes and peak shoulder changes; cutting off the propagation of the slow change area, opening a one-way propagation entrance; opening the candidate peak state through the same position closed verification; activating the peak shape competition path according to the peak shoulder span and the peak top relative amplitude; searching and reviewing the integral boundary from the peak shape competition path, calculating the peak area; and returning the review offset to the peak shape competition path to generate spatial distribution imaging results and intensity evolution trend sequence results after stabilization. The application relates to the technical field of chemical in-situ and working condition characterization and intelligent data processing and analysis, effectively improves the peak recognition accuracy and integral stability, and improves the efficiency of batch imaging and trend analysis.

Description

Technical Field

[0001] This invention relates to the field of chemical in-situ and operational condition characterization and intelligent data processing and analysis technology, and in particular to a method for multi-module sequence data identification and parsing based on machine learning. Background Technology

[0002] In research fields such as chemistry, materials science, electrochemistry, and catalysis, researchers often use Raman spectroscopy, infrared spectroscopy, ultraviolet-visible spectroscopy, chromatography, and mass spectrometry to monitor the composition, structure, and dynamic evolution of chemical systems during reactions. These tests generate chemical characterization sequences arranged by wavenumber, wavelength, retention time, mass-to-charge ratio, sampling time, or spatial sampling location. Taking in-situ Raman spectroscopy of lithium-sulfur batteries as an example, the testing process collects spectra at different times and spatial locations, resulting in a massive dataset encompassing time, space, and spectral intensity. Subsequent steps require background subtraction, smoothing and denoising, characteristic peak identification, peak area calculation, spatial imaging, batch aggregation, and trend analysis to elucidate the evolutionary patterns of chemical components such as sulfur species.

[0003] Current processing methods largely rely on segmented processing using spectrometer-specific software, scientific plotting software, spreadsheet software, and scripting tools. Data conversion between different software programs is difficult, processing parameters are hard to save uniformly, and peak positions, peak areas, and spatial coordinates require manual processing. When dealing with batch folders, continuous spatiotemporal conditions, and multimodal data, the processing chain is long, repetitive operations are numerous, and result verification costs are high.

[0004] Furthermore, existing baseline correction and peak integration methods often employ fixed rules, which have limitations when processing chemical characterization sequence data. Firstly, background misjudgment: When background drift is strong and the signal-to-noise ratio is low (e.g., fluorescence residue in Raman spectroscopy, water vapor absorption interference in infrared spectroscopy), non-characteristic background is easily misjudged as a true signal peak under a fixed threshold. Secondly, significant peak shape differences: Narrow, strong peaks (e.g., crystalline components) and broad, weak peaks (e.g., amorphous components) often coexist in the same spectrum. Fixed integration boundaries cannot account for the shoulder continuity of both types of peaks, leading to increased quantitative analysis errors. Thirdly, severe peak overlap: In in-situ reactions, the peaks of intermediate products and main products often overlap. Fixed rules lack peak shape discrimination and integration boundary verification mechanisms, making it difficult to accurately attribute contributions from each component. Additionally, when writing coordinate information into the filename, manual extraction of coordinates, grid construction, and correction of the scanning direction are required. The root cause of these problems lies in the lack of closed-loop constraints in existing schemes for peak shape recognition, boundary verification, and integrated output, making it difficult to meet the requirements for accurate identification and integrated analysis of multi-module chemical characterization sequence data.

[0005] Therefore, providing an identification and analysis method capable of automatic identification, integration, imaging, and batch trend analysis for multi-module sequence data in chemical characterization is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to propose a multi-module sequence data identification and analysis method based on machine learning for processing chemical characterization sequence data generated in spectroscopy, chromatography, and mass spectrometry. Addressing issues such as residual fluorescence in Raman spectroscopy, background absorption interference in infrared spectroscopy, discrete peak signals in mass spectrometry sequences, and common chemical reaction phenomena including slowly varying backgrounds, narrow strong peaks, wide weak peaks, overlapping peaks, and low signal-to-noise ratio weak peaks, this invention employs a peak-gated structured state-space sequence model. This model integrates background suppression, peak shoulder tracking, peak competition, integral boundary verification and reinjection, and imaging trend output into a single processing link. This allows different chemical spectral peaks and signal morphologies to be identified, integrated, imaged, and batch trend analyzed along corresponding paths, offering advantages such as accurate peak identification, stable integration, and high batch processing efficiency.

[0007] The multi-module sequence data identification and parsing method based on machine learning according to embodiments of the present invention includes the following steps: Read chemical characterization sequence data files from spectroscopy, chromatography, and mass spectrometry tests, form an ordered measurement sequence, and after verification, background subtraction, smoothing and denoising, send it into a peak-gated structured state-space sequence model; Within the peak-gated structured state-space sequence model, a background suppression channel and a peak continuation channel are set up. The background suppression channel divides the gradual change region and the abrupt bulge region according to the difference before and after background subtraction, driving state locking. The peak continuation channel drives the peak shoulder continuation according to the adjacent sampling changes after smoothing and denoising. The background suppression channel is blocked in the gradual change region and the propagation is cut off in the sudden change region. The unidirectional propagation entrance is opened in the sudden change convex region. The peak continuation channel continuously tracks the peak shoulder continuation state along both sides of the unidirectional propagation entrance. Perform a co-position closure check on the unidirectional propagation inlet and the continuous state of the two side peaks, and only open the candidate peak state when the co-position closure is successful. When a candidate peak enters the peak competition path, the narrow peak path, medium peak path, and wide peak path are activated mutually exclusively according to the peak shoulder span, and the strong peak path and weak peak path are activated mutually exclusively according to the relative amplitude of the peak. The integration boundary search is driven by the activated peak competition path. The integration boundary is locked by triple verification of attenuation, width and local baseline. The integration boundary is then checked and the peak area is calculated after subtracting the local baseline. The search position offset of the integral boundary verification is reversed and fed back to the peak competition path. When the preset offset threshold is reached, the peak competition path is closed and the search is restarted. When the preset offset threshold is not reached, the spatial distribution imaging results and intensity evolution trend sequence results are generated according to the file name coordinate rules.

[0008] Optionally, the formation process of the ordered measurement sequence specifically includes: The chemical characterization sequence data file is derived from at least one of the following: Raman spectroscopy, infrared spectroscopy, ultraviolet-visible spectroscopy, gas chromatography, liquid chromatography, and mass spectrometry, obtained from in-situ and operational tests of chemical research objects and chemical reaction processes. The chemical characterization sequence data file is parsed according to the file format, and the order of independent variables and the order of measurement amplitudes are read. The order of the independent variables corresponds to one of wavenumber, wavelength, retention time, mass-to-charge ratio, sampling time, and spatial sampling coordinates; the order of the measurement amplitudes corresponds to one of Raman scattering intensity, infrared absorption intensity, absorbance, chromatographic detection signal, and mass spectrometry ion count value. The independent variable order is paired with the measurement amplitude order point by point according to the original sampling order. Sampling positions that are not paired are not written into the ordered measurement sequence. The paired sampling locations are sequentially numbered to form an ordered measurement sequence.

[0009] Optionally, the background suppression channel is divided into a gradual change region and a sudden abrupt change region based on the difference before and after background subtraction, specifically including: The difference between the background before and after subtraction is read point by point along the direction of the ordered measurement sequence, and the difference between the background before and after subtraction at adjacent sampling positions is compared. Sampling locations where continuous differences fall within a preset gradual variation range are marked as gradual variation regions, and sampling locations where differences deviate from the preset gradual variation range and the background subtracts the difference to form a local bulge are marked as abrupt bulge regions. The background suppression channel writes a state lockout flag to the gradually changing region and a unidirectional propagation entrance open flag to the abruptly changing protruding region.

[0010] Optionally, the background suppression channel writes a state lockout flag to the gradually changing region and a unidirectional propagation entrance open flag to the abruptly changing protrusion region, specifically including: Background suppression channel read status lock flag, sets the propagation status of the sampling position in the gradual change zone to the lock state, and cuts off the propagation of the lock state to the adjacent sampling position; The background suppression channel reads the unidirectional propagation entrance open marker, sets the sampling position with the largest difference before and after background subtraction in the mutation protrusion region as the unidirectional propagation entrance, retains the propagation link from the unidirectional propagation entrance to the co-position closure check, and closes the propagation link from the sampling positions in the mutation protrusion region other than the unidirectional propagation entrance to the co-position closure check.

[0011] Optionally, the peak-shaped continuity channel continuously tracks the peak shoulder continuity state along both sides of the unidirectional propagation entrance, specifically including: The peak-shaped continuation channel starts from the unidirectional propagation entrance and reads the smoothed and denoised adjacent sampling changes at the forward and backward sampling positions along the direction of the ordered measurement sequence. Sampling positions where the direction of change remains consistent and the amplitude of change decreases are written as peak shoulder continuation state, and sampling positions where the direction of change reverses and the amplitude of change reaches the preset abrupt change threshold are written as peak shoulder termination state. At the peak shoulder termination state, the propagation of the peak shoulder continuation state to the outer sampling position is interrupted.

[0012] Optionally, the step of performing a co-location closure check on the unidirectional propagation entrance and the continuous state of the two side peaks specifically includes: Using the sampling position where the unidirectional propagation entrance is located as the verification center, read the peak and shoulder continuation status of the continuous sampling positions on the front side and the peak and shoulder continuation status of the continuous sampling positions on the back side. When the peak shoulder continuity state of the current continuous sampling position, the unidirectional propagation entrance, and the peak shoulder continuity state of the subsequent continuous sampling position are continuously connected in the direction of the ordered measurement sequence, it is determined that the co-position closure is established and the candidate peak state is opened. If any of the peak shoulder continuation states at the front continuous sampling position or the rear continuous sampling position is disconnected from the unidirectional propagation entrance, the co-position closure is determined to be invalid, and the candidate peak state is closed.

[0013] Optionally, the mutual exclusion activation specifically includes: The peak-shaped competition path reads the peak shoulder continuation state of the consecutive sampling positions on the front side and the peak shoulder continuation state of the consecutive sampling positions on the back side, and calculates the sampling interval between the peak shoulder continuation state of the consecutive sampling positions on the front side and the peak shoulder continuation state of the consecutive sampling positions on the back side as the peak shoulder span. Based on the peak shoulder span, select one path from the narrow peak path, medium peak path, and wide peak path to open, and close the unselected path; The peak-shaped competition path reads the amplitude corresponding to the unidirectional propagation entrance, the amplitude corresponding to the peak shoulder continuation state at the front continuous sampling position, and the amplitude corresponding to the peak shoulder continuation state at the rear continuous sampling position. The difference between the amplitude corresponding to the unidirectional propagation entrance and the average of the amplitude corresponding to the peak shoulder continuation state at the front continuous sampling position and the amplitude corresponding to the peak shoulder continuation state at the rear continuous sampling position is taken as the peak relative amplitude. Based on the relative amplitude of the peak, one path is selected from the strong peak path and the weak peak path to be opened, and the unselected path is closed.

[0014] Optionally, the integral boundary search driven by the activated peak-shaped competition path specifically includes: The peak position is taken as the sampling position corresponding to the unidirectional propagation entrance. The boundary search range is determined by one of the open paths among the narrow peak path, medium peak path, and wide peak path. The attenuation verification threshold is determined by one of the open paths among the strong peak path and weak peak path. Search from the peak position along the direction of the ordered measurement sequence to the front and back sampling positions, and perform attenuation verification, width verification, and local baseline verification point by point; The front sampling position that simultaneously passes attenuation verification, width verification, and local baseline verification is locked as the left integration boundary, and the rear sampling position that simultaneously passes attenuation verification, width verification, and local baseline verification is locked as the right integration boundary.

[0015] Optionally, the step of performing a verification search on the integral boundary and calculating the peak area after subtracting local baselines specifically includes: Using the left and right integration boundaries as initial boundaries, a verification search is performed outside the left and right integration boundaries along the direction of the ordered measurement sequence to obtain the positions of the left and right verification boundaries. The position of the left verification boundary is compared with the left integration boundary, and the position of the right verification boundary is compared with the right integration boundary to obtain the offset of the integration boundary verification search position. Select the continuous sampling positions inside the left integration boundary and the continuous sampling positions inside the right integration boundary as local baseline fitting points, and fit the local baseline; The local baseline is subtracted from the measured amplitude between the left and right integration boundaries. The area of ​​the peak is obtained by summing the measured amplitude after subtraction and combining the independent variable intervals of adjacent sampling positions.

[0016] Optionally, the step of reverse-feeding the integral boundary verification search position offset to the peak-shaped competition path specifically includes: The peak-shaped competition path reads the integral boundary verification search position offset and compares the integral boundary verification search position offset with the preset offset threshold. When the offset of the search position for the integral boundary verification reaches the preset offset threshold, the peak competition path for generating the left and right integral boundaries is closed, the candidate peak state is retained, and the mutual exclusion activation of the peak competition path and the integral boundary search are re-executed. When the offset of the search position for the integral boundary verification does not reach the preset offset threshold, the peak competition path for generating the left and right integral boundaries is retained, and the peak area is used to generate spatial distribution imaging results and intensity evolution trend sequence results according to the file name coordinate rules.

[0017] The beneficial effects of this invention are: (1) The present invention processes slowly changing background, fluorescence residue, infrared background absorption interference and low frequency drift through the background suppression channel; the background suppression channel performs state locking in the slowly changing region based on the difference before and after background subtraction, and opens a one-way propagation entrance in the abrupt convex region, making it difficult for non-characteristic background and slowly drifting signals to enter the candidate peak state, reducing misidentification caused by spectral line background fluctuations in Raman spectrum, infrared spectrum and ultraviolet-visible spectrum.

[0018] (2) This invention processes narrow strong peaks, wide weak peaks and low signal-to-noise ratio weak peaks through peak shape continuity channel and co-position closure check; the peak shape continuity channel tracks the peak shoulder continuity status along both sides of the unidirectional propagation entrance, and the co-position closure check determines whether the peak entrance and the two sides of the peak shoulder are continuously connected; narrow strong peaks can enter the narrow peak path and strong peak path through the peak shoulder span and the relative amplitude of the peak top, while wide weak peaks and low signal-to-noise ratio weak peaks enter the wide peak path and weak peak path when they have a continuous peak shoulder structure; this processing avoids judging the peak signal based solely on local peak height and local peak width, and reduces misjudgments caused by isolated noise protrusions and chemical matrix interference.

[0019] (3) This invention processes overlapping peaks, boundary-shifted peaks, and discrete mass spectrometry peaks through peak competition paths and integral boundary verification. The peak competition path opens narrow peak paths, medium peak paths, and wide peak paths according to the peak shoulder span, and opens strong peak paths and weak peak paths according to the relative amplitude of the peak top. The search position shift of the integral boundary verification is reversed and fed back to the peak competition path. When the connection between adjacent peak shoulders is unclear or the integral boundary shift is large, the current path is closed and the search is restarted. For discrete peaks with prominent adjacent sampling changes in the spectrum sequence, the candidate peak state is retained when there is a unidirectional propagation entrance and the peak shoulders on both sides are in a continuous state, thereby improving the integral stability of overlapping peaks, discrete peaks, and complex peak shapes.

[0020] (4) This invention can be used for data analysis in in-situ and working condition studies of chemical research objects, chemical reaction processes or diverse systems. It is applicable to scenarios such as the evolution of surface structure of two-dimensional materials, the evolution of intermediates in electrochemical reactions, the transformation of adsorbed species at catalytic interfaces, material phase change processes, the migration of components at the water-oil interface, the spatiotemporal distribution of reaction products and the detection of trace components. The chemical characterization sequence data generated in the above scenarios can be formed into an ordered measurement sequence according to this invention. Peak identification, peak integration, spatial distribution imaging results and intensity evolution trend sequence results are completed through background suppression channels, peak continuity channels, peak competition paths and integral boundary verification, thereby improving the accurate identification and integrated analysis capabilities of multi-module sequence data. Attached Figure Description

[0021] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the multi-module sequence data recognition and parsing method based on machine learning proposed in this invention; Figure 2 This is a schematic diagram of the candidate peak state activation of the peak-gated structured state-space sequence model proposed in this invention. Figure 3 This is a schematic diagram of the peak-shaped competition path, integral boundary search, and reverse backfeed proposed in this invention; Figure 4This is a schematic diagram illustrating the entire process of applying the present invention in in-situ chemical characterization and working condition characterization scenarios; Figure 5 Raman spectra of sites Y6 and X1 on the surface of the MXene sample; Figure 6 Raman spectra of sites Y4 and X3 on the surface of the MXene sample; Figure 7 The image shows the intensity distribution of the characteristic peak at 410 cm⁻¹ in the characteristic Raman spectrum of the sample. Detailed Implementation

[0022] The invention will now be described in further detail with reference to the accompanying drawings. Figures 1-3 For the sake of simplification, the basic structure of the present invention is illustrated schematically. Figure 4 This paper demonstrates the overall processing flow of the present invention in chemical in-situ and operational condition characterization scenarios, and provides a schematic diagram of the entire process from raw data files to final imaging and trend output. Figures 5-7 This is a diagram of an example of Raman spectroscopy analysis; therefore, it only shows the components relevant to this invention.

[0023] refer to Figures 1-7 A machine learning-based method for multi-module sequence data recognition and parsing includes the following steps: Read chemical characterization sequence data files from spectroscopy, chromatography, and mass spectrometry tests, form an ordered measurement sequence, and after verification, background subtraction, smoothing and denoising, send it into a peak-gated structured state-space sequence model; In this invention, the chemical characterization sequence data file is a file that records chemical characterization data in the sampling order, and the file format includes txt, csv, and xlsx. When reading, the txt and csv files read the independent variable column and the measurement amplitude column according to the delimiter, and the xlsx file reads the table columns according to the worksheet row order to obtain the independent variable column and the measurement amplitude column. The independent variable column records one of wavenumber, wavelength, retention time, mass-to-charge ratio, sampling time, and spatial sampling coordinates, and the measurement amplitude column records one of Raman scattering intensity, infrared absorption intensity, absorbance, chromatographic detection signal, and mass spectrometry ion count value. The independent variable column and the measurement amplitude column are paired point by point according to the original sampling order and sequential numbering is performed to form an ordered measurement sequence. The validation process excludes data files unsuitable for subsequent processing. Validation includes checks for consistent sampling length, non-zero measurement amplitude, and completeness of data. The consistent sampling length check verifies the number of sampling points in the independent variable column and the measurement amplitude column. The non-zero measurement amplitude check excludes data files with all zeros in the measurement amplitude column. The completeness of data check excludes data files with missing rows, unparseable characters, or blank data segments. Successful validation sequences proceed to background subtraction and smoothing / denoising. Successful validation sequences are recorded with their file names and reasons for failure. Background subtraction is used to remove low-frequency background drift in chemical characterization sequence data. The difference between the measured amplitude before and after background subtraction at the same sampling position is obtained to obtain the difference before and after background subtraction, which is used by the background suppression channel to identify gradually changing background and abrupt protrusions. Smoothing and denoising is used to suppress high-frequency noise. The measured amplitudes of adjacent sampling positions after smoothing and denoising are subtracted to obtain the changes between adjacent sampling positions, which is used by the peak continuity channel to identify the peak shoulder continuity state and peak shoulder termination state. The peak-gated structured state-space sequence model is a machine learning sequence model for ordered measurement sequences, used to carry the state transmission relationship between candidate peak states, peak competition paths, integration boundaries, and peak areas. The peak-gated structured state-space sequence model is derived from the structured state-space sequence model by modifying the sampling order of the ordered measurement sequence as the state transmission order. Background suppression, peak continuity, candidate peak state activation, peak competition, and inverse backfeeding of the integration boundary verification search position offset required for chemical feature peak identification are embedded as gating constraints in the state transmission process. This allows the model output to adapt to peak identification, peak integration, chemical imaging analysis, and batch trend analysis of chemical characterization sequence data such as spectrophotometers, chromatograms, and mass spectrometers. The preset gradual variation range, preset abrupt change threshold, attenuation verification threshold, boundary search range, preset baseline fluctuation range, and preset offset threshold in the peak-gated structured state-space sequence model can be calibrated based on ordered measurement sequences that have undergone manual verification. Manually verified ordered measurement sequences include manually confirmed peak positions, manually confirmed left integration boundaries, manually confirmed right integration boundaries, and manually confirmed peak areas. In specific processing, for gradually varying backgrounds corresponding to sampling positions where continuous differences fall within a preset gradually varying range, a state blocking flag is written into the background suppression channel to prevent them from entering the same-position closure check; for narrow strong peaks corresponding to candidate peak states with small shoulder spans and relatively high peak amplitudes, the narrow peak path and strong peak path are opened by the peak shape competition path; for wide weak peaks corresponding to candidate peak states with large shoulder spans and relatively low peak amplitudes, the wide peak path and weak peak path are opened by the peak shape competition path; for overlapping peaks corresponding to peak shapes with unclear connections between adjacent shoulder spans and easy shifts in integration boundaries, the shift in the integration boundary verification search position is reversed and fed back to the peak shape competition path, closing the peak shape competition path. The current peak type competition path is re-searched; discrete peaks correspond to peak signals with prominent changes in adjacent sampling in the mass spectrometry sequence, and are retained as candidate peaks when they have a unidirectional propagation entrance and a continuous peak shoulder state on both sides; isolated noise bulges lack continuous peak shoulder states on both sides, and do not pass the same-position closure verification; low signal-to-noise ratio weak peaks are retained as candidate peaks when they have continuous peak shoulder states, and the attenuation verification threshold and integration boundary are determined through the weak peak path; the above processing enables different chemical spectral peaks and signal morphologies in Raman spectroscopy, infrared spectroscopy, ultraviolet-visible spectroscopy, chromatography and mass spectrometry to enter the peak identification, peak integration and imaging analysis process according to the corresponding path; Within the peak-gated structured state-space sequence model, a background suppression channel and a peak continuation channel are set up. The background suppression channel divides the gradual change region and the abrupt bulge region according to the difference before and after background subtraction, driving state locking. The peak continuation channel drives the peak shoulder continuation according to the adjacent sampling changes after smoothing and denoising. In this invention, the background suppression channel is a state propagation path within a peak-gated structured state space sequence model used to process the difference before and after background subtraction. The background suppression channel is obtained by modifying the state propagation path of the structured state space sequence model. The input source is the difference between the measured amplitude before and after background subtraction at the same sampling position, and the output is used for subsequent division of the gradual change region and the abrupt bulge region. The peak continuation channel is a state propagation path within the peak-gated structured state-space sequence model used to process the changes in adjacent samples after smoothing and denoising. The peak continuation channel is obtained by modifying the state propagation path of the structured state-space sequence model. The input source is the amplitude change of adjacent sampling positions after smoothing and denoising, and the output is used for subsequent peak shoulder continuation state tracking. The gradual change region and the abrupt change region are divided by the background suppression channel according to the difference before and after background subtraction. The background suppression channel calculates the difference before and after background subtraction of adjacent sampling positions along the direction of the ordered measurement sequence. When the absolute value of the difference corresponding to three consecutive sampling positions does not exceed 3% of the maximum value of the difference before and after background subtraction in the same ordered measurement sequence, the corresponding sampling position is written into the gradual change region. When the absolute value of the difference corresponding to any sampling position exceeds 3% of the maximum value of the difference before and after background subtraction in the same ordered measurement sequence, and the difference before and after background subtraction is greater than the difference before and after background subtraction of the previous and next sampling positions, the corresponding sampling position is written into the abrupt change region. The background suppression channel is blocked in the gradual change region and the propagation is cut off in the sudden change region. The unidirectional propagation entrance is opened in the sudden change convex region. The peak continuation channel continuously tracks the peak shoulder continuation state along both sides of the unidirectional propagation entrance. In this invention, the truncation propagation is a propagation stop operation performed by the background suppression channel after the read state is locked; the background suppression channel sets the propagation state of the sampling position in the gradual change region to the locked state, stops outputting the propagation result to the adjacent sampling position, so that the gradual change region does not enter the same position closure check. The open unidirectional propagation entry is the operation of the background suppression channel to retain candidate propagation links in the mutation protrusion region; the background suppression channel retains the propagation link from the unidirectional propagation entry in the mutation protrusion region to the co-position closure verification, and closes the propagation link from other sampling positions in the mutation protrusion region to the co-position closure verification. Continuous tracking is the operation of reading adjacent sampling changes along the ordered measurement sequence direction in the peak-shaped continuation channel. Starting from the unidirectional propagation entrance, the peak-shaped continuation channel continuously reads adjacent sampling changes at the forward and backward sampling positions, and writes the sampling positions with consistent change direction and decreasing change amplitude into the peak shoulder continuation state. When the change direction of adjacent sampling changes is reversed, and the change amplitude at the reversed position reaches 10% of the amplitude corresponding to the unidirectional propagation entrance, the reversed position is written into the peak shoulder termination state. The peak shoulder termination state is the state where the peak shoulder continuation state stops propagating to the outer sampling positions. The amplitude corresponding to the unidirectional propagation entrance is the measured amplitude of the sampling position where the unidirectional propagation entrance is located after smoothing and denoising. Perform a co-position closure check on the unidirectional propagation inlet and the continuous state of the two side peaks, and only open the candidate peak state when the co-position closure is successful. In this invention, the candidate peak state is an intermediate state in the peak-gated structured state space sequence model that is allowed to enter the peak-competition path; the candidate peak state is generated by the same-position closure check, and an open flag is written when the same-position closure is successful, and a closed flag is written when the same-position closure is unsuccessful; the sampling position corresponding to the open flag enters the peak-competition path, and the sampling position corresponding to the closed flag does not enter the peak-competition path. When a candidate peak enters the peak competition path, the narrow peak path, medium peak path, and wide peak path are activated mutually exclusively according to the peak shoulder span, and the strong peak path and weak peak path are activated mutually exclusively according to the relative amplitude of the peak. In this invention, the peak-shaped competition path is the path selection structure before the candidate peak state enters the integral boundary search, established by the peak-shaped gated structured state-space sequence model based on the candidate peak state; the peak-shaped competition path includes narrow peak path, medium peak path, wide peak path, strong peak path, and weak peak path. Narrow peak path, medium peak path, and wide peak path are used to limit the boundary search range, while strong peak path and weak peak path are used to limit the attenuation verification threshold; Mutually exclusive activation is a single-path opening operation within a peak-shaped competitive path; only one path is opened in narrow-peak, medium-peak, and wide-peak paths, and only one path is opened in strong-peak and weak-peak paths. Unopened paths do not participate in the integral boundary search. The integration boundary search is driven by the activated peak competition path. The integration boundary is locked by triple verification of attenuation, width and local baseline. The integration boundary is then checked and the peak area is calculated after subtracting the local baseline. In this invention, attenuation verification is used as the boundary stopping judgment in the integral boundary search; attenuation verification uses the measured amplitude at the peak position as a reference and compares the degree of decrease of the measured amplitude along the front sampling position and the rear sampling position; the sampling position where the degree of decrease reaches the attenuation verification threshold is entered into the candidate position of the integral boundary. Width verification is used to determine the boundary span in the integral boundary search. Width verification is centered on the peak position and calculates the sampling interval from the front sampling position and the back sampling position to the peak position. Sampling positions whose sampling intervals fall within the boundary search range are included in the integral boundary candidate positions. Local baseline verification is used to determine the stability of the boundary in the integral boundary search. Local baseline verification reads the measurement amplitude of continuous sampling positions near the candidate position of the integral boundary and compares the changes of adjacent measurement amplitudes. Sampling positions where the changes of adjacent measurement amplitudes fall within the preset baseline fluctuation range are locked as the left integral boundary and the right integral boundary, respectively, according to the search direction. The attenuation verification threshold is set according to strong peak path and weak peak path. The measured amplitude at the peak position is the measured amplitude at the peak position after smoothing and denoising. The attenuation verification threshold for strong peak path is 20% of the measured amplitude at the peak position, and the attenuation verification threshold for weak peak path is 35% of the measured amplitude at the peak position. The boundary search range is set according to narrow peak path, medium peak path, and wide peak path. The narrow peak path corresponds to 3 sampling intervals on each side of the peak position, the medium peak path corresponds to 6 sampling intervals on each side of the peak position, and the wide peak path corresponds to 10 sampling intervals on each side of the peak position. The preset baseline fluctuation range is that the difference between the maximum and minimum measured amplitudes of three consecutive sampling positions does not exceed 5% of the measured amplitude at the peak position. The search position offset of the integral boundary verification is reversed and fed back to the peak competition path. When the preset offset threshold is reached, the peak competition path is closed and the search is restarted. When the preset offset threshold is not reached, the spatial distribution imaging results and intensity evolution trend sequence results are generated according to the file name coordinate rules.

[0024] In this invention, the preset offset threshold is the boundary value used to determine whether the offset of the integral boundary verification search position triggers a re-search. The offset of the integral boundary verification search position is determined by the larger of the sampling position offset of the left verification boundary position relative to the left integral boundary and the sampling position offset of the right verification boundary position relative to the right integral boundary. The preset offset threshold is set according to the peak-shaped competition path: 2 sampling intervals for narrow peak paths, 4 sampling intervals for medium peak paths, and 6 sampling intervals for wide peak paths; strong peak paths use the preset offset thresholds corresponding to the already opened narrow peak paths, medium peak paths, and wide peak paths, and weak peak paths are based on the preset offset thresholds corresponding to the already opened narrow peak paths, medium peak paths, and wide peak paths with an additional 1 sampling interval. When the offset of the integral boundary verification search position reaches the corresponding preset offset threshold, the current peak-shaped competition path is closed, and the integral boundary search is re-executed; when the offset of the integral boundary verification search position does not reach the corresponding preset offset threshold, the current peak-shaped competition path is retained.

[0025] The file name coordinate rule is the rule for extracting spatial coordinates from the names of chemical characterization sequence data files; the horizontal coordinate marker adopts the format of "X + integer" and the vertical coordinate marker adopts the format of "Y + integer"; when the file name contains both horizontal and vertical coordinate markers, the integer after X is read as the horizontal coordinate value and the integer after Y is read as the vertical coordinate value; the read horizontal and vertical coordinate values ​​are used to establish a spatial grid, and the peak area is written into the spatial distribution imaging results according to the spatial grid position. After the peak area is arranged in one of the following orders: folder order, acquisition time order, or file name order, the intensity evolution trend sequence results are generated.

[0026] In this embodiment, the process of forming an ordered measurement sequence specifically includes: The chemical characterization sequence data file is derived from at least one of the following: Raman spectroscopy, infrared spectroscopy, ultraviolet-visible spectroscopy, gas chromatography, liquid chromatography, and mass spectrometry, obtained from in-situ and operational tests of chemical research objects and chemical reaction processes. The chemical characterization sequence data file is parsed according to the file format, and the order of independent variables and the order of measurement amplitudes are read. The order of the independent variables corresponds to one of wavenumber, wavelength, retention time, mass-to-charge ratio, sampling time, and spatial sampling coordinates; the order of the measurement amplitudes corresponds to one of Raman scattering intensity, infrared absorption intensity, absorbance, chromatographic detection signal, and mass spectrometry ion count value. The independent variable order is paired with the measurement amplitude order point by point according to the original sampling order. Sampling positions that are not paired are not written into the ordered measurement sequence. In this invention, point-by-point pairing means forming a set of sampled data by combining one independent variable value in the independent variable sequence with one measurement amplitude in the measurement amplitude sequence according to the same sampling position; during pairing, the original sampling order is taken as the standard, and the independent variable value and measurement amplitude are read sequentially starting from the first sampling position. When both the independent variable value and the measurement amplitude exist and can be resolved into numerical values, they are written into an ordered measurement sequence. Unpaired sampling locations are those with missing independent variable values, missing measurement amplitudes, independent variable values ​​that cannot be parsed into numerical values, or measurement amplitudes that cannot be parsed into numerical values. Unpaired sampling locations are not written into the ordered measurement sequence and do not participate in subsequent background subtraction, smoothing and denoising, and peak identification. The paired sampling locations are sequentially numbered to form an ordered measurement sequence.

[0027] In this invention, the sequence number is an integer number set for the paired sampling positions according to the original sampling order; the sequence number increases sequentially from 1, and the numbering order is consistent with the original sampling order; the paired independent variable value, the measurement amplitude, and the sequence number together form a record in the ordered measurement sequence.

[0028] In this embodiment, the background suppression channel is divided into a gradual change region and a sudden abrupt change region based on the difference before and after background subtraction, specifically including: The difference between the background before and after subtraction is read point by point along the direction of the ordered measurement sequence, and the difference between the background before and after subtraction at adjacent sampling positions is compared. In this invention, the differential comparison is an adjacent difference calculation operation performed by the background suppression channel along the direction of the ordered measurement sequence; starting from the second sampling position, the background suppression channel subtracts the difference before and after background subtraction at the current sampling position from the difference before and after background subtraction at the previous sampling position to obtain the difference corresponding to the current sampling position; the absolute value of the difference corresponding to the current sampling position is taken to obtain the absolute value of the difference; Sampling locations where continuous differences fall within a preset gradual variation range are marked as gradual variation regions, and sampling locations where differences deviate from the preset gradual variation range and the background subtracts the difference to form a local bulge are marked as abrupt bulge regions. In this invention, the preset gradual change range is the range in which the absolute value of the difference does not exceed 3% of the maximum value of the difference before and after the background subtraction within the same ordered measurement sequence; when the absolute values ​​of the differences at three consecutive sampling positions all fall within the preset gradual change range, the three sampling positions are written into the gradual change region. Local bulges are bulges where the difference between the background subtraction values ​​at the current sampling position and the background subtraction values ​​at the previous and next sampling positions are greater than the difference between the background subtraction values ​​at the previous and next sampling positions. When the absolute value of the difference at the current sampling position deviates from the preset gradual change range and a local bulge is formed at the current sampling position, the current sampling position is written into the abrupt bulge region. The background suppression channel writes a state lockout flag to the gradually changing region and a unidirectional propagation entrance open flag to the abruptly changing protruding region.

[0029] In this invention, the state locking mark is a propagation restriction mark written to the sampling position of the gradually changing region; after the background suppression channel reads the state locking mark, it sets the propagation state of the corresponding sampling position to the locked state; the unidirectional propagation entrance opening mark is an entrance candidate mark written to the sampling position of the abrupt protrusion region; after the background suppression channel reads the unidirectional propagation entrance opening mark, it sets the sampling position with the largest difference before and after background subtraction in the abrupt protrusion region as the unidirectional propagation entrance.

[0030] In this embodiment, the background suppression channel writes a state lockout flag to the gradually changing region and a unidirectional propagation entrance open flag to the abruptly protruding region, specifically including: Background suppression channel read status lock flag, sets the propagation status of the sampling position in the gradual change zone to the lock state, and cuts off the propagation of the lock state to the adjacent sampling position; In this invention, the propagation state is the state record written to the sampling position by the background suppression channel, used to indicate whether the sampling position outputs the state within the background suppression channel to adjacent sampling positions; the blocking state is a value of the propagation state, used to block the sampling position in the gradual change region from continuing to participate in propagation; the blocking state is obtained by the state blocking flag; the blocking state does not output the state to adjacent sampling positions and does not enter the same position closure check. The background suppression channel reads the unidirectional propagation entrance open marker, sets the sampling position with the largest difference before and after background subtraction in the mutation protrusion region as the unidirectional propagation entrance, retains the propagation link from the unidirectional propagation entrance to the co-position closure check, and closes the propagation link from the sampling positions in the mutation protrusion region other than the unidirectional propagation entrance to the co-position closure check.

[0031] In this invention, the unidirectional propagation entry point is the sampling position within the abrupt protrusion region that is allowed to enter the co-position closure verification; the unidirectional propagation entry point is obtained by the unidirectional propagation entry point opening mark; when there are two or more sampling positions with the same maximum value of the difference before and after background subtraction, the sampling position with the smallest sequential number is selected as the unidirectional propagation entry point; the propagation link is the connection record of the unidirectional propagation entry point to the output state of the co-position closure verification; the propagation link is generated by the unidirectional propagation entry point.

[0032] In this embodiment, the peak-shaped continuity channel continuously tracks the peak shoulder continuity state along both sides of the unidirectional propagation entrance, specifically including: The peak-shaped continuation channel starts from the unidirectional propagation entrance and reads the smoothed and denoised adjacent sampling changes at the forward and backward sampling positions along the direction of the ordered measurement sequence. In this invention, the adjacent sampling change is the difference in measurement amplitude between two adjacent sampling positions after smoothing and denoising; during calculation, the measurement amplitude of the previous sampling position is subtracted from the measurement amplitude of the subsequent sampling position to obtain the adjacent sampling change; when the adjacent sampling change is greater than 0, it is in the upward direction, and when the adjacent sampling change is less than 0, it is in the downward direction, and the absolute value of the adjacent sampling change is the change amplitude; Sampling positions where the direction of change remains consistent and the amplitude of change decreases are written as peak shoulder continuation state, and sampling positions where the direction of change reverses and the amplitude of change reaches the preset abrupt change threshold are written as peak shoulder termination state. In this invention, "consistent change direction" means that the positive and negative directions of the change in the current adjacent sample are the same as those of the change in the previous adjacent sample; "decreasing change amplitude" means that the absolute value of the change in the current adjacent sample is less than the absolute value of the change in the previous adjacent sample; and "reversed change direction" means that the positive and negative directions of the change in the current adjacent sample are opposite to those of the change in the previous adjacent sample. The preset mutation threshold is the amplitude threshold for determining whether the peak shoulder continuation state has changed to the peak shoulder termination state; the preset mutation threshold is 10% of the amplitude corresponding to the unidirectional propagation entrance; the amplitude corresponding to the unidirectional propagation entrance is the measured amplitude of the sampling position where the unidirectional propagation entrance is located after smoothing and denoising; the peak shoulder termination state is the state in which the peak shoulder continuation state stops propagating to the outer sampling position; At the peak shoulder termination state, the propagation of the peak shoulder continuation state to the outer sampling position is interrupted.

[0033] In this invention, the front sampling position, rear sampling position, and outer sampling position are all determined according to the sequential numbering of the unidirectional propagation entrance; the sampling position with a sequential number less than the sequential number of the unidirectional propagation entrance is the front sampling position, and the peak-shaped continuation channel reads the front sampling position from the unidirectional propagation entrance along the direction of decreasing sequential number; the sampling position with a sequential number greater than the sequential number of the unidirectional propagation entrance is the rear sampling position, and the peak-shaped continuation channel reads the rear sampling position from the unidirectional propagation entrance along the direction of increasing sequential number; the outer sampling position is the sampling position in the direction away from the unidirectional propagation entrance, the sampling position in the front sampling position where the sequential number continues to decrease is the front outer sampling position, and the sampling position in the rear sampling position where the sequential number continues to increase is the rear outer sampling position.

[0034] In this embodiment, a co-location closure check is performed on the unidirectional propagation entrance and the continuous state of the two side peaks, specifically including: Using the sampling position where the unidirectional propagation entrance is located as the verification center, read the peak and shoulder continuation status of the continuous sampling positions on the front side and the peak and shoulder continuation status of the continuous sampling positions on the back side. In this invention, the front continuous sampling positions are sampling positions with continuous peak and shoulder states written continuously in the direction of decreasing sequential numbering, starting from the front sampling position adjacent to the unidirectional propagation entrance; the rear continuous sampling positions are sampling positions with continuous peak and shoulder states written continuously in the direction of increasing sequential numbering, starting from the rear sampling position adjacent to the unidirectional propagation entrance. When the peak shoulder continuity state of the current continuous sampling position, the unidirectional propagation entrance, and the peak shoulder continuity state of the subsequent continuous sampling position are continuously connected in the direction of the ordered measurement sequence, it is determined that the co-position closure is established and the candidate peak state is opened. In this invention, co-position closure is the verification result that the unidirectional propagation entrance and the continuous state of the two side peaks are connected without any gaps in the sequential numbering; continuous connection is the result that the sequential number of the sampling position adjacent to the unidirectional propagation entrance in the continuous sampling position on the front side is 1 less than the sequential number of the unidirectional propagation entrance, and the sequential number of the sampling position adjacent to the unidirectional propagation entrance in the continuous sampling position on the rear side is 1 greater than the sequential number of the unidirectional propagation entrance. If any of the peak shoulder continuation states at the front continuous sampling position or the rear continuous sampling position is disconnected from the unidirectional propagation entrance, the co-position closure is determined to be invalid, and the candidate peak state is closed.

[0035] In this invention, "disconnection" refers to either the state where the adjacent sampling positions before the unidirectional propagation entrance are not written into the peak shoulder continuation state, or the state where the adjacent sampling positions after the unidirectional propagation entrance are not written into the peak shoulder continuation state.

[0036] In this embodiment, mutual exclusion activation specifically includes: The peak-shaped competition path reads the peak shoulder continuation state of the consecutive sampling positions on the front side and the peak shoulder continuation state of the consecutive sampling positions on the back side, and calculates the sampling interval between the peak shoulder continuation state of the consecutive sampling positions on the front side and the peak shoulder continuation state of the consecutive sampling positions on the back side as the peak shoulder span. In this invention, the peak shoulder span is the sampling interval between the peak shoulder continuation state of the continuous sampling positions on the front side and the peak shoulder continuation state of the continuous sampling positions on the rear side; the sampling interval is determined by the absolute value of the difference between the sequential number of the sampling position farthest from the unidirectional propagation entrance in the continuous sampling positions on the front side and the sequential number of the sampling position farthest from the unidirectional propagation entrance in the continuous sampling positions on the rear side. Based on the peak shoulder span, select one path from the narrow peak path, medium peak path, and wide peak path to open, and close the unselected path; In this invention, narrow peak path, medium peak path, and wide peak path are selected according to the peak shoulder span; when the peak shoulder span does not exceed 6 sampling intervals, the narrow peak path is opened; when the peak shoulder span is greater than 6 sampling intervals but not more than 12 sampling intervals, the medium peak path is opened; when the peak shoulder span is greater than 12 sampling intervals, the wide peak path is opened. The peak-shaped competition path reads the amplitude corresponding to the unidirectional propagation entrance, the amplitude corresponding to the peak shoulder continuation state at the front continuous sampling position, and the amplitude corresponding to the peak shoulder continuation state at the rear continuous sampling position. The difference between the amplitude corresponding to the unidirectional propagation entrance and the average of the amplitude corresponding to the peak shoulder continuation state at the front continuous sampling position and the amplitude corresponding to the peak shoulder continuation state at the rear continuous sampling position is taken as the peak relative amplitude. In this invention, the relative amplitude of the peak is the difference between the amplitude corresponding to the unidirectional propagation entrance and the average amplitude corresponding to the continuous state of the two side peaks; the amplitude corresponding to the continuous state of the two side peaks is the amplitude measured at the sampling position farthest from the unidirectional propagation entrance in the continuous sampling positions on the front side and the amplitude measured at the sampling position farthest from the unidirectional propagation entrance in the continuous sampling positions on the rear side. The measured amplitudes are all taken from the ordered measurement sequence after smoothing and denoising. Based on the relative amplitude of the peak, one path is selected from the strong peak path and the weak peak path to be opened, and the unselected path is closed.

[0037] In this invention, the strong peak path and the weak peak path are selected according to the relative amplitude of the peak; when the relative amplitude of the peak reaches 30% of the amplitude corresponding to the unidirectional propagation entrance, the strong peak path is opened; when the relative amplitude of the peak does not reach 30% of the amplitude corresponding to the unidirectional propagation entrance, the weak peak path is opened.

[0038] In this embodiment, the integral boundary search is driven by the activated peak-shaped competition path, specifically including: The peak position is taken as the sampling position corresponding to the unidirectional propagation entrance. The boundary search range is determined by one of the open paths among the narrow peak path, medium peak path, and wide peak path. The attenuation verification threshold is determined by one of the open paths among the strong peak path and weak peak path. Search from the peak position along the direction of the ordered measurement sequence to the front and back sampling positions, and perform attenuation verification, width verification, and local baseline verification point by point; In this invention, the point-by-point execution of attenuation verification, width verification, and local baseline verification is an operation to determine the candidate positions of the integration boundary in sequence according to the sampling position; when the degree of decrease in the measured amplitude of the current sampling position relative to the peak position reaches the attenuation verification threshold, the current sampling position is within the boundary search range, and the change in the measured amplitude of the continuous sampling positions near the current sampling position falls within the preset baseline fluctuation range, it is determined that the current sampling position passes the attenuation verification, width verification, and local baseline verification. The front sampling position that simultaneously passes attenuation verification, width verification, and local baseline verification is locked as the left integration boundary, and the rear sampling position that simultaneously passes attenuation verification, width verification, and local baseline verification is locked as the right integration boundary.

[0039] In this invention, the left integral boundary is the first sampling position that passes the attenuation check, width check, and local baseline check when searching from the peak position to the front sampling position; the right integral boundary is the first sampling position that passes the attenuation check, width check, and local baseline check when searching from the peak position to the rear sampling position.

[0040] In this embodiment, a verification search is performed on the integral boundary, and the peak area is calculated after subtracting the local baseline. Specifically, this includes: Using the left and right integration boundaries as initial boundaries, a verification search is performed outside the left and right integration boundaries along the direction of the ordered measurement sequence to obtain the positions of the left and right verification boundaries. In this invention, the verification search is an operation of re-finding the stable position of the boundary outside the initial boundary; the left verification boundary position is obtained by searching the sampling position to the front and outer side of the left integral boundary, and the right verification boundary position is obtained by searching the sampling position to the rear and outer side of the right integral boundary; when the difference between the maximum and minimum values ​​of the measured amplitude of three consecutive sampling positions outside the left integral boundary does not exceed 5% of the measured amplitude at the peak position, the sampling position closer to the left integral boundary is taken as the left verification boundary position; when the difference between the maximum and minimum values ​​of the measured amplitude of three consecutive sampling positions outside the right integral boundary does not exceed 5% of the measured amplitude at the peak position, the sampling position closer to the right integral boundary is taken as the right verification boundary position. The position of the left verification boundary is compared with the left integration boundary, and the position of the right verification boundary is compared with the right integration boundary to obtain the offset of the integration boundary verification search position. In this invention, the offset of the integral boundary verification search position is the larger of the left boundary offset and the right boundary offset; the left boundary offset is the number of sampling intervals between the left verification boundary position and the left integral boundary, and the right boundary offset is the number of sampling intervals between the right verification boundary position and the right integral boundary. Select the continuous sampling positions inside the left integration boundary and the continuous sampling positions inside the right integration boundary as local baseline fitting points, and fit the local baseline; In this invention, the continuous sampling positions inside the left integration boundary are three consecutive sampling positions from the left integration boundary to the right integration boundary, and the continuous sampling positions inside the right integration boundary are three consecutive sampling positions from the right integration boundary to the left integration boundary; the local baseline fitting points are composed of three consecutive sampling positions from the left integration boundary to the right integration boundary and three consecutive sampling positions from the right integration boundary to the left integration boundary; the local baseline is a linear baseline obtained through the local baseline fitting points. The local baseline is subtracted from the measured amplitude between the left and right integration boundaries. The area of ​​the peak is obtained by summing the measured amplitude after subtraction and combining the independent variable intervals of adjacent sampling positions.

[0041] In this invention, the interval of the independent variable is the difference between the values ​​of the corresponding independent variables at two adjacent sampling positions; when the surface is accumulated, the average value of the measured amplitude after deducting the local baseline at two adjacent sampling positions is taken, and then multiplied by the corresponding interval of the independent variable to obtain the segmented area; the entire segmented surface between the left integration boundary and the right integration boundary is accumulated to obtain the peak area.

[0042] In this embodiment, the offset of the integral boundary verification search position is reversed and fed back to the peak-shaped competition path, specifically including: The peak-shaped competition path reads the integral boundary verification search position offset and compares the integral boundary verification search position offset with the preset offset threshold. In this invention, reverse backfeeding is a correction operation that feeds back the offset of the integral boundary verification search position to the peak-shaped competition path; the offset of the integral boundary verification search position is used to determine whether the peak-shaped competition paths corresponding to the left integral boundary and the right integral boundary are stable; When the offset of the search position for the integral boundary verification reaches the preset offset threshold, the peak competition path for generating the left and right integral boundaries is closed, the candidate peak state is retained, and the mutual exclusion activation of the peak competition path and the integral boundary search are re-executed. In this invention, closing the peak competition path that generates the left integration boundary and the right integration boundary means closing one of the currently open narrow peak path, medium peak path, and wide peak path, and closing one of the currently open strong peak path and weak peak path; retaining the candidate peak state means retaining the candidate peak state opened by the co-position closure check, without redetermining the unidirectional propagation entry point, and without re-executing the co-position closure check. When the offset of the search position for the integral boundary verification does not reach the preset offset threshold, the peak competition path for generating the left and right integral boundaries is retained, and the peak area is used to generate spatial distribution imaging results and intensity evolution trend sequence results according to the file name coordinate rules.

[0043] In this invention, the peak-shaped competition path that generates the left integration boundary and the right integration boundary is retained, which means retaining one of the currently open narrow peak path, medium peak path, and wide peak path, and retaining one of the currently open strong peak path and weak peak path.

[0044] The spatial distribution imaging result is a two-dimensional distribution formed by arranging peak areas according to spatial coordinates. The spatial distribution imaging result is obtained jointly by the filename coordinate rule and the peak area. The filename coordinate rule reads the horizontal and vertical coordinate values, which determine the writing positions in the spatial grid. The peak areas are then written to the corresponding positions to form a peak area matrix. Each position in the peak area matrix corresponds to a chemical characterization sequence data file, and the value at each position is the peak area calculated from the corresponding chemical characterization sequence data file.

[0045] Example 1: To verify the feasibility of this invention in practice, it was applied to a chemical characterization sequence data processing scenario. In this scenario, the data sources to be processed include Raman spectral sequences, infrared spectral sequences, chromatographic sequences, and mass spectrometric sequences obtained from in-situ and operational tests of chemical research objects and chemical reaction processes, involving chemical systems such as electrochemical reactions, catalytic interfaces, material phase transitions, or water-oil interfaces. Each chemical characterization sequence data file records the independent variable values ​​and measurement amplitudes in the sampling order, and the file name contains horizontal and vertical coordinate markers. The problems to be solved mainly focus on three aspects: First, when background drift is strong, the fixed threshold method easily identifies slowly changing background protrusions as peaks; second, when narrow strong peaks, wide weak peaks, and overlapping peaks are mixed in the same batch of data, the fixed integral boundary is difficult to stably cover the peak shoulders; third, when batch files enter spatial distribution imaging processing, peak area errors are directly transmitted to the spatial grid, causing patchy breaks in the spatial distribution imaging results and abnormal fluctuations in the intensity evolution trend sequence results.

[0046] In this scenario, chemical characterization sequence data files in txt, csv, and xlsx formats are first read. The first and second numerical columns are parsed to form the order of independent variables and the order of measurement amplitudes, respectively. Paired sampling positions are written into the ordered measurement sequence. Sampling positions corresponding to missing values, unparseable characters, and blank data segments are not processed further. After successful validation, background subtraction is performed on the measurement amplitude order to obtain the difference before and after background subtraction. Five-point convolution smoothing is then performed on the background-subtracted ordered measurement sequence to obtain the smoothed and denoised ordered measurement sequence and the changes in adjacent samples.

[0047] After the ordered measurement sequence enters the peak-gated structured state-space sequence model, the background suppression channel reads the difference before and after background subtraction and divides the region into a gradual change region and abrupt bulge region based on the difference change. The gradual change region is not entered into the same-position closure check after state locking. The sampling position with the largest difference before and after background subtraction in the abrupt bulge region is used as the unidirectional propagation entrance. The peak continuation channel reads the changes of adjacent samples along both sides of the unidirectional propagation entrance. Sampling positions with consistent change direction and decreasing change amplitude are written into the peak continuation state, while positions with reversed change direction and change amplitude reaching the preset abrupt threshold are written into the peak termination state.

[0048] The unidirectional propagation entrance and the continuous state of the peak shoulders on both sides enter the same-position closure verification. When the continuous state of the peak shoulders at the front continuous sampling position, the unidirectional propagation entrance, and the continuous state of the peak shoulders at the rear continuous sampling position are connected without any gaps in the sequential numbering, the candidate peak state is opened. After the candidate peak state enters the peak shape competition path, a path is opened according to the peak shoulder span among the narrow peak path, medium peak path, and wide peak path, and a path is opened according to the relative amplitude of the peak top among the strong peak path and weak peak path. The opened peak shape competition path drives the integration boundary search, and the left integration boundary and right integration boundary are locked according to the attenuation verification, width verification, and local baseline verification.

[0049] After the integration boundary is locked, the system performs a verification search outside the left and right integration boundaries to obtain the positions of the left and right verification boundaries. The offsets of the left and right verification boundary positions from the left and right integration boundaries, respectively, together form the integration boundary verification search position offsets. When the offset reaches a preset offset threshold, the current peak competition path is closed, the candidate peak state is retained, and mutual exclusion activation and integration boundary search are re-executed. When the offset does not reach the preset offset threshold, the current peak competition path is retained, and the peak area is calculated after subtracting the local baseline. The peak area is written into the spatial grid according to the file name coordinate rules to form a spatial distribution imaging result; the peak area is arranged in one of the following orders: folder order, acquisition time order, or file name order to form an intensity evolution trend sequence result.

[0050] refer to Figures 5-7 This embodiment uses Raman spectral data from the surface of an MXene sample as a specific application to illustrate the complete processing procedure of this invention in batch file reading, background subtraction, smoothing and denoising, automatic peak identification, peak area calculation, spatial imaging, and batch trend output. The MXene material is a Ti3C2Tx-type two-dimensional transition metal carbide material. Its Raman characteristic spectrum is easily affected by factors such as differences in layered structure, uneven distribution of surface end groups, differences in local oxidation degree, and residual fluorescence background, resulting in significant differences in background intensity, peak shoulder width, peak height, and local noise level in Raman spectra at different spatial locations. Furthermore, considering that infrared spectroscopy, chromatography, and mass spectrometry data also have measurement amplitude sequences arranged in the order of independent variables, their data processing also suffers from problems such as background drift, peak broadening, overlapping peaks, weak peak identification, and unstable integration boundaries. Therefore, the same processing link is further applied to isomorphic chemical characterization data scenarios such as infrared broad peak sequences, chromatographic overlapping peak sequences, mass spectrometry discrete peak sequences, and low signal-to-noise ratio sequences to evaluate the extended applicability of this invention to multiple types of sequence data.

[0051] Specifically, in the Raman spectral sequence processing of MXene sample surfaces, the order of independent variables in the sequence data file is the order of Raman shifts (wavenumbers), and the order of measured amplitudes is the order of Raman scattering intensities. The file names use coordinate notation "Y + integer, X + integer" to record the spatial sampling location, and the same batch of files corresponds to multiple sampling points on the same sample surface.

[0052] After feeding this type of Raman spectral sequence into the peak-gated structured state-space sequence model of this invention, the background suppression channel first truncates the propagation of the slowly varying region based on the difference before and after background subtraction, so that the slowly changing fluorescence background or baseline rise does not directly enter the candidate peak state; the peak continuation channel then tracks the peak shoulder continuation state along both sides of the unidirectional propagation entrance based on the changes in adjacent samples after smoothing and denoising; when the unidirectional propagation entrance and the peak shoulder continuation states on both sides are consecutively connected in sequence number, the co-location closure check is established and the candidate peak state is opened. Subsequently, the candidate peak state enters the peak competition path, and the system selects the corresponding narrow peak path, medium peak path, or wide peak path, as well as strong peak path or weak peak path, according to the peak shoulder span and the relative amplitude of the peak top, and the integrated boundary search is driven by the opened peak competition path. After the integrated boundary is checked and searched, if the check and search position offset does not reach the preset offset threshold, the system retains the current peak competition path and calculates the peak area; if the check and search position offset reaches the preset offset threshold, the system closes the current peak competition path and re-executes the mutual exclusion activation and integrated boundary search. Through this closed-loop processing, misidentification and integration bias caused by background drift, isolated noise, weak peaks and peak shoulder variations in MXene Raman data can be suppressed.

[0053] After calculating the peak area of ​​a single Raman spectrum, the system reads the corresponding horizontal and vertical coordinate values ​​of the spectrum according to the filename coordinate rules, and writes the calculated peak area into the corresponding spatial grid position. Repeating the above process for all valid sequence data files in the same batch of MXene data, the system generates a two-dimensional peak area matrix with spatial coordinates as row and column indices and peak areas as values. This peak area matrix can be used to generate the spatial distribution results of the target Raman peaks, thus visually displaying the differences in Raman signal intensity at different locations on the MXene sample surface. Simultaneously, the peak areas are arranged in one of the following orders: folder order, acquisition time order, or filename order, forming an intensity evolution trend sequence result. Therefore, this invention can transform Raman data, which originally required manual processing of each spectrum, integration of each peak, and manual coordinate arrangement, into a continuous and automated batch analysis process.

[0054] This embodiment uses manually verified peaks as a reference to compare the fixed threshold plus fixed integration method with the present invention. Statistical results show that in 1504 sequence data files and 11602 manually verified peaks, the fixed threshold plus fixed integration method generated 988 falsely identified peaks, while the present invention generated only 203 falsely identified peaks, showing a significant reduction in the number of falsely identified peaks. The average deviation of the fixed integration method is concentrated between 7.9% and 17.6%, while the average deviation of the present invention is concentrated between 1.8% and 4.6%, indicating that peak shape competition path and integral boundary verification search can reduce the impact of erroneous peak shapes on peak area. In terms of processing time, manual processing and fixed rule verification take 1311 minutes, while the present invention takes 154 minutes to complete the same batch of data processing, showing a significant improvement in batch processing efficiency. In the imaging sequence, the present invention reduces the number of missing peak area matrices, improves the continuity of coordinate writing, and reduces discrete outliers in the spatial distribution map.

[0055] Table 1: Comparison of Implementation Data for Multi-Module Sequence Data Identification and Parsing

[0056] Table 1 shows that background drift and low signal-to-noise ratio (SNR) significantly affect the fixed threshold method. In sequences with slow background drift, the number of false peaks identified by the fixed threshold was 126, while this invention reduced it to 19. In sequences with low SNR, the number of false peaks identified by the fixed threshold was 115, while this invention reduced it to 29. This difference stems from the combination of the background suppression channel and the isotopic closure check; after the gradual change region is truncated during propagation, background fluctuations no longer directly form candidate peak states. In Raman fluorescence residual sequences, the average deviation of the fixed integral was 15.2%, while the average deviation of this invention was 3.3%, indicating that the continued state of the peak shoulders on both sides of the unidirectional propagation entrance can help the model avoid false peak shoulders caused by residual background.

[0057] Table 1 also reflects the integration advantages of this invention in the broad peak and overlapping peak scenarios. In infrared broad peak sequences, the average deviation of this invention is 2.9%, and in chromatographic overlapping peak sequences, the average deviation is 3.5%. Fixed rules are prone to having boundaries that are too narrow or too wide in these two types of data. This invention determines the search range through narrow peak paths, medium peak paths, and broad peak paths, and then performs reverse backfeeding by verifying the search position offset through integration boundary verification. When the boundary is unstable, the search is restarted, reducing the situation where erroneous boundaries directly enter the peak area calculation. In batch folder mixed sequences and coordinate imaging sequences, the processing time of this invention is 24 minutes and 20 minutes, respectively, which is significantly shorter than manual verification. Spatial distribution imaging results and intensity evolution trend sequence results can be generated within the same processing link.

[0058] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A multi-module sequence data recognition and parsing method based on machine learning, characterized in that, Includes the following steps: Read chemical characterization sequence data files from spectroscopy, chromatography, and mass spectrometry tests, form an ordered measurement sequence, and after verification, background subtraction, smoothing and denoising, send it into a peak-gated structured state-space sequence model; Within the peak-gated structured state-space sequence model, a background suppression channel and a peak continuation channel are set up. The background suppression channel divides the gradual change region and the abrupt bulge region according to the difference before and after background subtraction, driving state locking. The peak continuation channel drives the peak shoulder continuation according to the adjacent sampling changes after smoothing and denoising. The background suppression channel is blocked in the gradual change region and the propagation is cut off in the sudden change region. The unidirectional propagation entrance is opened in the sudden change convex region. The peak continuation channel continuously tracks the peak shoulder continuation state along both sides of the unidirectional propagation entrance. Perform a co-position closure check on the unidirectional propagation inlet and the continuous state of the two side peaks, and only open the candidate peak state when the co-position closure is successful. When a candidate peak enters the peak competition path, the narrow peak path, medium peak path, and wide peak path are activated mutually exclusively according to the peak shoulder span, and the strong peak path and weak peak path are activated mutually exclusively according to the relative amplitude of the peak. The integration boundary search is driven by the activated peak competition path. The integration boundary is locked by triple verification of attenuation, width and local baseline. The integration boundary is then checked and the peak area is calculated after subtracting the local baseline. The search position offset of the integral boundary verification is reversed and fed back to the peak competition path. When the preset offset threshold is reached, the peak competition path is closed and the search is restarted. When the preset offset threshold is not reached, the spatial distribution imaging results and intensity evolution trend sequence results are generated according to the file name coordinate rules.

2. The method for multi-module sequence data recognition and parsing based on machine learning according to claim 1, characterized in that, The formation process of the ordered measurement sequence specifically includes: The chemical characterization sequence data file is derived from at least one of the following: Raman spectroscopy, infrared spectroscopy, ultraviolet-visible spectroscopy, gas chromatography, liquid chromatography, and mass spectrometry, obtained from in-situ and operational tests of chemical research objects and chemical reaction processes. The chemical characterization sequence data file is parsed according to the file format, and the order of independent variables and the order of measurement amplitudes are read. The order of the independent variables corresponds to one of wavenumber, wavelength, retention time, mass-to-charge ratio, sampling time, and spatial sampling coordinates; the order of the measurement amplitudes corresponds to one of Raman scattering intensity, infrared absorption intensity, absorbance, chromatographic detection signal, and mass spectrometry ion count value. The independent variable order is paired with the measurement amplitude order point by point according to the original sampling order. Sampling positions that are not paired are not written into the ordered measurement sequence. The paired sampling locations are sequentially numbered to form an ordered measurement sequence.

3. The method for multi-module sequence data recognition and parsing based on machine learning according to claim 1, characterized in that, The background suppression channel is divided into a gradual change region and a sudden abrupt change region based on the difference before and after background subtraction, specifically including: The difference between the background before and after subtraction is read point by point along the direction of the ordered measurement sequence, and the difference between the background before and after subtraction at adjacent sampling positions is compared. Sampling locations where continuous differences fall within a preset gradual variation range are marked as gradual variation regions, and sampling locations where differences deviate from the preset gradual variation range and the background subtracts the difference to form a local bulge are marked as abrupt bulge regions. The background suppression channel writes a state lockout flag to the gradually changing region and a unidirectional propagation entrance open flag to the abruptly changing protruding region.

4. The multi-module sequence data recognition and parsing method based on machine learning according to claim 3, characterized in that, The background suppression channel writes a state lockout flag to the gradually changing region and a unidirectional propagation entrance open flag to the abruptly changing protrusion region, specifically including: Background suppression channel read status lock flag, sets the propagation status of the sampling position in the gradual change zone to the lock state, and cuts off the propagation of the lock state to the adjacent sampling position; The background suppression channel reads the unidirectional propagation entrance open marker, sets the sampling position with the largest difference before and after background subtraction in the mutation protrusion region as the unidirectional propagation entrance, retains the propagation link from the unidirectional propagation entrance to the co-position closure check, and closes the propagation link from the sampling positions in the mutation protrusion region other than the unidirectional propagation entrance to the co-position closure check.

5. The method for multi-module sequence data recognition and parsing based on machine learning according to claim 1, characterized in that, The peak-shaped continuity channel continuously tracks the peak shoulder continuity state along both sides of the unidirectional propagation entrance, specifically including: The peak-shaped continuation channel starts from the unidirectional propagation entrance and reads the smoothed and denoised adjacent sampling changes at the forward and backward sampling positions along the direction of the ordered measurement sequence. Sampling positions where the direction of change remains consistent and the amplitude of change decreases are written as peak shoulder continuation state, and sampling positions where the direction of change reverses and the amplitude of change reaches the preset abrupt change threshold are written as peak shoulder termination state. At the peak shoulder termination state, the propagation of the peak shoulder continuation state to the outer sampling position is interrupted.

6. The method for multi-module sequence data recognition and parsing based on machine learning according to claim 5, characterized in that, The process of performing a co-location closure check on the unidirectional propagation entrance and the continuous state of the two side peaks specifically includes: Using the sampling position where the unidirectional propagation entrance is located as the verification center, read the peak and shoulder continuation status of the continuous sampling positions on the front side and the peak and shoulder continuation status of the continuous sampling positions on the back side. When the peak shoulder continuity state of the current continuous sampling position, the unidirectional propagation entrance, and the peak shoulder continuity state of the subsequent continuous sampling position are continuously connected in the direction of the ordered measurement sequence, it is determined that the co-position closure is established and the candidate peak state is opened. If any of the peak shoulder continuation states at the front continuous sampling position or the rear continuous sampling position is disconnected from the unidirectional propagation entrance, the co-position closure is determined to be invalid, and the candidate peak state is closed.

7. The method for multi-module sequence data recognition and parsing based on machine learning according to claim 6, characterized in that, The mutual exclusion activation specifically includes: The peak-shaped competition path reads the peak shoulder continuation state of the consecutive sampling positions on the front side and the peak shoulder continuation state of the consecutive sampling positions on the back side, and calculates the sampling interval between the peak shoulder continuation state of the consecutive sampling positions on the front side and the peak shoulder continuation state of the consecutive sampling positions on the back side as the peak shoulder span. Based on the peak shoulder span, select one path from the narrow peak path, medium peak path, and wide peak path to open, and close the unselected path; The peak-shaped competition path reads the amplitude corresponding to the unidirectional propagation entrance, the amplitude corresponding to the peak shoulder continuation state at the front continuous sampling position, and the amplitude corresponding to the peak shoulder continuation state at the rear continuous sampling position. The difference between the amplitude corresponding to the unidirectional propagation entrance and the average of the amplitude corresponding to the peak shoulder continuation state at the front continuous sampling position and the amplitude corresponding to the peak shoulder continuation state at the rear continuous sampling position is taken as the peak relative amplitude. Based on the relative amplitude of the peak, one path is selected from the strong peak path and the weak peak path to be opened, and the unselected path is closed.

8. The method for multi-module sequence data recognition and parsing based on machine learning according to claim 7, characterized in that, The integral boundary search driven by the activated peak-shaped competition path specifically includes: The peak position is taken as the sampling position corresponding to the unidirectional propagation entrance. The boundary search range is determined by one of the open paths among the narrow peak path, medium peak path, and wide peak path. The attenuation verification threshold is determined by one of the open paths among the strong peak path and weak peak path. Search from the peak position along the direction of the ordered measurement sequence to the front and back sampling positions, and perform attenuation verification, width verification, and local baseline verification point by point; The front sampling position that simultaneously passes attenuation verification, width verification, and local baseline verification is locked as the left integration boundary, and the rear sampling position that simultaneously passes attenuation verification, width verification, and local baseline verification is locked as the right integration boundary.

9. The method for multi-module sequence data recognition and parsing based on machine learning according to claim 8, characterized in that, The process of performing a verification search on the integral boundary and calculating the peak area after subtracting local baselines specifically includes: Using the left and right integration boundaries as initial boundaries, a verification search is performed outside the left and right integration boundaries along the direction of the ordered measurement sequence to obtain the positions of the left and right verification boundaries. The position of the left verification boundary is compared with the left integration boundary, and the position of the right verification boundary is compared with the right integration boundary to obtain the offset of the integration boundary verification search position. Select the continuous sampling positions inside the left integration boundary and the continuous sampling positions inside the right integration boundary as local baseline fitting points, and fit the local baseline; The local baseline is subtracted from the measured amplitude between the left and right integration boundaries. The area of ​​the peak is obtained by summing the measured amplitude after subtraction and combining the independent variable intervals of adjacent sampling positions.

10. The method for multi-module sequence data recognition and parsing based on machine learning according to claim 9, characterized in that, The step of reverse-feeding the search position offset of the integral boundary verification to the peak-shaped competition path specifically includes: The peak-shaped competition path reads the integral boundary verification search position offset and compares the integral boundary verification search position offset with the preset offset threshold. When the offset of the search position for the integral boundary verification reaches the preset offset threshold, the peak competition path for generating the left and right integral boundaries is closed, the candidate peak state is retained, and the mutual exclusion activation of the peak competition path and the integral boundary search are re-executed. When the offset of the search position for the integral boundary verification does not reach the preset offset threshold, the peak competition path for generating the left and right integral boundaries is retained, and the peak area is used to generate spatial distribution imaging results and intensity evolution trend sequence results according to the file name coordinate rules.

Images