Automatic analysis system and method for chemical composition of petroleum products based on deep learning

Abstract

The application discloses a deep learning-based automatic analysis system and method for chemical composition of petroleum products, and relates to the field of petroleum product analysis.The system comprises the following steps: obtaining chromatography-mass spectrometry full-scan data of a petroleum sample, detecting peak vertices and extracting parameters such as retention time, peak height and peak width; calculating the ratio of the interval between adjacent peaks and the sum of the widths of the two peaks as a relative interval ratio; judging the degree of peak overlap according to the ratio, and dividing the peaks into three categories: highly overlapped, moderately overlapped or separable; separating the overlapped peaks by adjusting the peak width parameter, re-verifying the interval ratio after separation; and if the interval ratio does not meet the standard, updating the parameters or calculating a correction coefficient to optimize the separation result; the application can automatically identify and process peak overlap problems, can perform secondary optimization on peaks that cannot be successfully separated, and is suitable for chemical composition analysis of complex petroleum samples.

Description

Technical Field

[0001] This invention relates to the field of petroleum product analysis, specifically to an automated analysis system and method for the chemical composition of petroleum products based on deep learning. Background Technology

[0002] In fields such as petrochemicals, environmental monitoring, and biomedicine, the chemical composition analysis of complex mixtures relies heavily on chromatography-mass spectrometry (GC-MS). This technique achieves preliminary separation by utilizing the differences in retention times of compounds within the chromatographic column, followed by qualitative and quantitative detection via mass spectrometry.

[0003] However, complex systems such as petroleum products often contain hundreds of components with highly similar physicochemical properties (boiling point, polarity), leading to severe "co-elution" in chromatographic separation—that is, multiple compounds fail to separate sufficiently, forming overlapping peaks on the spectrum. Especially when high and low concentration components coexist, the broad peaks of high-concentration components often mask the weak signals of low-concentration components, significantly reducing quantitative accuracy. Traditional methods rely heavily on manual judgment when handling overlapping peaks, resulting in high subjectivity and low efficiency. Furthermore, current technologies lack a hierarchical processing mechanism for peak overlap of varying degrees, failing to adaptively select the optimal separation strategy based on the degree of overlap, leading to unstable separation results.

[0004] To address the aforementioned issues, existing technologies urgently need improvement. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an automated analysis system and method for the chemical composition of petroleum products based on deep learning.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] In a first aspect, this invention discloses an automated method for analyzing the chemical composition of petroleum products based on deep learning, comprising the following steps:

[0008] Obtain raw GC-MS full scan data of petroleum samples;

[0009] Peak detection is performed on the GC-MS full scan raw data to locate all peak vertices and extract peak feature parameters; the peak feature parameters include the retention time, peak height, peak width and retention time interval of each peak;

[0010] Calculate the relative spacing ratio between adjacent peaks based on the peak characteristic parameters;

[0011] If the relative spacing ratio is less than a preset first distance threshold, the adjacent peaks are marked as highly overlapping peaks; otherwise, if the relative spacing ratio is less than a preset second distance threshold, the adjacent peaks are marked as moderately overlapping peaks; otherwise, they are marked as separable peaks.

[0012] The peak width is adjusted based on the initial feature parameters, the relative spacing ratio, and the preset initial compression factor. Peak separation is then performed on highly overlapping peaks and moderately overlapping peaks based on the adjusted peak width.

[0013] After peak separation, peak feature parameters are re-extracted and the relative spacing ratio is calculated. It is determined whether the relative spacing ratio is greater than a preset first distance threshold. If it is, the separation result is accepted. Otherwise, it is further determined whether the adjacent peaks are highly overlapping peaks. If they are, the peak correction coefficient is calculated based on the peak feature parameters. Otherwise, the initial compression factor is updated and peak separation is performed again. If the number of updates exceeds the preset number of iterations, the peak correction coefficient is calculated based on the peak height.

[0014] Peak separation is performed on overlapping peaks based on the peak correction coefficient.

[0015] Secondly, this invention discloses an automated chemical composition analysis system for petroleum products based on deep learning, comprising:

[0016] The data acquisition module is used to acquire raw GC-MS full scan data of petroleum samples;

[0017] The peak detection module is used to perform peak detection on the GC-MS full scan raw data, locate all peak vertices and extract peak feature parameters; the peak feature parameters include the retention time, peak height, peak width and retention time interval of each peak;

[0018] The peak overlap determination module is used to calculate the relative spacing ratio between adjacent peaks based on the peak feature parameters; determine whether the relative spacing ratio is less than a preset first distance threshold; if so, mark the adjacent peak as a highly overlapping peak; otherwise, further determine whether the relative spacing ratio is less than a preset second distance threshold; if so, mark the adjacent peak as a moderately overlapping peak; otherwise, mark it as a separable peak.

[0019] The first separation module is used to adjust the peak width of highly overlapping peaks and moderately overlapping peaks based on the initial feature parameters, the relative spacing ratio, and the preset initial compression factor, and to perform peak separation and re-extract peak feature parameters according to the optimized peak widths.

[0020] The separation verification module is used to calculate the relative spacing ratio after peak separation based on the re-extracted peak feature parameters, determine whether the relative spacing ratio is greater than a preset first distance threshold, and if so, accept the separation result; otherwise, further determine whether the adjacent peaks are highly overlapping peaks. If so, calculate the peak correction coefficient based on the peak feature parameters; otherwise, update the initial compression factor and perform peak separation again. If the number of updates exceeds the preset number of iterations, calculate the peak correction coefficient based on the peak height.

[0021] The second separation module is used to separate overlapping peaks according to the peak correction coefficient.

[0022] The beneficial effects of this invention are as follows: By employing techniques such as peak detection and feature extraction, relative spacing ratio calculation, peak separation methods, peak correction, peak separation result verification, and the application of deep learning models, this invention effectively solves the problem of separating complex overlapping peaks in GC-MS data, greatly improving the accuracy of peak separation. The method of this invention can automatically identify and handle peak overlap problems without manual intervention, improving the automation level of analysis and adapting to the chemical composition analysis of complex petroleum samples. Optimizing the separation effect of moderately overlapping peaks through a deep learning model ensures the accuracy of the analytical results and adapts to complex background information during peak separation. Through an iterative update mechanism of the peak correction coefficient and the initial compression factor, this invention can perform secondary optimization on peaks that failed to be separated, further improving the stability and accuracy of the separation effect. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is an overall block diagram of the method in Embodiment 1 of the present invention;

[0025] Figure 2 This is a flowchart of the method according to Embodiment 1 of the present invention;

[0026] Figure 3 This is an overall block diagram of the system in Embodiment 3 of the present invention. Detailed Implementation

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

[0028] Application Overview: In existing technologies, the chemical composition analysis of complex mixtures mainly relies on chromatography-mass spectrometry (GC-MS), which separates compounds based on differences in retention times. However, complex systems such as petroleum products often contain hundreds of components with similar physicochemical properties, leading to co-elution and overlapping peaks during chromatographic separation. When high- and low-concentration components coexist, the broad peaks of the high-concentration components often mask the weak signals of the low-concentration components, resulting in decreased quantitative accuracy. For example, when analyzing crude oil samples at a petrochemical company, multiple isomers formed overlapping peaks due to their similar retention times, making it impossible to accurately separate them using traditional peak identification methods, thus affecting the quantitative results.

[0029] To address these issues, researchers discovered that the difficulty of separating overlapping peaks is related to the ratio of peak spacing to peak width, but existing methods lack quantitative evaluation metrics. Analysis revealed that using the ratio of the retention time interval between adjacent peaks to the sum of their peak widths as an indicator of separation difficulty allows for the establishment of a tiered processing mechanism. Further investigation showed that dynamically adjusting the peak width parameter in conjunction with an iterative optimization strategy can improve separation performance. For highly overlapping peaks, a model fitting optimization boundary is used; for moderately overlapping peaks, a deep learning model is introduced to establish a separation baseline, ultimately forming a technical approach combining tiered processing and dynamic adjustment.

[0030] Example 1:

[0031] like Figure 1-2 As shown, the automated chemical composition analysis method for petroleum products based on deep learning includes the following steps:

[0032] Obtain raw GC-MS full scan data of petroleum samples;

[0033] Peak detection is performed on the GC-MS full scan raw data to locate all peak vertices and extract peak feature parameters; the peak feature parameters include the retention time, peak height, peak width and retention time interval of each peak;

[0034] Calculate the relative spacing ratio between adjacent peaks based on the peak characteristic parameters;

[0035] If the relative spacing ratio is less than a preset first distance threshold, the adjacent peaks are marked as highly overlapping peaks; otherwise, if the relative spacing ratio is less than a preset second distance threshold, the adjacent peaks are marked as moderately overlapping peaks; otherwise, they are marked as separable peaks.

[0036] The peak width is adjusted based on the initial feature parameters, the relative spacing ratio, and the preset initial compression factor. Peak separation is then performed on highly overlapping peaks and moderately overlapping peaks based on the adjusted peak width.

[0037] After peak separation, peak feature parameters are re-extracted and the relative spacing ratio is calculated. It is determined whether the relative spacing ratio is greater than a preset first distance threshold. If it is, the separation result is accepted. Otherwise, it is further determined whether the adjacent peaks are highly overlapping peaks. If they are, the peak correction coefficient is calculated based on the peak feature parameters. Otherwise, the initial compression factor is updated and peak separation is performed again. If the number of updates exceeds the preset number of iterations, the peak correction coefficient is calculated based on the peak height. Peak separation is performed on the overlapping peaks based on the peak correction coefficient.

[0038] The GC-MS full-scan raw data refers to three-dimensional data containing time, mass-to-charge ratio, and ion intensity information acquired through gas chromatography-mass spectrometry (GC-MS). Specifically, it can be achieved by acquiring mass spectrometry signals in full-scan mode, used to comprehensively record the chromatographic separation and mass spectrometry characteristics of each compound in petroleum samples. Peak detection refers to identifying and extracting the chromatographic peaks corresponding to compounds from the GC-MS data. Specifically, peak apex locations can be determined using derivative change point detection algorithms, and peak shape characteristics can be described by combining peak height and peak width parameters to determine the location and signal intensity of each component in the mixture. The relative spacing ratio is the ratio of the retention time interval between adjacent peaks to the sum of their peak widths. Specifically, it can be calculated by dividing the retention time difference by the sum of the peak widths, used to quantify the degree of peak overlap and classify the separation difficulty level. Peak width adjustment refers to dynamically correcting the chromatographic peak width parameter based on the relative spacing ratio. Specifically, the peak width is adjusted by establishing a linear relationship between the compression factor and the reciprocal of the relative spacing ratio, used to optimize the separation boundary of overlapping peaks and improve subsequent separation accuracy. Peak separation verification refers to recalculating the relative spacing ratio of the adjusted peaks to evaluate the separation effect. Specifically, it involves comparing the separated parameters with a preset threshold to determine whether to accept the result or trigger a correction mechanism, in order to ensure that the separation process converges and avoid invalid iterations.

[0039] This application further proposes a process for locating all peaks, which includes: calculating the first derivative of the GC-MS full scan raw data, and taking the point where the first derivative changes from a positive value to a negative value as the peak.

[0040] The calculation of the first derivative is implemented using a difference algorithm, taking the signal strength difference between adjacent data points as the basis for derivative calculation. The identification of change points employs a sign-based method; when the sign of the derivative of two consecutive data points changes from positive to negative, the preceding data point is marked as the peak. This process filters out minute fluctuations through a preset noise threshold, retaining only change points exceeding the threshold as valid peaks.

[0041] Specifically, on the time-signal intensity curve of the raw GC-MS full-scan data, the peak corresponds to the inflection point where the signal intensity changes from rising to falling. The signal change trend can be quantified by calculating the first derivative; a positive value indicates the rising phase, and a negative value indicates the falling phase. When a change in the derivative sign from positive to negative is detected, that time point corresponds to a local maximum point of the signal intensity curve. For example, within the retention time window, if the derivative of the nth data point is +0.5 and the derivative of the (n+1)th data point is -0.3, then the nth data point is determined to be the peak. This process replaces manual visual judgment with numerical calculation, eliminating subjective errors and achieving automated and precise peak location.

[0042] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0043] When performing peak detection on the raw GC-MS full-scan data, the process of locating peak vertices includes calculating the first derivative. Specifically, the intensity values ​​of the raw data are numerically differentiated to obtain the first derivative curve. On the first derivative curve, each point where a value changes from positive to negative is identified as a peak vertices. For example, the central difference method can be used to calculate the first derivative.

[0044] f'(x)≈[f(x+h)-f(xh)] / (2h), where f(x) is the intensity value of the original data point and h is the step size. By traversing the first derivative curve, the peak position is marked at each positive and negative change point. Furthermore, a signal-to-noise ratio threshold can be used to filter out false peaks caused by noise.

[0045] Through the above technical solution, this application can accurately locate the apex position of chromatographic peaks, laying the foundation for subsequent peak characteristic parameter extraction and overlapping peak separation. This improves the accuracy and reliability of peak detection and reduces errors caused by human judgment. Specifically, the peak apex localization method based on the first derivative has good adaptability to chromatographic peaks of different shapes and widths, and can effectively address the challenge of detecting co-elution peaks in complex samples.

[0046] This application further proposes a relative spacing ratio, which is the ratio of the retention time interval between adjacent peaks to the sum of the peak widths of the two peaks. This relative spacing ratio is used to quantify the difficulty of separating adjacent peaks.

[0047] The retention time interval refers to the difference in residence time of two adjacent peak apexes in the chromatographic column, which can be calculated by the time coordinate difference on the chromatogram. This parameter directly reflects the absolute value of the distance between peaks. The peak width sum refers to the sum of the base widths of adjacent peaks, which can be measured by the half-width at half-maximum (WHM) or integral width of the chromatographic peak. This parameter reflects the degree to which peak broadening occupies the separation space. The relative spacing ratio, calculated by comparing the retention time interval with the peak width sum, can eliminate absolute value fluctuations caused by differences in instrument conditions and establish a unified evaluation index for separation difficulty.

[0048] Specifically, in chromatographic separation, the retention time interval and peak width of adjacent peaks jointly determine their degree of overlap. When the retention time interval is fixed, increasing the peak width will lead to an increased probability of overlap in the peak base region; when the peak width is fixed, shortening the retention time interval will directly cause overlap in the peak apex region. By combining these two factors to calculate the relative spacing ratio, the type of peak overlap can be accurately determined: when this ratio is small, it indicates that the peak spacing is insufficient to offset the peak width superposition effect, and forced separation is required; when this ratio is large, it indicates that the peak spacing is sufficient to achieve baseline separation. This quantitative method provides an objective basis for subsequent selection of peak separation algorithms.

[0049] Traditional methods typically rely solely on retention time interval or peak width as a single parameter to determine separation status, neglecting the synergistic effect of both on peak overlap. For example, judging solely based on the retention time interval threshold may misclassify a broad peak as an overlapping peak or a narrow peak as a separable peak. This approach, by establishing a ratio between retention time interval and peak width, more accurately reflects the actual separation difficulty and avoids misjudgments caused by fluctuations in instrument parameters or differences in sample matrix.

[0050] Through the above technical solution, this application effectively solves the problem of misjudgment caused by the single standard of traditional peak separation judgment. It accurately identifies highly overlapping peaks and moderately overlapping peaks through quantitative indicators, providing a reliable basis for subsequent selection of exponentially modified Gaussian model or deep learning separation model, thereby improving the accuracy and analytical efficiency of chemical component separation in complex petroleum samples.

[0051] This application further proposes a method to determine whether the relative spacing ratio is less than a preset first distance threshold. If so, the adjacent peaks are marked as highly overlapping peaks; otherwise, it further determines whether the relative spacing ratio is less than a preset second distance threshold. If so, the adjacent peaks are marked as moderately overlapping peaks; otherwise, they are marked as separable peaks. Specifically, the first distance threshold is set to 0.5, and the second distance threshold is set to 1.2. When the relative spacing ratio is less than 0.5, it indicates that the two peaks severely overlap and are marked as highly overlapping peaks; when the relative spacing ratio is between 0.5 and 1.2, it is marked as a moderately overlapping peak; and when the relative spacing ratio is greater than 1.2, it is marked as a separable peak.

[0052] This application further proposes the following formula for calculating the adjusted peak width:

[0053]

[0054] in The initial compression factor, The initial peak width, The adjusted peak width, This represents the ratio of the relative spacing between adjacent peaks.

[0055] The initial compression factor refers to a preset coefficient used to control the peak width adjustment range. It can be determined empirically or through experimental optimization, for example, a value within the range of 0.1 to 0.3. Its function is to balance the peak width adjustment range with the stability of the separation effect. The initial peak width refers to the peak width parameter extracted from the original data using a peak detection algorithm. It can be calculated using half-peak width or Gaussian fitting width, reflecting the extent of peak expansion along the time axis. The adjusted peak width refers to the optimized peak width used for peak separation after calculation using a formula. It can be dynamically adjusted to adapt to peak separation requirements with different degrees of overlap.

[0056] Traditional methods typically adjust peak width using fixed coefficients, failing to dynamically optimize parameters based on peak spacing, leading to unstable separation results for highly overlapping peaks. This approach introduces dynamic calculations of the relative spacing ratio and the initial compression factor, enabling peak width adjustment to precisely match the actual separation difficulty of adjacent peaks, thereby improving peak separation accuracy.

[0057] Through the above technical solution, this application solves the problem that traditional peak width adjustment methods cannot meet the separation requirements of peaks with different degrees of overlap. By dynamically calculating and optimizing the peak width parameters, the separation accuracy of highly overlapping peaks and moderately overlapping peaks is significantly improved, providing a reliable data foundation for subsequent quantitative analysis of chemical composition.

[0058] This application further proposes a specific process for peak separation in an automated analysis method for the chemical composition of petroleum products based on deep learning, including: for highly overlapping peaks, determining the overlapping region by extending the adjusted peak width by a factor of 2 to the left and right of the center of the two peak vertices of the overlapping peak; fitting the left and right peaks respectively with an exponentially modified Gaussian model within the overlapping region, optimizing the peak boundary with the goal of minimizing the fitting residual; for moderately overlapping peaks, extracting a subset of the GC-MS full-scan raw data containing the adjusted peak width, retention time interval, peak height, and corresponding time window of the current moderately overlapping peak, and inputting it along with the GC-MS full-scan raw data into a pre-trained deep learning separation model; the deep learning separation model outputs a separation baseline, which characterizes the background contribution of the overlapping peak region; subtracting the separation baseline from the GC-MS full-scan raw data to obtain the separated single-peak profile; and re-extracting the peak vertices and peak boundaries based on the single-peak profile.

[0059] The processing methods for highly overlapping peaks involve determining the boundary of the overlapping region by expanding and adjusting the peak width range. This can be achieved by symmetrically expanding the peak apex, ensuring the overlapping region covers the area where the two peaks interfere with each other. The exponentially modified Gaussian model is a hybrid model combining a Gaussian distribution and an exponential decay function. Specifically, it achieves asymmetric peak fitting by adjusting peak shape parameters, thus adapting to peak tailing phenomena. The deep learning separation model for moderately overlapping peaks is a baseline prediction model based on a convolutional neural network. This can be implemented using a network structure including a time-series feature extraction module and a regression output layer, learning the distribution pattern of the background signal of overlapping peaks through training data. Baseline subtraction involves eliminating interference signals from co-eluting components from the original data, specifically by extracting the true contour of the target peak through numerical subtraction.

[0060] Specifically, in the separation of highly overlapping peaks, the overlap region is first determined based on the adjusted peak width parameters to ensure coverage of the chromatographic region where the two peaks interfere with each other. Then, an exponentially modified Gaussian model is used to independently fit the left and right peaks. By minimizing the sum of squared residuals between the fitted curve and the original data, the peak boundary positions and peak shape parameters are dynamically adjusted, ultimately achieving mathematical separation of the overlapping peaks. For moderately overlapping peaks, the adjusted peak parameters and a subset of the original data are input into a pre-trained model. The model automatically identifies the background signal pattern in the overlapping region and generates a baseline characterizing the contribution of the co-eluting components. After subtracting this baseline from the original data, the residual signal reflects the true shape of the target single peak, and the peak apex and boundary parameters are then recalculated.

[0061] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0062] The peak separation process includes processing highly overlapping peaks and moderately overlapping peaks. For highly overlapping peaks, the overlapping region is determined by extending the peak width by a factor of 2 to the left and right of the center of the two peak vertices of the overlapping peak. Within the overlapping region, the left and right peaks are fitted using an exponentially modified Gaussian model, and the peak boundaries are optimized with the goal of minimizing the fitting residuals.

[0063] For moderately overlapping peaks, a subset of the GC-MS full-scan raw data with adjusted peak width, retention time interval, peak height, and corresponding time window is extracted and input along with the GC-MS full-scan raw data into a pre-trained deep learning separation model. The deep learning separation model outputs a separation baseline, which represents the background contribution of the overlapping peak region.

[0064] The separation baseline is subtracted from the GC-MS full-scan raw data to obtain the separated single-peak profile. The peak apex and peak boundary are then extracted based on the single-peak profile.

[0065] Specifically, for highly overlapping peaks, the following steps can be taken: First, determine the extent of the overlap region. Using the midpoint between the two peak vertices as the center, extend the overlap region to the left and right by twice the adjusted peak width, defining the boundaries of the overlap region. Second, apply an exponentially modified Gaussian model within the overlap region. This model considers peak asymmetry and can more accurately describe the shape of the chromatographic peak. Third, use an iterative optimization algorithm, such as the Levenberg-Marquardt algorithm, to minimize the fitting residuals and obtain the optimal peak boundaries.

[0066] The specific formula for the exponentially modified Gaussian model is as follows:

[0067]

[0068] Where A is the peak height. To retain the time, is the standard deviation, and k is the exponential correction coefficient. For example, k=0.05 can effectively correct the peak trailing shadow.

[0069] For moderately overlapping peaks, a deep learning separation model can employ a convolutional neural network (CNN) structure. This model is pre-trained using a large amount of data on known overlapping peaks to learn peak shape characteristics under different degrees of overlap. In practical applications, the model input includes: adjusted peak width, retention time interval between adjacent peaks, peak height, and the original GC-MS full-scan data for the corresponding time window. The model output is a separation baseline, representing the background signal of the overlapping region.

[0070] Finally, by subtracting the separation baseline from the original data, the contours of the separated individual peaks are obtained. Based on these contours, the peak apex positions and peak boundaries can be redefined, thus completing the peak separation.

[0071] Through the above technical solutions, this application achieves adaptive separation of peaks with varying degrees of overlap. For highly overlapping peaks, the exponentially modified Gaussian model more accurately describes the peak shape, improving separation precision. For moderately overlapping peaks, a deep learning model is introduced to learn and utilize peak shape features from a large amount of historical data, enhancing the robustness of separation. Furthermore, the concept of a separation baseline effectively eliminates interference from background signals, improving the detection sensitivity of low-concentration components. Overall, this method significantly improves the separation effect of co-eluting components in complex mixtures, laying the foundation for subsequent qualitative and quantitative analysis.

[0072] This application further proposes a process for calculating peak correction coefficients based on peak characteristic parameters, which includes extracting a subset of full-scan mass spectrometry data within the retention time window corresponding to highly overlapping peaks. This subset contains three-dimensional information on the time-mass-charge ratio-ion intensity of the co-eluting components. Based on the characteristic ions of the target compound, characteristic ion clusters of the co-eluting components are identified in the mass spectrometry data subset. The full-scan mass spectrometry data subset and the characteristic ion clusters are analyzed using a deconvolution algorithm to obtain the peak correction coefficients of the two peaks in the highly overlapping peaks.

[0073] The full-scan mass spectrometry data subset refers to a three-dimensional data set extracted from the raw GC-MS data, corresponding to the retention time window of highly overlapping peaks. This subset includes time axis, mass-to-charge ratio axis, and ion intensity information, and can be implemented using a sliding time window extraction method. It provides a complete signal distribution of the co-eluting components in both chromatographic and mass spectrometric dimensions. Characteristic ion clusters refer to unique and highly abundant combinations of characteristic ions in the mass spectrometer for the target compound. This can be achieved through matching with a standard mass spectrometry library of compounds, and is used to distinguish the chemical composition of the co-eluting components. The deconvolution algorithm is a mathematical method for separating the contributions of each component from the mixed signal. It can be implemented using alternating least squares multivariate curve resolution to eliminate signal interference between overlapping peaks.

[0074] Specifically, in the processing of highly overlapping peaks, the time window range is first determined based on the peak retention time. Ion intensity data for all mass-to-charge ratio channels within this window are extracted, forming a data subset containing three-dimensional information of time, mass-to-charge ratio, and ion intensity. Then, based on the known characteristic ions of the target compound, corresponding characteristic ion clusters are selected from the mass spectrometry data subset to determine the mass spectrometry characteristics of the co-eluting components. Furthermore, the full-scan mass spectrometry data subset and the characteristic ion clusters are input into a deconvolution algorithm. Through iterative optimization of the signal distribution of each component in the time and mass-to-charge ratio dimensions, the contribution ratio of each peak is analyzed, and this ratio is the peak correction coefficient. This achieves accurate separation of highly overlapping peaks.

[0075] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0076] Extract a subset of full-scan mass spectrometry data within the retention time window corresponding to highly overlapping peaks. This subset contains three-dimensional information on the time-mass-charge ratio-ion intensity of the co-eluted components. For example, for a highly overlapping peak with a retention time of 10.5–11.5 minutes, extract the full-scan mass spectrometry data within this time window.

[0077] Based on the characteristic ions of the target compound, characteristic ion clusters of co-eluting components are identified in a subset of mass spectrometry data. Specifically, using known characteristic ion information of the target compound, such as characteristic ions of alkanes at m / z 57 and 85, and characteristic ions of aromatics at m / z 91 and 105, the intensity variation trends of these characteristic ions are searched in the mass spectrometry data to identify the characteristic ion clusters of each co-eluting component.

[0078] The full-scan mass spectrometry data subset and characteristic ion clusters are analyzed using a deconvolution algorithm to obtain the peak correction coefficients for each peak in highly overlapping peaks. Further, deconvolution algorithms such as nonnegative matrix factorization are used to decompose the full-scan mass spectrometry data into the mass spectra and chromatographic profiles of each component. From this, the relative contribution of each component in the overlapping peaks, i.e., the peak correction coefficient, can be calculated.

[0079] Traditional methods rely solely on retention time or simple peak shape models for overlapping peak separation, failing to effectively distinguish the differences in mass spectrometry characteristics among co-eluting components. This proposed method integrates the three-dimensional information of full-scan mass spectrometry data, combined with the chemical specificity of characteristic ion clusters, and utilizes a deconvolution algorithm to simultaneously resolve signal overlap in both chromatographic and mass spectrometric dimensions, significantly improving the chemical resolution of peak separation. Furthermore, existing correction methods based on single ion channels are susceptible to matrix interference, while this method enhances its anti-interference capabilities through the synergistic analysis of multiple characteristic ion clusters.

[0080] Through the above technical solution, this application can accurately identify the chemical characteristics of co-eluting components, eliminate signal confusion caused by insufficient chromatographic separation, and improve the quantitative analysis accuracy of homologues and isomers in complex petroleum samples. At the same time, it avoids subjective errors from human experience-based judgment and achieves automated analysis of highly overlapping peaks.

[0081] This application further proposes a process for updating the initial compression factor, which includes setting a fixed update step size and updating the initial compression factor by summing the initial compression factor with the update step size.

[0082] The initial compression factor is a calculation parameter used to adjust the peak width. It can be implemented using a preset initial value, such as a range of 0.1 to 0.3. Its function is to optimize peak separation by dynamically adjusting the peak width. The update step size refers to the adjustment range of the compression factor in each iteration. It can be implemented using a fixed value, such as 0.05. Its function is to enhance the convergence of peak width adjustment by gradually increasing the compression factor, avoiding peak separation failure due to excessively large single adjustment ranges.

[0083] Specifically, during peak separation, if the relative spacing ratio after separation still does not meet the preset threshold and the current peak is not marked as a highly overlapping peak, the initial compression factor is updated incrementally with a fixed step size. For example, if the initial compression factor is 0.2, it can be increased by 0.05 with each update, and after multiple iterations, the compression factor gradually increases to 0.35. This process gradually expands the peak width adjustment range in a linear increment, thereby balancing peak separation accuracy and computational efficiency.

[0084] Through the above technical solution, this application can effectively avoid the problem of excessive peak width compression or insufficient separation caused by improper adjustment of compression factor, thereby improving the stability and repeatability of overlapping peak separation. In particular, when dealing with moderately overlapping peaks with similar retention time intervals, it can significantly reduce the need for manual intervention and improve the efficiency of automated analysis.

[0085] This application further proposes a process for calculating peak correction coefficients based on peak heights, which includes: calculating the peak correction coefficients of each of the two peaks in a moderately overlapping peak by comparing their respective peak heights with the sum of their peak heights.

[0086] The peak correction coefficient is a weighting parameter used to adjust the distribution ratio of overlapping peak areas. It can be calculated using the peak height ratio, which reflects the relative concentration contribution of each component in the mixed signal. The ratio calculation uses the quotient of a single peak height to the total peak height as the allocation basis. This can be achieved through normalization, which eliminates the influence of absolute intensity differences on the allocation results.

[0087] Specifically, when the separation verification module determines that the number of updates for moderately overlapping peaks exceeds the preset number of iterations, the peak height data is extracted as the calculation benchmark. The original peak height values ​​of the two peaks are input into the calculation unit, and a correction coefficient ranging from 0 to 1 is generated by dividing the peak height of each peak by the sum of the peak heights of the two peaks. This coefficient is passed to the second separation module, which, after calculating the total area of ​​the overlapping peaks by integration, splits the total area into the areas of two independent peaks according to the proportion of the correction coefficient. For example, if the peak height of the left peak accounts for 70% of the total height of the two peaks, then 70% of the overlapping peak area is allocated to the left peak, and the remaining 30% is allocated to the right peak.

[0088] Traditional methods for handling moderately overlapping peaks that are not completely separated typically employ fixed-ratio segmentation or direct background baseline subtraction, which cannot adapt to the complex dynamic changes in concentration in real samples. This approach dynamically generates correction coefficients based on peak height ratios, making the area distribution results more closely match the actual concentration ratios. Especially in scenarios where high-concentration components mask low-concentration components, it can effectively restore the quantitative information of weak signals.

[0089] Through the above technical solution, this application solves the problem of inaccurate quantification of low-content components caused by concentration differences in the chromatographic co-elution region. Through the peak height dynamic allocation mechanism, the chemical composition of overlapping peaks can still be accurately resolved when physical separation cannot be achieved, which significantly improves the detection sensitivity of trace components in complex petroleum samples.

[0090] This application further proposes to calculate the area of ​​overlapping peaks by integrating the original GC-MS full scan data, and to allocate the area of ​​overlapping peaks proportionally according to the peak correction coefficients of the two peaks in the overlapping peaks.

[0091] Among these, integral calculation refers to the quantitative calculation of the area under the chromatographic peak curve using mathematical methods, specifically the trapezoidal integral method or Simpson's integral method, to accurately characterize the total amount of compounds in the chromatographic separation process. Peak correction coefficient is a weighted parameter reflecting the true contribution ratio of each component in overlapping peaks, specifically calculated using deconvolution algorithms or peak height ratios, and used to correct area distribution deviations caused by peak overlap.

[0092] Specifically, after identifying overlapping peak regions in the GC-MS full-scan raw data, the total peak area of ​​these regions is first calculated using an integral algorithm. Then, based on pre-generated peak correction coefficients, the total peak area is allocated according to the coefficient ratios of the two peaks. For example, when the correction coefficients for the two peaks are 0.6 and 0.4, the total peak area will be split into 60% and 40% parts, corresponding to the quantitative results of the two components, respectively. This allocation process is accomplished by establishing a linear relationship model between peak area and correction coefficients, making the quantitative results of each component in the overlapping peaks closer to the true values.

[0093] In some specific implementations, the integral calculation can be limited to the retention time window of the overlapping peaks, which can be the time interval covered by twice the adjusted peak width before and after the peak apex. The peak correction coefficients can be generated using an iterative optimization method, for example, by alternately adjusting the coefficient values ​​to minimize the residual between the deconvolutioned single-peak profile and the original data.

[0094] Traditional methods for processing co-eluent peaks typically use peak height ratio or retention time interval for area allocation, which is prone to errors due to peak asymmetry or baseline drift. This approach introduces a peak correction coefficient to establish a dynamic allocation mechanism, which can adaptively adjust the area allocation ratio, making it particularly suitable for scenarios with overlapping peaks exhibiting significant differences in peak width or concentration.

[0095] Through the above technical solution, this application effectively solves the problem of quantitative distortion caused by chromatographic co-elution, making the calculated area of ​​each component in the overlapping peak more accurately reflect the actual component ratio, significantly improving the quantitative analysis accuracy of trace components in complex petroleum samples, and providing a reliable data basis for subsequent chemical composition analysis.

[0096] Example 2:

[0097] This embodiment provides an automated chemical composition analysis method for petroleum products based on deep learning, applied to the Agilent 7890-5977 GC-MS analysis platform.

[0098] In this embodiment, the analyzed petroleum sample had an API gravity of 32 and a sulfur content of 1.2%. Its chemical composition included complex components such as light alkanes, aromatic hydrocarbons, cycloalkanes, and polycyclic aromatic hydrocarbons. The analytical instrument used was an Agilent 7890 gas chromatograph coupled with a 5977 mass spectrometer detector (GC-MS), equipped with an HP-5MS column (30m × 0.25mm × 0.25μm). The GC program was set as follows: initial temperature 40℃ held for 3 minutes, then increased to 300℃ at a rate of 10℃ / min and held for 10 minutes; electron impact ionization (EI) was used; and the scan range was m / z = 30–500.

[0099] After diluting the crude oil sample 10-fold, 1 mg of the sample was injected, and the system acquired its full-scan raw GC-MS data, obtaining a three-dimensional data matrix containing retention time (t), mass-to-charge ratio (m / z), and ionic strength (I), providing basic data for subsequent data processing. This matrix comprehensively records the chemical characteristic signals of each component in the C5–C40 range.

[0100] The system performs peak detection on the full scan data, automatically locating peak apexes and extracting peak characteristic parameters, including retention time, peak height, peak width, and the retention time interval between adjacent peaks. In the light component region (such as the C7–C9 alkane region), the detected peak width is approximately 0.12–0.15 minutes, and the peak height is approximately... to It can accurately describe the signal intensity and retention characteristics of light fractions in crude oil.

[0101] The peak is identified by calculating the first derivative D(t) of the full scan data and identifying the point where the derivative changes from positive to negative. For example, for C8 alkanes, the derivative sign changes at a retention time of 12.35 minutes, thus determining that point as the peak position. This method can effectively eliminate spurious peaks caused by noise fluctuations.

[0102] For adjacent peaks, the system calculates the relative spacing ratio according to the following formula:

[0103]

[0104] in, To preserve time intervals between peaks, These represent the widths of the two peaks, respectively.

[0105] Taking C8 alkanes and isoalkanes as examples, if This value is used to determine the difficulty of separation.

[0106] The system sets the first distance threshold to 0.5 and the second distance threshold to 1.2. Peaks with R < 0.5 are marked as highly overlapping; peaks with R ≤ 1.2 are marked as moderately overlapping; otherwise, they are marked as separable peaks. In the aromatic hydrocarbon region (e.g., after a retention time of 50 minutes), the peak spacing of polycyclic aromatic hydrocarbons is relatively small, and the system automatically identifies them as highly overlapping peaks.

[0107] Subsequently, the peak widths of overlapping peaks were optimized according to a preset peak width adjustment formula. Taking a pair of highly overlapping polycyclic aromatic hydrocarbons (such as benzo[a]pyrene and fluoranthene) as an example, their initial peak width was approximately 0.15 min, the measured relative spacing ratio was 0.4, and the initial compression factor was set to 0.2. The adjusted peak widths were then... .

[0108] For highly overlapping peaks, the center point of the two peak apexes is selected, and the peak width is adjusted by a factor of 2 on both sides to determine the overlapping region. Then, an exponentially modified Gaussian model is used for fitting, and the peak boundary is optimized by minimizing the residual. For example, the overlapping region of benzo[a]pyrene and fluoranthene is [45.3, 46.1] minutes, and the fitting residual is optimized to 1.2%.

[0109] For moderately overlapping peaks, the adjusted peak width, retention time interval, peak height, and corresponding original data subset are extracted and input into a deep learning separation model. After the model outputs a baseline, the background signal is subtracted from the original data to generate a single-peak profile.

[0110] After peak separation is completed, feature parameters are re-extracted and R-values ​​are calculated. If R < 0.5 and the number of iterations exceeds 5, the system calculates correction coefficients based on peak heights to ensure stable separation results. For example, if peak heights are h1 = 2000 and h2 = 1000, then correction coefficients C1 = 0.67 and C2 = 0.33.

[0111] For highly overlapping peaks, the system extracts a subset of the full-scan mass spectrometry data within the retention time window and uses the MCR-ALS deconvolution algorithm to decompose the peak contribution of each component. For example, the correction coefficient for the characteristic ion of benzo[a]pyrene with m / z=252 is 0.68. For moderately overlapping peaks, the correction coefficient is calculated directly based on the peak height ratio. The initial compression factor is updated sequentially with a fixed step size (0.05), for example, the initial compression factor is increased from 0.2 to 0.3 to further optimize the separation effect.

[0112] Finally, the integral area of ​​the overlapping peaks is proportionally allocated based on the peak correction coefficient. For example, the total area of ​​the overlapping peaks is... After being allocated to 0.6 and 0.4, the quantitative results of the two single peaks were as follows: and .

[0113] Example 3:

[0114] like Figure 3 As shown, the deep learning-based automated chemical composition analysis system for petroleum products includes:

[0115] The data acquisition module is used to acquire raw GC-MS full scan data of petroleum samples;

[0116] The peak detection module is used to perform peak detection on the GC-MS full scan raw data, locate all peak vertices and extract peak feature parameters; the peak feature parameters include the retention time, peak height, peak width and retention time interval of each peak;

[0117] The peak overlap determination module is used to calculate the relative spacing ratio between adjacent peaks based on the peak feature parameters; determine whether the relative spacing ratio is less than a preset first distance threshold; if so, mark the adjacent peak as a highly overlapping peak; otherwise, further determine whether the relative spacing ratio is less than a preset second distance threshold; if so, mark the adjacent peak as a moderately overlapping peak; otherwise, mark it as a separable peak.

[0118] The first separation module is used to adjust the peak width of highly overlapping peaks and moderately overlapping peaks based on the initial feature parameters, the relative spacing ratio, and the preset initial compression factor, and to perform peak separation and re-extract peak feature parameters according to the optimized peak widths.

[0119] The separation verification module is used to calculate the relative spacing ratio after peak separation based on the re-extracted peak feature parameters, determine whether the relative spacing ratio is greater than a preset first distance threshold, and if so, accept the separation result; otherwise, further determine whether the adjacent peaks are highly overlapping peaks. If so, calculate the peak correction coefficient based on the peak feature parameters; otherwise, update the initial compression factor and perform peak separation again. If the number of updates exceeds the preset number of iterations, calculate the peak correction coefficient based on the peak height.

[0120] The second separation module is used to separate overlapping peaks according to the peak correction coefficient.

[0121] The above description is merely an example and illustration of the structure of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the structure of the invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.

[0122] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0123] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for automatic analysis of chemical composition of petroleum products based on deep learning, characterized in that, The method comprises the following steps: obtaining GC-MS full scan raw data of a petroleum sample; performing peak detection on the GC-MS full scan raw data, locating all peak apexes and extracting peak characteristic parameters; the peak characteristic parameters include the retention time, peak height, peak width and adjacent peak retention time interval of each peak; calculating the relative interval ratio between adjacent peaks according to the peak characteristic parameters; the relative interval ratio is the ratio of the retention time interval of adjacent peaks to the sum of the peak widths of the two peaks, which is used to quantify the separation difficulty of adjacent peaks; determining whether the relative interval ratio is less than a preset first distance threshold; if yes, marking the adjacent peak as a highly overlapped peak, otherwise further determining whether the relative interval ratio is less than a preset second distance threshold; if yes, marking the adjacent peak as a moderately overlapped peak, otherwise marking it as a separable peak; adjusting the peak width based on the peak characteristic parameters, the relative interval ratio and a preset initial compression factor, and performing peak separation on the highly overlapped peak and the moderately overlapped peak respectively according to the adjusted peak width; after peak separation, reextracting the peak characteristic parameters and calculating the relative interval ratio, determining whether the relative interval ratio is greater than the preset first distance threshold; if yes, accepting the separation result, otherwise further determining whether the adjacent peak is marked as a highly overlapped peak before separation; if yes, calculating a peak correction coefficient according to the peak characteristic parameters, otherwise updating the initial compression factor and reperforming peak separation; if the number of updates exceeds a preset iteration number, calculating a peak correction coefficient according to the peak height; performing peak separation on the overlapped peak according to the peak correction coefficient.

2. The deep learning-based automatic analysis method of chemical composition of petroleum products according to claim 1, characterized in that: The process of locating all peak apexes comprises: calculating the first derivative of the GC-MS full scan raw data, and taking each point where the first derivative changes from positive to negative as a peak apex.

3. The deep learning-based automatic analysis method of chemical composition of petroleum products according to claim 1, characterized in that: The calculation formula of the adjusted peak width is: ； wherein is the initial compression factor, is the initial peak width, is the adjusted peak width, is the relative spacing ratio between adjacent peaks.

4. The deep learning-based automatic analysis method of chemical composition of petroleum products according to claim 1, characterized in that: The process of peak separation comprises: for a highly overlapped peak: expanding 2 times the adjusted peak width to the left and right of the center of the two peak apexes of the overlapped peak to determine the overlapping region; fitting the left and right peaks in the overlapping region with an exponential modified Gaussian model to optimize the peak boundary with the minimum fitting residual as the target; for a moderately overlapped peak: extracting the adjusted peak width, retention time interval, peak height of the current moderately overlapped peak and the GC-MS full scan raw data subset of the corresponding time window, inputting them into a pre-trained deep learning separation model together with the GC-MS full scan raw data; the deep learning separation model outputs a separation baseline, which represents the background contribution of the overlapped peak region; subtracting the separation baseline from the GC-MS full scan raw data to obtain the single peak profile after separation; reextracting the peak apex and peak boundary based on the single peak profile.

5. The deep learning-based automatic analysis method of chemical composition of petroleum products according to claim 1, characterized in that: The process of calculating the peak correction coefficient according to the peak characteristic parameters comprises: extracting the full scan mass spectrum data subset in the corresponding retention time window of the highly overlapped peak, which contains the time-mass-to-charge ratio-ion intensity three-dimensional information of the co-elution component; identifying the characteristic ion cluster of the co-elution component in the mass spectrum data subset based on the characteristic ion of the target compound; performing a deconvolution algorithm on the full scan mass spectrum data subset and the characteristic ion cluster to obtain the peak correction coefficient of each peak in the highly overlapped peak.

6. The deep learning-based automatic analysis method of chemical composition of petroleum products according to claim 1, characterized in that: The process of updating the initial compression factor comprises: setting a fixed update step, updating the initial compression factor by summing the initial compression factor and the update step.

7. The deep learning-based automatic analysis method of chemical composition of petroleum products according to claim 1, characterized in that: The process of calculating the peak correction coefficient according to the peak height comprises: calculating the peak correction coefficient of each of the two peaks in the moderately overlapped peak by taking the ratio of the peak height of each of the two peaks to the sum of the peak heights of the two peaks.

8. The deep learning-based automatic analysis method of chemical composition of petroleum products according to claim 1, characterized in that: The process of performing peak separation on the overlapped peak according to the peak correction coefficient comprises: calculating the overlapped peak area by integrating the overlapped peak according to the GC-MS full scan raw data, and proportionally distributing the overlapped peak area according to the peak correction coefficient of each of the two peaks in the overlapped peak.

9. An automatic analysis system for chemical composition of petroleum products based on deep learning, characterized by, It comprises: a data acquisition module configured to acquire GC-MS full scan raw data of a petroleum sample; a peak detection module configured to perform peak detection on the GC-MS full scan raw data, locate all peak vertices, and extract peak characteristic parameters; the peak characteristic parameters comprise the retention time, peak height, and peak width of each peak, and the retention time interval of adjacent peaks; a peak overlap judgment module configured to calculate the relative interval ratio between adjacent peaks according to the peak characteristic parameters; the relative interval ratio is the ratio of the retention time interval of adjacent peaks to the sum of the peak widths of the two peaks, and is used to quantify the separation difficulty of adjacent peaks; it is judged whether the relative interval ratio is less than a preset first distance threshold value, and if yes, the adjacent peak is marked as a highly overlapped peak, otherwise it is further judged whether the relative interval ratio is less than a preset second distance threshold value, and if yes, the adjacent peak is marked as a moderately overlapped peak, otherwise it is marked as a separable peak; a first separation module configured to adjust the peak width of the highly overlapped peak and the moderately overlapped peak based on the peak characteristic parameters, the relative interval ratio, and a preset initial compression factor, perform peak separation according to the optimized peak width, and re-extract peak characteristic parameters; a separation verification module configured to calculate the relative interval ratio after peak separation according to the re-extracted peak characteristic parameters, judge whether the relative interval ratio is greater than a preset first distance threshold value, and if yes, accept the separation result, otherwise further judge whether the adjacent peak before separation is a highly overlapped peak, and if yes, calculate a peak correction coefficient according to the peak characteristic parameters, otherwise update the initial compression factor and re-perform peak separation, and if the number of updates exceeds a preset iteration number, calculate a peak correction coefficient according to the peak height; a second separation module configured to perform peak separation on the overlapped peak according to the peak correction coefficient.

Images