A knowledge graph-based and large-scale model reasoning-based intelligent prediction system for A-level blue value

Abstract

This application relates to an intelligent prediction system for methylene blue value based on knowledge graphs and large-scale model reasoning. The system includes a data acquisition module, a multi-source feature construction module, a knowledge graph construction module, a multimodal unified coding module, a graph-enhanced reasoning module, a prediction and evaluation module, an interpretation and generation module, and a re-inspection decision module. The system collects data such as aggregate particle size distribution, particle morphology, stone powder content, mud powder composition, parent rock type, water absorption rate, titration halo image, and experimental conditions. It constructs a knowledge graph of "aggregate characteristics—experimental process—halo characteristics—methylene blue value—quality grade," and combines similar case retrieval and large-scale model reasoning to achieve methylene blue value interval prediction, endpoint sensitivity assessment, stability analysis, quality grade determination, and re-inspection suggestion output. This invention can improve detection efficiency, consistency, and result interpretability.

Description

Technical Field

[0001] This application relates to the field of building material testing technology, and in particular to an intelligent prediction system for Methylene Blue Value based on knowledge graphs and large model reasoning. Background Technology

[0002] Methylene Blue Value (MBV) is an important indicator for evaluating the clay content and activity of fine aggregates. It is widely used in the evaluation of manufactured sand quality, control of concrete raw materials, and quality testing of highway engineering construction. Current testing methods mainly rely on relevant standards, determining the endpoint of the test and calculating the Methylene Blue Value by titrating a methylene blue solution and observing the diffusion of the color halo on filter paper.

[0003] In practical engineering applications, the methylene blue test typically involves multiple steps, including sample preparation, solution titration, filter paper diffusion observation, and endpoint determination. In existing technologies, endpoint determination mainly relies on manual observation of whether a distinct blue ring forms at the boundary of the filter paper halo. This process is highly experience-dependent and easily affected by factors such as operator habits, differences in visual judgment, and changes in environmental conditions. Furthermore, aggregates from different sources exhibit significant differences in particle size distribution, particle morphology, stone powder content, mud mineral composition, and water absorption rate. These factors all influence the adsorption behavior and diffusion process of methylene blue, leading to increased dispersion in the test results.

[0004] Furthermore, existing methylene blue testing methods typically require a complete titration process to obtain results, lacking the ability to predict the methylene blue value of the sample beforehand. They also struggle to provide supplementary assessments of endpoint sensitivity and result stability during the testing process. When samples exhibit complex mud-powder composition, significant differences in particle surface characteristics, or large fluctuations in environmental conditions, multiple repeated tests are often necessary to obtain reliable results, leading to extended testing cycles and reduced efficiency.

[0005] In recent years, some studies have attempted to identify filter paper color halos using image processing or machine vision methods to aid in endpoint determination. However, these methods typically focus on single image information, failing to adequately utilize multi-source information such as aggregate physical properties, experimental conditions, and historical test data, making it difficult to establish a systematic correlation between sample characteristics and methylene blue values. Furthermore, existing methods generally lack interpretability analysis of test results, failing to clearly identify the key factors influencing methylene blue value changes and providing a basis for determining whether retesting is necessary.

[0006] Furthermore, in the existing technical system, the relationship between aggregate characteristics, test process parameters, color halo characteristics and test results is mostly scattered in the form of experience, and a unified structured knowledge expression method has not yet been formed. There is also a lack of technical means for reasoning and analysis based on knowledge association, which makes it difficult to accumulate and reuse test experience.

[0007] Therefore, the existing technology has at least the following problems:

[0008] 1. The methylene blue test relies on manual judgment, and the results are greatly affected by subjective factors, resulting in insufficient consistency;

[0009] 2. The testing process is mainly based on post-event judgment, lacking the ability to predict test results in advance;

[0010] 3. The multiple influencing factors have not been systematically integrated, making it difficult to accurately reflect the intrinsic relationship between sample characteristics and methylene blue value;

[0011] 4. Lack of quantitative assessment methods for the stability of test results and the sensitivity of endpoints;

[0012] 5. The test results lack interpretability, making it difficult to provide an effective basis for decision-making in engineering quality control.

[0013] To address the aforementioned issues, we have designed an intelligent prediction system for the blue value of the sub-blue spectrum based on knowledge graphs and large-scale model reasoning. Summary of the Invention

[0014] This application provides an intelligent prediction system for methylene blue values ​​based on knowledge graphs and large model reasoning, in order to solve the problems in related technologies, such as the reliance on human experience in determining the endpoint of methylene blue tests, the significant impact of sample differences and operational factors on test results, the lack of effective fusion of multi-source information, the lack of pre-prediction capabilities for test results, and the lack of stability assessment and interpretable analysis methods.

[0015] Firstly, it provides an intelligent prediction system for the blue value of sub-blue based on knowledge graphs and large-scale model reasoning, including:

[0016] The data acquisition module is used to acquire multi-source data of the entire methylene blue test process, and to perform time alignment of structured parameter data, image data and test process data based on timestamps to construct sample data units;

[0017] The multi-source feature construction module is used to standardize the multi-source data, construct structured feature vectors, and extract geometric features and temporal variation features of color halo diffusion based on image sequences to form a multimodal feature set.

[0018] The knowledge graph construction module is used to construct a graph structure based on the multimodal feature set, which includes sample nodes, aggregate characteristic nodes, test condition nodes, color halo feature nodes and methylene blue value nodes, and to establish semantic relationships between nodes to form sample-level subgraphs.

[0019] The multimodal unified coding module is used to vectorize structured features, image features, and knowledge graph subgraphs, and generate unified feature representations through fusion operations.

[0020] The graph-enhanced reasoning module is used to calculate the similarity with historical samples based on the unified feature representation, extract a set of similar cases, extract semantic paths from the knowledge graph, and input the unified feature representation, the set of similar cases, and the semantic paths into the large model for joint reasoning to obtain the sub-blue value prediction result.

[0021] The prediction and evaluation module is used to convert the inference results into methylene blue value prediction intervals, endpoint sensitivity indicators, stability indicators, and quality levels.

[0022] The explanation generation module is used to generate explanatory information for prediction results based on feature contribution analysis, map paths, and similar cases.

[0023] The re-inspection decision module is used to construct re-inspection criteria based on prediction confidence, endpoint sensitivity, and stability, and output the decision result of whether to re-inspect.

[0024] In some embodiments, the data acquisition module includes a structured parameter acquisition unit, an image acquisition unit, and an experimental process recording unit. The structured parameter acquisition unit is used to acquire aggregate particle size distribution, particle morphology parameters, stone powder content, mud powder mineral composition, parent rock type, and water absorption rate, and stores them through standardized fields. The image acquisition unit includes an industrial camera, a constant light source module, and a trigger control unit. The trigger control unit triggers image acquisition according to the titration time interval or titration event to form a time-series image sequence of filter paper color halo diffusion, and achieves image consistency by fixing the shooting angle, spatial scale, and focal length constraints. The experimental process recording unit is used to record the single titration volume, cumulative titration amount, number of titrations, stirring time, settling time, and ambient temperature and humidity, and records them synchronously with the image data, thereby forming a unified acquisition system for multi-source heterogeneous data.

[0025] In some embodiments, the data acquisition module performs synchronous processing of multi-source data based on a unified time calibration mechanism. By adding timestamps to structured parameters, image sequences, and experimental process data, a unified timeline is constructed, and data from different sources are mapped to the same time dimension to form a unified data representation. ;in, This represents the static structured parameters of the sample. This represents the filter paper image captured at the corresponding time. This represents the experimental process parameters at that moment. This time synchronization mechanism enables a precise correspondence between image changes and titration operations, and ensures the alignment consistency between multimodal data during subsequent feature construction.

[0026] In some embodiments, the multi-source feature construction module normalizes and encodes the structured parameters and constructs feature vectors:

[0027] ;

[0028] Simultaneously, the filter paper halo image sequence was subjected to grayscale conversion, noise reduction, and boundary enhancement. Geometric and grayscale features of the halo region were extracted, including the halo diameter. Area of ​​color halo and boundary grayscale gradient And calculate dynamic change characteristics based on time series:

[0029] ;

[0030] This results in a time-series feature set that can characterize the adsorption and diffusion process of methylene blue, and the structural features, image features, and experimental process features are fused to obtain a unified multi-source feature expression.

[0031] In some embodiments, the knowledge graph construction module constructs a graph structure as follows: Among them, the set of nodes This includes nodes for sample type, particle size distribution, particle morphology, stone powder content, mud powder composition, parent rock type, water absorption rate, test conditions, color halo characteristics, and methylene blue value, as well as edge sets. Represents the semantic relationships between nodes, including relationship types such as "has parameters", "contains components", "generates features", and "corresponds to results", and attribute sets. Numerical attributes used to describe nodes and relationships are used to realize a structured expression of sample characteristics, experimental behavior and test results through the above structure, and are stored and managed based on a graph database.

[0032] In some embodiments, the knowledge graph construction module constructs a semantic path from aggregate properties to detection results within a graph structure: ;in, Indicates particle size distribution. Indicates particle morphology, Indicates the composition of clay powder. Indicates color halo characteristics, Indicates the Asian League Blue Value. It represents the quality level, describes the causal relationship between material properties, test behavior and test results through this path, and participates in path development and result generation as a constraint in the reasoning process.

[0033] In some embodiments, the multimodal unified coding module performs vectorized encoding on structured features, image features, and knowledge graph subgraphs, respectively, wherein the structured features are encoded through a multi-layer mapping function. Image features are encoded using a temporal coding function. Encoding, knowledge graph subgraphs via graph embedding functions Encode the data and generate a unified feature representation through vector concatenation or weighted fusion:

[0034] ;

[0035] During the fusion process, the features of different modalities are normalized, and the contribution ratio of each modality is adjusted by weighting, thereby achieving a unified semantic expression of multi-source data.

[0036] In some embodiments, the graph-enhanced inference module constructs a similarity function based on a unified feature representation:

[0037] ;

[0038] Furthermore, several samples with the highest similarity are selected from the historical sample library as reference cases. At the same time, multi-hop semantic paths are extracted from the knowledge graph. The unified feature representation, reference cases, and semantic paths are jointly input into the large model for joint reasoning. During the reasoning process, the reasoning space is restricted through the path constraint mechanism, so that the generated results satisfy the semantic relationships and rule constraints in the knowledge graph, thereby improving the accuracy and stability of the prediction results.

[0039] In some embodiments, the prediction and evaluation module quantifies the inference results to construct a prediction range for the sub-blue value:

[0040] ;

[0041] And based on the color halo diffusion characteristics, an endpoint sensitivity index was constructed: ;

[0042] And stability indicators: ;

[0043] At the same time, the samples are classified into quality grades according to the testing standards, and the prediction confidence level is calculated to form a multi-dimensional evaluation index system.

[0044] In some embodiments, the re-inspection decision module constructs re-inspection criteria based on the methylene blue value range, endpoint sensitivity, stability, and prediction confidence output by the prediction evaluation module: ;

[0045] When the criterion exceeds the preset threshold, a re-examination suggestion is output, and the reasons for triggering the re-examination are analyzed to generate a set of reasons, including insufficient prediction confidence, insignificant changes in the endpoint, and discrete sample distribution. At the same time, operation optimization suggestions are output, including adjusting the titration step size or increasing the image sampling frequency, so as to achieve closed-loop quality control of the detection results.

[0046] The beneficial effects of the technical solution provided in this application include:

[0047] 1. This invention transforms the detection of methylene blue from traditional post-event judgment to pre-event prediction by introducing multi-source data fusion and large model inference mechanism. By jointly modeling aggregate particle size distribution, stone powder content, mud powder composition and color halo diffusion characteristics, the range of methylene blue values ​​can be predicted before the test begins, thereby effectively reducing the blindness of the test, reducing unnecessary repeated tests, significantly shortening the overall detection cycle and improving detection efficiency.

[0048] 2. This invention constructs an endpoint sensitivity evaluation index by performing time-series modeling on the color halo diffusion behavior during titration, thereby achieving a quantitative analysis of the difficulty in determining the experimental endpoint. Compared with the traditional method that relies on manual experience observation, this technology can objectively reflect the color halo change trend, effectively reduce subjective differences between different testing personnel, and improve the consistency and repeatability of test results.

[0049] 3. This invention constructs a multimodal feature space by uniformly encoding and fusing structured parameters, image features, and experimental process data, thereby realizing comprehensive modeling of factors affecting methylene blue value. Compared with existing methods based on only a single image or single parameter, this method can comprehensively reflect the coupling relationship between material properties and experimental behavior, thus significantly improving prediction accuracy and model generalization ability.

[0050] 4. This invention constructs a knowledge graph for methylene blue detection, which expresses the relationship between sample attributes, test conditions, color halo characteristics and test results in a structured way. This allows the detection knowledge that originally relied on experience to be systematically stored and managed. This technology enables the continuous accumulation and reuse of detection experience, and gives the system the ability to evolve knowledge, continuously optimizing prediction performance as data increases.

[0051] 5. This invention introduces knowledge graph path constraints during the reasoning process, which restricts the reasoning of large models to a semantic space that conforms to domain knowledge, forming a joint reasoning mechanism of "data-driven + knowledge constraint". Compared with pure data-driven models, this method can effectively avoid misjudgment problems caused by abnormal samples or data bias, and improve the stability and credibility of prediction results.

[0052] 6. By constructing an interpretation generation and re-examination decision mechanism, this invention realizes the transformation of test results from "numerical output" to "interpretable decision". The system can automatically output the key factors affecting the methylene blue value, reasoning path and similar cases, and generate re-examination suggestions based on endpoint sensitivity and stability indicators, thereby realizing closed-loop control of test results and improving the scientific nature and engineering application value of the test work. Attached Figure Description

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

[0054] Figure 1 This is a flowchart illustrating the overall architecture of the intelligent prediction system for sub-blue light value provided in this embodiment of the invention.

[0055] Figure 2 This is a flowchart of multi-source data acquisition and time alignment provided in an embodiment of the present invention;

[0056] Figure 3 A flowchart for constructing multi-source features provided in an embodiment of the present invention;

[0057] Figure 4 A flowchart for knowledge graph construction provided in an embodiment of the present invention;

[0058] Figure 5 A flowchart of graph-enhanced reasoning provided in an embodiment of the present invention;

[0059] Figure 6 This is a schematic diagram of the knowledge graph structure for methylene blue detection provided in an embodiment of the present invention;

[0060] Figure 7 This is a schematic diagram of multimodal unified coding and computing power allocation provided in an embodiment of the present invention;

[0061] Figure 8 The endpoint sensitivity and color halo diffusion change data diagram provided in the embodiments of the present invention;

[0062] Figure 9 This is an example diagram showing the methylene blue value prediction results and stability evaluation provided in an embodiment of the present invention;

[0063] Figure 10 This is a statistical chart of the methylene blue value prediction interval provided in an embodiment of the present invention;

[0064] Figure 11 A graph showing the relationship between endpoint sensitivity and color halo change rate provided in an embodiment of the present invention;

[0065] Figure 12 Stability indices and dispersion statistics of similar cases provided for embodiments of the present invention;

[0066] Figure 13 This is a statistical chart showing the ranking of feature contributions provided in an embodiment of the present invention.

[0067] Figure 14 A statistical chart of model inference resource consumption and response time provided for embodiments of the present invention. Detailed Implementation

[0068] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0069] This application provides an intelligent prediction system for methylene blue values ​​based on knowledge graphs and large model reasoning. It can solve the problems in related technologies, such as the reliance on human experience to determine the endpoint of methylene blue tests, the significant impact of sample differences and operational factors on test results, the lack of effective fusion of multi-source information, the lack of pre-prediction capabilities for test results, and the lack of stability assessment and interpretable analysis methods.

[0070] Please see Figures 1-14 This invention proposes an intelligent prediction system for sub-blue values ​​based on knowledge graphs and large-scale model reasoning. The system employs an eight-module collaborative architecture: data acquisition, feature construction, knowledge modeling, multimodal coding, graph-enhanced reasoning, prediction evaluation, interpretation generation, and review decision-making. This forms a closed-loop processing flow from raw data to intelligent decision-making, specifically including:

[0071] 1. Data Acquisition Module

[0072] The data acquisition module is used to acquire multi-source raw data throughout the methylene blue experiment. Through unified time calibration and structured organization, it achieves precise alignment between structured parameter data, image data and experimental process data, providing consistent input for subsequent feature construction and inference analysis.

[0073] The data acquisition module constructs a multi-source data acquisition system through laboratory testing equipment interfaces, industrial vision acquisition systems, and test recording terminals. Structured parameters include aggregate particle size distribution (sieve curves and passing rates for each particle size), particle morphology parameters (roundness, angularity, and surface roughness), stone powder content, mudstone mineral composition, parent rock type, and water absorption rate. This type of data is automatically acquired through the experimental equipment interface or entered by the operating terminal and uniformly converted into standardized field formats for storage.

[0074] During the titration experiment, the diffusion process of the filter paper's color halo was continuously captured using an industrial camera. The image acquisition system included an industrial camera, a constant light source module, and a trigger control unit. The constant light source was used to eliminate the influence of ambient light fluctuations on image quality. The trigger control unit triggered image acquisition based on titration time intervals or titration events (such as the completion of a single titration), thereby forming a set of image data with time-series characteristics. The acquired images met the conditions of uniform shooting angle, fixed spatial scale, and stable focal length, ensuring that images acquired at different times were within a consistent spatial reference frame.

[0075] Meanwhile, key operating parameters and environmental information were recorded during the experiment, including single titration volume, cumulative titration amount, number of titrations, stirring time, settling time, and ambient temperature and humidity. This type of information was obtained in real time through sensor interfaces or manual recording.

[0076] To address the inconsistency of multi-source data in the time dimension, this module introduces a unified time calibration mechanism, which adds timestamps to all collected data and constructs a unified timeline to synchronously map structured parameters, image sequences, and experimental process data.

[0077] For the One sample, at time The data is uniformly represented as: ;

[0078] in, This represents the static structured parameters of the sample. This represents the filter paper image captured at the corresponding time. This indicates the parameters of the test process at that moment.

[0079] This time alignment mechanism enables a precise correlation between image changes and experimental operations, providing a foundation for subsequent extraction of dynamic features of color halo diffusion.

[0080] At the data organization level, the aforementioned multi-source data are uniformly encapsulated into sample data units:

[0081] ;

[0082] in, This represents the methylene blue value obtained from historical testing.

[0083] This data structure centers on the sample, unifying the management of static attributes, dynamic processes, and historical results. It is organized using a hierarchical storage method, including a raw data layer (images and records), a feature data layer (generated through subsequent calculations), and a label data layer (detection results), thereby ensuring the traceability and consistency of data in subsequent modules.

[0084] Through the above methods, the data acquisition module not only completes the acquisition of multi-source information, but also realizes the time alignment and structural unification of cross-modal data, and establishes a conversion mechanism from raw experimental data to standardized input data, providing a stable and reliable data foundation for subsequent multimodal feature construction, knowledge graph modeling and large model inference.

[0085] 2. Multi-source feature construction module

[0086] The multi-source feature construction module is used to uniformly process and feature-express the multi-source raw data output by the data acquisition module. It transforms discrete structured parameters, time-series image data and experimental process records into a feature set with unified semantics and computability, thereby realizing the expression and association of different types of data in the same feature space.

[0087] In terms of structured parameter processing, data such as particle size distribution, particle morphology, stone powder content, mud powder composition, parent rock type, and water absorption rate of the samples were standardized and coded. For continuous variables, normalization or interval mapping was used to eliminate dimensional differences, while for categorical variables (such as parent rock type and mud powder type), discrete coding or embedded vector representation was used to construct a unified structured feature vector.

[0088] ;

[0089] in, Indicates particle size distribution characteristics, Indicates the morphological characteristics of the particles. Indicates the stone powder content, This represents a vector representing the composition of clay minerals. Indicates the code for the type of parent rock. Indicates water absorption rate. This indicates the characteristics of the test conditions.

[0090] In image data processing, specifically for time-series images First, image preprocessing is performed, including grayscale conversion, filtering and noise reduction, background correction and boundary enhancement, to improve the stability of color halo region recognition.

[0091] Subsequently, a threshold-based segmentation or edge detection method was used to extract the halo region, and its geometric and grayscale features, including the halo diameter, were calculated. Area of ​​color halo and boundary grayscale gradient .

[0092] Based on this, and combined with time series information, we can further calculate the dynamic change characteristics: ;

[0093] This forms a set of temporal features that can characterize the color halo diffusion process: ;

[0094] In terms of data processing during the experiment, time series reconstruction was performed on titration volume, number of titrations, stirring time, and environmental parameters, mapping them to a time axis consistent with the image sequence, and constructing a process feature sequence. It is used to characterize the coupling relationship between operational behavior and color halo diffusion.

[0095] To achieve a unified representation of multi-source information, this module fuses structured features, image features, and process features based on the time-synchronized dataset to construct a multimodal feature representation of the sample: ;

[0096] Furthermore, the aforementioned feature set undergoes dimensionality normalization and feature filtering. Redundant features are removed through correlation analysis or statistical filtering methods, while retaining a subset of key features that significantly influence the methylene blue value. Simultaneously, features at different scales are uniformly normalized to ensure the numerical stability of each feature in subsequent modeling processes.

[0097] Ultimately, this module outputs standardized multi-source feature representations, providing an input foundation for subsequent knowledge graph modeling and multimodal coding. Through the processing of this module, the transformation from "multi-source heterogeneous raw data" to "unified structured feature space" is realized, solving the problem of difficulty in directly integrating different data types, and providing high-quality feature support for MDMA blue value prediction.

[0098] 3. Knowledge Graph Construction Module

[0099] The knowledge graph construction module is used to perform structured modeling of the relationships between aggregate characteristics, test process parameters, color halo characteristics and test results involved in the methylene blue test. It transforms the implicit knowledge scattered in historical test data and empirical rules into explicit graph structure expressions, thereby providing computable and traceable knowledge support for the subsequent reasoning module.

[0100] This module constructs a knowledge graph for methylene blue detection based on the feature data output by the multi-source feature construction module: ;

[0101] In this set, node set 𝑉 represents the entity set, edge set 𝐸 represents the set of relationships between entities, and attribute set 𝐴 is used to describe the numerical characteristics of nodes and relationships.

[0102] The node types include at least sample nodes, particle size distribution nodes, particle morphology nodes, stone powder content nodes, mud powder composition nodes, parent rock type nodes, water absorption rate nodes, test condition nodes, color halo characteristic nodes, methylene blue value nodes, and quality grade nodes. Different types of nodes are connected by semantic relationships, which include "has parameters", "contains components", "produced under test conditions", "characteristic features", "corresponding values", and "judgment grade".

[0103] During the graph structure construction process, the sample node is taken as the center, and its corresponding structured feature vector is... Image feature set and characteristics of the test process Each node is mapped to an associated child node, and multi-layer semantic connections are established through edge relationships to form a sample-level subgraph structure.

[0104] For each sample 𝑖, its corresponding subplot can be represented as: ;

[0105] in, Includes the set of all entity nodes related to the sample. It represents the set of semantic relations within it.

[0106] Furthermore, this module constructs the core semantic path from aggregate properties to detection results: ;

[0107] in, Indicates particle size distribution. Indicates particle morphology, Indicates the composition of clay powder. Indicates color halo characteristics, Indicates the Asian League Blue Value. This indicates the quality level. This path is used to describe the causal chain from material properties to test results, providing a structured basis for subsequent reasoning.

[0108] In the process of graph construction, a rule constraint mechanism is introduced to transform testing standards, engineering experience and statistical laws into executable rules, which are then embedded into the graph in the form of attribute or relation constraints.

[0109] For example, when the content of stone powder and the proportion of clay minerals are high, their influence on the methylene blue value node can be enhanced by relational weights or rule markings; when the gradient change rate of the color halo boundary is low, the risk of decreased endpoint sensitivity can be identified by constraint relations.

[0110] The above rules are expressed in functional form as follows: ;

[0111] Each variable originates from the multi-source feature construction module, and the function... This indicates the coupling effect of multiple factors.

[0112] At the data storage level, this module uses a graph database to store nodes and relationships, and constructs node tables, relationship tables, and attribute tables. The node table records unique entity identifiers and attribute information, the relationship table records connections and relationship types between nodes, and the attribute table stores numerical features and statistical information. Furthermore, by establishing a mapping relationship between node indexes and feature vectors, the module achieves a connection between the knowledge graph and the vector space, providing support for subsequent retrieval and reasoning.

[0113] In addition, this module supports a dynamic update mechanism for knowledge graphs. When new sample data or test results are generated, new nodes and their relationships are inserted into the graph through incremental construction, and relevant statistical attributes and relationship weights are updated, thereby realizing the continuous evolution and adaptive expansion of the knowledge system.

[0114] Through the above construction method, this module realizes the transformation from "multi-source feature set" to "structured knowledge network", which unifies the originally scattered detection experience, sample characteristics and result information in the graph structure. This not only provides path constraints for subsequent graph enhancement reasoning, but also significantly improves the interpretability and knowledge reuse capability of the system, which is one of the core technical foundations of this invention.

[0115] 4. Multimodal unified coding module

[0116] The multimodal unified coding module is used to perform unified vectorization representation of the structured features, image temporal features and knowledge graph subgraphs output by the multi-source feature construction module, mapping data of different modalities and scales to the same semantic space, thereby providing a consistent and computable input representation for the subsequent graph-enhanced inference module.

[0117] This module processes sample-level data units, focusing on structured feature vectors. Image feature set and knowledge graph subgraphs They are encoded separately and then a unified feature representation is generated through a fusion mechanism.

[0118] In terms of structured feature encoding, the normalized feature vectors... The input is fed into a multi-layer mapping network, where linear transformations and non-linear activation functions map the original features to a low-dimensional embedding space, resulting in a structured feature representation. ;

[0119] in, This represents a structured feature encoding function that uses a weight matrix to weight and combine features of different dimensions while maintaining the relative relationships between key physical quantities.

[0120] In terms of image feature coding, the color halo temporal feature set The input is fed into a time-series coding network to dynamically model the color halo diffusion process.

[0121] First, the image features at each time step are spatially encoded. Then, temporal evolution features are extracted through temporal aggregation to obtain the image modal representation. ;in, This represents the image feature encoding function, used to characterize the diffusion speed of color halo, boundary change trend, and stability characteristics.

[0122] In terms of knowledge graph encoding, sample-related subgraphs are extracted from the knowledge graph construction module. And based on the graph structure, the nodes and their adjacency relationships are embedded and represented.

[0123] By aggregating node attributes and features of adjacent nodes, a graph modal representation is obtained: ;

[0124] in, This represents a graph encoding function used to characterize the semantic association structure between sample features and their positional relationship in the overall knowledge network.

[0125] After completing the encoding of each modality, this module integrates the representations of different modalities through a feature fusion mechanism. The fusion methods include vector concatenation, weighted summation, or dynamic weighting based on an attention mechanism, to obtain a unified representation vector of the sample.

[0126] ;

[0127] To avoid inconsistencies in numerical scales among different modal features, the modal vectors are normalized before fusion, and the contribution ratio of each modality in the overall representation is adjusted by weighting coefficients, thereby achieving the collaborative expression of structured information, image information, and knowledge semantic information.

[0128] In addition, this module introduces a feature consistency constraint mechanism to align the representations of the same sample in different modalities, ensuring that their distribution remains consistent in a unified semantic space. By minimizing the differences between modalities, the stability and generalization ability of the fused representation are improved.

[0129] Finally, this module outputs a unified feature representation. This representation simultaneously includes material property information, dynamic information of the experimental process, and semantic information of the knowledge graph, realizing the transformation from "multi-source heterogeneous features" to "unified semantic vectors" and providing high-quality input for the subsequent graph-enhanced reasoning module.

[0130] Through the above methods, this module effectively solves the problem of the difficulty in directly integrating different data modalities, enabling the system to perform reasoning analysis in a unified feature space. It is a key technical link connecting the data layer and the reasoning layer.

[0131] 5. Graph-enhanced reasoning module

[0132] The graph-enhanced reasoning module is used to perform joint reasoning on the methylene blue value and related indicators of the test sample based on a unified feature representation and combining knowledge graph structure information and historical sample data. It is the core technical link of this invention to realize the integration of "data-driven, knowledge-constrained, and intelligent reasoning".

[0133] This module uses the feature vector output by the multimodal unified coding module. The system takes the graph structure data and historical sample database from the knowledge graph construction module as input and completes the prediction process through a three-stage mechanism of "similar case retrieval - graph path expansion - large model reasoning".

[0134] First, based on the unified feature vector A similarity metric function between samples is constructed to calculate the similarity of historical sample sets. The similarity function comprehensively considers structured features, image features, and spectral structure information, and is defined as follows:

[0135] ;

[0136] in, Represents cosine similarity. This represents a similarity measure based on graph structure. , , The weighting coefficients are satisfied. .

[0137] Based on this similarity function, the samples with the highest similarity are selected from the historical sample database. Each sample constitutes a reference case set. This is for reference in subsequent reasoning.

[0138] Based on similar cases, semantic subgraphs related to the current sample are extracted from the knowledge graph, and multi-hop path expansion is performed to form a set of candidate inference paths from sample features to detection results: ;

[0139] The node sequence in the path represents the correlation process from aggregate characteristics, test conditions, and color halo features to methylene blue value and quality grade. By filtering the path length, relationship type, and node attributes, several representative key paths are retained as inference constraints.

[0140] Subsequently, the unified feature representation will be used. Similar Case Collection and path set Common input large model inference engine.

[0141] During the inference process, the large model uses structured input as a condition, and by fusing sample features, historical case distribution, and semantic paths of the graph, it performs interval prediction of Methylene Blue values ​​and generates corresponding inference results. ;

[0142] The output results This includes the methylene blue value prediction range, endpoint sensitivity assessment, stability analysis results, and quality grade determination.

[0143] To ensure the controllability of the reasoning process, this module introduces a graph constraint mechanism during the large model reasoning process. This mechanism restricts the model reasoning space through path information, ensuring that the generated results satisfy the logical relationships and constraint rules in the knowledge graph. This avoids the bias that may occur in a purely data-driven model. Specifically, during the reasoning process, feature weights related to key path nodes are prioritized, and reasoning results that are inconsistent with the graph constraints are suppressed.

[0144] In addition, this module performs consistency verification on the inference results. By comparing the difference between the prediction results and the statistical distribution of similar cases, abnormal outputs are corrected or marked. When the prediction results deviate from the historical sample distribution range or violate the constraints of the graph rules, an uncertainty flag will be triggered, providing input basis for the subsequent review decision module.

[0145] Through the above mechanism, this module realizes multimodal joint reasoning based on knowledge graph constraints, which can not only improve the accuracy of AMI value prediction, but also combine the reasoning process with structured knowledge, so that the results have clear semantic support and logical path, thereby effectively solving the problems of "difficult to interpret results" and "black box model" in traditional methods. This is one of the core innovations that distinguishes this invention from the existing technology.

[0146] 6. Prediction and Evaluation Module: The prediction and evaluation module is used to express the inference results output by the graph enhancement inference module in a structured and quantitative manner. It transforms the prediction information generated by the large model into evaluation indicators with clear physical meaning and engineering usability, thereby realizing the standardized output and quantitative analysis of methylene blue detection results.

[0147] This module uses the reasoning results Using this as input, a unified calculation and expression of the methylene blue value prediction interval, endpoint sensitivity, detection result stability, and quality level are performed to construct a multi-dimensional evaluation index system.

[0148] Regarding the prediction of Asian League Blue Value, the predicted values ​​in the inference results are converted into interval form: ;

[0149] in, and These represent the lower and upper limits of the prediction, respectively. This interval reflects the possible range of methylene blue values ​​under the current sample characteristics. By introducing an interval expression, the uncertainty of the prediction results can be effectively characterized, and tolerance space can be provided for subsequent decision-making.

[0150] Regarding endpoint sensitivity assessment, an endpoint sensitivity index is constructed based on the dynamic characteristics of halo diffusion to describe the degree of response to changes in halo near the titration endpoint. Its expression is: ;

[0151] Where 𝐷 represents the diameter of the halo, and 𝐺 represents the boundary grayscale gradient. This indicates a crucial moment as one approaches the finish line. , These are the weighting coefficients.

[0152] This indicator reflects the rate of change of color halo at the final stage, when When the value is small, it indicates that the endpoint change is not significant, and there is a risk of difficulty in judgment.

[0153] Regarding the stability assessment of test results, stability indicators were constructed by statistically analyzing similar case sets and historical repeated test data.

[0154] ;

[0155] in, This represents the average value of the Asian League Blues. Indicates standard deviation, To prevent tiny constants with a denominator of zero.

[0156] This metric is used to measure the dispersion of prediction results within the historical sample distribution. The higher the value, the better the stability of the result.

[0157] In terms of quality grade assessment, based on the predicted methylene blue value range and related indicators, combined with testing specifications or preset grading rules, samples are classified into corresponding quality grades. This process can be achieved through interval mapping or classification models, allowing the prediction results to directly correspond to engineering quality control standards.

[0158] In addition, to improve the reliability of the evaluation results, this module introduces a confidence calculation mechanism to quantify the credibility of the prediction results. The confidence score can be calculated based on the distribution density of similar cases, the consistency of the map path, and the stability of the model output, and is used to reflect the reliability of the current prediction results.

[0159] Finally, this module outputs standardized evaluation results, including the methylene blue value prediction range, endpoint sensitivity index, stability index, quality grade, and prediction confidence level, forming a complete quantitative evaluation result set. This result set can not only be directly used to generate test reports, but also provide key inputs for the subsequent interpretation generation module and retesting decision module.

[0160] Through the above processing, this module realizes the transformation from "reasoning results" to "engineering indicator system", transforming the methylene blue detection results from unstructured reasoning output into standardized indicators that are quantifiable, comparable, and verifiable, providing a reliable basis for engineering applications.

[0161] 7. Explanation and generation module

[0162] The interpretation generation module is used to perform semantic parsing and evidence construction on the results output by the prediction and evaluation module. It transforms quantitative indicators such as methylene blue value prediction range, endpoint sensitivity, stability and quality level into explanatory information with clear evidence, thereby achieving the traceability and understandability of the test results.

[0163] This module takes the prediction and evaluation results and intermediate information from the graph-enhanced inference module as input, including unified feature representation. Similar Case Collection Knowledge graph path set The predictive index results are used to generate explanatory content through three mechanisms: feature contribution analysis, path mapping, and case matching.

[0164] In terms of feature contribution analysis, the importance of each dimension of features in the unified feature representation is evaluated, and their influence on the prediction results is calculated. By weighted decomposition of structured features, image features, and process features, a set of key influencing factors is obtained:

[0165] ;

[0166] in, Indicates the first One characteristic factor, This indicates its corresponding weight.

[0167] By sorting the weights, we can extract several dominant factors that have the greatest impact on the change of methylene blue value, such as stone powder content, mud powder mineral composition, particle morphology roughness, water absorption rate, and color halo boundary gradient change rate.

[0168] Regarding the interpretation of graph paths, the set of paths extracted from the inference module will be used. It is converted into a readable causal chain expression, where each path represents the semantic derivation process from sample attributes to detection results. By mapping the nodes and relationships in the path, a structured evidence chain is generated.

[0169] For example, the path: ;

[0170] The process transforms the graph structure into a natural language interpretation while preserving the causal relationships between nodes. This is achieved by converting the graph structure into a natural language interpretation.

[0171] In terms of similar case matching, from a set of similar cases Representative samples are selected, and their feature distribution and detection results are compared and analyzed. By calculating the distance between the current sample and historical samples in the feature space, the closest reference case is identified, and its methylene blue value and stability characteristics are extracted as auxiliary interpretation basis. This process is used to explain the position and rationality of the current prediction result in historical data.

[0172] To enhance the completeness of the interpretation, this module integrates feature contribution results, map path evidence, and similar case information to generate a unified interpretation output: ;

[0173] Simultaneously, based on the indicator results output by the prediction and evaluation module, the explanatory content is semantically enhanced to form an explanatory text containing the following information: main influencing factors, explanation of causal relationships, historical case support, and explanation of the credibility of the results.

[0174] In addition, this module supports the extraction and annotation of key image frames, and associates representative color halo change images during the titration process with the explanatory content to intuitively show the endpoint change characteristics and their impact on the prediction results.

[0175] Through the above mechanism, this module realizes the transformation from "numerical prediction results" to "structured interpretation information", so that the methylene blue detection results not only have quantitative expression, but also have clear sources of evidence and logical derivation process, thereby effectively solving the problem of uninterpretable results in traditional detection methods and improving the credibility and acceptability of the system in engineering applications.

[0176] 8. Re-inspection decision module

[0177] The retesting decision module is used to comprehensively determine the reliability of the current methylene blue test results based on the obtained predictive evaluation results and interpretation information, and to provide decision suggestions on whether retesting is necessary, thereby realizing the transformation of the testing process from "single judgment" to "risk control and quality assurance".

[0178] This module takes the methylene blue value range, endpoint sensitivity index, stability index and prediction confidence level output by the prediction and evaluation module as input, and combines the key influencing factors and case distribution information output by the interpretation and generation module to quantify the uncertainty of the test results and construct a retest criterion function.

[0179] First, the sources of uncertainty in the prediction results are uniformly expressed, including factors such as insufficient confidence in model predictions, low sensitivity to the endpoint, and large dispersion in the distribution of historical samples. Based on these factors, a comprehensive re-examination criterion is constructed:

[0180] ;

[0181] in, Indicates the prediction confidence level. Indicates stability index, Indicates endpoint sensitivity. For sensitivity threshold, , , These are the weighting coefficients. This is an indicative function; when the endpoint sensitivity is below the threshold, it indicates that the endpoint change is not obvious, increasing the risk of re-inspection.

[0182] Based on this, the criteria for re-examination will be... With preset threshold Compare;

[0183] When satisfied If the system determines that the current test result has an uncertainty risk, it will output "Retest recommended".

[0184] when If the test result is deemed reliable, "No retesting required" will be output.

[0185] Furthermore, to improve the engineering interpretability of the decision-making process, this module analyzes the reasons for triggering re-inspection, decomposes the contribution of each influencing factor to the re-inspection criteria, and forms a set of reasons: ;

[0186] Each of these For specific triggering conditions, such as "low prediction confidence", "no significant change in the endpoint", "large dispersion of similar cases" or "abnormal fluctuations in key features", etc.

[0187] In addition, this module combines similar case information from the interpretation and generation module to evaluate the position of the current sample in the historical sample space. When the sample characteristics are located in the boundary area of ​​the historical data distribution or have obvious deviations, this information is used as one of the important bases for re-examination suggestions, thereby avoiding prediction bias caused by sample extrapolation.

[0188] To further enhance the system's engineering adaptability, this module supports the refinement of re-inspection strategy output, including suggestions to adjust test operation parameters (such as reducing the titration step size, increasing the image sampling frequency in key stages), supplementing detection indicators, or conducting repeated tests, thereby expanding the re-inspection decision from a simple judgment to an actionable operational suggestion.

[0189] Finally, this module outputs the re-inspection decision results and corresponding explanations, achieving a closed-loop connection from prediction results to quality control decisions. Through this module, the present invention can issue timely warnings when test results are uncertain or risky, effectively reducing the probability of misjudgment and improving the reliability of methylene blue test results and the safety of engineering applications.

[0190] The working principle of this application is as follows:

[0191] By unifying the multi-source information on aggregate characteristics, test process, and color halo diffusion behavior, discrete testing experience is transformed into structured knowledge. Under the constraints of knowledge graphs and combined with large-scale model reasoning, the results of methylene blue testing can be predicted in advance, assisted in the process, and interpreted and output as a decision.

[0192] First, at the data level, the system acquires multi-source raw data of the entire methylene blue test process through the data acquisition module, including structured parameters (particle size distribution, stone powder content, mud powder composition, parent rock type, water absorption rate, etc.), time-series images of filter paper color halo diffusion during titration, as well as test operation parameters and environmental information. Through a unified time calibration mechanism, data from different sources are mapped to the same time axis to form sample data units containing static characteristics and dynamic process information, providing a consistent input basis for subsequent analysis.

[0193] In the feature construction stage, the system processes the collected data, transforms the structured parameters into standardized feature vectors, and extracts geometric and grayscale features reflecting diffusion behavior from the color halo image sequence. At the same time, it calculates the diffusion change rate by combining time series information, thereby forming a multi-source feature set that can describe the adsorption and diffusion process of methylene blue. Through this process, the transformation from "raw data" to "computable features" is realized.

[0194] In the knowledge modeling stage, the system constructs a knowledge graph for methylene blue detection based on multi-source features. It expresses the relationship between sample characteristics, test conditions, color halo characteristics and detection results in the form of a graph structure. By constructing a semantic path of "aggregate characteristics - color halo behavior - methylene blue value - quality grade", the original implicit knowledge that relied on experience is transformed into an explicit structured knowledge network. The knowledge is stored and managed through a graph database, thereby providing semantic constraints and path basis for the reasoning process.

[0195] In the multimodal coding stage, the system maps structured features, image features, and knowledge graph subgraphs to a unified semantic space. Through vectorization, it achieves the fusion expression of information from different modalities and obtains a unified feature representation. This process eliminates the expression differences between different data types, enabling subsequent reasoning to be carried out in a unified feature space.

[0196] In the core reasoning stage, the system predicts the test sample through a graph-enhanced reasoning mechanism. First, it calculates the similarity with historical samples based on a unified feature representation and selects several similar cases as references. Second, it extracts multi-hop semantic paths related to the current sample from the knowledge graph and constructs causal association chains. Finally, it uses the sample features, similar cases, and graph paths as joint inputs and performs reasoning using a large model to generate Methylene Blue value ranges and related evaluation results. In this process, the knowledge graph constrains the reasoning process, ensuring that the model output satisfies existing knowledge relationships, thereby avoiding the instability caused by purely data-driven approaches.

[0197] During the results evaluation phase, the system quantifies the inference results and transforms the prediction output into engineering-usable indicators, including the methylene blue value prediction range, endpoint sensitivity, detection stability, and quality level. Among these, endpoint sensitivity is characterized by the rate of color halo change, and stability is evaluated by the dispersion of historical sample distribution, thereby achieving a quantitative description of the detection reliability.

[0198] In the interpretation generation stage, the system interprets and outputs the prediction results based on feature contribution analysis, graph path mapping and similar case matching. By identifying key influencing factors, restoring semantic reasoning paths and providing historical case support, the detection results have clear sources of evidence and logical processes, thereby enhancing the interpretability and credibility of the system.

[0199] During the retesting decision-making stage, the system comprehensively predicts confidence, endpoint sensitivity and stability and constructs a retesting criterion function to assess the uncertainty of the test results. When there is a risk in the test results, the system automatically outputs retesting suggestions and explanations of the reasons, and can further provide operational optimization suggestions, thereby achieving closed-loop control of the testing process.

[0200] In summary, this invention constructs a complete technical system from data acquisition to result decision-making by combining multi-source data fusion, knowledge graph modeling, and large-scale model reasoning. It realizes the transformation of methylene blue detection from traditional experience-driven to "data-knowledge-reasoning" collaborative driving. This system not only improves detection efficiency and result consistency, but also significantly enhances the interpretability and engineering application value of the results.

[0201] The above description is merely a specific embodiment of this application, enabling those skilled in the art to understand or implement this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A knowledge graph and large-scale model reasoning-based intelligent prediction system for sub-blue value, characterized in that, include: The data acquisition module is used to acquire multi-source data of the entire methylene blue test process, and to perform time alignment of structured parameter data, image data and test process data based on timestamps to construct sample data units; The multi-source feature construction module is used to standardize the multi-source data, construct structured feature vectors, and extract geometric features and temporal variation features of color halo diffusion based on image sequences to form a multimodal feature set. The knowledge graph construction module is used to construct a graph structure based on the multimodal feature set, which includes sample nodes, aggregate characteristic nodes, test condition nodes, color halo feature nodes and methylene blue value nodes, and to establish semantic relationships between nodes to form sample-level subgraphs. The multimodal unified coding module is used to vectorize structured features, image features, and knowledge graph subgraphs, and generate unified feature representations through fusion operations. The graph-enhanced reasoning module is used to calculate the similarity with historical samples based on the unified feature representation, extract a set of similar cases, extract semantic paths from the knowledge graph, and input the unified feature representation, the set of similar cases, and the semantic paths into the large model for joint reasoning to obtain the sub-blue value prediction result. The prediction and evaluation module is used to convert the inference results into methylene blue value prediction intervals, endpoint sensitivity indicators, stability indicators, and quality levels. The explanation generation module is used to generate explanatory information for prediction results based on feature contribution analysis, map paths, and similar cases. The re-inspection decision module is used to construct re-inspection criteria based on prediction confidence, endpoint sensitivity, and stability, and output the decision result of whether to re-inspect.

2. The intelligent prediction system for sub-blue value based on knowledge graph and large model reasoning as described in claim 1, characterized in that: The data acquisition module includes a structured parameter acquisition unit, an image acquisition unit, and an experimental process recording unit. The structured parameter acquisition unit is used to acquire aggregate particle size distribution, particle morphology parameters, stone powder content, mud powder mineral composition, parent rock type, and water absorption rate, and stores them through standardized fields. The image acquisition unit includes an industrial camera, a constant light source module, and a trigger control unit. The trigger control unit triggers image acquisition according to the titration time interval or titration event, forming a time-series image sequence of filter paper color halo diffusion, and achieves image consistency by fixing the shooting angle, spatial scale, and focal length constraints. The experimental process recording unit is used to record the single titration volume, cumulative titration amount, number of titrations, stirring time, settling time, and ambient temperature and humidity, and records them synchronously with the image data, thereby forming a unified acquisition system for multi-source heterogeneous data.

3. The intelligent prediction system for sub-blue value based on knowledge graph and large model reasoning as described in claim 1, characterized in that: The data acquisition module synchronously processes multi-source data based on a unified time calibration mechanism. By adding timestamps to structured parameters, image sequences, and experimental process data, a unified time axis is constructed, and data from different sources are mapped to the same time dimension, forming a unified data representation. ;in, This represents the static structured parameters of the sample. This represents the filter paper image captured at the corresponding time. This represents the experimental process parameters at that moment. This time synchronization mechanism enables a precise correspondence between image changes and titration operations, and ensures the alignment consistency between multimodal data during subsequent feature construction.

4. The intelligent prediction system for sub-blue value based on knowledge graph and large model reasoning as described in claim 1, characterized in that: The multi-source feature construction module normalizes and encodes the structured parameters, and constructs feature vectors: ； Simultaneously, the filter paper halo image sequence was subjected to grayscale conversion, noise reduction, and boundary enhancement. Geometric and grayscale features of the halo region were extracted, including the halo diameter. Area of ​​color halo and boundary grayscale gradient And calculate dynamic change characteristics based on time series: ； This results in a time-series feature set that can characterize the adsorption and diffusion process of methylene blue, and the structural features, image features, and experimental process features are fused to obtain a unified multi-source feature expression.

5. The intelligent prediction system for sub-blue value based on knowledge graph and large model reasoning as described in claim 1, characterized in that: The knowledge graph construction module constructs the graph structure as follows: Among them, the set of nodes This includes nodes for sample type, particle size distribution, particle morphology, stone powder content, mud powder composition, parent rock type, water absorption rate, test conditions, color halo characteristics, and methylene blue value, as well as edge sets. Represents the semantic relationships between nodes, including relationship types such as "has parameters", "contains components", "generates features", and "corresponds to results", and attribute sets. Numerical attributes used to describe nodes and relationships are used to realize a structured expression of sample characteristics, experimental behavior and test results through the above structure, and are stored and managed based on a graph database.

6. The intelligent prediction system for sub-blue value based on knowledge graph and large model reasoning as described in claim 1, characterized in that: The knowledge graph construction module constructs a semantic path from aggregate characteristics to detection results within a graph structure: ;in, Indicates particle size distribution. Indicates particle morphology, Indicates the composition of clay powder. Indicates color halo characteristics, Indicates the Asian League Blue Value. It represents the quality level, describes the causal relationship between material properties, test behavior and test results through this path, and participates in path development and result generation as a constraint in the reasoning process.

7. The intelligent prediction system for sub-blue value based on knowledge graph and large model reasoning as described in claim 1, characterized in that: The multimodal unified coding module performs vectorized encoding on structured features, image features, and knowledge graph subgraphs, respectively, where structured features are encoded through multi-layer mapping functions. Image features are encoded using a temporal coding function. Encoding, knowledge graph subgraphs via graph embedding functions Encode the data and generate a unified feature representation through vector concatenation or weighted fusion: ; During the fusion process, the features of different modalities are normalized, and the contribution ratio of each modality is adjusted by weighting, thereby achieving a unified semantic expression of multi-source data.

8. The intelligent prediction system for sub-blue value based on knowledge graph and large model reasoning as described in claim 1, characterized in that: The graph-enhanced inference module constructs a similarity function based on a unified feature representation: ； Furthermore, several samples with the highest similarity are selected from the historical sample library as reference cases. At the same time, multi-hop semantic paths are extracted from the knowledge graph. The unified feature representation, reference cases, and semantic paths are jointly input into the large model for joint reasoning. During the reasoning process, the reasoning space is restricted through the path constraint mechanism, so that the generated results satisfy the semantic relationships and rule constraints in the knowledge graph, thereby improving the accuracy and stability of the prediction results.

9. The intelligent prediction system for sub-blue value based on knowledge graph and large model reasoning as described in claim 1, characterized in that: The prediction and evaluation module quantifies the inference results and constructs the MCA blue value prediction range: ; And based on the color halo diffusion characteristics, an endpoint sensitivity index was constructed: ; And stability indicators: ; At the same time, the samples are classified into quality grades according to the testing standards, and the prediction confidence level is calculated to form a multi-dimensional evaluation index system.

10. The intelligent prediction system for sub-blue value based on knowledge graph and large model reasoning as described in claim 1, characterized in that: The re-inspection decision module constructs re-inspection criteria based on the methylene blue value range, endpoint sensitivity, stability, and prediction confidence output by the prediction and evaluation module: ; When the criterion exceeds the preset threshold, a re-examination suggestion is output, and the reasons for triggering the re-examination are analyzed to generate a set of reasons, including insufficient prediction confidence, insignificant changes in the endpoint, and discrete sample distribution. At the same time, operation optimization suggestions are output, including adjusting the titration step size or increasing the image sampling frequency, so as to achieve closed-loop quality control of the detection results.

Images