A knowledge graph-based decision support method and system for tea evaluation experts

Abstract

This invention discloses a knowledge graph-based expert decision support method and system for tea evaluation, relating to the field of tea quality evaluation technology. The method includes collecting tea sample data and expert review data; calculating the expert scoring consistency index; constructing an expert decision knowledge graph; analyzing the distribution patterns of expert scores to obtain rule mapping weights; mapping standardized features to the knowledge graph and generating feature transmission paths; correcting the transmission results based on rule applicability and conflict levels to obtain tea sample quality scores and feature analysis graphs; finally, converting the feature analysis graphs into a process control rule sequence; determining execution priorities based on quality scores; and generating process parameter adjustment schemes. This invention achieves intelligent optimization of tea production processes, improving the stability and consistency of tea quality.

Description

Technical Field

[0001] This invention relates to the field of tea quality evaluation technology, specifically to a knowledge graph-based decision support method and system for tea evaluation experts. Background Technology

[0002] Tea quality evaluation is a crucial step in ensuring the quality of tea products. Traditional tea evaluation methods mainly rely on expert experience for sensory assessment, which suffers from strong subjectivity, poor consistency, and difficulty in quantification. With the development of computer technology, machine learning-based tea quality evaluation methods have gradually emerged. However, due to the lack of full utilization of expert knowledge, the evaluation results deviate from actual experience.

[0003] The main problems in tea evaluation include: difficulty in standardizing the expression of expert experience, conflicts between evaluation rules, unreasonable allocation of feature weights, and lack of interpretability in quality scores. Although some studies have attempted to standardize the evaluation process by establishing an evaluation index system, they still cannot effectively capture the implicit knowledge in the expert decision-making process and are unable to cope with the dynamic changes in tea quality characteristics.

[0004] Knowledge graphs, as a knowledge representation method, can describe complex relationships between concepts. However, existing knowledge graph construction methods lack in-depth analysis of expert decision-making processes, make it difficult to accurately reflect the transmission relationships between features, and fail to fully consider the resolution mechanism of rule conflicts. Summary of the Invention

[0005] The purpose of this invention is to provide a knowledge graph-based decision support method and system for tea evaluation experts. By providing a tea evaluation method that can formalize expert experience and knowledge, dynamically adjust feature mapping, and resolve rule conflicts, the accuracy and interpretability of tea quality evaluation can be improved.

[0006] This invention provides a knowledge graph-based decision support method for tea evaluation experts, comprising the following steps:

[0007] Acquire tea sample data and perform standardization processing to obtain standardized features;

[0008] Data on the evaluation process of tea samples by multiple experts were collected, the consistency index of expert scores was calculated, expert evaluation rules were selected based on the consistency index, and an expert decision-making knowledge graph was constructed.

[0009] The distribution pattern of expert scores is analyzed to obtain the rule mapping weights. The standardized features are then mapped to the expert decision knowledge graph according to the rule mapping weights. The information transmission probability is adjusted according to the rule mapping weights to generate feature transmission paths.

[0010] Calculate the rule applicability and rule conflict degree along the feature transfer path, correct the transfer results based on the rule applicability and rule conflict degree, and obtain the quality score and feature analysis map of the tea sample;

[0011] The feature analysis diagram is converted into a process control rule sequence. The execution priority of the process control rule sequence is determined based on the quality score, and a process parameter adjustment plan is generated according to the rule execution priority.

[0012] The standardized features obtained by acquiring tea sample data and performing standardization processing include:

[0013] The spectral reflectance of tea samples was collected under different illumination angles, and the variation trend of reflectance in adjacent bands was calculated to obtain the variation curve.

[0014] Extract the inflection point of the change curve, map the reflectance at the inflection point to the manual score to obtain the correction coefficient, and correct the spectral reflectance data according to the correction coefficient;

[0015] The feature enhancement factor is obtained by extracting the ratio of the peak range to the steady-state range of the sensor response signal, and the corrected reflectivity data is enhanced based on the feature enhancement factor.

[0016] Wavelet decomposition is performed on the enhanced reflectivity data, and the scale coefficient with the highest energy is selected for reconstruction. The reconstructed feature data is then normalized to obtain standardized features.

[0017] This process involves collecting data from multiple experts on the evaluation of tea samples, calculating the consistency index of expert scores, selecting expert evaluation rules based on the consistency index, and constructing an expert decision-making knowledge graph, including:

[0018] Collect behavioral data and scoring data of multiple experts reviewing tea samples, extract expert attention area sequences from the behavioral data, and extract expert evaluation features from the scoring data;

[0019] Identify the stage boundaries of the review process based on changes in the expert attention area sequence, segment the review process based on the stage boundaries, and obtain the expert review rules for each segment.

[0020] The scoring consistency index is obtained by calculating the dispersion of the scoring data of different experts within each segment. Based on the scoring consistency index, the scoring consistency segments are screened, and the expert review rules corresponding to the scoring consistency segments are obtained.

[0021] The expert evaluation characteristics within the consistent scoring segment are mapped to the expert review rules to obtain the review mapping rules. The frequency of occurrence of the review mapping rules in the consistent scoring segment is calculated to obtain the rule credibility. The rule correlation is obtained based on the transformation relationship of the consistent scoring segment.

[0022] By using review mapping rules as nodes, rule credibility as node weight, and rule relevance as node connection strength, an expert decision-making knowledge graph is constructed.

[0023] The process involves analyzing the distribution patterns of expert scores to obtain rule-mapping weights, mapping standardized features to the expert decision-making knowledge graph according to these weights, adjusting information transmission probabilities based on the rule-mapping weights, and generating feature transmission paths, including:

[0024] The frequency distribution of expert ratings in different rating intervals is statistically analyzed and normalized to obtain the rating distribution pattern. Based on the rating distribution pattern, the conditional probability of adjacent rating intervals is calculated to obtain the rule mapping weight.

[0025] Standardized features are mapped to hierarchical nodes of the expert decision knowledge graph according to rule mapping weights. The rule mapping weights guide the distribution ratio of standardized features at different hierarchical nodes, resulting in a feature node distribution matrix.

[0026] Based on the rule mapping weights and feature node distribution matrix, the connection relationships between nodes in the expert decision knowledge graph are weighted and adjusted to obtain the information transmission probability matrix between nodes.

[0027] A state transition matrix is ​​constructed based on node distribution and information transmission probability. The state transition sequence of features in the expert decision knowledge graph is calculated, and the sequence with the highest probability in the state transition sequence is determined as the feature transmission path.

[0028] The process involves calculating the rule applicability and rule conflict degree along the feature transfer path, correcting the transfer results based on these factors, and obtaining the quality score and feature analysis diagram for the tea sample, including:

[0029] Historical sample data evaluated by experts is used as training samples. The Euclidean distance between the node rule features and the training sample features on the feature transfer path is calculated to obtain the rule applicability on the feature transfer path.

[0030] Calculate the semantic similarity of rule vectors of adjacent nodes on the feature transmission path, and normalize the semantic similarity to obtain the rule conflict degree on the feature transmission path;

[0031] An initial weight matrix is ​​constructed using the rule applicability as the diagonal element of the node. The rule conflict degree is normalized and used as the weight decay coefficient. The weight is then calculated based on the product of the initial weight matrix and the weight decay coefficient.

[0032] An exponential correction function is constructed based on the transmission weight to adaptively adjust the scores of each node in the transmission path, thereby obtaining the corrected transmission result.

[0033] The corrected transmission results are fused according to the transmission weights to obtain the quality score of the tea sample, and a feature analysis map is generated based on the correction process of the node scores.

[0034] The process involves converting the feature analysis diagram into a sequence of process control rules, determining the execution priority of the process control rules based on quality scores, and generating process parameter adjustment schemes according to the rule execution priority.

[0035] Obtain quality feature information from the feature analysis map, extract the feature response curve corresponding to the quality feature information, and analyze the fluctuation period and inflection point position of the feature response curve.

[0036] Calculate the rate of change of the characteristic response curve at the inflection point, determine the ratio of the rate of change to the fluctuation period as the weight adjustment factor, and generate dynamic characteristic weights based on the weight adjustment factor.

[0037] The quality feature information is mapped and transformed based on the dynamic feature weights to generate a process control rule sequence;

[0038] The execution sensitivity of rules in the process control rule sequence is calculated, the coupling relationship between rules is identified based on the execution sensitivity, and the constraint strength of rules is calculated based on the coupling relationship.

[0039] Constraints are constructed by combining rule constraint strength with quality score, and dynamic programming is used to determine the execution priority of the process control rule sequence.

[0040] Extract the adjustment direction of each rule in the process control rule sequence, determine the parameter adjustment amount by multiplying the adjustment direction by the execution priority, and generate a process parameter adjustment scheme based on the parameter adjustment amount.

[0041] Specifically, the rate of change of the characteristic response curve at the inflection point is calculated, and the ratio of the rate of change to the fluctuation period is determined as the weight adjustment factor. Dynamic characteristic weights are generated based on the weight adjustment factor, including:

[0042] Obtain the inflection point sequence of the characteristic response curve, determine the sampling window size based on the distribution density of the inflection point sequence, and calculate the rate of change of the characteristic response curve within the sampling window;

[0043] Perform time-frequency analysis on the characteristic response curve to identify the main frequency components, and determine the fluctuation period based on the main frequency components;

[0044] The ratio of the rate of change to the fluctuation period is determined as the weight adjustment factor. A mapping function is established based on the numerical distribution of the weight adjustment factor, and the weight adjustment factor is converted into dynamic feature weights through the mapping function.

[0045] This invention provides a knowledge graph-based decision support system for tea evaluation experts, the system comprising:

[0046] The sample processing module is used to acquire tea sample data and perform standardization processing to obtain standardized features.

[0047] The knowledge construction module is used to collect data on the review process of tea samples by multiple experts, calculate the consistency index of expert scores, select expert review rules based on the consistency index, and construct an expert decision knowledge graph.

[0048] The feature mapping module is used to analyze the distribution pattern of expert scores to obtain rule mapping weights, map standardized features to the expert decision knowledge graph according to the rule mapping weights, adjust the information transmission probability according to the rule mapping weights, and generate feature transmission paths.

[0049] The scoring generation module is used to calculate the rule applicability and rule conflict degree on the feature transmission path, correct the transmission results based on the rule applicability and rule conflict degree, and obtain the quality score and feature analysis map of the tea sample.

[0050] The control rules module is used to convert feature analysis diagrams into process control rule sequences, determine the execution priority of the process control rule sequences based on quality scores, and generate process parameter adjustment schemes according to the rule execution priority.

[0051] This invention constructs an expert decision-making knowledge graph, enabling the formal expression and quantitative analysis of expert review rules, effectively solving the problem of standardizing expert experience in traditional tea evaluation processes. The expert review rule selection mechanism based on a consistency index improves the reliability and stability of evaluation results. By analyzing the distribution patterns of expert scores to obtain rule mapping weights, the feature mapping process better aligns with expert decision-making logic, enhancing the accuracy of evaluation results. Introducing rule applicability and rule conflict degrees for result correction effectively alleviates conflicts between evaluation rules. Converting the feature analysis graph into a sequence of process control rules with execution priorities improves the practicality and operability of the evaluation results. This invention not only achieves intelligent and standardized tea quality evaluation but also ensures continuous optimization of the evaluation system through a dynamic adjustment mechanism of the knowledge graph, providing effective decision support for quality control in the tea production process.

[0052] The above description of the invention is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0053] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0054] Figure 1 A flowchart illustrating a knowledge graph-based decision support method for tea evaluation experts, provided as an embodiment of the present invention;

[0055] Figure 2 This is the process for generating an adaptive adjustment scheme for process parameters based on quality characteristic analysis, as described in this embodiment of the invention.

[0056] Figure 3 This is a schematic diagram of the structure of a tea evaluation expert decision support system based on knowledge graph, provided in an embodiment of the present invention. Detailed Implementation

[0057] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. In the following description relating to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements.

[0058] The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with one or more embodiments of this specification. It should be noted that the steps of the corresponding methods are not necessarily performed in the order shown and described in this specification in other embodiments. In some other embodiments, the methods may include more or fewer steps than described in this specification. Furthermore, a single step described in this specification may be broken down into multiple steps in other embodiments; and multiple steps described in this specification may be combined into a single step in other embodiments.

[0059] like Figure 1 As shown, Figure 1 A flowchart of a knowledge graph-based decision support method for tea evaluation experts is provided for embodiments of the present invention. The method includes the following steps:

[0060] Acquire tea sample data and perform standardization processing to obtain standardized features;

[0061] Data on the evaluation process of tea samples by multiple experts were collected, the consistency index of expert scores was calculated, expert evaluation rules were selected based on the consistency index, and an expert decision-making knowledge graph was constructed.

[0062] The distribution pattern of expert scores is analyzed to obtain the rule mapping weights. The standardized features are then mapped to the expert decision knowledge graph according to the rule mapping weights. The information transmission probability is adjusted according to the rule mapping weights to generate feature transmission paths.

[0063] Calculate the rule applicability and rule conflict degree along the feature transfer path, correct the transfer results based on the rule applicability and rule conflict degree, and obtain the quality score and feature analysis map of the tea sample;

[0064] The feature analysis diagram is converted into a process control rule sequence. The execution priority of the process control rule sequence is determined based on the quality score, and a process parameter adjustment plan is generated according to the rule execution priority.

[0065] Furthermore, the standardized features obtained from acquiring tea sample data and performing standardization processing include:

[0066] The spectral reflectance of tea samples was collected under different illumination angles, and the variation trend of reflectance in adjacent bands was calculated to obtain the variation curve.

[0067] Extract the inflection point of the change curve, map the reflectance at the inflection point to the manual score to obtain the correction coefficient, and correct the spectral reflectance data according to the correction coefficient;

[0068] The feature enhancement factor is obtained by extracting the ratio of the peak range to the steady-state range of the sensor response signal, and the corrected reflectivity data is enhanced based on the feature enhancement factor.

[0069] Wavelet decomposition is performed on the enhanced reflectivity data, and the scale coefficient with the highest energy is selected for reconstruction. The reconstructed feature data is then normalized to obtain standardized features.

[0070] First, the spectral reflectance of tea samples was collected under different illumination angles. During the collection process, a spectrometer was used to scan the tea samples, with the spectral range set to 400-2500 nm and a sampling interval of 1 nm. The tea samples were placed on a standard measurement platform and illuminated at three different angles: 0°, 45°, and 90°. Measurements were repeated five times at each angle, and the average value was taken as the spectral reflectance data for that angle.

[0071] After the spectral data acquisition is completed, the trend of reflectance variation in adjacent bands is calculated. For wavelength λ i and λ i+1 The reflectance difference is calculated and the change curve is plotted. For example, for a tea sample under a 45° illumination angle, adjacent wavelength bands in the range of 500-600nm are selected, and the reflectance difference at 1nm intervals is calculated to obtain the reflectance change curve.

[0072] Extracting inflection points from the spectral curve is a crucial step. Inflection points represent significant turning points in the trend of spectral curve changes and are usually related to the absorption or reflection characteristics of specific chemical components in tea. The inflection point locations are determined by detecting changes in the curve slope using the sliding window method. For example, a tea sample typically shows 3-7 inflection points, and the wavelengths at these inflection point locations often correspond to the characteristic absorption bands of key components such as tea polyphenols and caffeine.

[0073] Once the inflection point is determined, the reflectance at that point is mapped to the human score to obtain a correction coefficient. The human score is determined by five professional tea tasters who evaluate the samples based on sensory indicators, including appearance, aroma, taste, and liquor color, with a maximum score of 100. The correction coefficient is obtained by establishing a correspondence between the reflectance at the inflection point and the human score. For example, for a certain variety of tea, the reflectance at the inflection point at a wavelength of 540nm is 0.35, corresponding to a human score of 85, and the calculated correction coefficient for that wavelength is 1.08.

[0074] The original spectral reflectance data is corrected using correction factors. The reflectance value at each wavelength is multiplied by the corresponding correction factor to obtain the corrected reflectance data. The corrected data more accurately reflects the relationship between tea quality and spectral characteristics.

[0075] Next, the feature enhancement factor is obtained by extracting the ratio of the peak range to the steady-state range of the sensor response signal. The peak range refers to the region in the spectral curve where the reflectance reaches a local maximum, while the steady-state range refers to the region where the reflectance is relatively stable. The feature enhancement factor is obtained by calculating the ratio of the average reflectance in the peak range to the average reflectance in the steady-state range. For green tea samples, the peak range is typically 550-570 nm, and the steady-state range is typically 700-750 nm. If the average reflectances in these two ranges for a certain sample are 0.48 and 0.23 respectively, the calculated feature enhancement factor is 2.09.

[0076] The corrected reflectance data is enhanced using a feature enhancement factor. Specifically, the reflectance data in the peak region is multiplied by the feature enhancement factor, while other regions remain unchanged. This step highlights characteristic band information highly correlated with tea quality, enhancing the accuracy of subsequent analyses.

[0077] Wavelet decomposition was performed on the enhanced reflectivity data. The Db4 wavelet basis was selected, and a 5-level decomposition was performed to obtain a series of wavelet coefficients. Energy analysis was used to calculate the energy distribution of the coefficients at each scale, and the scale coefficients with the highest energy were selected for reconstruction. For most tea samples, the 3rd and 4th level wavelet coefficients typically contain the most significant energy; therefore, these coefficients were selected for reconstruction.

[0078] The reconstructed feature data was normalized to obtain standardized features. The normalization method used was maximum-minimum normalization, adjusting the data range to between 0 and 1. The standardized feature data was then used to construct a tea evaluation knowledge graph, establishing a mapping relationship between tea quality and spectral characteristics.

[0079] By employing spectral analysis techniques and data processing algorithms, an objective evaluation of tea quality has been achieved. This method effectively overcomes the subjectivity and instability of traditional manual evaluation methods, significantly improving the accuracy and repeatability of tea quality assessment.

[0080] Furthermore, data from the review process of multiple experts on tea samples was collected, the consistency index of expert scores was calculated, expert review rules were selected based on the consistency index, and an expert decision-making knowledge graph was constructed, including:

[0081] Collect behavioral data and scoring data of multiple experts reviewing tea samples, extract expert attention area sequences from the behavioral data, and extract expert evaluation features from the scoring data;

[0082] Identify the stage boundaries of the review process based on changes in the expert attention area sequence, segment the review process based on the stage boundaries, and obtain the expert review rules for each segment.

[0083] The scoring consistency index is obtained by calculating the dispersion of the scoring data of different experts within each segment. Based on the scoring consistency index, the scoring consistency segments are screened, and the expert review rules corresponding to the scoring consistency segments are obtained.

[0084] The expert evaluation characteristics within the consistent scoring segment are mapped to the expert review rules to obtain the review mapping rules. The frequency of occurrence of the review mapping rules in the consistent scoring segment is calculated to obtain the rule credibility. The rule correlation is obtained based on the transformation relationship of the consistent scoring segment.

[0085] By using review mapping rules as nodes, rule credibility as node weight, and rule relevance as node connection strength, an expert decision-making knowledge graph is constructed.

[0086] First, behavioral and scoring data of multiple experts reviewing tea samples were collected to construct an expert decision-making knowledge graph, providing objective and systematic decision support for tea quality evaluation.

[0087] Collecting behavioral and scoring data from experts reviewing tea samples is fundamental to this method. Behavioral data collection utilizes eye-tracking devices to record the experts' fixation points, duration, and sequence during the review process. For example, when an expert reviews a green tea sample, the eye-tracking device records the expert's fixation behavior on areas such as the tea's appearance, liquor color, and infused leaves, including the coordinate sequence of fixation points and the corresponding dwell time. Scoring data is collected using a dedicated tea scoring sheet, which includes scores for five dimensions: tea appearance, aroma, taste, liquor color, and infused leaves. Each dimension has a maximum score of 20 points, for a total of 100 points. The scoring sheet also includes a comment section, requiring experts to provide written descriptions of the key quality characteristics for each dimension. For example, for the appearance dimension, experts must provide specific evaluations of factors such as the tenderness of the raw material, its shape characteristics, color type, oiliness, uniformity, and purity, such as descriptive comments like "tightly rolled, thin, and straight leaves" or "bright green and vibrant color."

[0088] Extracting expert attention region sequences from behavioral data requires processing the eye-tracking data. First, the tea sample images are divided into different regions of interest, such as the shape area, the liquor color area, and the leaf residue area. Then, the expert's gaze points are mapped onto these regions, generating an attention region sequence in chronological order. Extracting expert evaluation features from the rating data involves transforming the expert rating data into a structured evaluation feature vector, including scores for each dimension and comment keywords. Comment keywords are obtained through word segmentation and keyword extraction of the expert's written comments.

[0089] The review process is divided into stages based on changes in the sequence of areas of focus observed by experts. Key transition points in the review process are identified by analyzing the transition patterns within the sequence of areas of focus. Specifically, the rate of change of the areas of focus is calculated over two consecutive time windows; when the rate of change exceeds a preset threshold, that time point is marked as a stage boundary. For example, when experts shift their focus from the shape area to the color area, a significant change occurs, and this time point is marked as a stage boundary. Based on these boundaries, the entire review process is divided into multiple stages, such as the sensory evaluation stage, the quality analysis stage, and the comprehensive evaluation stage.

[0090] The expert review rules for each segment are obtained by analyzing the behavioral patterns and scoring characteristics of experts in each stage. For each stage, the distribution characteristics of the areas of interest of experts, the distribution of their dwell time, and the pattern of area switching are extracted. Combined with the scoring data, the review rules for that stage are summarized. First, the samples within this stage are grouped according to their scoring ranges. For example, samples with an appearance score of 18-20 are grouped together. Then, quality characteristic descriptive words are extracted from the expert comments of this group of samples. The expert comments record the experts' specific evaluations of various quality elements of tea. For example, comments on appearance include "uniform tenderness" for the tenderness of the raw material, "tight, thin, and straight strips" for the shape characteristics, "bright green and vibrant color" for the color type, "good oiliness" for the oiliness, "high uniformity" for the uniformity, and "excellent purity" for the purity. The frequency of each quality characteristic descriptive word in the comments of this group of samples is counted, and the proportion of each descriptive word's frequency to the total number of times all descriptive words in this group appear is calculated. The quality characteristic descriptive words with the highest frequency proportions are identified as the main quality characteristics of this scoring range. The quality elements corresponding to these high-frequency descriptive words are the dominant elements in this scoring range. Finally, the main quality characteristic descriptive words are combined to form the evaluation rules for this scoring range in this stage. For example, in the sensory evaluation stage, an analysis was conducted on the sample group with an appearance score of 18-20. Quality characteristic descriptive words were extracted from the expert comments of this sample group. Statistics showed that "tightly rolled, fine, and straight strips" appeared 32 times, "bright green and vibrant color" appeared 30 times, "uniform tenderness" appeared 8 times, "good oiliness" appeared 5 times, "high uniformity" appeared 4 times, and "excellent purity" appeared 3 times. The frequency ratio of each descriptive word was calculated: "tightly rolled, fine, and straight strips" accounted for 39%, "bright green and vibrant color" accounted for 37%, and other descriptive words all accounted for less than 10%. Based on this, "tightly rolled, fine, and straight strips" and "bright green and vibrant color" were determined to be the main quality characteristics of this scoring range. The corresponding quality elements "shape characteristics" and "color type" were the dominant elements of this scoring range. Therefore, the evaluation rule was stated as "high-quality green tea should have tightly rolled, fine, and straight strips, and a bright green and vibrant color."

[0091] The scoring consistency index is obtained by calculating the dispersion of scores from different experts within each segment. For each review stage, the standard deviation of scores from different experts on each dimension is calculated, and the average of the standard deviations of each dimension is taken as the scoring dispersion for that stage. The scoring consistency index is equal to 1 minus the ratio of the scoring dispersion to the maximum probable dispersion. For example, in the sensory evaluation stage of a green tea sample, the appearance scores of 5 experts are 18, 19, 18, 17, and 18. The calculated standard deviation is 0.63, and the maximum probable standard deviation is 8.94. The scoring consistency index for this stage is 0.93, indicating a high degree of consistency in expert scores.

[0092] Scoring consistency segments are selected based on a scoring consistency index, with a threshold of 0.85. When the scoring consistency index for a given stage exceeds this threshold, that stage is marked as a scoring consistency segment. A scoring consistency segment indicates a high degree of similarity between the experts' review behavior and scoring characteristics at that stage. For each scoring consistency segment, the corresponding expert review rules are extracted as candidate rules. These candidate rules require comprehensive verification through subsequent indicators such as rule credibility, rule applicability, and rule conflict to ensure the reliability of the rules.

[0093] It should be noted that the scoring consistency index is only used to identify segments where expert review behavior is relatively concentrated for extracting candidate rules, and does not represent the reliability of the rules. The reliability of a rule is verified through subsequent rule credibility (frequency of occurrence), rule applicability (Euclidean distance from historical standard samples), and rule conflict degree. This mechanism can effectively handle different situations: for cases with low consistency, if it is due to a mismatch in the level of individual experts, the rule will be excluded due to its low frequency; if it is due to reasonable disagreement, different candidate rules will be selected based on applicability verification to select the one closest to the standard. For cases with high consistency but with systematic bias, rule applicability verification will identify the bias and reduce or exclude the weight of the rule.

[0094] The expert evaluation features within the consistent scoring range are mapped to the expert review rules to obtain the review mapping rules. The mapping process is achieved by analyzing the correspondence between expert evaluation features and review rules. In establishing the mapping, the expert review rules are first semantically parsed, decomposing the rule text into standardized quality feature entities and evaluation attribute entities. For example, the rule text "fresh and sweet taste with a lasting aftertaste" is decomposed using natural language processing technology into the quality dimension entity "taste," the sensory feature entities "fresh" and "sweet," and the persistence attribute entity "lasting aftertaste." Then, semantic relationships between entities are established, including the descriptive relationship between "taste" and "fresh," the descriptive relationship between "taste" and "sweet," and the modifying relationship between "aftertaste" and "lasting." Further, each entity is mapped to the conceptual level of the tea review system, where "taste" is mapped to the sensory evaluation dimension level, and "fresh" and "sweet" are mapped to... At the flavor feature subclass level, "lingering sweetness" is mapped to the flavor quality attribute level. Based on the above semantic parsing and entity mapping, a structured review mapping rule is constructed. This rule includes entity nodes, entity attributes, and semantic relationships between entities. For example, when experts score the flavor of tea between 18 and 20 points, the corresponding review rule is "fresh, sweet, and mellow flavor with a lingering sweetness." When establishing the mapping, the rule text is parsed into structured entities: "flavor" as a quality dimension entity, "fresh" and "mellow" as flavor feature entities, and "lingering sweetness" as a persistence attribute entity. Semantic relationships between entities are established, thus establishing a mapping relationship between the evaluation feature "flavor score ≥ 18" and the structured review rule "fresh, sweet, and mellow flavor with a lingering sweetness."

[0095] The confidence level of a review mapping rule is calculated by determining its frequency of occurrence in consistent scoring segments. For each review mapping rule, the number of times it appears in all consistent scoring segments is counted, and this count is divided by the total number of consistent scoring segments to obtain the rule's confidence level. For example, if a review mapping rule appears 8 times in 10 consistent scoring segments, its confidence level is 0.8.

[0096] The correlation degree of rules is obtained based on the transformation relationship of scoring consistency segments. The transformation patterns between different scoring consistency segments are analyzed, and the frequency of co-occurrence of two segments is calculated as the correlation degree between the corresponding review rules. For example, if the frequency of co-occurrence of the rules in the sensory evaluation stage and the rules in the quality analysis stage in the sample set is 0.75, then the correlation degree between these two rules is 0.75.

[0097] An expert decision-making knowledge graph is constructed by using review mapping rules as nodes, rule credibility as node weight, and rule relevance as node connection strength. Each node in the knowledge graph represents a review mapping rule, the size of the node indicates the credibility of the rule, the lines between nodes represent the relationships between rules, and the thickness of the lines indicates the strength of the relationships.

[0098] By constructing an expert decision-making knowledge graph, this method effectively captures key decision-making rules and the relationships between rules in the expert review process, providing a scientific basis for tea quality evaluation. It significantly improves the consistency and reliability of tea evaluation, reduces the influence of subjective factors, lowers evaluation costs, and provides an effective learning tool for novice evaluators, accelerating the transmission of professional knowledge. The knowledge graph implemented by this method can also be further integrated with intelligent evaluation systems to achieve automated tea quality evaluation, promoting the digital and intelligent development of the tea industry and enhancing its overall competitiveness.

[0099] Furthermore, the distribution patterns of expert ratings are analyzed to obtain rule-mapping weights. Standardized features are then mapped onto the expert decision-making knowledge graph according to these weights. Information transmission probabilities are adjusted based on these weights to generate feature transmission paths, including:

[0100] The frequency distribution of expert ratings in different rating intervals is statistically analyzed and normalized to obtain the rating distribution pattern. Based on the rating distribution pattern, the conditional probability of adjacent rating intervals is calculated to obtain the rule mapping weight.

[0101] Standardized features are mapped to hierarchical nodes of the expert decision knowledge graph according to rule mapping weights. The rule mapping weights guide the distribution ratio of standardized features at different hierarchical nodes, resulting in a feature node distribution matrix.

[0102] Based on the rule mapping weights and feature node distribution matrix, the connection relationships between nodes in the expert decision knowledge graph are weighted and adjusted to obtain the information transmission probability matrix between nodes.

[0103] A state transition matrix is ​​constructed based on node distribution and information transmission probability. The state transition sequence of features in the expert decision knowledge graph is calculated, and the sequence with the highest probability in the state transition sequence is determined as the feature transmission path.

[0104] The first step in this method is to statistically analyze and normalize the frequency distribution of expert ratings across different rating intervals to obtain the distribution pattern. The expert rating range is divided into multiple intervals, such as five intervals for appearance ratings: 0-60, 61-70, 71-80, 81-90, and 91-100. For the collected rating data from multiple experts on the same batch of tea samples, the frequency of each rating within each interval is calculated. For example, for the appearance ratings of a batch of green tea samples, the frequency in the 81-90 interval is 32 times, and the frequency in the 91-100 interval is 18 times. The frequencies of each interval are normalized by dividing the frequency of each interval by the total number of ratings to obtain the rating probability for that interval. For example, if the frequency of a certain interval is 32 times and the total number of ratings is 100, then the rating probability for that interval is 0.32. In this way, the distribution pattern of ratings across each interval is obtained.

[0105] The rule mapping weights are obtained by calculating the conditional probabilities of adjacent rating intervals based on the rating distribution patterns. For any two adjacent rating intervals A and B, the probability that a rating exists in interval A and therefore also exists in interval B is calculated, and this probability serves as the rule mapping weight from interval A to interval B. For example, if 80% of the samples with appearance ratings of 81-90 also have aroma ratings in the 81-90 range, then the rule mapping weight from the appearance rating interval of 81-90 to the aroma rating interval of 81-90 is 0.8. In this way, the rule mapping weights between all adjacent rating intervals are calculated, forming a rule mapping weight matrix.

[0106] Standardized features are mapped to hierarchical nodes of the expert decision-making knowledge graph according to rule-based mapping weights, resulting in a feature node distribution matrix. Standardized features refer to normalized tea spectral reflectance data, with values ​​ranging from 0 to 1. Hierarchical nodes in the expert decision-making knowledge graph correspond to different evaluation dimensions (such as appearance, aroma, and taste) and scoring ranges. During the mapping process, the allocation ratio of standardized features to different hierarchical nodes is determined according to the rule-based mapping weights. For example, for a wavelength point with a standardized feature value of 0.85, according to the rule-based mapping weights, it is allocated as follows: 0.6 to nodes in the appearance scoring range of 91-100 points, 0.3 to nodes in the aroma scoring range of 81-90 points, and 0.1 to nodes in the taste scoring range of 71-80 points. In this way, all standardized features are mapped to the hierarchical nodes of the knowledge graph, forming a feature node distribution matrix.

[0107] The connection relationships between nodes in the expert decision-making knowledge graph are weighted based on rule mapping weights and feature node distribution matrices to obtain an information transmission probability matrix. The initial connection weights between nodes are derived from the relevance of the expert review rules and are adjusted using rule mapping weights and feature node distributions. During the adjustment process, for any two connected nodes i and j, the information transmission probability from node i to node j is calculated. This probability equals the feature distribution value of node i multiplied by the rule mapping weight from node i to node j, then divided by the sum of the rule mapping weights of all out-degree connections of node i. For example, if the feature distribution value of node i is 0.6, the rule mapping weight from node i to node j is 0.8, and the sum of the rule mapping weights of all out-degree connections of node i is 2.0, then the information transmission probability from node i to node j is 0.6 × 0.8 ÷ 2.0 = 0.24. In this way, the information transmission probability between all connected node pairs in the knowledge graph is calculated, forming an information transmission probability matrix.

[0108] A state transition matrix is ​​constructed based on node distribution and information transmission probabilities to calculate the state transition sequence of features in the expert decision-making knowledge graph. The state transition matrix describes the probability distribution of feature transitions between different nodes in the knowledge graph. During construction, the rows and columns of the matrix correspond to nodes in the knowledge graph, and the matrix elements represent the probability of transitioning from a row node to a column node; this probability is the information transmission probability between nodes. For example, if the probability of transitioning from appearance rating node A to aroma rating node B is 0.24, then the corresponding element in the state transition matrix has a value of 0.24. Based on the constructed state transition matrix, an iterative calculation method is used to generate multiple possible state transition sequences starting from the initial node based on the transition probabilities. In each iteration, starting from the current node, the next node is selected based on the transition probability, and the transition path and its probability are recorded.

[0109] The sequence with the highest probability among the state transition sequences is determined as the feature transfer path. The probabilities of all generated state transition sequences are compared, and the sequence with the highest probability is selected as the optimal feature transfer path in the knowledge graph. If, among multiple generated state transition sequences, the sequence "Appearance Node A → Aroma Node B → Taste Node C → Tea Color Node D" has a probability of 0.021, which is higher than the probabilities of other sequences, then this sequence is determined as the feature transfer path. The final output feature transfer path reflects the most likely flow mode of standardized features in the expert decision-making knowledge graph, representing the decision-making and focus shift patterns of experts when evaluating tea.

[0110] By analyzing the distribution patterns of expert scores and constructing feature transmission paths, this method achieves precise modeling and simulation of expert review logic. By organically combining tea spectral data with expert review knowledge, a mapping mechanism from objective characteristics to subjective evaluation is established, providing scientific and objective decision support for tea quality evaluation. Through the determination of feature transmission paths, this method can explain the thought processes and decision-making basis in the expert review process, enhancing the interpretability of the evaluation results.

[0111] Furthermore, the rule applicability and rule conflict degree along the feature transfer path are calculated. Based on these factors, the transfer results are corrected to obtain the quality score and feature analysis diagram of the tea sample, including:

[0112] Historical sample data evaluated by experts is used as training samples. The Euclidean distance between the node rule features and the training sample features on the feature transfer path is calculated to obtain the rule applicability on the feature transfer path.

[0113] Calculate the semantic similarity of rule vectors of adjacent nodes on the feature transmission path, and normalize the semantic similarity to obtain the rule conflict degree on the feature transmission path;

[0114] An initial weight matrix is ​​constructed using the rule applicability as the diagonal element of the node. The rule conflict degree is normalized and used as the weight decay coefficient. The weight is then calculated based on the product of the initial weight matrix and the weight decay coefficient.

[0115] An exponential correction function is constructed based on the transmission weight to adaptively adjust the scores of each node in the transmission path, thereby obtaining the corrected transmission result.

[0116] The corrected transmission results are fused according to the transmission weights to obtain the quality score of the tea sample, and a feature analysis map is generated based on the correction process of the node scores.

[0117] Obtaining historical sample data from expert evaluations as training samples is fundamental to the implementation of this method. Historical samples include tea sample data that has been reviewed and scored by experts. Each sample contains standardized feature data and a corresponding expert score. For example, spectral reflectance data of 100 green tea samples and expert scores for these samples across five dimensions—appearance, aroma, taste, liquor color, and infused leaf appearance—were collected. Each sample has a standardized feature dimension of 400, corresponding to spectral reflectance in the 400-2500 nm wavelength range, and a scoring dimension of 5, corresponding to the scores across the five evaluation dimensions.

[0118] The Euclidean distance between the node rule features and the training sample features along the feature propagation path is calculated to obtain the rule applicability along the feature propagation path. Each node along the feature propagation path corresponds to a review rule, such as the feature vector corresponding to "tight, uniform, and straight appearance, and vibrant green color". For the tea sample to be evaluated, the Euclidean distance between its standardized features and the rule features of each node is calculated. The smaller the Euclidean distance, the more similar the sample features are to the rule features, and the higher the rule applicability. The rule applicability is calculated by subtracting the normalized Euclidean distance from 1.

[0119] The semantic similarity of rule vectors between adjacent nodes along the feature propagation path is calculated, and the semantic similarity is normalized to obtain the rule conflict degree along the feature propagation path. A node rule vector refers to the textual representation of the review rule, such as the vector representation of "the shape is tight, thin, uniform, and straight" after text encoding. For any two adjacent nodes along the feature propagation path, the cosine similarity of their rule vectors is calculated to obtain the semantic similarity. The cosine similarity is calculated by dividing the dot product of the two vectors by the product of their Euclidean norms. The higher the semantic similarity, the more similar the content of the two rules. The rule conflict degree is calculated by subtracting the semantic similarity from 1. The lower the rule conflict degree, the higher the consistency between adjacent rules.

[0120] An initial weight matrix is ​​constructed using the rule applicability as the diagonal elements of each node. The rule conflict degree, after normalization, is used as the weight decay coefficient. The transferred weights are calculated based on the product of the initial weight matrix and the weight decay coefficient. The initial weight matrix is ​​a diagonal matrix; the diagonal elements represent the rule applicability of each node, and the off-diagonal elements are 0. For example, for a transfer path containing three nodes—appearance, aroma, and taste—if the rule applicability of each node is 0.88, 0.92, and 0.85 respectively, then the diagonal elements of the initial weight matrix will be these three values. The weight decay coefficient reflects the degree of influence of rule conflict on the weights. It is calculated by dividing the reciprocal of the rule conflict degree by the maximum value of the reciprocal of the rule conflict degree. For example, if the rule conflict degree from the appearance node to the aroma node is 0.05, and the rule conflict degree from the aroma node to the taste node is 0.08, then the corresponding weight decay coefficients are 0.98 and 0.96 respectively. The transferred weights are calculated by multiplying the diagonal elements of the initial weight matrix by the corresponding weight decay coefficient. For the above example, the transfer weight of the shape node is 0.88 × 0.98 = 0.86, the transfer weight of the aroma node is 0.92 × 0.96 = 0.88, and the transfer weight of the taste node is 0.85 × 1 = 0.85.

[0121] An exponential correction function is constructed based on the transmission weights to adaptively adjust the scores of each node in the transmission path, resulting in a corrected transmission result. The exponential correction function is an exponential function, where the exponent is the product of the transmission weights and the adjustment coefficients. The adjustment coefficients are determined based on the node's position in the transmission path; typically, nodes further forward have smaller adjustment coefficients, and nodes further back have larger ones. For example, for a transmission path with three nodes, the adjustment coefficients can be set to 0.5, 0.7, and 0.9. For each node, its initial score is substituted into the exponential correction function to obtain the corrected score. For example, the initial score of the appearance node is 85 points, the transmission weight is 0.86, and the adjustment coefficient is 0.5; substituting these values ​​into the correction function yields a corrected score of 88 points. Similarly, the corrected scores for the aroma and flavor nodes are calculated to be 90 and 87 points, respectively.

[0122] The corrected transmission results are fused according to transmission weights to obtain the quality score of the tea sample, and a feature analysis chart is generated based on the correction process of the node scores. The quality score fusion uses a weighted average method, with the transmission weight of each node as the weight, to calculate the weighted average of the corrected scores. For example, for the three nodes mentioned above, their corrected scores are 88, 90, and 87 points respectively, with transmission weights of 0.86, 0.88, and 0.85 respectively. The calculated quality score is (88×0.86+90×0.88+87×0.85) / (0.86+0.88+0.85)=88.4 points. The feature analysis chart is used to visualize the correction process of the node scores and the contribution of each node to the final score. The feature analysis chart consists of two parts: a radar chart and a contribution bar chart. The radar chart shows the changes in the scores of each node before and after correction. Each axis of the radar chart represents a different evaluation dimension, such as appearance, aroma, and taste, and the position of the point represents the score value. The contribution bar chart shows the percentage of each node's contribution to the final score. The calculation method is the proportion of the node's transmitted weight to the total weight. For example, the contribution of the shape node is 0.86 / (0.86+0.88+0.85)=0.33, or 33%.

[0123] By assessing rule applicability and conflict levels, calculating transfer weights, adaptively adjusting scores, and integrating quality scores, a precise evaluation of tea quality was achieved. The system fully considers the mutual influence and conflict relationships between expert review rules, and effectively solves the scoring bias problem caused by rule conflicts in traditional evaluation methods by adaptively adjusting scores through an exponential correction function. The generation of feature analysis diagrams makes the evaluation process more transparent and interpretable, facilitating the understanding of the decision-making basis behind the scoring results.

[0124] like Figure 2 As shown, the process for generating an adaptive adjustment scheme for process parameters based on quality feature analysis according to an embodiment of the present invention is illustrated.

[0125] Furthermore, the feature analysis diagram is converted into a sequence of process control rules. The execution priority of the process control rules sequence is determined based on the quality score. A process parameter adjustment scheme is generated according to the rule execution priority, including:

[0126] Obtain quality feature information from the feature analysis map, extract the feature response curve corresponding to the quality feature information, and analyze the fluctuation period and inflection point position of the feature response curve.

[0127] Calculate the rate of change of the characteristic response curve at the inflection point, determine the ratio of the rate of change to the fluctuation period as the weight adjustment factor, and generate dynamic characteristic weights based on the weight adjustment factor.

[0128] The quality feature information is mapped and transformed based on the dynamic feature weights to generate a process control rule sequence;

[0129] The execution sensitivity of rules in the process control rule sequence is calculated, the coupling relationship between rules is identified based on the execution sensitivity, and the constraint strength of rules is calculated based on the coupling relationship.

[0130] Constraints are constructed by combining rule constraint strength with quality score, and dynamic programming is used to determine the execution priority of the process control rule sequence.

[0131] Extract the adjustment direction of each rule in the process control rule sequence, determine the parameter adjustment amount by multiplying the adjustment direction by the execution priority, and generate a process parameter adjustment scheme based on the parameter adjustment amount.

[0132] The quality feature information in the feature analysis graph is obtained, and the corresponding feature response curves are extracted. The fluctuation period and inflection point position of the feature response curves are analyzed. The feature analysis graph is a visualization result generated during the aforementioned tea evaluation process, containing quality feature information for each evaluation dimension. The quality feature information includes feature descriptions and scoring data for five dimensions: appearance, aroma, taste, liquor color, and leaf residue. For each dimension of quality feature, its corresponding feature response curve is extracted. The feature response curve is the curve showing the change of spectral reflectance with wavelength corresponding to that quality feature. For example, for the "fragrant and lasting" feature in the aroma dimension, the spectral reflectance curve with a wavelength range of 600-900nm is extracted as its feature response curve. The fluctuation period of the feature response curve is analyzed. The peak point in the curve is identified using a peak detection method, and the wavelength interval between adjacent peak points is calculated as the fluctuation period. For example, if the response curve of a certain aroma feature has peaks at 650nm, 725nm, and 800nm, then the fluctuation period is 75nm. The inflection point position refers to the position where the curve slope changes significantly; the point with the largest rate of change of the curve slope is used as the inflection point. For example, by calculating the first and second differences of the curve, inflection points are determined to exist at 678 nm and 763 nm.

[0133] The rate of change of the characteristic response curve at the inflection point is calculated, and the ratio of the rate of change to the fluctuation period is determined as the weight adjustment factor. Dynamic feature weights are then generated based on this weight adjustment factor. The rate of change refers to the slope of the curve at the inflection point, calculated by dividing the change in reflectance within a certain wavelength range before and after the inflection point by the change in wavelength. For example, at the inflection point of 678nm, the change in reflectance within a 5nm range before and after the inflection point is 0.12, while the change in wavelength is 10nm, resulting in a rate of change of 0.012. The ratio of the rate of change to the fluctuation period is used as the weight adjustment factor: 0.012 ÷ 75 = 0.00016. The dynamic feature weights are obtained by adjusting the initial feature weights based on the weight adjustment factor. The initial feature weights are derived from the node weights corresponding to the quality features in the expert review knowledge graph. For example, the initial weight for the aroma dimension is 0.25, which, after multiplying by the weight adjustment factor 1.08 (calculated by combining the adjustment factors of multiple inflection points), yields a dynamic feature weight of 0.27.

[0134] Quality characteristic information is mapped and transformed based on dynamic feature weights to generate a sequence of process control rules. The mapping and transformation process is based on a pre-established correlation model between quality characteristics and process parameters. This correlation model, constructed using extensive historical production data and expert experience, describes the quantitative relationship between quality characteristics and tea-making process parameters. For example, the "fragrant and lasting aroma" characteristic is related to process parameters such as fixation temperature, fixation time, and rolling intensity. Based on the dynamic feature weights, the degree of influence of each process parameter is determined, generating process control rules. For example, for the aroma characteristic with a dynamic feature weight of 0.27, the generated process control rule is "fixation temperature controlled at 220-230℃, fixation time controlled at 4-5 minutes, and rolling intensity maintained at a moderate level." After mapping and transforming multiple quality characteristics, a series of process control rules are formed, constituting a process control rule sequence.

[0135] The execution sensitivity of rules in the process control rule sequence is calculated. Based on this execution sensitivity, the coupling relationship between rules is identified, and the rule constraint strength is calculated based on the coupling relationship. Execution sensitivity refers to the degree to which changes in process parameters affect quality characteristics. It is calculated by dividing the change in quality characteristics by the change in process parameters. For example, if a 10°C change in fixation temperature results in an average change of 5 points in aroma score, then the execution sensitivity of fixation temperature to aroma is 0.5. The coupling relationship between rules refers to the mutual influence between different process control rules. This is identified by analyzing the joint change patterns of quality characteristics when multiple process parameters change simultaneously. For example, an increase in fixation temperature combined with a prolonged fixation time has a different effect on aroma than when the parameters change individually, indicating a coupling relationship between these two parameters. Rule constraint strength is used to quantify the strength of the coupling relationship. It is calculated by dividing the joint sensitivity of the coupled parameters by the sum of their individual sensitivities. For example, if the joint sensitivity of fixation temperature and fixation time is 0.8, and the sum of their individual sensitivities is 0.5 + 0.4 = 0.9, then the rule constraint strength is 0.8 ÷ 0.9 = 0.89.

[0136] Constraints are constructed by combining rule constraint strength and quality score, and dynamic programming is used to determine the execution priority of the process control rule sequence. The constraints consist of two parts: a rule constraint strength matrix and a quality score vector. The rule constraint strength matrix describes the mutual constraints between rules, and the quality score vector describes the importance of the corresponding quality characteristics of each rule. A dynamic programming algorithm is used to maximize the overall quality score, considering rule constraints to determine the execution priority of the process control rules. The dynamic programming process is conducted in stages, with each stage corresponding to the selection of one process control rule. The state transition equation considers the quality contribution of the current rule and its constraint influence with the selected rules. For example, for a rule sequence containing three process steps—fixing, rolling, and drying—the determined execution priority after dynamic programming is: fixing temperature adjustment 0.85 → rolling intensity adjustment 0.72 → drying time adjustment 0.63 → fixing time adjustment 0.58 → drying temperature adjustment 0.50.

[0137] The adjustment direction of each rule in the process control rule sequence is extracted. The product of the adjustment direction and the execution priority is determined as the parameter adjustment amount. A process parameter adjustment scheme is generated based on the parameter adjustment amount. The adjustment direction refers to whether the process parameter needs to be increased or decreased, which is determined by the difference between the quality characteristics and the target value. For example, if the current aroma score is lower than the target value, and the aroma is positively correlated with the fixation temperature, then the adjustment direction of the fixation temperature is to increase it, quantified as +1; if it is negatively correlated with the fixation time, then the adjustment direction of the fixation time is to decrease it, quantified as -1. The parameter adjustment amount is equal to the product of the adjustment direction and the execution priority, multiplied by the baseline adjustment step size. For example, if the adjustment direction of the fixation temperature is +1, the execution priority is 0.85, and the baseline adjustment step size is 5℃, then the parameter adjustment amount is +1 × 0.85 × 5 = +4.25℃. Based on the adjustment amount of each process parameter, a complete process parameter adjustment scheme is generated, such as "increase the fixation temperature by 4℃, decrease the fixation time by 30s, increase the rolling intensity by 10%, decrease the drying temperature by 2℃, and extend the drying time by 2 minutes".

[0138] By organically combining expert evaluation knowledge with tea production processes, a mapping mechanism between quality characteristics and process parameters was established, providing a precise process control scheme for tea production. Through dynamic feature weight generation and rule priority determination, this method can adaptively adjust process parameters based on quality evaluation results, effectively improving tea quality.

[0139] Furthermore, the rate of change of the characteristic response curve at the inflection point is calculated, and the ratio of the rate of change to the fluctuation period is determined as the weight adjustment factor. Dynamic characteristic weights are generated based on the weight adjustment factor, including:

[0140] Obtain the inflection point sequence of the characteristic response curve, determine the sampling window size based on the distribution density of the inflection point sequence, and calculate the rate of change of the characteristic response curve within the sampling window;

[0141] Perform time-frequency analysis on the characteristic response curve to identify the main frequency components, and determine the fluctuation period based on the main frequency components;

[0142] The ratio of the rate of change to the fluctuation period is determined as the weight adjustment factor. A mapping function is established based on the numerical distribution of the weight adjustment factor, and the weight adjustment factor is converted into dynamic feature weights through the mapping function.

[0143] Obtaining the inflection point sequence of the characteristic response curve is the first step in generating dynamic feature weights. The characteristic response curve refers to the curve showing the change in spectral reflectance of a tea sample with wavelength at a specific quality dimension. Taking aroma as an example, the characteristic response curve of a high-quality green tea exhibits multiple peaks and troughs in the near-infrared and visible light wavelength ranges. An inflection point is a location where the curve slope changes significantly, indicating a key turning point in the spectral characteristics. The inflection point sequence is obtained using the curve slope change rate analysis method. Specifically, the first and second differences of the response curve are calculated. When the absolute value of the second difference exceeds a preset threshold, the corresponding position is marked as an inflection point. For example, the aroma characteristic response curve of a certain green tea sample detects inflection points at multiple wavelength positions, forming a complete inflection point sequence. These inflection points often correspond to the spectral characteristics of key chemical components in the tea.

[0144] The sampling window size is determined based on the distribution density of the inflection point sequence, and the rate of change of the characteristic response curve is calculated within the sampling window. The inflection point distribution density refers to the number of inflection points per unit wavelength range, calculated by dividing the total number of inflection points by the wavelength range width. The sampling window size is inversely proportional to the inflection point distribution density and is determined by dividing the baseline window size by the product of the inflection point distribution density and the adjustment coefficient. For ease of practical operation, the sampling window size is typically set to tens of nanometers. The rate of change of the characteristic response curve is calculated within the determined sampling window; the rate of change is the ratio of the maximum change in reflectance within the window to the window width. For each inflection point in the inflection point sequence, the rate of change within its corresponding sampling window is calculated, forming a rate of change sequence. These rate of change values ​​reflect the spectral sensitivity of the tea sample at different wavelengths; positions with larger rates of change typically correspond to key wavelengths for tea quality.

[0145] Time-frequency analysis is performed on the characteristic response curve to identify the main frequency components, and the fluctuation period is determined based on these main frequency components. The time-frequency analysis uses the Discrete Fourier Transform (DFT) method to transform the characteristic response curve from the wavelength domain to the frequency domain, obtaining a spectrum. The spectrum shows the intensity distribution of each frequency component in the curve; the larger the frequency value, the more intense the curve fluctuation. By analyzing the peak values ​​in the spectrum, frequency components with significant intensity are identified, and these are considered the main frequency components. The fluctuation period is the reciprocal of the frequency; for each identified main frequency component, its corresponding fluctuation period is calculated. The combined influence of multiple main frequency components also needs to be considered; a weighted average method can be used to calculate the comprehensive fluctuation period, with the weights being the relative intensities of each frequency component. The final comprehensive fluctuation period reflects the overall variation law of the tea's spectral characteristics and is an important parameter for subsequent weight adjustment.

[0146] The ratio of the rate of change to the fluctuation period is determined as the weight adjustment factor. A mapping function is established based on the numerical distribution of the weight adjustment factor, and this function converts the weight adjustment factor into dynamic feature weights. For each inflection point in the inflection point sequence, the ratio of its rate of change to the fluctuation period is calculated as the weight adjustment factor. This calculation is performed on all inflection points to obtain the weight adjustment factor sequence. The numerical distribution of the weight adjustment factor reflects the importance of the feature response curve at different wavelength positions; a larger value indicates a more significant impact of that position on the quality characteristic. A mapping function is established based on the numerical distribution of the weight adjustment factor, which converts the weight adjustment factor into dynamic feature weights. The mapping function adopts a piecewise linear function form, dividing the numerical range of the weight adjustment factor into multiple intervals, each interval corresponding to a different slope and intercept. Through the mapping function, the weight adjustment factor sequence is converted into a dynamic feature weight sequence. The final dynamic feature weight is the maximum value in the sequence, which serves as the basis for adjusting the weight of that quality characteristic in the knowledge graph. This dynamic weight adjustment mechanism can adaptively adjust the evaluation weights according to the actual spectral characteristics of the tea sample, making the evaluation results more objective and accurate.

[0147] This implementation method achieves adaptive generation of dynamic feature weights through refined analysis of feature response curves, solving the evaluation bias problem caused by fixed weights in traditional tea quality evaluation. It fully considers the fluctuation patterns and trends of tea spectral characteristics, accurately capturing key information about quality characteristics through inflection point sequence analysis and time-frequency characteristic extraction. The established mapping function realizes a scientific conversion from spectral characteristics to weight values, providing objective basis and theoretical support for weight adjustment.

[0148] like Figure 3 As shown, Figure 3 A schematic diagram of a knowledge graph-based tea evaluation expert decision support system provided in this embodiment of the invention, the system comprising:

[0149] The sample processing module is used to acquire tea sample data and perform standardization processing to obtain standardized features.

[0150] The knowledge construction module is used to collect data on the review process of tea samples by multiple experts, calculate the consistency index of expert scores, select expert review rules based on the consistency index, and construct an expert decision knowledge graph.

[0151] The feature mapping module is used to analyze the distribution pattern of expert scores to obtain rule mapping weights, map standardized features to the expert decision knowledge graph according to the rule mapping weights, adjust the information transmission probability according to the rule mapping weights, and generate feature transmission paths.

[0152] The scoring generation module is used to calculate the rule applicability and rule conflict degree on the feature transmission path, correct the transmission results based on the rule applicability and rule conflict degree, and obtain the quality score and feature analysis map of the tea sample.

[0153] The control rules module is used to convert feature analysis diagrams into process control rule sequences, determine the execution priority of the process control rule sequences based on quality scores, and generate process parameter adjustment schemes according to the rule execution priority.

[0154] The specific embodiments described above are preferred embodiments of the present invention and are not intended to limit the specific scope of the present invention. The scope of the present invention includes, but is not limited to, these specific embodiments. All equivalent changes made in accordance with the shape and structure of the present invention are within the protection scope of the present invention.

Claims

1. A knowledge graph-based decision support method for tea evaluation experts, characterized in that, Includes the following steps: Acquire tea sample data and perform standardization processing to obtain standardized features; Data on the evaluation process of tea samples by multiple experts were collected, the consistency index of expert scores was calculated, expert evaluation rules were selected based on the consistency index, and an expert decision-making knowledge graph was constructed. The distribution pattern of expert scores is analyzed to obtain the rule mapping weights. The standardized features are then mapped to the expert decision knowledge graph according to the rule mapping weights. The information transmission probability is adjusted according to the rule mapping weights to generate feature transmission paths. Calculate the rule applicability and rule conflict degree along the feature transfer path, correct the transfer results based on the rule applicability and rule conflict degree, and obtain the quality score and feature analysis map of the tea sample; The feature analysis diagram is converted into a process control rule sequence. The execution priority of the process control rule sequence is determined based on the quality score. A process parameter adjustment plan is generated according to the rule execution priority. Historical sample data evaluated by experts is used as training samples. The Euclidean distance between the node rule features and the training sample features on the feature transfer path is calculated to obtain the rule applicability on the feature transfer path. Calculate the semantic similarity of rule vectors of adjacent nodes on the feature transmission path, and normalize the semantic similarity to obtain the rule conflict degree on the feature transmission path; An initial weight matrix is ​​constructed using the rule applicability as the diagonal element of the node. The rule conflict degree is normalized and used as the weight decay coefficient. The weight is then calculated based on the product of the initial weight matrix and the weight decay coefficient. An exponential correction function is constructed based on the transmission weight to adaptively adjust the scores of each node in the transmission path, thereby obtaining the corrected transmission result. The corrected transmission results are fused according to the transmission weights to obtain the quality score of the tea sample, and a feature analysis map is generated based on the correction process of the node scores. Obtain quality feature information from the feature analysis map, extract the feature response curve corresponding to the quality feature information, and analyze the fluctuation period and inflection point position of the feature response curve. Calculate the rate of change of the characteristic response curve at the inflection point, determine the ratio of the rate of change to the fluctuation period as the weight adjustment factor, and generate dynamic characteristic weights based on the weight adjustment factor. The quality feature information is mapped and transformed based on the dynamic feature weights to generate a process control rule sequence; The execution sensitivity of rules in the process control rule sequence is calculated, the coupling relationship between rules is identified based on the execution sensitivity, and the constraint strength of rules is calculated based on the coupling relationship. Constraints are constructed by combining rule constraint strength with quality score, and dynamic programming is used to determine the execution priority of the process control rule sequence. Extract the adjustment direction of each rule in the process control rule sequence, determine the parameter adjustment amount by multiplying the adjustment direction by the execution priority, and generate a process parameter adjustment scheme based on the parameter adjustment amount.

2. The method according to claim 1, characterized in that, The standardized features obtained from acquiring tea sample data and performing standardization processing include: The spectral reflectance of tea samples was collected under different illumination angles, and the variation trend of reflectance in adjacent bands was calculated to obtain the variation curve. Extract the inflection point of the change curve, map the reflectance at the inflection point to the manual score to obtain the correction coefficient, and correct the spectral reflectance data according to the correction coefficient; The feature enhancement factor is obtained by extracting the ratio of the peak range to the steady-state range of the sensor response signal, and the corrected reflectivity data is enhanced based on the feature enhancement factor. Wavelet decomposition is performed on the enhanced reflectivity data, and the scale coefficient with the highest energy is selected for reconstruction. The reconstructed feature data is then normalized to obtain standardized features.

3. The method according to claim 1, characterized in that, Data from the review process of multiple experts on tea samples was collected, the consistency index of expert scores was calculated, expert review rules were selected based on the consistency index, and an expert decision-making knowledge graph was constructed, including: Behavioral and scoring data of multiple experts reviewing tea samples were collected. Expert attention area sequences were extracted from the behavioral data, and expert evaluation features were extracted from the scoring data. Identify the stage boundaries of the review process based on changes in the expert attention area sequence, segment the review process based on the stage boundaries, and obtain the expert review rules for each segment. The scoring consistency index is obtained by calculating the dispersion of the scoring data of different experts within each segment. Based on the scoring consistency index, the scoring consistency segments are screened, and the expert review rules corresponding to the scoring consistency segments are obtained. The expert evaluation characteristics within the consistent scoring segment are mapped to the expert review rules to obtain the review mapping rules. The frequency of occurrence of the review mapping rules in the consistent scoring segment is calculated to obtain the rule credibility. The rule correlation is obtained based on the transformation relationship of the consistent scoring segment. By using review mapping rules as nodes, rule credibility as node weight, and rule relevance as node connection strength, an expert decision-making knowledge graph is constructed.

4. The method according to claim 1, characterized in that, The distribution patterns of expert ratings are analyzed to obtain rule-mapping weights. Standardized features are then mapped onto the expert decision-making knowledge graph according to these weights. Information transmission probabilities are adjusted based on these weights to generate feature transmission paths, including: The frequency distribution of expert ratings in different rating intervals is statistically analyzed and normalized to obtain the rating distribution pattern. Based on the rating distribution pattern, the conditional probability of adjacent rating intervals is calculated to obtain the rule mapping weight. Standardized features are mapped to hierarchical nodes of the expert decision knowledge graph according to rule mapping weights. The rule mapping weights guide the distribution ratio of standardized features at different hierarchical nodes, resulting in a feature node distribution matrix. Based on the rule mapping weights and feature node distribution matrix, the connection relationships between nodes in the expert decision knowledge graph are weighted and adjusted to obtain the information transmission probability matrix between nodes. A state transition matrix is ​​constructed based on node distribution and information transmission probability. The state transition sequence of features in the expert decision knowledge graph is calculated, and the sequence with the highest probability in the state transition sequence is determined as the feature transmission path.

5. The method according to claim 1, characterized in that, Calculate the rate of change of the characteristic response curve at the inflection point, and determine the ratio of the rate of change to the fluctuation period as the weight adjustment factor. Based on the weight adjustment factor, generate dynamic characteristic weights, including: Obtain the inflection point sequence of the characteristic response curve, determine the sampling window size based on the distribution density of the inflection point sequence, and calculate the rate of change of the characteristic response curve within the sampling window; Perform time-frequency analysis on the characteristic response curve to identify the main frequency components, and determine the fluctuation period based on the main frequency components; The ratio of the rate of change to the fluctuation period is determined as the weight adjustment factor. A mapping function is established based on the numerical distribution of the weight adjustment factor, and the weight adjustment factor is converted into dynamic feature weights through the mapping function.

6. A knowledge graph-based tea evaluation expert decision support system, used to implement the method described in any one of claims 1-5, characterized in that, The system includes: The sample processing module is used to acquire tea sample data and perform standardization processing to obtain standardized features. The knowledge construction module is used to collect data on the review process of tea samples by multiple experts, calculate the consistency index of expert scores, select expert review rules based on the consistency index, and construct an expert decision knowledge graph. The feature mapping module is used to analyze the distribution pattern of expert scores to obtain rule mapping weights, map standardized features to the expert decision knowledge graph according to the rule mapping weights, adjust the information transmission probability according to the rule mapping weights, and generate feature transmission paths. The scoring generation module is used to calculate the rule applicability and rule conflict degree on the feature transmission path, correct the transmission results based on the rule applicability and rule conflict degree, and obtain the quality score and feature analysis map of the tea sample. The control rules module is used to convert feature analysis diagrams into process control rule sequences, determine the execution priority of the process control rule sequences based on quality scores, and generate process parameter adjustment schemes according to the rule execution priority.

Images