Intelligent analysis method and system for liquid chromatogram peak based on deep learning

Abstract

The invention discloses a liquid chromatographic peak intelligent analysis method and system based on deep learning, and relates to the technical field of data processing. The method comprises the following steps: constructing a chromatographic peak intelligent analysis network based on deep learning, and training the network by using a sample chromatogram until convergence; inputting a to-be-analyzed chromatographic signal into the chromatographic peak intelligent analysis network, and evaluating the reliability of a prediction result based on the probability distribution of the key physical parameters output by the chromatographic peak intelligent analysis network to obtain a result confidence coefficient; dynamically adjusting the number and composition of the sample chromatograms based on the result confidence and a preset confidence threshold, and continuously updating the chromatographic peak intelligent analysis network; and recording an attention mechanism of the chromatographic peak intelligent analysis network as a weight fraction allocated to each data point, mapping the weight fractions back to a time sequence of an original chromatographic signal, carrying out visualization through a color gradient, and superposing the weight fractions on an original chromatographic curve in a semitransparent layer form to generate a chromatographic peak thermodynamic diagram. According to the invention, the efficiency and precision of liquid chromatography analysis are effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a liquid chromatography peak intelligent analysis method and system based on deep learning. BACKGROUND

[0002] As a core means of separation and detection, chromatographic analysis is widely used in chemical industry, environment, medicine and other fields. Key physical parameters such as peak position and peak area need to be analyzed from chromatographic signals to realize component quantification and qualification. However, with the increasing complexity of samples, the demand for analysis efficiency and accuracy has increased significantly.

[0003] In liquid chromatography analysis, traditional chromatographic peak analysis methods have the advantages of high calculation efficiency and clear principle when dealing with simple samples with good separation degree. However, when facing the common problems of severe overlapping peaks, baseline drift and noise interference in complex samples, the analysis accuracy is low and the efficiency is limited, which usually relies on manual setting of parameters and thresholds. SUMMARY

[0004] The present application provides a liquid chromatography peak intelligent analysis method and system based on deep learning, aiming to solve the technical problems of low analysis accuracy and limited efficiency in the prior art.

[0005] In view of the above problems, the present application provides a liquid chromatography peak intelligent analysis method and system based on deep learning.

[0006] In a first aspect, the present application provides a liquid chromatography peak intelligent analysis method based on deep learning, comprising: Based on deep learning, a chromatographic peak intelligent analysis network is constructed, and the chromatographic peak intelligent analysis network is trained using sample chromatograms until convergence, wherein the chromatographic peak intelligent analysis network includes a feature extractor, a physical parameter probability prediction output head, and a physical signal generator; The chromatographic signal to be analyzed is input into the chromatographic peak intelligent analysis network, and based on the probability distribution of the key physical parameters of the output of the chromatographic peak intelligent analysis network, the reliability of the prediction result is evaluated to obtain a result confidence; Based on the result confidence and a preset confidence threshold, the number and composition of sample chromatograms are dynamically adjusted, and the chromatographic peak intelligent analysis network is continuously updated; The weight score assigned to each data point by the attention mechanism of the chromatographic peak intelligent analysis network is recorded, the weight score is mapped back to the time series of the original chromatographic signal, and visualized through color gradient, superimposed on the original chromatographic curve in the form of a semi-transparent layer, to generate a chromatographic peak heat map.

[0007] In a second aspect, the present application provides a liquid chromatography peak intelligent analysis system based on deep learning, comprising: The network construction training module is configured to construct a chromatographic peak intelligent analysis network based on deep learning, and train the chromatographic peak intelligent analysis network using sample chromatograms until the network converges, wherein the chromatographic peak intelligent analysis network comprises a feature extractor, a physical parameter probability prediction output head, and a physical signal generator. The result confidence evaluation module is configured to input a chromatographic signal to be analyzed into the chromatographic peak intelligent analysis network, and evaluate the reliability of a predicted result based on a probability distribution of a key physical parameter of an output of the chromatographic peak intelligent analysis network, to obtain a result confidence. The dynamic adaptive updating module is configured to dynamically adjust the number and composition of sample chromatograms based on the result confidence and a preset confidence threshold, and continuously update the chromatographic peak intelligent analysis network. The heat map visualization module is configured to record a weight score assigned to each data point by an attention mechanism of the chromatographic peak intelligent analysis network, map the weight score back to a time sequence of an original chromatographic signal, and visualize the weight score through a color gradient, to generate a chromatographic peak heat map in the form of a semi-transparent layer superimposed on the original chromatographic curve.

[0008] The one or more technical solutions provided in the present application have at least the following technical effects or advantages: The present application provides a deep learning-based liquid chromatographic peak intelligent analysis method and system, which realizes accurate analysis and reliable evaluation of complex liquid chromatographic signals. By constructing an intelligent analysis network comprising a feature extractor, a physical parameter probability prediction head, and a physical signal generator, key physical parameters can be robustly extracted from overlapping peaks and noise interference, and the results can be output in the form of a probability distribution. A dynamic sample optimization mechanism based on prediction confidence can continuously improve network performance and adaptability. The heat map visualization technology of attention weights provides an intuitive decision basis for the analysis process, enhances the explainability and credibility, and effectively improves the accuracy and efficiency of the liquid chromatographic analysis method. BRIEF DESCRIPTION OF DRAWINGS

[0009] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0010] Figure 1 A flowchart of a deep learning-based liquid chromatographic peak intelligent analysis method according to an embodiment of the present application is shown in the figure. Figure 2 A structural diagram of a deep learning-based liquid chromatographic peak intelligent analysis system according to an embodiment of the present application is shown in the figure. Figure 3A chromatogram containing one symmetric peak provided by an embodiment of the present application; Figure 4 A heat map of chromatographic peaks provided by an embodiment of the present application; In the drawings, the components represented by each reference numeral are described as follows: The network construction training module 11, the result confidence evaluation module 12, the dynamic adaptive updating module 13, and the heat map visualization module 14. DETAILED DESCRIPTION

[0011] The present application provides a liquid chromatographic peak intelligent analysis method and system based on deep learning, which is used to solve the technical problems of low resolution and limited efficiency of existing liquid chromatographic analysis.

[0012] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0013] It should be noted that the terms "comprising" and "having" are intended to cover non-exclusive inclusion, for example, a process, method, system, product or server containing a series of steps or units does not have to be limited to those clearly listed steps or units, but can include other steps or modules that are not clearly listed or inherent to these processes, methods, products or devices.

[0014] Embodiment one, as shown in the present application provides a liquid chromatographic peak intelligent analysis method based on deep learning, which comprises: Figure 1 S100: based on deep learning, constructing a chromatographic peak intelligent analysis network, and training the chromatographic peak intelligent analysis network using sample chromatograms until convergence, wherein the chromatographic peak intelligent analysis network comprises a feature extractor, a physical parameter probability prediction output head, and a physical signal generator.

[0015] ​In the embodiments of the present application, based on deep learning, a chromatographic peak intelligent analysis network is constructed, and the chromatographic peak intelligent analysis network is trained using sample chromatograms until convergence, wherein the chromatographic peak intelligent analysis network comprises a feature extractor, a physical parameter probability prediction output head, and a physical signal generator. Complex chromatographic features are automatically extracted by a one-dimensional CNN, avoiding the limitations of manually designed features; the parameter uncertainty is quantified by the probability prediction branch, improving the analysis reliability; the differentiable generator constrains the prediction by a physical model, ensuring that the output conforms to the physical law of the chromatographic peak. The parameter prediction accuracy and signal reconstruction effect are simultaneously optimized by using labeled samples through a composite loss, so that the network learns the correlation between the peak shape features and the physical parameters from the data, and end-to-end optimization is realized through back propagation.

[0016] The step S100 in the method provided by the embodiments of the present application comprises: In the method provided by the embodiments of the present application, based on deep learning, a chromatographic peak intelligent analysis network is constructed, and the chromatographic peak intelligent analysis network at least comprises a feature extractor, a physical parameter probability prediction output head, and a physical signal generator, comprising: The feature extractor is constructed based on a one-dimensional CNN, and the input chromatogram can identify the feature tensor of the chromatographic peak, wherein the feature tensor at least comprises the rising edge, the falling edge, the vertex, and the trend of baseline drift of the chromatographic peak; After the feature extractor, a plurality of independent probability prediction branches are connected in parallel, wherein each probability prediction branch corresponds to a parameter prediction result of a potential chromatographic peak, the parameter prediction result comprises a probability distribution of a key physical parameter, and the probability distribution comprises a predicted mean and a predicted variance; The physical signal generator is constructed based on a differentiable function.

[0017] Firstly, the feature extractor is constructed based on a one-dimensional CNN, and the input chromatogram can identify the feature tensor of the chromatographic peak, wherein the feature tensor at least comprises the rising edge, the falling edge, the vertex, and the trend of baseline drift of the chromatographic peak. The one-dimensional CNN is a convolutional neural network suitable for processing one-dimensional signals such as time series, and features are extracted from local continuous data by sliding a convolution kernel. The feature extractor is constructed by using a one-dimensional convolutional neural network, and the network structure comprises a plurality of convolution layers, batch normalization layers, and pooling layers. The input is the original chromatogram, i.e. a one-dimensional sequence of time-response intensity, such as 1000 time points of response values. Local features such as the response rate of adjacent time points are extracted by the convolution layer, the dimension is reduced and key features are retained by the pooling layer, and finally the feature tensor is output. The feature tensor comprises: the rising edge, i.e. the trend of the response increasing with time, such as the change in the slope; the falling edge, i.e. the trend of the response decreasing with time; the vertex, i.e. the time point and intensity corresponding to the maximum response; and the baseline drift, i.e. the slow change trend of the background response with time. For example, as shown in FIG. 1, the chromatogram is input into the feature extractor, and the feature tensor is output, which comprises the rising edge, the falling edge, the vertex, and the baseline drift. Figure 3As shown, the input chromatogram contains one symmetric peak: time 0-100 s, vertex at 50 s, rising edge 0-50 s response from 100 to 5000, falling edge 50-100 s to 100, baseline drift 50; the feature extractor outputs feature tensor: rising edge: 4900 / 50=98; falling edge: 98; vertex: 50 s, 5000; baseline drift: 0.5.

[0018] Secondly, after the feature extractor, a plurality of independent probability prediction branches are connected in parallel, wherein each probability prediction branch corresponds to a parameter prediction result of a potential chromatographic peak, and the parameter prediction result includes a probability distribution of a key physical parameter, and the probability distribution includes a predicted mean and a predicted variance. The key physical parameters include: peak height, retention time, peak width, and asymmetry factor. At the output end of the feature extractor, N independent probability prediction branches are arranged in parallel, N being a preset maximum number of potential chromatographic peaks, such as 5, and each branch is composed of a fully connected layer and a probability output layer. After the feature tensor is mapped through the fully connected layer, the probability output layer is parameterized by a normal distribution, and outputs the probability distribution of the key physical parameters of each potential chromatographic peak, including the probability distribution of the peak height, the retention time, the peak width, and the asymmetry factor, which describes the statistical distribution of the possible values of the parameters, the mean μ represents the most likely value, and the variance represents the uncertainty. Each branch corresponds to a potential peak, and the peak that does not appear is filtered by the variance tending to 0 or a probability threshold. For example, for a chromatogram containing two overlapping peaks, three probability prediction branches are set. Branch 1 outputs the retention time , peak height , half-peak width , and asymmetry factor ; branch 2 outputs the retention time , peak height , half-peak width , and asymmetry factor ; and branch 3 outputs the maximum variance, such as , indicating that there is no corresponding peak.

[0019] Finally, the physical signal generator is constructed based on a differentiable function. The physical signal generator is constructed by using a differentiable function, and the function form is based on the physical model of the chromatographic peak. Taking a Gaussian peak model as an example, the signal function of a single peak is: where H is the peak height, T is the retention time, W is the half-peak width related parameter, and B(t) is the baseline function, such as a linear function The generator input is the physical parameters of each branch sample: H, T, W, a, b, etc. The predicted chromatogram signal, i.e. the time series response value, is calculated by the function superposition output, and because the function is differentiable, it supports collaborative back propagation with other parts of the network. For example, the input: branch 1 sampling parameters: H1=3000, T1=10.2s, W1=0.5s, baseline parameters: a=0.1, b=50; branch 2 sampling parameters: H2=2500, T2=10.8s, W2=0.4s, baseline parameters: a=0.1, b=50. The generator calculates and superimposes two Gaussian peaks and the baseline, and outputs the predicted response sequence of time 0-20s.

[0020] In the method provided by the embodiments of the present application, a chromatogram labeled with chromatographic peak information is obtained as a sample chromatogram, and a plurality of sample chromatograms are used to train the chromatographic peak intelligent analysis network until convergence, comprising: A chromatogram labeled with chromatographic peak information is obtained as a sample chromatogram. The preprocessed sample chromatogram is input into the feature extractor for feature tensor extraction. The extracted feature tensor is input into the physical parameter probability prediction output head to obtain the probability distribution of the key physical parameters. Based on the probability distribution, sampling is performed, and the sampling value is input into the physical signal generator to generate a predicted chromatographic signal. The compound loss between the predicted chromatographic signal and the true chromatographic signal is calculated, and the network weight is updated through back propagation based on the compound loss until the chromatographic peak intelligent analysis network converges.

[0021] First, a chromatogram labeled with chromatographic peak information is obtained as a sample chromatogram. Chromatograms from actual detection scenarios are collected, such as signals of mixed solvents detected by gas chromatography. By manual annotation or high-precision instrument assisted annotation, the key physical parameters and true chromatographic signals of each chromatographic peak are recorded to form a sample set. The samples need to cover simple peaks, complex peaks and other scenarios to improve the network generalization. For example, the sample set contains 1000 chromatograms, of which 300 are single peaks, 200 are double peaks, 300 are multi-peaks, and 200 are complex peaks with strong noise. Each graph is labeled with parameters such as retention time (10.2s±0.05s) and peak height (3000±50).

[0022] Secondly, the pretreated sample chromatogram is input into the feature extractor to extract the feature tensor. The sample chromatogram is pretreated, such as removing high-frequency noise by wavelet threshold denoising and correcting baseline drift by polynomial fitting. The pretreated signal is input into the feature extractor, and the feature tensor is output through convolution, pooling and other operations. For example, for a sample chromatogram containing noise, such as random noise superimposed in the original signal with a response fluctuation of ±200, after wavelet threshold denoising, the noise fluctuation is reduced to ±50; for a sample with baseline drift, such as a sample with a baseline rising from 100 to 300 within 0-100s, the baseline is fitted by a 3rd order polynomial and subtracted to obtain the corrected signal, which is input into the feature extractor to output a tensor containing the peak shape characteristics after denoising.

[0023] Further, the extracted feature tensor is input into the physical parameter probability prediction output head to obtain the probability distribution of the key physical parameters. The feature tensor output by the feature extractor is simultaneously input into each probability prediction branch, and after mapping through a fully connected layer, each branch outputs the probability distribution of the physical parameters of the corresponding potential peak. For example, for the retention time parameter, the branch outputs , indicating that the peak retention time is most likely 10.2s with less uncertainty.

[0024] Then, sampling is performed based on the probability distribution, and the sampled value is input into the physical signal generator to generate a predicted chromatogram signal. Random sampling is performed from the probability distribution of each branch to obtain specific values of the physical parameters, such as sampling 10.18s from the retention time distribution N(10.2, 0.01). The sampled parameters are input into the physical signal generator, and each peak signal and baseline are calculated and superimposed through a differentiable function to generate a predicted chromatogram signal, i.e., a time series with the same length as the sample chromatogram. For example, sampling 2980 from the peak height distribution N(3000, 100) of branch 1 and 10.19s from the retention time distribution N(10.2, 0.01) of branch 1 are input into the generator to generate a predicted signal sequence containing two overlapping peaks together with the sampled parameters of branch 2.

[0025] Finally, a composite loss between the predicted chromatogram signal and the true chromatogram signal is calculated, and the network weights are updated based on the composite loss by back propagation until the chromatographic peak intelligent analysis network converges. The composite loss of the predicted chromatogram signal and the true chromatogram signal is calculated, such as a reconstruction loss and a probability loss. The reconstruction loss, such as mean square error (MSE), measures the degree of signal fitting; the probability loss, such as KL divergence, measures the deviation of the predicted distribution from the labeled parameters. Back propagation refers to updating the network weights in the direction of loss reduction based on the gradient descent algorithm. The composite loss is optimized by back propagation, and the iteration is repeated until the loss value tends to be stable, that is, the network converges. For example, the MSE of the true signal and the predicted signal of a certain sample is 5000, the KL divergence is 0.1, and the composite loss is set to 0.8*MSE+0.2*KL=4000.02; by back propagation, the convolution kernel weights of the feature extractor, the full connection layer weights of the prediction branch, etc. are adjusted, so that the loss of the next iteration is reduced to below 4000.02, and finally after 800 iterations, the loss is stable at about 100, and the chromatographic peak intelligent analysis network converges.

[0026] In the embodiments of the present application, by constructing the chromatographic peak intelligent analysis network, simple and complex chromatographic peaks can be automatically analyzed, and the output probability distribution can quantitatively analyze the uncertainty; by the physical generator constraint, the physical rationality of parameter prediction is improved; after sufficient training, the analysis accuracy and efficiency are significantly improved, and the ability to generalize to various sample scenarios is possessed.

[0027] S200: inputting the chromatogram signal to be analyzed into the chromatographic peak intelligent analysis network, and evaluating the reliability of the prediction result based on the probability distribution of the key physical parameters output by the chromatographic peak intelligent analysis network to obtain a result confidence.

[0028] In the embodiments of the present application, the chromatogram signal to be analyzed is input into the chromatographic peak intelligent analysis network, and the reliability of the prediction result is evaluated based on the probability distribution of the key physical parameters output by the chromatographic peak intelligent analysis network to obtain a result confidence. The quantitative evaluation of the reliability of the analysis result is realized, and the result confidence degree is intuitively reflected through the confidence value; the peak area confidence interval provides an error range reference for quantitative analysis; the parameter uncertainty and signal matching degree are dynamically combined to avoid the one-sidedness of a single index, and finally the credibility and application value of the chromatographic peak analysis result are improved.

[0029] The step S200 in the method provided in the embodiments of the present application comprises: extracting the prediction variance of each physical parameter output by the chromatographic peak intelligent analysis network, and calculating the normalized weighted average value of the prediction variance of each physical parameter; based on the error propagation principle, deriving the confidence interval of the peak area from the normalized weighted average value of the prediction variance of each physical parameter to obtain an interval width; The signal matching degree is obtained by comparing the degree of matching between the predicted chromatographic signal reconstructed by the intelligent peak analysis network and the chromatographic signal to be analyzed. The confidence level of the obtained result is evaluated based on the interval width and the signal matching degree.

[0030] First, the predicted variances of each physical parameter output by the intelligent peak analysis network are extracted, and the normalized weighted average of the predicted variances of each physical parameter is calculated. Each variance is then divided by the maximum value of all parameter variances to obtain a normalized value between 0 and 1. This is to eliminate differences in the units of different parameters, such as the unit of retention time variance. The peak height variance is expressed in response values. 2 Direct comparison is not possible. Weights are assigned to each parameter based on their importance to peak resolution. For example, peak width and retention time have a greater impact on peak identification, so their weight is set to 0.3. Peak height and asymmetry factor are weighted at 0.2. The weighted sum of the normalized variances is then calculated, resulting in the normalized weighted average. For instance, the predicted variance of a certain chromatographic peak to be resolved is: peak height variance... Response value 2 Retention time variance Peak width variance Asymmetric factor variance The maximum variance is 2500, after normalization. Weighted calculation: The normalized weighted average is mainly contributed by the normalized variance of the peak height.

[0031] Secondly, based on the error propagation principle, the confidence interval of the peak area is derived from the normalized weighted average of the prediction variances of each physical parameter, thus obtaining the interval width.

[0032] The method provided in this application embodiment, based on the error propagation principle, derives the confidence interval of the peak area from the normalized weighted average of the prediction variances of each physical parameter, and obtains the interval width, including: Obtain the probability distribution of the prediction of each physical parameter by the intelligent analysis network of chromatographic peaks, and use the prediction variance of the probability distribution as a quantitative indicator of the prediction uncertainty of that parameter. Based on the prediction variance of the probability distribution, the optimal parameter estimate of the physical parameter is obtained, and the sensitivity of the peak area to the change of each physical parameter near the optimal parameter estimate is calculated as the sensitivity coefficient of the physical parameter. Multiply the prediction variance of each physical parameter by the square of its corresponding sensitivity coefficient to obtain multiple uncertainty contribution values; The summation of multiple uncertainty contributions and the square root are used as the standard uncertainty of the peak area. Configure a multiplier factor, and determine the interval width based on the multiplier factor and the standard uncertainty.

[0033] Firstly, the probability distribution predicted by the chromatographic peak intelligent analysis network for each physical parameter is obtained, and the prediction variance of the probability distribution is taken as the quantitative indicator of the parameter prediction uncertainty. The prediction variance of each parameter is extracted from the probability distribution output by the network, which is directly taken as the quantitative indicator of the parameter prediction uncertainty. The greater the variance, the higher the possibility that the true value of the parameter deviates from the predicted value. In the above example, the variances of the parameters are: peak height , retention time , peak width , and asymmetry factor .

[0034] Secondly, based on the prediction variance of the probability distribution, the optimal parameter estimate value of the physical parameter is obtained, and the sensitivity coefficient of the physical parameter is calculated by calculating the sensitivity of the peak area to the change of each physical parameter near the optimal parameter estimate value. The peak area is the core indicator for quantification in chromatographic analysis and is a function of the key physical parameters, denoted as , where H is the peak height, T is the retention time, W is the peak width, and S is the asymmetry factor. The sensitivity coefficient is defined as the partial derivative of the peak area with respect to each parameter, i.e. , which is used to measure the degree of influence of a small change in the parameter on the peak area. The greater the absolute value of the sensitivity coefficient, the more significant the influence of the parameter change on the peak area. For example, assuming that the peak area function is , k is a constant related to the asymmetry factor S, which is calculated as S = 1.2, and the optimal parameter estimate value is . The sensitivity coefficient of the peak height is: ; the sensitivity coefficient of the peak width is: ; the sensitivity coefficient of the retention time is: ; and the sensitivity coefficient of the asymmetry factor is: .

[0035] Further, the prediction variance of each physical parameter is multiplied by the square of its corresponding sensitivity coefficient to obtain multiple uncertainty contribution values. The contribution value of each parameter to the uncertainty of the peak area is the product of the prediction variance of the parameter and the square of its sensitivity coefficient, i.e. , which is used to quantify the influence of the uncertainty of a single parameter on the total uncertainty of the peak area. For example, the contribution value of the peak height is: ; the contribution value of the peak width is: ; the contribution value of the asymmetry factor is: ; and the contribution value of the retention time is: .

[0036] Then, the sum of the multiple uncertainty contribution values is taken and the square root is taken as the standard uncertainty of the peak area. The standard uncertainty is the square root of the sum of the uncertainty contribution values of the parameters, i.e. , for quantifying the overall dispersion of the peak area prediction values when they are subject to a normal distribution, is the standard deviation. For example, .

[0037] Finally, a multiple factor is configured, and the interval width is determined based on the multiple factor and the standard uncertainty.

[0038] In the method provided by the embodiments of the present application, a multiple factor is configured, and the interval width is determined based on the multiple factor and the standard uncertainty, including: According to the risk situation of the application scenario, a confidence level is pre-set, wherein the confidence level is proportional to the risk level; Based on the confidence level, a corresponding Z value is queried from a standard normal distribution table as the multiple factor; Based on the multiple factor, the optimal parameter estimate value, and the standard uncertainty, a confidence interval of the peak area is obtained; Based on the confidence interval of the peak area, the interval width is obtained.

[0039] First, according to the risk situation of the application scenario, a confidence level is pre-set, wherein the confidence level is proportional to the risk level. The confidence level refers to the probability that the true value of the peak area falls within the confidence interval, and is set according to the risk level of the application scenario. The higher the risk, the higher the confidence level required. For example, the detection of environmental pollutants needs to strictly control the risk of misjudgment, and the confidence level is pre-set to 95%.

[0040] Secondly, based on the confidence level, a corresponding Z value is queried from a standard normal distribution table as the multiple factor. The Z value is the quantile of the standard normal distribution corresponding to the confidence level, such as 95% confidence level corresponding to Z=1.96, indicating that 95% of the probability density of the normal distribution is concentrated in the interval [-1.96, 1.96], which is used as a multiple factor to expand the standard uncertainty. For example, the Z value corresponding to the 95% confidence level is 1.96, so the multiple factor .

[0041] Then, based on the multiple factor, the optimal parameter estimate value, and the standard uncertainty, a confidence interval of the peak area is obtained. The confidence interval is: optimal parameter estimate value ± (multiple factor × standard uncertainty). The optimal estimate value A of the peak area is calculated by the optimal parameter, and the confidence interval is , which indicates that the true peak area has a specified probability of falling within this interval. For example, the optimal estimate value A of the peak area is , and the confidence interval is [5000-1.96×255.15, 5000+1.96×255.15]≈[5000-500, 5000+500]=[4500, 5500].

[0042] Finally, based on the confidence interval of the peak area, the interval width is obtained. The interval width is the difference between the upper and lower limits of the confidence interval. The smaller the width, the lower the uncertainty of the peak area prediction. For example, the interval width = 5500-4500 = 1000.

[0043] Further, the matching degree of the predicted chromatographic signal of the chromatographic peak intelligent analysis network reconstruction and the chromatographic signal to be analyzed is compared, and the signal matching degree is obtained. The signal matching degree is used to measure the overall similarity between the predicted chromatographic signal of the network reconstruction and the chromatographic signal to be analyzed, and is obtained by calculating the similarity index of the two, with a value range of 0-1. The closer to 1, the higher the matching degree. For example, the normalized mean square error of the to-be-analyzed signal and the predicted signal is 0.05, and the signal matching degree = 1-0.05 = 0.95, and the matching degree is high.

[0044] Finally, based on the interval width and the signal matching degree, the result confidence is evaluated. The result confidence is a comprehensive index of the interval width and the signal matching degree, and is calculated by a weighted formula, confidence = α × signal matching degree + (1-α) × (1-interval width / maximum possible width), α is the weight, such as 0.6, with a value range of 0-1, and the closer to 1, the higher the reliability. For example, the interval width is 1000, and the maximum possible width is 2000, so the interval width score = 1-1000 / 2000 = 0.5; the signal matching degree is 0.95, α = 0.6, and the confidence = 0.6 × 0.95 + 0.4 × 0.5 = 0.57 + 0.2 = 0.77, and the reliability is medium to high.

[0045] In the embodiments of the present application, the normalized weighted average variance is calculated to comprehensively analyze the uncertainty of each parameter, to eliminate the dimensional difference, and to provide a basis for subsequent peak area uncertainty analysis; based on the error propagation principle, the peak area confidence interval is derived: the uncertainty of the key parameters is quantitatively transmitted to the core quantitative index, and the reliability is quantified by the confidence interval; the network reconstruction effect is verified from the overall signal level, and the limitations of the parameter level analysis are supplemented; the reliability of the analysis result is comprehensively quantified by combining the uncertainty and the signal consistency.

[0046] S300: Based on the result confidence and the preset confidence threshold, the number and composition of the sample chromatogram are dynamically adjusted, and the chromatographic peak intelligent analysis network is continuously updated.

[0047] In the embodiments of the present application, based on the result confidence and the preset confidence threshold, the number and composition of sample chromatograms are dynamically adjusted, and the chromatographic peak intelligent analysis network is continuously updated. The confidence threshold is used to screen weak chromatograms, accurately locate the short board of model analysis capability, extract common features to clarify the knowledge blind area of the model, generate targeted synthetic samples to supplement training data, solve the problem of scarcity of weak scene data in real samples, control the training endpoint based on the average confidence of the validation set, and ensure that the generalization ability of the chromatographic peak intelligent analysis network is significantly improved after updating.

[0048] The method provided in the embodiments of the present application comprises step S300: a preset confidence threshold is set; chromatograms that do not meet the confidence threshold are identified as weak chromatograms; common features of weak chromatograms are identified, including specific types of severely overlapping peaks and extreme baseline drifts; based on the common features of weak chromatograms, synthetic chromatograms with similar features are generated as targeted samples using a physical model; the newly generated targeted samples are added to the sample chromatograms, and the chromatographic peak intelligent analysis network is retrained until the average result confidence of the model on the validation set exceeds the preset high confidence threshold, wherein the validation set is labeled with accurate chromatographic peak information.

[0049] First, a confidence threshold is set. The confidence threshold is a critical value for judging whether the chromatographic peak analysis result meets the reliability requirement. The threshold is preset according to the accuracy requirement of the application scenario, such as a higher threshold in a medical detection scenario where the result reliability requirement is high, and a slightly lower threshold in a general chemical analysis scenario. The threshold value range is consistent with the result confidence, which is 0-1, and needs to be determined in combination with historical analysis errors and business requirements. For example, in the pharmaceutical industry, in order to ensure the quantitative accuracy of ingredients, the confidence threshold is set to 0.8, that is, the result confidence ≥ 0.8 is considered reliable; in the general environmental monitoring scenario, the confidence threshold is set to 0.7.

[0050] In addition, chromatograms that do not meet the confidence threshold are identified as weak chromatograms. The result confidence of all chromatograms to be analyzed in step S200 is traversed, and chromatograms with a confidence value lower than the preset threshold are screened out and defined as weak chromatograms. Weak chromatograms indicate that the analysis result is not reliable, and the model's learning of its features is not sufficient. For example, among 500 chromatograms to be analyzed, the result confidence of 80 chromatograms is lower than the preset confidence threshold of 0.8, so these 80 chromatograms are marked as weak chromatograms.

[0051] Secondly, identify common features of weak chromatograms, including specific types of severely overlapped peaks and extreme baseline drift. Feature analysis is performed on weak chromatograms to extract common signal features that contribute to the difficulty of resolution. Common features include: specific types of severely overlapped peaks: such as 3 or more peaks with a retention time interval < 0.3s, an overlap degree > 80%, and a peak height difference < 20%; extreme baseline drift: such as a baseline shift > 50% of the main peak height within 100s, or exhibiting non-linear drift; strong noise interference: such as a signal-to-noise ratio < 5:1, with periodic fluctuations in noise. Dominant common features are determined through statistical analysis, such as calculating the frequency of feature occurrence. For example, analysis of 80 weak chromatograms found that 65 of them had severe overlapping features: 3 peaks with a retention time interval of 0.2-0.3s and an overlap degree of 85%-90%; and 15 of them had extreme baseline drift features: non-linear baseline drift within 100s, from 100 to 2000 response values.

[0052] Further, based on the common features of weak chromatograms, synthetic chromatograms with similar features are generated using physical models as targeted samples. Based on the identified common features, physical models of chromatographic peaks such as Gaussian peak model, Lorentz peak model or hybrid model can be used to precisely control peak position, peak shape, baseline and other parameters to generate synthetic chromatograms with similar features. When generating, the weak features need to be accurately reproduced, such as controlling the retention time difference of overlapping peaks, the amplitude and trend of baseline drift, while accurately labeling the physical parameters as labels to form targeted samples for supplementing the lack of weak scene data in the training set. For example, for the 3 severe overlapping peak feature, 50 synthetic chromatograms are generated using the Gaussian peak model: the retention times are set to 10.0s, 10.2s, and 10.4s, with an interval of 0.2s, the peak heights are all 5000±500 response values, the difference is < 20%, the half-peak width is 0.3s, and the overlap degree is 88%; for the extreme non-linear baseline drift feature, 30 synthetic chromatograms are generated: the baseline is set to an exponential drift, , t is time, superimposed with a main peak with a retention time of 50s and a peak height of 3000, and all peak physical parameters are labeled.

[0053] Finally, the newly generated targeted samples are added to the sample chromatogram, and the chromatographic peak intelligent analysis network is retrained until the average result confidence of the model on the validation set exceeds the preset high confidence threshold, wherein the validation set is labeled with accurate information of chromatographic peaks. The newly generated targeted samples are added to the original sample chromatogram set to expand the training set and ensure that the proportion of new samples is reasonable, such as 10%-20%, to avoid data distribution imbalance. The validation set is a chromatogram set independent of the training set, used to objectively evaluate the generalization ability of the model, covering multiple scenes such as regular peaks and complex peaks, and each sample is accurately labeled with accurate information of chromatographic peaks: peak start and end points, the start time of the chromatographic peak rising from the baseline and the end time of the chromatographic peak descending to the baseline; vertex, the time point and response intensity value corresponding to the maximum response intensity of the chromatographic peak; peak area, the area enclosed by the chromatographic peak and the baseline. The expanded training set is used to retrain the chromatographic peak intelligent analysis network, and the original training process is followed, such as calculating the compound loss and updating the weights through back propagation. After each round of training, the average result confidence of the model is evaluated using the validation set, and when the value exceeds the preset high confidence threshold, the training is stopped and the network update is completed. For example, the preset high confidence threshold is 0.9; the original training set contains 1000 samples, and after adding 80 targeted samples, there are a total of 1080 samples; the validation set contains 200 samples, and the initial average confidence is 0.75. After retraining for 50 rounds, when the average confidence of the validation set rises to 0.92, which exceeds the high confidence threshold 0.9, the training is stopped, and the chromatographic peak intelligent analysis network is updated.

[0054] In the embodiments of the present application, the weak scene samples are supplemented in a targeted manner, so that the chromatographic peak intelligent analysis network can strengthen the learning of complex features in a targeted manner, and avoid the low efficiency caused by blind data expansion; the supervision of the validation set ensures that the updated model not only performs well on the training data, but also can adapt to complex signals in actual scenes stably, and finally continuously improves the reliability and generalization ability of chromatographic peak analysis, realizing dynamic iterative optimization of the chromatographic peak intelligent analysis network.

[0055] S400: The weight score allocated to each data point by the attention mechanism of the chromatographic peak intelligent analysis network is recorded, the weight score is mapped back to the time sequence of the original chromatographic signal, and is visualized through a color gradient, and is superimposed on the original chromatographic curve in the form of a semi-transparent layer to generate a chromatographic peak heat map.

[0056] In the embodiments of the present application, first, the weight score assigned to each data point by the attention mechanism of the recording chromatographic peak intelligent analysis network is utilized. The attention mechanism is a deep learning module that simulates human attention allocation, assigns a weight score to each input data point by calculating the contribution of the data point to the model decision, and the higher the weight, the greater the impact of the point on peak identification and parameter prediction. During network inference, the weight score output by the attention mechanism is recorded in real time. For a one-dimensional chromatographic signal [time series, containing N data points, each point corresponding to the response intensity at a specific time] input, the attention mechanism will generate a weight score sequence of length N, with a value range of 0-1, or after normalization processing, it is mapped to this range. For example, input a chromatographic signal containing 1 single peak [time 0-100s, a total of 1000 data points, peak at 40-60s], the attention mechanism assigns a high weight of 0.7-0.9 to the data points at 40-60s, and a low weight of 0.1-0.3 to the baseline regions at 0-30s and 70-100s. The recorded weight score sequence corresponds to the 1000 time points one by one.

[0057] Secondly, the weight score is mapped back to the time sequence of the original chromatographic signal. The original chromatographic signal has time as the horizontal axis, and each time point corresponds to a unique index. Although the weight score sequence output by the attention mechanism may have undergone dimension adjustment within the network, through the preset index mapping relationship, each weight score can be accurately associated with the time point of the original signal, ensuring that the weight score and the original time sequence are completely aligned in the time dimension. For example, during internal processing, 1000 original data points are temporarily reduced to 500 feature points by a step of 2 pooling, and the attention mechanism assigns weights to these 500 feature points. Through the mapping rule: original point index = feature point index x 2, the weight score of each feature point is copied to the corresponding 2 original data points, such as the weight of feature point 1 is assigned to original points 1 and 2, and finally a weight sequence completely aligned with the original 1000 time points is obtained.

[0058] Finally, as Figure 4As shown, the chromatographic peak heat map is generated by visualizing the weight scores as a semi-transparent layer superimposed on the original chromatogram curve. The weight scores are visualized using a color gradient scheme, such as from cool to warm tones: low weight, such as 0-0.3, corresponds to blue; medium weight, 0.3-0.7, corresponds to yellow; and high weight, 0.7-1.0, corresponds to red. The color depth increases with the weight, forming a continuous gradient. The weight scores need to be normalized to 0-1 to ensure matching with the color gradient range. Above the time axis of the original chromatogram curve, a color heat layer of the weight scores is plotted with the same horizontal axis range. The height of the heat layer can be set to 10%-20% of the vertical axis range of the original curve to avoid obscuring the curve, and the transparency can be set to 50%-70% to ensure that the heat color distribution can be clearly seen without hiding the original peak shape. Integrating the original curve and the heat layer, the final chromatographic peak heat map is formed, which intuitively displays the key time points of interest to the model. For example, for a certain overlapping peak chromatographic signal [time 10-20 s, containing two overlapping peaks, located at 12-14 s and 13-15 s], the attention mechanism allocates a high weight of 0.8-0.9 to the overlapping region of 12-15 s, which is displayed as red, and the deepest red color is at 13 s, which is the core of the two peaks. The baseline or non-overlapping region of 10-12 s and 15-20 s is assigned a low weight of 0.2-0.4, which is displayed as light blue. In the heat map, the red band is concentrated in 12-15 s, which completely coincides with the overlapping peak region of the original curve, and the semi-transparent layer clearly covers the curve below, showing the peak shape and highlighting the model's focus.

[0059] In the embodiments of the present application, the interpretability of the chromatographic peak analysis process is achieved, and the heat map clearly displays the peak regions of interest to the model, such as peak values, overlapping boundaries, and ignored baseline or noise regions, verifying the rationality of the analysis logic. At the same time, if the heat map shows that the weight is abnormally concentrated in the noise points or baseline drift region, it can prompt that there is a learning bias in the chromatographic peak intelligent analysis network, providing an intuitive clue for optimizing the chromatographic peak intelligent analysis network, and ultimately enhancing the user's trust in the analysis results and the practicality of the chromatographic peak intelligent analysis network.

[0060] Through the specific implementation method described above, the following technical effects are achieved: the deep learning network is used to accurately extract the key physical parameters of complex chromatographic peaks, and the probability distribution is used to quantify the reliability of the analysis results, providing an objective basis for subsequent decision-making. With the help of confidence feedback, targeted samples are dynamically supplemented to continuously optimize the model's adaptation to complex scenarios such as severely overlapping peaks and extreme baseline drift, significantly improving the generalization performance. At the same time, the heat map visualization intuitively presents the model's decision dependence on key regions such as peak shoulders and inflection points, breaking the black box nature of the analysis process, ensuring the accuracy and efficiency of chromatographic analysis, and enhancing the verifiability of the analysis logic, providing efficient and reliable technical support for qualitative and quantitative analysis of chromatographic detection.

[0061] Example 2, as Figure 2 As shown, this application provides a deep learning-based intelligent peak analysis system for liquid chromatography, the system comprising: The network construction and training module 11 is used to construct a chromatographic peak intelligent analysis network based on deep learning, and to train the chromatographic peak intelligent analysis network using sample chromatograms until convergence. The chromatographic peak intelligent analysis network includes a feature extractor, a physical parameter probability prediction output head, and a physical signal generator. The result confidence assessment module 12 is used to input the chromatographic signal to be analyzed into the intelligent chromatographic peak analysis network, and to assess the reliability of the prediction result based on the probability distribution of the key physical parameters output by the intelligent chromatographic peak analysis network, and obtain the result confidence. The dynamic adaptive update module 13 is used to dynamically adjust the number and composition of sample chromatograms based on the confidence level of the results and a preset confidence threshold, and to continuously update the intelligent analysis network of chromatographic peaks. The heatmap visualization module 14 is used to record the weight score assigned to each data point by the attention mechanism of the chromatographic peak intelligent analysis network, map the weight score back to the time series of the original chromatographic signal, and visualize it through color gradient. It is superimposed on the original chromatographic curve in the form of a semi-transparent layer to generate a chromatographic peak heatmap.

[0062] In one embodiment, the network construction and training module 11 is further configured to: Among them, a chromatographic peak intelligent analysis network is constructed based on deep learning. This network includes at least a feature extractor, a physical parameter probability prediction output head, and a physical signal generator, comprising: The feature extractor is constructed based on a one-dimensional CNN. When a chromatogram is input, the feature tensor of the chromatographic peak can be identified and output. The feature tensor includes at least the rising edge, falling edge, apex, and baseline drift trend of the chromatographic peak. Following the feature extractor, multiple independent probability prediction branches are connected in parallel, wherein each probability prediction branch corresponds to the parameter prediction result of a potential chromatographic peak, the parameter prediction result including the probability distribution of key physical parameters, the probability distribution including the prediction mean and prediction variance; The physical signal generator is constructed based on a differentiable function.

[0063] The process includes obtaining chromatograms labeled with chromatographic peak information as sample chromatograms, and training the intelligent peak analysis network using multiple sample chromatograms until convergence, including: Obtain the chromatogram labeled with chromatographic peak information as the sample chromatogram; The preprocessed sample chromatogram is input into the feature extractor for feature tensor extraction. inputting the extracted feature tensor into the physical parameter probability prediction output head to obtain a probability distribution of the key physical parameters; sampling based on the probability distribution and inputting a sampling value into the physical signal generator to generate a predicted chromatographic signal; calculating a compound loss between the predicted chromatographic signal and the true chromatographic signal, and updating network weights by back propagation based on the compound loss until the chromatographic peak intelligent resolution network converges.

[0064] In one embodiment, the result confidence evaluation module 12 is further configured to: extracting a prediction variance of each physical parameter output by the chromatographic peak intelligent resolution network, and calculating a normalized weighted average value of the prediction variance of each physical parameter; deriving a confidence interval of the peak area from the normalized weighted average value of the prediction variance of each physical parameter based on the error propagation principle to obtain an interval width; comparing the matching degree of the predicted chromatographic signal reconstructed by the chromatographic peak intelligent resolution network and the chromatographic signal to be resolved to obtain a signal matching degree; based on the interval width and the signal matching degree, evaluating the result confidence.

[0065] wherein the confidence interval of the peak area is derived from the normalized weighted average value of the prediction variance of each physical parameter based on the error propagation principle to obtain the interval width, comprising: obtaining the probability distribution predicted by the chromatographic peak intelligent resolution network for each physical parameter, and taking the prediction variance of the probability distribution as a quantitative indicator of the prediction uncertainty of the parameter; based on the prediction variance of the probability distribution, obtaining an optimal parameter estimate value of the physical parameter, and calculating the sensitivity coefficient of the physical parameter as the sensitivity of the peak area to the change of each physical parameter near the optimal parameter estimate value; multiplying the prediction variance of each physical parameter by the square value of its corresponding sensitivity coefficient to obtain a plurality of uncertainty contribution values; summing the plurality of uncertainty contribution values and taking the square root as the standard uncertainty of the peak area; configuring a multiplication factor, and determining the interval width based on the multiplication factor and the standard uncertainty.

[0066] wherein the interval width is determined based on the multiplication factor and the standard uncertainty by configuring the multiplication factor, comprising: according to the risk situation of the application scenario, pre-setting a confidence level, wherein the confidence level is proportional to the risk level; based on the confidence level, querying a corresponding Z value from a standard normal distribution table as the multiplication factor; Based on the fold factor, the optimal parameter estimate value, and the standard uncertainty, a confidence interval of the peak area is obtained. Based on the confidence interval of the peak area, the interval width is obtained.

[0067] In one embodiment, the dynamic adaptive updating module 13 is further configured to: pre-set a confidence threshold; identify chromatograms that do not meet the confidence threshold as weak chromatograms; identify common features of weak chromatograms, the common features including specific types of severe overlapping peaks and extreme baseline drifts; based on the common features of weak chromatograms, generate synthetic chromatograms with similar features as targeted samples using a physical model; add the newly generated targeted samples to the sample chromatograms, retrain the chromatographic peak intelligent analysis network until the average result confidence of the model on the validation set exceeds a pre-set high confidence threshold, wherein the validation set is labeled with chromatographic peak accurate information.

[0068] It should be noted that the above-mentioned sequence of the embodiments of the present application is only for description, and does not represent the advantages and disadvantages of the embodiments. Moreover, the above-mentioned embodiments of the present application are described. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multi-task processing and parallel processing are also possible or can be advantageous.

[0069] The above only describes the preferred embodiments of the present application, and does not limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

[0070] The present specification and drawings are merely exemplary of the present application, and any and all modifications, variations, combinations or equivalents that fall within the scope of the present application should be considered. Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the present application and its equivalents, the present application is intended to include these modifications and variations.

Claims

1. A deep learning-based intelligent peak analysis method for liquid chromatography, characterized in that, The method includes: Based on deep learning, a chromatographic peak intelligent analysis network is constructed, and the chromatographic peak intelligent analysis network is trained using sample chromatograms until convergence. The chromatographic peak intelligent analysis network includes a feature extractor, a physical parameter probability prediction output head, and a physical signal generator. The chromatographic signal to be analyzed is input into the intelligent chromatographic peak analysis network, and the reliability of the prediction results is evaluated based on the probability distribution of the key physical parameters output by the intelligent chromatographic peak analysis network, so as to obtain the confidence level of the results. Based on the confidence level of the results and the preset confidence threshold, the number and composition of sample chromatograms are dynamically adjusted, and the intelligent peak analysis network is continuously updated. The attention mechanism of the chromatographic peak intelligent analysis network records the weight score assigned to each data point. The weight score is then mapped back to the time series of the original chromatographic signal and visualized through color gradient. It is then superimposed on the original chromatographic curve in the form of a semi-transparent layer to generate a chromatographic peak heatmap.

2. The intelligent peak analysis method for liquid chromatography based on deep learning according to claim 1, characterized in that, Based on deep learning, a chromatographic peak intelligent analysis network is constructed. This network includes at least a feature extractor, a physical parameter probability prediction output head, and a physical signal generator, comprising: The feature extractor is constructed based on a one-dimensional CNN. When a chromatogram is input, the feature tensor of the chromatographic peak can be identified and output. The feature tensor includes at least the rising edge, falling edge, apex, and baseline drift trend of the chromatographic peak. Following the feature extractor, multiple independent probability prediction branches are connected in parallel, wherein each probability prediction branch corresponds to the parameter prediction result of a potential chromatographic peak, the parameter prediction result including the probability distribution of key physical parameters, the probability distribution including the prediction mean and prediction variance; The physical signal generator is constructed based on a differentiable function.

3. The intelligent peak analysis method for liquid chromatography based on deep learning according to claim 1, characterized in that, Obtain chromatograms labeled with chromatographic peak information as sample chromatograms, and train the intelligent peak analysis network using multiple sample chromatograms until convergence, including: Obtain the chromatogram labeled with chromatographic peak information as the sample chromatogram; The preprocessed sample chromatogram is input into the feature extractor for feature tensor extraction. The extracted feature tensor is input into the physical parameter probability prediction output head to obtain the probability distribution of key physical parameters; Based on this probability distribution, sampling is performed, and the sampled values ​​are input into the physical signal generator to generate a predictive chromatographic signal; The composite loss between the predicted chromatographic signal and the actual chromatographic signal is calculated, and the network weights are updated through backpropagation based on the composite loss until the intelligent chromatographic peak analysis network converges.

4. The intelligent peak analysis method for liquid chromatography based on deep learning according to claim 1, characterized in that, The chromatographic signal to be analyzed is input into the intelligent peak analysis network, and the reliability of the prediction results is evaluated based on the probability distribution of key physical parameters output by the network, thus obtaining the confidence level of the results, including: Extract the prediction variance of each physical parameter output by the intelligent analysis network for chromatographic peaks, and calculate the normalized weighted average of the prediction variance of each physical parameter. Based on the error propagation principle, the confidence interval of the peak area is derived from the normalized weighted average of the prediction variances of each physical parameter, and the interval width is obtained. The signal matching degree is obtained by comparing the degree of matching between the predicted chromatographic signal reconstructed by the intelligent peak analysis network and the chromatographic signal to be analyzed. The confidence level of the obtained result is evaluated based on the interval width and the signal matching degree.

5. The intelligent peak analysis method for liquid chromatography based on deep learning according to claim 4, characterized in that, Based on the error propagation principle, the confidence interval for the peak area is derived from the normalized weighted average of the prediction variances of each physical parameter, thus obtaining the interval width, including: Obtain the probability distribution of the prediction of each physical parameter by the intelligent analysis network of chromatographic peaks, and use the prediction variance of the probability distribution as a quantitative indicator of the prediction uncertainty of that parameter. Based on the prediction variance of the probability distribution, the optimal parameter estimate of the physical parameter is obtained, and the sensitivity of the peak area to the change of each physical parameter near the optimal parameter estimate is calculated as the sensitivity coefficient of the physical parameter. Multiply the prediction variance of each physical parameter by the square of its corresponding sensitivity coefficient to obtain multiple uncertainty contribution values; The summation of multiple uncertainty contributions and the square root are used as the standard uncertainty of the peak area. Configure a multiplier factor, and determine the interval width based on the multiplier factor and the standard uncertainty.

6. The intelligent peak analysis method for liquid chromatography based on deep learning according to claim 5, characterized in that, Configure the multiplier factor, and determine the interval width based on the multiplier factor and the standard uncertainty, including: Based on the risk profile of the application scenario, a pre-set confidence level is established, wherein the confidence level is proportional to the risk level. Based on the confidence level, the corresponding Z value is retrieved from the standard normal distribution table and used as the multiple factor; Based on the multiplication factor, the optimal parameter estimate, and the standard uncertainty, the confidence interval of the peak area is obtained; The width of the interval is obtained based on the confidence interval of the peak area.

7. The intelligent peak analysis method for liquid chromatography based on deep learning according to claim 1, characterized in that, Based on the confidence level of the results and a preset confidence threshold, the number and composition of sample chromatograms are dynamically adjusted, and the intelligent peak analysis network is continuously updated, including: Preset reliability threshold; Chromatograms that do not meet the confidence threshold are identified as weak chromatograms; Identify common features of weak chromatograms, including specific types of severe peak overlap and extreme baseline drift; Based on the common features of weak chromatograms, a physical model is used to generate synthetic chromatograms with similar features as targeted samples; The newly generated targeted samples are added to the sample chromatogram, and the intelligent chromatographic peak analysis network is retrained until the average confidence level of the model on the validation set exceeds the preset high confidence threshold. The validation set is labeled with accurate chromatographic peak information.

8. A deep learning-based intelligent peak analysis system for liquid chromatography, characterized in that, The system is used to implement the deep learning-based intelligent peak analysis method for liquid chromatography according to any one of claims 1-7, the system comprising: The network construction and training module is used to construct a chromatographic peak intelligent analysis network based on deep learning, and to train the chromatographic peak intelligent analysis network using sample chromatograms until convergence. The chromatographic peak intelligent analysis network includes a feature extractor, a physical parameter probability prediction output head, and a physical signal generator. The result confidence assessment module is used to input the chromatographic signal to be analyzed into the intelligent chromatographic peak analysis network, and to assess the reliability of the prediction results based on the probability distribution of the key physical parameters output by the intelligent chromatographic peak analysis network, thereby obtaining the result confidence level. The dynamic adaptive update module is used to dynamically adjust the number and composition of sample chromatograms based on the confidence level of the results and a preset confidence threshold, and to continuously update the intelligent peak analysis network. The heatmap visualization module is used to record the weight score assigned to each data point by the attention mechanism of the chromatographic peak intelligent analysis network. The weight score is mapped back to the time series of the original chromatographic signal and visualized through color gradient. It is superimposed on the original chromatographic curve in the form of a semi-transparent layer to generate a chromatographic peak heatmap.

Images