A method for analyzing components of a flavor synthesis product based on chromatographic data

Abstract

The present application relates to the technical field of chemical analysis data processing, and particularly relates to a flavor synthesis product component analysis method based on chromatographic data. The method comprises the following steps: obtaining chromatographic data of multiple historical qualified batches and a to-be-tested batch, calculating golden mean values and golden standard deviations of features of all historical qualified batches; constructing a perception stability mask, decomposing deviations of features of each target batch from corresponding golden mean values into perception deviations and amplitude deviations by using the perception stability mask; determining a dynamic process boundary; judging whether the perception deviations and the amplitude deviations of the to-be-tested batch are within the dynamic process boundary, so as to realize analysis on flavor synthesis product components. The scheme of the present application can accurately diagnose abnormal types in flavor synthesis product components, reduce false positive rates, and improve detection sensitivity of key trace risks.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of chemical analysis data processing. More particularly, the present application relates to a method for analyzing components of a perfume synthesis product based on chromatographic data. BACKGROUND

[0002] Gas chromatography or liquid chromatography is a standard method for analyzing product purity, main component content, byproducts and trace impurity distribution. Furthermore, by automatically comparing and analyzing chromatographic fingerprints of successive production batches, the stability of the production process can be monitored in real time.

[0003] At present, the industry generally uses multi-dimensional statistical process control technology, such as principal component analysis (PCA) combined with a box plot method, to achieve product component analysis and thus achieve the purpose of production process monitoring. The above method first collects a large amount of historical qualified batch chromatographic data to establish a PCA model, which compresses high-dimensional chromatographic data into several key principal components, which represent the directions with the largest variance in the data. Then, the scores of these principal components are statistically analyzed using a box plot, and a boundary is set, which is usually dependent on the upper quartile Q3 and the lower quartile Q1 of the data, the interquartile range IQR is calculated, and a fixed statistical multiplier (usually 1.5) is used to define the boundary. When the principal component scores of a new batch fall outside the boundary, it is determined to be a process anomaly.

[0004] However, the prior art has defects in the specific scenario of perfume quality control. First, the mathematical principle of the PCA model is to find the direction with the largest variance in the data. In perfume production, most of the variance may come from slight fluctuations in several major perfume components. However, those trace impurities with extremely low content and extremely small variance, such as allergens or odor components, often have a decisive influence on the quality of the aroma. The standard PCA, in the dimensionality reduction process, will preferentially extract the principal component variance, discarding these key trace impurity signals as noise, resulting in a model that is not sensitive to this perception-critical anomaly, causing false negatives. Second, the statistical basis of the 1.5 times IQR rule of the box plot is to assume that the data follows a symmetric normal distribution. However, in actual chemical production, due to physical or chemical limitations of the process, the data distribution often presents an asymmetric skewness. At this time, if a symmetric and fixed boundary is forcibly used, it will result in a large number of false positives at the long tail, while the boundary is too wide at the short tail, causing the problem of false negatives. SUMMARY

[0005] The purpose of the present application is to propose a method for analyzing components of a perfume synthesis product based on chromatographic data, to solve the problem that the principal component analysis in the prior art is not sensitive to key trace impurities and the statistical rule of the standard box plot is not suitable for skewed data, thus causing false negatives and false positives. To this end, the present application provides a solution in the following aspect.

[0006] The application provides a flavor synthesis product component analysis method based on chromatographic data, comprising:

[0007] Obtaining chromatographic data of a plurality of historical qualified batches and a to-be-tested batch, performing binning processing on all the chromatographic data to obtain a plurality of features, and calculating golden mean values and golden standard deviations of the features of all the historical qualified batches;

[0008] Constructing a perception stability mask, the perception stability mask being negatively correlated with a coefficient of variation corresponding to the features, the coefficient of variation being obtained from the golden mean values and the golden standard deviations of the features, and decomposing deviations of the features of each target batch from corresponding golden mean values into perception deviations and amplitude deviations by using the perception stability mask; the target batch being any batch in all the historical qualified batches and the to-be-tested batch;

[0009] Obtaining sequences of all the historical qualified batches in the perception deviations and the amplitude deviations, calculating robust skewnesses of the sequences, and constructing an adjustment factor based on the robust skewnesses to determine a dynamic process boundary;

[0010] Judging whether the perception deviations and the amplitude deviations of the to-be-tested batch are within the dynamic process boundary to realize analysis of components of a flavor synthesis product.

[0011] The above scheme replaces the variance-driven logic of traditional PCA by constructing a two-dimensional evaluation space of the perception deviations and the amplitude deviations, and replaces the fixed symmetric threshold of the standard box plot by constructing a dynamic process boundary, thereby solving the defects of PCA in distinguishing trace impurities and the defects of the standard box plot in misjudging skewed data, and achieving accurate diagnosis of abnormal types, while reducing the false positive rate and improving the detection sensitivity of key trace risks.

[0012] Optionally, a calculation formula of the perception stability mask is:

[0013] ;

[0014] In the formula, perception stability mask of the i-th feature is perception stability mask of the j-th feature is coefficient of variation of the i-th feature is median of the coefficients of variation of all the features is hyperbolic tangent function.

[0015] The above construction of the perception stability mask can nonlinearly lower the weight of the feature corresponding to the coefficient of variation (i.e., the key region with low signal and small fluctuation), and suppress the weight of the feature with high coefficient of variation (such as a fluctuating main peak), thereby providing a basis for subsequent separation of the perception deviations and the amplitude deviations. ​​​​​​

[0016] Optionally, the perception bias is:

[0017] ;

[0018] In the formula, For target batch Perceptual bias, For target batch The Middle The value of each feature, For the first The golden mean of each characteristic, For the first Perceptual stability mask for each feature The total number of features.

[0019] By calculating the sensing bias, the degree of deviation of the corresponding target batch in the critical trace area can be prioritized for evaluation, making the fluctuations of trace impurities that would be submerged in PCA significant, thus ensuring the sensitivity of sensing anomalies.

[0020] Optionally, the amplitude deviation is:

[0021] ;

[0022] In the formula, For target batch amplitude deviation, For target batch The Middle The value of each feature, For the first The golden mean of each characteristic, For the first Perceptual stability mask for each feature The total number of features.

[0023] Optionally, the adjustment factor includes an upper adjustment factor at the upper boundary and a lower adjustment factor at the lower boundary, wherein the upper adjustment factor and the lower adjustment factor are respectively:

[0024] ;

[0025] ;

[0026] In the formula, and These are the up-regulation factor and down-regulation factor of the target sequence, respectively. For the robustness skewness of the target sequence, Based on the multipliers, For the maximum adjustment range, For skewness sensitivity, is a maximum function, is a hyperbolic tangent function, and the target sequence is a sequence of perceived deviations or a sub-sequence of amplitude deviations.

[0027] By constructing the adjustment factor, the statistical boundary can be automatically adjusted according to the real skewness distribution of the data, for example, the boundary is relaxed on the long tail side of the data, thereby greatly reducing the false alarm caused by the skewness of the data.

[0028] Optionally, the robust skewness is obtained by using the median of the median.

[0029] Optionally, the dynamic process boundary includes an upper boundary and a lower boundary, and the upper boundary and the lower boundary are respectively:

[0030]

[0031]

[0032] wherein, , are an upper boundary and a lower boundary of the target sequence, and are a lower quartile and an upper quartile of the target sequence, is a quartile range of the target sequence, , are an upper adjustment factor and a lower adjustment factor of the target sequence.

[0033] The dynamic process boundary automatically adjusted according to the skewness of the data, that is, the boundary is automatically enlarged on the long tail side of the data, thereby significantly relaxing the boundary, so that the normal drift of the process is correctly determined as normal, and the false alarm rate is greatly reduced.

[0034] Optionally, the judgment of whether the perceived deviation and the amplitude deviation of the to-be-tested batch are within the dynamic process boundary to realize the analysis of the components of the synthesized product of the perfume comprises:

[0035] If the perceived deviation and the amplitude deviation of the to-be-tested batch are both outside the corresponding dynamic process boundary, it is determined as a composite anomaly; otherwise, it is determined that the to-be-tested batch is a process stable batch, and the components of the perfume are qualified.

[0036] If the perceived deviation of the to-be-tested batch is greater than the upper boundary of the sequence of perceived deviations, and the amplitude deviation of the to-be-tested batch is normal, it is determined as a perceived anomaly, and a new impurity in trace amount affecting the aroma or the absence of a key trace component occurs; if the amplitude deviation of the to-be-tested batch is greater than the upper boundary of the sequence of amplitude deviations, and the perceived deviation of the to-be-tested batch is normal, it is determined as an amplitude anomaly, and there is a fluctuation in the proportion or total yield of the components of the perfume.

[0037] ​​Optionally, the binning of all the chromatographic data obtains a plurality of features, including:

[0038] dividing each chromatographic data into a plurality of equal-width time intervals;

[0039] taking the peak area or average signal intensity in each time interval as the corresponding feature.

[0040] Optionally, further comprising: a step of pre-processing each chromatographic data, the pre-processing including: performing baseline correction of each chromatographic data using adaptive iteratively reweighted penalized least squares.

[0041] The beneficial effects of the present application are:

[0042] Firstly, the present application decouples batch bias into perception bias and amplitude bias by constructing a perception stability mask, which enables the monitoring system not only to determine whether there is an anomaly, but also to diagnose the type of anomaly, providing valuable decision guidance for quality control personnel (for example, perception bias anomaly may indicate that raw material pollution needs to be checked, and amplitude bias anomaly may indicate that reaction time needs to be adjusted). That is, the present application solves the defect in the prior art that all variances in the PCA model are mixed in the principal component, and cannot distinguish between trace key anomalies and major conventional fluctuations, and realizes accurate diagnosis of the type of anomaly.

[0043] Secondly, the present application solves the misjudgment defect of the standard box chart in the quality control process by constructing a regulation factor and further constructing a dynamic process boundary, thereby improving the sensitivity to key risks. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 a step flow chart of a flavor synthesis product component analysis method based on chromatographic data in the present embodiment is schematically shown;

[0045] Figure 2 a comparison effect diagram of a traditional boundary and an adaptive boundary obtained by the present application in flavor batch quality monitoring is schematically shown. DETAILED DESCRIPTION

[0046] 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.

[0047] As shown in the drawings, Figure 1 a flavor synthesis product component analysis method based on chromatographic data in the present embodiment includes the following steps:

[0048] S101: Collect and pre-process chromatographic data of a plurality of historical batches and chromatographic data of a batch to be tested, obtain chromatographic data of a plurality of historical qualified batches, and calculate a golden mean vector and a golden standard deviation vector.

[0049] The acquisition process of the golden mean vector and the golden standard deviation vector in this embodiment is as follows:

[0050] Firstly, the chromatographic data of a plurality of historical qualified batches is acquired.

[0051] In this embodiment, the chromatographic data, such as gas chromatography (GC) data, of historical batches of spice products is collected, and from which historical batches that are completely qualified in terms of artificial confirmation or laboratory physical and chemical indicators (such as aroma, purity, refractive index, etc.) are selected as historical qualified batches (i.e., golden batches), and the chromatographic data of the golden batches is obtained.

[0052] Illustratively, in this embodiment, the chromatographic data of historical batches is collected, and

[0053] In this embodiment, the chromatographic data of golden batches and all chromatographic data of the batch to be tested is also uniformly preprocessed. The preprocessing includes:

[0054] (1) Baseline correction: using adaptive iteratively reweighted penalized least squares (airPLS) or morphological method to subtract the drifting baseline.

[0055] (2) Noise filtering: using Savitzky-Golay smoothing filter to remove high-frequency noise while preserving peak shape characteristics.

[0056] (3) Retention time alignment: using correlation optimization algorithm (COW) or dynamic time warping (DTW) to eliminate retention time drift caused by column aging or flow rate fluctuations, and to ensure alignment of characteristic peaks of all chromatograms.

[0057] Secondly, a golden feature matrix is constructed.

[0058] In this embodiment, the preprocessed chromatographic data is binned at a fixed time interval, i.e., each preprocessed chromatographic data is divided into equal-width time intervals, and the peak area or average signal intensity in each time interval is calculated as a feature.

[0059] It should be noted that although the chromatographic data is a numerical sequence, it is generally displayed in the form of a chromatogram. Therefore, by integrating the signal intensity of the chromatogram, the peak area in each time interval can be obtained.

[0060] Illustratively, in this embodiment, when binning is performed, N binning operations can be performed, and thus a The gold feature matrix. Wherein, the feature represent the peak area of the th gold batch in the th time interval.

[0061] For example, if there are bins, the size of the gold feature matrix is .

[0062] The sequence number of the above time interval is the same as that of the feature.

[0063] Then, the gold mean vector and the gold standard deviation vector are calculated.

[0064] Based on the gold feature matrix, the two core reference vectors are the gold mean vector and the gold standard deviation vector, both of which are dimensional vectors, as follows:

[0065] (1) Gold mean vector: the gold feature matrix is averaged by column (feature), , is the gold mean value of the th feature in the gold mean vector.

[0066] The above gold mean vector represents the standard fingerprint of the spice chromatogram.

[0067] (2) Gold standard deviation vector: the gold feature matrix is calculated by column (feature) standard deviation, , is the gold standard deviation of the

[0068] th feature in the gold standard deviation vector.

[0069] Exemplarily, assuming that the gold batch , the number of features , the gold feature matrix at this time is : ; wherein, the 1st, 2nd and 4th columns represent the main components (large signal value), and the 3rd column represents the trace impurities (small signal value).

[0070] Based on the above gold feature matrix, the gold mean vector is , and the gold standard deviation vector is .

[0071] The above obtains the standard model for subsequent comparison and evaluation, i.e. the gold mean vector and the gold standard deviation vector, through the chromatogram data of all gold batches.

[0072] S102: Based on the golden mean vector and the golden standard deviation vector, the coefficient of variation of each data feature is calculated, and a perception stability mask is constructed. The deviation of the target batch feature from the corresponding golden mean is decomposed into a perception deviation and an amplitude deviation using the perception stability mask.

[0073] The acquisition process of the perception stability mask in the embodiment is as follows:

[0074] First, the coefficient of variation of each feature is calculated to reflect the relative volatility of the feature.

[0075] In perfume analysis, the most critical features are those with low signals (trace amounts) and high stability (small fluctuations). A feature with a mean and standard deviation close to 0 (i.e., a stable baseline region or trace impurities) is a major anomaly if it changes significantly, so the relative volatility of the feature needs to be obtained.

[0076] Specifically, the coefficient of variation of each feature is:

[0077] ;

[0078] In the formula, is the coefficient of variation of the i-th feature, is the golden standard deviation of the i-th feature, is the golden mean of the i-th feature, is a small offset constant to prevent division by zero error in the baseline region and ensure the stability of the calculated value. In one embodiment, it can be set that . Exemplarily, when the golden mean vector is

[0079] , the golden standard deviation vector is , , the four coefficients of variation can be obtained as follows:

[0080] , ,

[0081] , ,

[0082] , .

[0083] As can be seen from the above example, the coefficient of variation of the third feature belonging to the trace feature is , which has a relatively large relative volatility, while the coefficient of variation of the fourth feature belonging to the major feature is , which has a relatively small relative volatility, so the above trace feature is a feature that needs to be concerned.​​

[0084] Secondly, a perceptual stability mask is constructed to amplify "low CV" features (critical regions) and suppress "high CV" features (non-critical regions).

[0085] Specifically, the formula for calculating the perceived stability mask is:

[0086] ;

[0087] In the formula, For the first Perceptual stability mask for each feature For the first The coefficient of variation of each feature The median of the coefficients of variation for all features. It is the hyperbolic tangent function.

[0088] When the number of coefficients of variation is even, the mean of the two middle coefficients of variation can be used as the median.

[0089] The value range of the above-mentioned perceptual stability mask is [0,1]; This value serves as a normalization mechanism, exhibiting robustness and remaining unaffected by fluctuations in a few extreme cases of the coefficient of variation; when the coefficient of variation... Perceptual stability mask when much smaller than the median Approaching 1; when Perceptual stability mask when much larger than the median Close to 0.

[0090] In this embodiment, the perceptual stability masks of all features constitute the perceptual stability mask vector.

[0091] After obtaining the perceptual stability mask, the perceptual stability mask is used to obtain the following: Perception and amplitude deviation of the gold batch and the batch to be tested.

[0092] Specifically, the perception bias is used to assess the corresponding batch in the perception critical area ( The degree of deviation (of the region).

[0093] Among them, taking gold batch l as the target batch, the perceptual bias is calculated. The calculation formula is as follows:

[0094] ;

[0095] In the formula, Gold batch The Middle Values ​​on each feature Let j be the golden mean value corresponding to the j-th feature. For the first The perceptual stability mask for each feature, where N is the total number of features. The calculation of the perceptual bias described above is actually a weighted absolute bias.

[0096] For example, when a batch of gold The characteristic data of the chromatographic data after binning are as follows: And the constructed perceptual stability mask vector is So, gold batches The perceptual bias is approximately .

[0097] In the above embodiments, the weak signals of the sensing critical area are extracted separately and amplified onto the sensing deviation axis using a sensing stability mask, making the minute impurity fluctuations that would be submerged in PCA extremely significant, thus ensuring the detection sensitivity of high-risk sensing anomalies.

[0098] Among them, the amplitude deviation is used to assess the batch within the normal fluctuation range ( The degree of deviation (of the region). Specifically, the magnitude deviation. for:

[0099] ;

[0100] In the formula, Gold batch The Middle Values ​​on each feature Let j be the golden mean value corresponding to the j-th feature. For the first A perceptual stability mask for N features, where N is the total number of features.

[0101] The above calculation of amplitude deviation is actually a weighted root mean square error. It uses... As a weight, it complements the perceptual bias.

[0102] For example, when a batch of gold The characteristic data of the chromatographic data after binning are as follows: So, gold batches The amplitude deviation is approximately .

[0103] In the above embodiments, The standard dataset of gold batches from The 1D space is mapped to a more information-dense 2D space composed of perception bias and amplitude bias, resulting in a 2D distribution cloud, which is the normal process model; where a 2D vector in the 2D distribution cloud is... , the magnitude deviation of the gold batches, the perceptual deviation of the gold batches. the magnitude deviation of the gold batches, the perceptual deviation of the gold batches.

[0104] It should be noted that the above embodiment is used to replace the standard PCA, and the purpose is not to find the direction with the largest variance, but to separate the deviations according to the importance of the chemical components.

[0105] In this way, by constructing a two-dimensional space, the deviations of all gold batches are decomposed into the perceptual deviation reflecting the key trace area change and the magnitude deviation reflecting the conventional major area change, thereby solving the problem that the PCA cannot clearly distinguish the key impurities.

[0106] S103: Obtain the sequence of the perceptual deviation and the sequence of the magnitude deviation of all gold batches, calculate the robust skewness of each sequence, and construct an adjustment factor of the corresponding deviation based on the robust skewness, and then determine the asymmetric dynamic process boundary.

[0107] It should be noted that the fixed 1.5 times boundary is not used in the embodiment, but a dynamic process boundary is constructed for the axis where the magnitude deviation is located and the axis where the perceptual deviation is located .

[0108] Taking the sequence of the perceptual deviation or the sequence of the magnitude deviation as a target sequence, a dynamic process boundary of the target sequence is constructed, and the specific process is as follows:

[0109] First, the robust statistics of the target sequence are calculated, and the robust statistics include the upper quartile, the lower quartile, the interquartile range, and the robust skewness.

[0110] Obtain the target sequence of the gold batches.

[0111] Specifically, the upper quartile , the lower quartile , the interquartile range ( ) and the robust skewness of the above sequence are calculated.

[0112] The above robust skewness is calculated by Medcouple, the value range of which is [-1, 1], which can accurately describe the data skewness and is not affected by the abnormal value itself.

[0113] Since Medcouple is prior art, the specific process will not be described here.

[0114] ​Exemplarily, after obtaining the target sequence of the golden batch, it can be obtained by statistics that: , , , . Among them, indicates that the data is strongly right-skewed.

[0115] Since each bias (amplitude bias or perception bias) in the target sequence is theoretically always , the distribution is likely to be strongly right-skewed ( ).

[0116] Secondly, the adjustment factor is calculated.

[0117] For the right-skewed distribution of , the upper adjustment factor of the upper boundary of the axis where the target sequence is located is needed, while the lower adjustment factor of the lower boundary should be kept (no need to tighten on the short tail side). It should be noted that relaxing the upper boundary can prevent false positives.

[0118] Specifically, the calculation formulas of the adjustment factor are as follows:

[0119] ;

[0120] ;

[0121] In the formulas, and are the upper adjustment factor of the upper boundary and the lower adjustment factor of the lower boundary of the dynamic process boundary corresponding to the target sequence, is the robust skewness of the target sequence, is the basic multiplier, which is set to , representing the reference of the standard box plot, is the maximum adjustment amplitude, which is set to , is the skewness sensitivity, which is set to .

[0122] The above formulas use the hyperbolic tangent function to construct a one-way, non-linear adjustment relationship, which can dynamically enlarge the upper boundary, while keeping the lower boundary unchanged.

[0123] Then, the adaptive boundary is determined based on the adjustment factor.

[0124] The upper boundary and the lower boundary of the final dynamic process boundary of the axis (the axis or the axis) where the target sequence in the embodiment is located are as follows: ​​

[0125] ;

[0126] ;

[0127] wherein, , upper boundary, lower boundary of the dynamic process boundary corresponding to the target sequence respectively.

[0128] It should be noted that when is less than 0, the value of the lower boundary is 0, because the amplitude deviation score , the lower boundary will be usually truncated by 0, so the actual lower boundary is 0.

[0129] In this embodiment, a set of asymmetric two-dimensional rectangular boundaries is obtained, i.e. , the set of asymmetric two-dimensional rectangular boundaries is a dynamic boundary.

[0130] The above asymmetric boundary automatically stretched with data skew is constructed by calculating the robust skewness and the adjustment factor, which solves the problem that the traditional fixed boundary is prone to false positives on skewed data.

[0131] S104: Compare whether the perception deviation and the amplitude deviation of the to-be-tested batch are within the dynamic process boundary, to judge the to-be-tested batch, and realize the analysis of the components of the perfume synthesis product.

[0132] In this embodiment, the amplitude deviation and the perception deviation of the to-be-tested batch are obtained, i.e. , and compared with the dynamic process boundary to perform abnormality judgment and diagnosis.

[0133] Specifically, the process of abnormality judgment and diagnosis is as follows:

[0134] (1) Case one: if ( normal), it is determined that the perception is abnormal (high risk), and a high-priority alarm is triggered immediately. The diagnosis result is that the to-be-tested batch deviates in the perception key area, and there may be a new impurity of trace amount but affecting aroma or a missing key trace component.

[0135] (2) Case two: if ( normal), it is determined that the amplitude is abnormal (medium risk), and a medium-priority alarm is triggered. The diagnosis result is that the to-be-tested batch deviates in the conventional fluctuation area, indicating that the proportion of main perfume components or the total yield fluctuates significantly, and the process stability decreases.

[0136] (3) Case three: if and All exceeded the boundary, which was determined to be a composite anomaly, triggering the highest priority alarm.

[0137] (4) Case 4: If the scores of both axes are within their respective adaptive boundaries, the batch to be tested is determined to be a process-stable batch.

[0138] For example, such as Figure 2 As shown, it is a two-dimensional scatter plot, with the X-axis representing the amplitude deviation ( The Y-axis represents perceptual bias. ).

[0139] If the final adaptive boundary is Boundary is , The boundary is: So: For Figure 2 For batch A to be tested, ,at this time and If so, then the batch A to be tested is a batch with stable process.

[0140] for Figure 2 For batch B to be tested, ,at this time If so, then the batch B to be tested is of abnormal amplitude.

[0141] for Figure 2 For batch C to be tested, ,at this time If the batch C to be tested is found to be perceptually abnormal (high risk), then the batch C to be tested is considered to be abnormal (high risk).

[0142] in, Figure 2 The blue dots in the middle represent The distribution of gold batches in two-dimensional space is right-skewed on both axes. The red dashed box represents a fixed 1.5x. The rule, with its overly strict boundary, incorrectly classifies many gold batches (blue dots) and long-tail normal batches (such as the blue square test batch D) outside the bounding box, resulting in a large number of false alarms.

[0143] The green solid line box represents the dynamic boundary of the invention. Its upper and right boundaries extend significantly outward. The green box completely encompasses all gold batches and normal batch D, eliminating false alarms, while still correctly detecting batch B with X-axis anomalies (amplitude anomalies) and batch C with Y-axis anomalies (sensory anomalies). This demonstrates that the invention maintains high detection sensitivity for true anomalies while eliminating false alarms.

[0144] Therefore, from Figure 2As can be seen, by comparing the two-dimensional space deviation of the batch to be tested with the dynamically constructed boundary, not only can it be accurately judged whether the batch is abnormal, but also the specific type of abnormality can be diagnosed, thereby providing accurate decision basis for production control.

[0145] The scheme of the present application solves the defects that PCA is difficult to distinguish trace impurities and the defects that standard box chart misjudges skewed data, and can realize accurate diagnosis of abnormal types.

[0146] In the description of the present specification, the meaning of 'a plurality of' is at least two, such as two, three or more, etc., unless otherwise explicitly specifically limited.

[0147] Although the present specification has shown and described multiple embodiments of the present application, it is obvious to those skilled in the art that such embodiments are provided only in an exemplary manner. Those skilled in the art will think of many changes, changes and alternatives without departing from the idea and spirit of the present application.

Claims

1. A method for component analysis of fragrance synthesis products based on chromatographic data, characterized in that, include: Acquire chromatographic data from multiple historical qualified batches and batches to be tested, and perform binning on all chromatographic data to obtain multiple features, including: dividing each chromatographic data into multiple time intervals of equal width; Using the peak area or average signal strength within each time interval as the corresponding feature, calculate the gold mean and gold standard deviation of each feature for all historical qualified batches. Construct a perceptual stability mask: In the formula, For the first Perceptual stability mask for each feature For the first The coefficient of variation of each feature The median of the coefficients of variation for all features. The function is a hyperbolic tangent function. The perceptual stability mask is negatively correlated with the coefficient of variation of the corresponding feature. The coefficient of variation is obtained from the golden mean and golden standard deviation of each feature. The perceptual stability mask is used to decompose the deviation between the feature of each target batch and the corresponding golden mean into perceptual deviation and amplitude deviation. The target batch is any batch among all historical qualified batches and the batch to be tested. Obtain the sequences of all historical qualified batches in terms of perception deviation and amplitude deviation, calculate the robust skewness of each sequence, and construct an adjustment factor based on the robust skewness to determine the dynamic process boundary. Determine whether the perception deviation and amplitude deviation of the batch to be tested are within the dynamic process boundary in order to analyze the composition of the fragrance synthesis product.

2. The method for analyzing the components of fragrance synthesis products based on chromatographic data according to claim 1, characterized in that, The perception bias is: ； In the formula, For target batch Perceptual bias, For target batch The Middle The value of each feature, For the first The golden mean of each characteristic, For the first Perceptual stability mask for each feature The total number of features.

3. The method for analyzing the components of fragrance synthesis products based on chromatographic data according to claim 2, characterized in that, The amplitude deviation is: ； In the formula, For target batch amplitude deviation, For target batch The Middle The value of each feature, For the first The golden mean of each characteristic, For the first Perceptual stability mask for each feature The total number of features.

4. The method for analyzing the components of fragrance synthesis products based on chromatographic data according to claim 1, characterized in that, The adjustment factor includes an upper boundary adjustment factor and a lower boundary adjustment factor, wherein the upper and lower adjustment factors are respectively: ； ； In the formula, and These are the up-regulation factor and down-regulation factor of the target sequence, respectively. For the robustness skewness of the target sequence, Based on the multipliers, For the maximum adjustment range, For skewness sensitivity, To find the maximum value function, The function is a hyperbolic tangent, and the target sequence is either a sequence with perceptual bias or a subsequence with amplitude bias.

5. The method for analyzing the components of fragrance synthesis products based on chromatographic data according to claim 1, characterized in that, The robustness skewness is obtained using the Meadow median.

6. The method for analyzing the components of fragrance synthesis products based on chromatographic data according to claim 4, characterized in that, The dynamic process boundary includes an upper boundary and a lower boundary, wherein the upper boundary and the lower boundary are respectively: ； ； In the formula, , These are the upper and lower boundaries of the target sequence, respectively. and These are the lower quartile and upper quartile of the target sequence, respectively. The interquartile range of the target sequence. , These are the up-regulation factor and down-regulation factor of the target sequence, respectively.

7. The method for analyzing the components of fragrance synthesis products based on chromatographic data according to claim 1, characterized in that, The step of determining whether the perception deviation and amplitude deviation of the batch to be tested are within the dynamic process boundary, in order to achieve the analysis of the components of the fragrance synthesis product, includes: If both the perception deviation and amplitude deviation of the batch to be tested are outside the corresponding dynamic process boundary, it is determined to be a composite anomaly; otherwise, the batch to be tested is determined to be a process-stable batch and the fragrance components are qualified. If the perceptual deviation of the batch to be tested is greater than the upper boundary of the perceptual deviation sequence, and the amplitude deviation of the batch to be tested is normal, it is judged as perceptual abnormality, and there is a trace amount of new impurities or missing key trace components that affect the aroma; if the amplitude deviation of the batch to be tested is greater than the upper boundary of the amplitude deviation sequence, and the perceptual deviation of the batch to be tested is normal, it is judged as amplitude abnormality, and there is fluctuation in the proportion or total yield of flavor components.

8. The method for analyzing the components of fragrance synthesis products based on chromatographic data according to claim 1, characterized in that, Also includes: The preprocessing steps for each chromatographic data include: baseline correction of each chromatographic data using an adaptive iterative reweighted penalized least squares method.

Images