A knowledge graph construction method and system based on traditional ceramic techniques

Abstract

The application relates to the technical field of knowledge graph construction, and discloses a knowledge graph construction method and system based on traditional ceramic techniques, which comprises the following steps: dividing a kiln into M sub-regions along the longitudinal direction at fixed intervals, and obtaining observation data of each sub-region; determining a zero point region in the M sub-regions and a corresponding mid-section oxygen partial pressure fluctuation value of the zero point region according to the observation data; taking the zero point region as the center, compensating the preset position weight of each sub-region according to the mid-section oxygen partial pressure fluctuation value to obtain an updated position weight of each sub-region; constructing a knowledge graph data group based on the zero point region, the mid-section oxygen partial pressure fluctuation value and the updated position weight; calculating a mid-section feature constraint loss, combining the knowledge graph data group to construct a target knowledge graph; obtaining an inference answer of an updated LLM question and answer model, and determining a mid-section reference rate of the inference answer; and generating a quality evaluation result of the target knowledge graph according to the mid-section reference rate of the inference answer.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph construction technology, and more specifically, to a method and system for constructing a knowledge graph based on traditional ceramic techniques. Background Technology

[0002] Blue and white porcelain is an underglaze ceramic. Its color relies on cobalt-containing pigments painted on the body of the porcelain, followed by high-temperature firing to create unique blue patterns. The firing temperature is generally between 1280 and 1330 degrees Celsius, and it must be done in a relatively stable reducing atmosphere to avoid quality problems such as yellowing of the glaze and graying of the blue. Therefore, the temperature distribution and oxygen concentration (atmosphere) distribution during the firing stage directly determine the final product's appearance. In actual production, to ensure the stability of the ceramic surface color and structure, potters control the kiln loading density, use saggers for protection, and adjust the flow of fire channels or flues to achieve the optimal balance of firing conditions. These parameters form the core relationship chain in the knowledge graph of blue and white porcelain techniques.

[0003] The "mantou" kiln is a common ancient kiln structure in traditional kilns like Jingdezhen. It features an arched top, with flames deflected at the top and expelled from the bottom. Its rear wall often has suction holes or long, narrow flues, and some even connect to a smoke chamber and chimney to enhance draft. It possesses the following characteristics:

[0004] Near the kiln tail (rear section), the smoke exhaust is smooth and the suction is stronger, creating an atmosphere that is more conducive to reduction.

[0005] The area near the kiln head (front section) has a sufficient flame supply and a higher temperature, which may create an oxidizing environment;

[0006] The middle section of the kiln is prone to atmospheric fluctuations and unstable areas due to structural folds and changes in kiln loading density.

[0007] These characteristics result in significant differences in the firing environment of different sections in the front, middle, and rear spatial dimensions.

[0008] While modern kilns can collect temperature and atmosphere data using high-temperature thermocouples and oxygen probes, traditional steamed bun kilns present significant limitations in sensor placement due to their enclosed structure, high loading density, and severe obstruction from saggers. Particularly in the middle section of the kiln, sensor placement is difficult due to factors such as upper flame backflow, lower flue restrictions, and dense arrangement of artifacts, leading to unstable data acquisition and inherent gaps or unreliability in the data collected from this section. When constructing a knowledge graph, these missing or unreliable data points prevent the accurate establishment of semantic nodes and data group relationships related to this section, or result in a lack of highly credible sources.

[0009] When processing structured or graphed knowledge, large language models often adopt a strategy of first retrieval and then generation, that is, first retrieving relevant evidence from the knowledge graph, and then reasoning or answering through the language generation module. However, existing studies have shown that LLM language models have defects: weak perception of intermediate information in the context, especially in long texts or graph structures, and tend to focus on information at the beginning and end. This defect is also reflected in the selection of graph evidence nodes, that is, if there is a lack of additional constraint mechanism, the large language model will tend to select entities, relationships and attribute nodes in the graph that are related to the beginning and end, while ignoring the evidence information in the middle section. When the Man-tou kiln itself causes difficulty in observing the middle section, the defects of the LLM language model will further amplify the imbalance of information, ultimately leading to deviation or loss of the generated answer.

[0010] In summary, the existing knowledge graph construction method of blue and white porcelain technology has the following technical problems:

[0011] The Man-tou kiln exhaust structure enhances the rear section suction, the front section heating is concentrated, and the middle section is prone to form a thermodynamic unstable zone due to the structure and loading density. The middle section of the Man-tou kiln is limited in distribution, blocked by the sagger, and has large temperature / air fluctuation, so that the data in the middle section is weakened or even lost at the perception level. The constructor of the knowledge graph tends to model the structure of the front and rear sections with sufficient data during the establishment of the original knowledge graph, and the middle section is automatically weakened due to data scarcity. During the retrieval and generation process, since the large language model is more sensitive to the information at the beginning and end, it further leads to the omission of key middle section information in the decision-making.

[0012] In summary, the technical problem is a systematic evidence weakening and knowledge loss problem caused by physical limitations, data sparsity and algorithm defects in the multi-layer system of "blue and white porcelain → Man-tou kiln → knowledge graph → language model", which seriously affects the integrity and accuracy of the knowledge graph in supporting the understanding and generation tasks of blue and white porcelain technology. SUMMARY

[0013] The present application provides a knowledge graph construction method and system based on traditional ceramic technology, which solves the technical problems raised in the background art.

[0014] In a first aspect, the present application provides a knowledge graph construction method based on traditional ceramic technology, comprising:

[0015] S1, dividing the kiln along the longitudinal direction into M sub-regions at a fixed interval, and obtaining observation data of each sub-region; wherein the observation data includes normalized temperature value and oxygen partial pressure value;

[0016] S2, determining a zero point region in the M sub-regions according to the observation data, and a middle section oxygen partial pressure fluctuation value corresponding to the zero point region;

[0017] S3, with the zero point region as the center, the preset position weight of each sub-region is compensated according to the mid-section oxygen partial pressure fluctuation value to obtain the updated position weight of each sub-region;

[0018] S4, based on the zero-point region, the mid-section oxygen partial pressure fluctuation value, and the updated position weight, to construct a knowledge graph data set;

[0019] S5, calculate the mid-segment feature constraint loss and combine it with the knowledge graph data set to construct the target knowledge graph;

[0020] S6. Based on the target knowledge graph, obtain the updated LLM question answering model, acquire the reasoning answer of the updated LLM question answering model, and determine the mid-section citation rate of the reasoning answer;

[0021] S7 generates a quality evaluation result of the target knowledge graph based on the mid-section citation rate of the reasoning answer.

[0022] Furthermore, based on the observation data, the zero-point regions within the M sub-regions are determined, along with the corresponding mid-range oxygen partial pressure fluctuation values, including:

[0023] S21, Determine N within the observation period. t Each sampling time;

[0024] S22, at time t during the observation period, calculate the temperature difference ΔT between the m-th sub-region and the (m+1)-th sub-region. m，t The difference between oxygen partial pressure and ΔpO2 m，t ;

[0025] S23, for the m-th sub-region and the (m+1)-th sub-region, calculate the statistics by combining the temperature difference and oxygen partial pressure difference at each time step:

[0026]

[0027] Among them, κ m Let be the statistic for the m-th sub-region during the observation period, and Δz represent the fixed interval;

[0028] S24. Along the front to the back of the kiln, calculate the product of the statistics of adjacent sub-regions in the order of sub-regions, so as to extract the first sub-region with a product ≤ 0 and take that sub-region as the zero point region.

[0029] S25, Calculate the oxygen partial pressure value in the zero-point region at N. t The variance at each sampling time point is used to obtain the fluctuation value of oxygen partial pressure in the middle section.

[0030] Furthermore, taking the zero-point region as the center, the preset position weights of each sub-region are compensated based on the mid-section oxygen partial pressure fluctuation value to obtain the updated position weights of each sub-region, including:

[0031] S31, extracting a middle section sub-region set from the M sub-regions based on the zero point region;

[0032] S32, determining interval distance d k of the kth sub-region in the middle section sub-region set to the zero point region;

[0033] S33, calculating spatial proximity of the kth sub-region to the zero point region according to the interval distance d k , as follows:

[0034]

[0035] wherein G(k) represents the spatial proximity of the kth sub-region to the zero point region, and η represents a preset bandwidth parameter, and η is greater than 0;

[0036] S34, performing normalization processing on the middle section oxygen partial pressure fluctuation value to obtain a compensation weight;

[0037] S35, taking the product of the compensation weight and the spatial proximity as a compensation factor S(k) of the kth sub-region;

[0038] S36, weighting and compensating a preset position weight of the kth sub-region by using the compensation factor to obtain an initial updated position weight of the kth sub-region;

[0039] S37, keeping the position weight of the sub-region in the non-middle section sub-region set unchanged, determining the initial updated position weight of each sub-region in the middle section sub-region set, and performing normalization processing on the position weight and the initial updated position weight to obtain an updated position weight of each sub-region.

[0040] Further, based on the zero point region, the middle section oxygen partial pressure fluctuation value and the updated position weight, a knowledge graph data set is constructed, including:

[0041] S41, defining a kiln run entity corresponding to an observation time period as KilnRun, and a mth sub-region entity as KilnSection(m); and constructing attribute triples of the kiln run entity, including:

[0042] (KilnRun, hZPP, z * )

[0043] (KilnRun, hZPR, Ω0)

[0044] (KilnRun, hMOV, V mid )

[0045] wherein z * represents the zero point region, Ω0 represents the middle section sub-region set, and V midhZPP represents the first predicate, hZPR represents the second predicate, and hMOV represents the third predicate;

[0046] S42, an attribute triple of the mth sub-region entity is constructed, including:

[0047] (KilnSection(m), hUPW, ω upd (m))

[0048] wherein ω upd (m) represents the update position weight of the mth sub-region, and hUPW represents the fourth predicate;

[0049] S43, the attribute triple of the kiln section entity and the attribute triple of the sub-region entity are combined to obtain a knowledge graph data set.

[0050] Further, a middle section feature constraint loss is calculated, and the knowledge graph data set is combined to obtain a target knowledge graph, including:

[0051] S51, a region importance distribution of the mth sub-region is generated, as follows:

[0052]

[0053] wherein τ(m) represents the region importance distribution of the mth sub-region, and S(m) is 0 if the mth sub-region does not belong to the middle section sub-region set;

[0054] S52, an embedding vector e m is defined for the mth sub-region entity;

[0055] S53, a corresponding parameter matrix W p is defined for each predicate;

[0056] S54, the mth attribute triple is evaluated and processed, including:

[0057] f(m) = sigmoid(e m T W p e0)

[0058] wherein f(m) represents the rationality score of the mth attribute triple, sigmoid represents a sigmoid activation function, e k T represents the transpose vector of e t , e f , and e t 0 is e tis a first object vector, e0 takes a value of e f f is a second object vector;

[0059] S55, obtaining rationality scores of the M sub-regions and performing normalization processing to obtain normalized scores of each sub-region; wherein a sum value of the normalized scores of the M sub-regions is 1;

[0060] S56, calculating a mid-section feature constraint loss, including:

[0061]

[0062] wherein L mid represents the mid-section feature constraint loss, and R(m) represents the normalized score of the mth sub-region;

[0063] S57, based on the mid-section feature constraint loss, iteratively updating the embedding vector e m , the parameter matrix W p , the first object vector e t and the second object vector e f by the gradient descent method until a preset iteration number is reached, to obtain a target embedding vector and a target parameter matrix;

[0064] S58, solidifying the target embedding vector, the target parameter matrix and the attribute triple of each sub-region to form a full target knowledge graph.

[0065] Further, according to the target knowledge graph, an updated LLM question and answer model is obtained, an inference answer of the updated LLM question and answer model is obtained, and a mid-section reference rate of the inference answer is determined, including:

[0066] S61, loading the target knowledge graph into the LLM question and answer model to obtain an updated LLM question and answer model;

[0067] S62, inputting a preset question to the updated LLM question and answer model; wherein the preset question includes a prompt word, and the prompt word requires the updated LLM question and answer model to give a problem cause and a cause region of the preset question;

[0068] S63, repeating S62 for Q times to obtain Q groups of inference answers;

[0069] S64, determining a number q of cause regions belonging to the mid-section sub-region set in the Q groups of inference answers to obtain a mid-section reference rate U, and the mid-section reference rate is

[0070] Further, according to the mid-section reference rate of the inference answer, a quality evaluation result of the target knowledge graph is generated, including:

[0071] ​S71, taking the sum value of the updated position weight of the sub-region in the middle sub-region set as the expected reference rate;

[0072] S72, calculating the reference rate difference between the middle reference rate of the inference answer and the expected reference rate;

[0073] S73, if the absolute value of the reference rate difference is greater than the preset quality threshold, the target knowledge graph is unqualified; if the absolute value of the reference rate difference is less than or equal to the preset quality threshold, the target knowledge graph is qualified.

[0074] Further, if the target knowledge graph is unqualified, a reference rate loss is constructed, and S56 is returned; the reference rate loss is combined with the middle segment feature constraint loss to form a comprehensive loss, and S57 is executed with the comprehensive loss until the target knowledge graph is qualified.

[0075] In a second aspect, a knowledge graph construction system based on traditional ceramic techniques is applied to any of the knowledge graph construction methods based on traditional ceramic techniques, and includes:

[0076] A data acquisition module is configured to divide a kiln into M sub-regions along the longitudinal direction at fixed intervals, and obtain observation data of each sub-region; wherein the observation data includes normalized temperature values and oxygen partial pressure values;

[0077] A zero point extraction module is configured to determine a zero point region in the M sub-regions and a middle segment oxygen partial pressure fluctuation value corresponding to the zero point region based on the observation data;

[0078] A weight compensation module is configured to compensate the preset position weight of each sub-region based on the middle segment oxygen partial pressure fluctuation value with the zero point region as the center, to obtain an updated position weight of each sub-region;

[0079] A data construction module is configured to construct a knowledge graph data set based on the zero point region, the middle segment oxygen partial pressure fluctuation value, and the updated position weight;

[0080] A graph construction module is configured to calculate a middle segment feature constraint loss, and combine the knowledge graph data set to construct a target knowledge graph;

[0081] A graph test module is configured to obtain an updated LLM question and answer model based on the target knowledge graph, obtain an inference answer of the updated LLM question and answer model, and determine a middle reference rate of the inference answer;

[0082] A graph evaluation module is configured to generate a quality evaluation result of the target knowledge graph based on the middle reference rate of the inference answer.

[0083] The present application has the beneficial effect that by guiding the weight compensation of the middle section oxygen partial pressure fluctuation value centered on the zero point area, a knowledge graph with complete structure and balanced semantics is further constructed, and on this basis, the middle section embedding expression is strengthened by the middle section feature constraint loss, which significantly improves the attention ability of the large language model to the middle section knowledge in the question and answer reasoning. At the same time, a graph quality evaluation mechanism based on the middle section reference rate is proposed to realize closed-loop correction of the graph generation process, effectively solving the problems of missing middle section knowledge, reasoning bias and graph imbalance in traditional technologies, and enhancing the reliability, accuracy and verifiability of the knowledge graph in the auxiliary understanding and generation task. BRIEF DESCRIPTION OF DRAWINGS

[0084] Figure 1 is a flowchart of the present application. DETAILED DESCRIPTION

[0085] The subject matter described herein will now be discussed with reference to example implementations. It should be understood that the discussion of these implementations is merely meant to provide a better understanding of the subject matter described herein and can be changed in function and arrangement without departing from the scope of the present specification. Various processes or components can be omitted, substituted, or added according to desired implementations. Additionally, features described with respect to some examples can be combined in other examples.

[0086] As Figure 1 shown, a knowledge graph construction method based on traditional ceramic technology includes:

[0087] S1, the kiln is divided into M sub-regions along the longitudinal direction at a fixed interval, and observation data of each sub-region is obtained; wherein the observation data includes normalized temperature value and oxygen partial pressure value;

[0088] S2, according to the observation data, determine the zero point region in the M sub-regions, and the middle section oxygen partial pressure fluctuation value corresponding to the zero point region;

[0089] S3, taking the zero point region as the center, compensating the preset position weight of each sub-region according to the middle section oxygen partial pressure fluctuation value, to obtain the updated position weight of each sub-region;

[0090] S4, based on the zero point region, the middle section oxygen partial pressure fluctuation value and the updated position weight, to construct a knowledge graph data set;

[0091] S5, calculate the middle section feature constraint loss, and combine the knowledge graph data set to construct a target knowledge graph;

[0092] S6, according to the target knowledge graph, obtain an updated LLM question and answer model, obtain the reasoning answer of the updated LLM question and answer model, and determine the middle section reference rate of the reasoning answer;

[0093] S7, generating a quality evaluation result of the target knowledge graph according to the middle section reference rate of the reasoning answer.

[0094] In an embodiment of the present application, according to the observation data, a zero point region in the M sub-regions is determined, and a middle section oxygen partial pressure fluctuation value corresponding to the zero point region is determined, comprising:

[0095] S21, determining N t sampling time points in the observation time period;

[0096] S22, at the tth time point in the observation time period, calculating a temperature value difference ΔT m,t and an oxygen partial pressure value difference ΔpO2 m,t of the mth sub-region and the m+1th sub-region;

[0097] S23, for the mth sub-region and the m+1th sub-region, integrating the temperature value difference and the oxygen partial pressure value difference at each time point to calculate a statistic quantity:

[0098]

[0099] wherein, κ m is the statistic quantity of the mth sub-region in the observation time period, and ΔZ represents a fixed interval;

[0100] S24, along the front section to the rear section of the kiln, calculating the product of the statistic quantities of adjacent sub-regions in the order of the sub-regions to extract the first sub-region with the product ≤0, and taking the sub-region as the zero point region;

[0101] S25, calculating the variance of the oxygen partial pressure value at N t sampling time points for the zero point region to obtain the middle section oxygen partial pressure fluctuation value.

[0102] It should be noted that the observation time period needs to cover the entire stage of the firing of blue and white porcelain, usually including the heating, holding and cooling stages. The number of sampling time points needs to be determined according to the firing period, and generally needs to be dense enough to capture the dynamic changes of temperature and oxygen partial pressure, so as to convert the continuous changes of the kiln environment into discrete time point data.

[0103] It should be noted that temperature and oxygen partial pressure are the core factors affecting the color of blue and white porcelain: temperature determines the degree of glaze melting and pigment reaction, and oxygen partial pressure determines the reduction state of cobalt material. For example, blue and white porcelain is prone to gray in an oxidizing environment, and blue and white porcelain is prone to pure blue in a reducing environment. Calculate the temperature difference and oxygen partial pressure difference between the mth sub-region and the m+1th sub-region at the tth moment. The temperature difference and oxygen partial pressure difference can directly reflect the environmental difference between adjacent sub-regions at the tth moment. The positive and negative and size of the difference can reflect the environmental gradient: if the temperature difference is positive, it means that the temperature of the m+1th sub-region is higher than that of the mth sub-region; if the oxygen partial pressure difference is negative, it means that the oxygen partial pressure of the m+1th sub-region is lower than that of the mth sub-region, and it may be closer to the reducing environment.

[0104] It should be noted that the core role of the statistical quantity is to integrate the time dimension of environmental change and quantify the thermodynamic characteristics of the mth sub-region. Specifically, it includes: taking the product of the temperature difference and the oxygen partial pressure difference at each moment. This product can reflect the joint change trend of the two. For example, when the temperature difference and the oxygen partial pressure difference are of the same sign, the product is positive, indicating that the temperature of the adjacent region increases while the oxygen partial pressure also increases (possibly simultaneously in an oxidizing environment change); when they are of different signs, the product is negative, indicating that the temperature increases while the oxygen partial pressure decreases (possibly transitioning to a reducing environment). Taking the average of the products of all moments eliminates accidental fluctuations at a single moment, and obtains the average joint change trend of the sub-region within the observation period. Multiply by a fixed interval to associate the environmental difference with the spatial distance, so that the numerical size of the statistical quantity can reflect the intensity of the joint environmental change per unit space. The final statistical quantity can quantify the stability of the thermodynamic characteristics of the adjacent interval of the mth sub-region: the larger the absolute value of the statistical quantity, the more significant the joint change of temperature and oxygen partial pressure in the adjacent interval of the region; the sign reflects the overall trend of the change (positive or negative corresponds to a specific environmental change pattern).

[0105] It should be noted that the zero point region is the core marker of the middle section, which is used to reflect the region where the environmental characteristics in the kiln change. Along the front to the back of the kiln, calculate the product of the statistical quantities of adjacent sub-regions in order. If the product is greater than 0, it means that the thermodynamic characteristics of adjacent sub-regions are consistent (same positive or same negative), and the environmental characteristics have not changed essentially; if the product is less than or equal to 0, it means that the sign of the statistical quantity of adjacent sub-regions changes (from positive to negative or from negative to positive), indicating that the environmental characteristics have changed (for example, from oxidation dominated to reduction dominated, or the joint change trend of temperature and oxygen partial pressure is reversed). Extract the first sub-region with a product less than or equal to 0 as the zero point region, because this region is the first turning point of the environmental characteristics in the kiln, which corresponds to the unstable starting point of the environment in the middle section of the dumpling kiln due to the folding of the flame and the structure of the flue, which is the key dividing point for extracting the zero point region.

[0106] It should be noted that the mid-section oxygen partial pressure fluctuation value is a core index reflecting the stability of the zero point region environment. The stability of the oxygen partial pressure directly affects the consistency of the cobalt material reduction reaction: the greater the fluctuation, the more uneven the reduction degree of the cobalt material, and the more prone to deviation of the blue and white porcelain color. For the zero point region, the variance of the oxygen partial pressure value at all sampling times is calculated. The variance quantifies the dispersion degree of the data: the greater the variance, the more intense the change of the oxygen partial pressure of the zero point region in the time dimension, and the more unstable the environment. This fluctuation value will be a key basis for subsequent weight compensation, that is, the greater the fluctuation of the region, the more it needs to be focused on in the knowledge graph to avoid neglecting the natural scarcity of data or model preferences.

[0107] In an embodiment of the present application, the preset position weight of each sub-region is compensated according to the mid-section oxygen partial pressure fluctuation value centered on the zero point region, to obtain an updated position weight of each sub-region, including:

[0108] S31, extracting a mid-section sub-region set from the M sub-regions based on the zero point region;

[0109] S32, determining the interval distance d k of the kth sub-region in the mid-section sub-region set to the zero point region;

[0110] S33, calculating the spatial proximity of the kth sub-region to the zero point region according to the interval distance d k , as follows:

[0111]

[0112] Wherein, G(k) represents the spatial proximity of the kth sub-region to the zero point region, and η represents a preset bandwidth parameter, η being greater than 0;

[0113] S34, normalizing the mid-section oxygen partial pressure fluctuation value to obtain a compensation weight;

[0114] S35, multiplying the compensation weight and the spatial proximity as a compensation factor S(k) of the kth sub-region;

[0115] S36, weighting and compensating the preset position weight of the kth sub-region with the compensation factor to obtain an initial updated position weight of the kth sub-region;

[0116] S37, keeping the position weight of the sub-region in the non-mid-section sub-region set unchanged, and determining the initial updated position weight of each sub-region in the mid-section sub-region set, normalizing the position weight and the initial updated position weight to obtain an updated position weight of each sub-region.

[0117] It should be noted that the zero point area is the core of the turning point of the kiln environment, and the surrounding sub-areas are also thermodynamically unstable due to the influence of flame turning and flue structure. When extracting the middle sub-area set, the zero point area should be taken as the center, and the typical space range of the middle section of the steamed bun kiln should be combined to cover the two adjacent sub-areas before and after the zero point area, so as to merge these sub-areas into the middle sub-area set.

[0118] It should be noted that the interval distance is used to quantify the spatial correlation degree between the middle sub-area and the zero point area. The distance calculation is based on the fixed interval of sub-area division. For example, the kth sub-area is 2 fixed intervals away from the zero point area, and the interval distance is 2Δz. Since the closer to the zero point area, the more directly affected by the middle section environment fluctuation (oxygen partial pressure fluctuation transmission is more significant), the more critical it is to the blue and white porcelain color rendering, therefore, it needs higher compensation priority.

[0119] It should be noted that the spatial proximity is calculated by Gaussian function to realize the quantization relationship that the closer to the zero point area, the higher the proximity. The numerator of the Gaussian function is the square of the interval distance, and the denominator is the square of 2 times the preset bandwidth parameter η, which needs to be preset according to the kiln length and the number of sub-areas, for example, η takes 0.1L, L is the kiln length. The attenuation characteristics of the Gaussian function are consistent with the spatial distribution of the middle section environment influence: the environmental fluctuation (oxygen partial pressure fluctuation) of the zero point area has a weakening effect on the surrounding sub-areas with the increase of distance. Through this function, the interval distance can be converted into a spatial proximity between 0 and 1. For example, when the distance is 0, the proximity is 1, and when the distance increases, it gradually approaches 0.

[0120] It should be noted that the middle section oxygen partial pressure fluctuation value reflects the instability degree of the middle section environment: the larger the fluctuation value, the more violent the change of the oxygen partial pressure in the middle section of the kiln, which is a high-risk area of blue and white porcelain color deviation (needs to be paid attention to by knowledge graph and LLM). Normalization processing needs to convert the fluctuation value into a compensation weight between 0 and 1: if the fluctuation value is the maximum fluctuation of the middle section, the compensation weight is 1; if it is the minimum, it is 0. Thus, the fluctuation intensity is converted into a weight value that can be directly used for calculation, ensuring that the more unstable the middle section area, the higher the compensation weight. The compensation factor is the product of the spatial proximity and the compensation weight, to comprehensively consider the two dimensions of spatial correlation and environmental importance. Specifically: spatial proximity ensures that sub-areas close to the zero point get more attention; compensation weight ensures that sub-areas with more unstable environment get more attention. After multiplication, the compensation factor can reflect the differentiation of space (near high and far low) and the differentiation of environmental importance (fluctuation large high).

[0121] It should be noted that the preset position weight is an original weight without compensation, and the same weight has a lower actual attention due to the defects of the LLM model. By weighting the preset weight by the compensation factor, the weight of the middle section can be compensated, that is, the higher the compensation factor of the sub-region (near zero point and large fluctuation), the more significant the weight improvement. Thus, the low weight of the middle section caused by the defects of the LLM model is corrected, and the LLM bias is preliminarily corrected. The initial update position weight of the middle section sub-region and the original weight of the non-middle section are normalized as a whole, so that the overall distribution is reasonable after the weight adjustment, which not only improves the weight of the middle section, but also avoids excessive suppression of information in other regions due to the excessively high local weight. The final update position weight can accurately reflect the key areas that need to be focused on in the middle section, and provide a balanced weight basis for subsequent knowledge graph construction.

[0122] In an embodiment of the present application, based on the zero point region, the middle section oxygen partial pressure fluctuation value and the update position weight, a knowledge graph data set is constructed, comprising:

[0123] S41, define the kiln run entity corresponding to the observation time period as KilnRun, and the mth sub-region entity as KilnSection(m); and construct the attribute triple of the kiln run entity, including:

[0124] (KilnRun, hZPP, z * )

[0125] (KilnRun, hZPR, Omega0)

[0126] (KilnRun, hMOV, V mid )

[0127] Wherein, Z * represents the zero point region, Omega0 represents the middle section sub-region set, V mid represents the middle section oxygen partial pressure fluctuation value, hZPP represents the first predicate, hZPR represents the second predicate, and hMOV represents the third predicate;

[0128] S42, construct the attribute triple of the mth sub-region entity, including:

[0129] (KilnSection(m), hUPW, omega upd (m))

[0130] Wherein, omega upd (m) represents the update position weight of the mth sub-region, and bUPW represents the fourth predicate;

[0131] S43, combine the attribute triples of the kiln run entity and the sub-region entity to obtain a knowledge graph data set.

[0132] It should be noted that the KilnRun kiln entity represents a complete firing cycle, which is the core node in the graph that carries the overall state of the kiln; the KilnSection(m) mth sub-region entity corresponds to the sub-region divided longitudinally in the kiln, which is a node that carries the local space state. The combination of the two can clearly associate the firing process in the time dimension and the regional differences in the space dimension in the graph.

[0133] The essence of the triple is to bind the core physical parameters affecting the color of blue and white porcelain to the KilnRun entity through the structure of "entity → predicate → attribute", ensuring that key information is not lost.

[0134] (KilnRun, hZPP, z * ): The first predicate hZPP represents having a zero-point region position, and the zero-point region z* is the core of the environmental transition in the kiln. This triple associates the spatial coordinates of the environmental transition with the kiln, and clearly defines the core position of the middle section;

[0135] (KilnRun, hZPR, Ω0): The second predicate hZPR represents containing a middle section sub-region set, and the middle section sub-region set Ω0 is the middle section range that needs to be focused on. This triple clearly defines the spatial range that needs to be compensated in the kiln, and locks the key region for subsequent reasoning;

[0136] (KilnRun, hMOV, V mid ): The third predicate hMOV represents having a middle section oxygen partial pressure fluctuation value, and the middle section oxygen partial pressure fluctuation value V mid reflects the degree of instability of the middle section environment. This triple associates the coloration risk indicator with the kiln, ensuring that the graph can identify high-risk areas.

[0137] It should be noted that the triple is a semantic relationship that the graph can recognize.

[0138] (KilnSection(m), hUPW, ω upd (m)) where the fourth predicate hUPW represents having an updated position weight, and ω upd (m) is the updated position weight of the mth sub-region, reflecting the weight compensation of the middle section region, and the weight of the non-middle section region remains stable. This triple binds the importance of the sub-region in the graph to the sub-region entity, allowing the graph to distinguish the priority of different regions through the weight.

[0139] It should be noted that the core of the knowledge graph is a structured network of entity → relationship → attribute, and a single triple cannot fully express the knowledge association. It needs to be combined to form an organic whole.

[0140] The triples of the kiln sub-entity (overall state) and the triples of the sub-region entity (local state) are combined to form a complete link of "kiln sub-region key attribute": for example, the kiln sub-region has a zero point region, and the sub-region where the zero point region is located has a high update position weight. This combination enables the knowledge graph to reflect both the firing characteristics of the kiln as a whole and the importance of specific sub-regions, achieving macro-to-micro knowledge correlation. The final knowledge graph data set converts physical observations (temperature, oxygen partial pressure), spatial features (zero point region, middle section range), and weight information (update position weight) into structured semantic relationships, thereby solving the technical problem of the inability to effectively utilize middle section key information.

[0141] In an embodiment of the present application, the middle section feature constraint loss is calculated, and the knowledge graph data set is combined to construct a target knowledge graph, including:

[0142] S51, the region importance distribution of the mth sub-region is generated, as follows:

[0143]

[0144] Wherein, τ(m) represents the region importance distribution of the mth sub-region, and if the mth sub-region does not belong to the middle section sub-region set, S(m) is 0;

[0145] S52, the corresponding embedding vector e m of the mth sub-region entity is defined.

[0146] S53, the corresponding parameter matrix W p of each predicate is defined.

[0147] S54, the mth attribute triple is evaluated and processed, including:

[0148] f(m) = sigmoid(e m T W p e0)

[0149] Wherein, f(m) represents the rationality score of the mth attribute triple, sigmoid represents the sigmoid activation function, e k T represents the transpose vector of e k , e0∈{e t ,e f}, if the mth sub-region belongs to the middle section sub-region set, e0 takes the value of e t , and e t is the first object vector, if the mth sub-region does not belong to the middle section sub-region set, e0 takes the value of e f , and e f is the second object vector.

[0150] S55, obtain the rationality scores of the M sub-regions and perform normalization processing to obtain the normalized score of each sub-region; wherein the sum of the normalized scores of the M sub-regions is 1;

[0151] S56, calculate the middle section feature constraint loss, including:

[0152]

[0153] wherein, L mid represents the middle section feature constraint loss, and R(m) represents the normalized score of the mth sub-region;

[0154] S57, based on the middle section feature constraint loss, iteratively update the embedding vector e m , the parameter matrix W p , the first object vector e t and the second object vector e r by the gradient descent method until a preset number of iterations is reached, to obtain a target embedding vector and a target parameter matrix;

[0155] S58, solidify the target embedding vector, the target parameter matrix and the attribute triple of each sub-region to form a full target knowledge graph.

[0156] It should be noted that the region importance distribution defines the degree to which each sub-region should be concerned by the knowledge graph. wherein the numerator is the product of a compensation factor and a middle section oxygen partial pressure fluctuation value: the compensation factor reflects the spatial correlation and compensation demand of the sub-region with the zero point region (non-zero in the middle section and 0 in the non-middle section), and the middle section oxygen partial pressure fluctuation value reflects the instability degree of the middle section environment (the greater the fluctuation, the more critical the influence on color rendering). The product of the two means that: in the middle section sub-region, the closer to the zero point region and the greater the oxygen partial pressure fluctuation, the higher the region importance distribution value; the non-middle section sub-region has a region importance distribution of 0 due to the compensation factor being 0, indicating that it does not need to be concerned first. The denominator is the sum of the numerators of all sub-regions, used to normalize the region importance to the range of 0 to 1, ensuring that the importance of each sub-region can be directly compared. Thus, by quantifying the theoretical importance of the middle section key region, a target benchmark is provided for subsequent evaluation of whether the knowledge graph sufficiently focuses on the middle section, i.e., the actual expression of the subsequent knowledge graph needs to be close to the region importance distribution.

[0157] It should be noted that the embedding vector is to convert the sub-region entity KilnSection(m) into a computer-processable low-dimensional vector, and its core role is to express the semantic features of the sub-region through vector values. Through vectorization processing, the spatial features and importance of the sub-region can be recognized by the knowledge graph through vector operation. The parameter matrix is used to quantify the association strength of the predicate and the entity vector. Different predicates correspond to different matrices. The introduction of the parameter matrix makes the triple relationship of "entity predicate attribute" quantifiable.

[0158] It should be noted that the rationality score is used to judge whether the association of the sub-region entity and the attribute meets the expectation, and the calculation logic is differentiated for the middle section and the non-middle section. The formula sigmoid(e m T W p e0) is that the sigmoid function maps the vector operation result to 0 to 1 (0 represents unreasonable, and 1 represents reasonable); e m T W p e0 is the core operation: e m is the sub-region embedding vector, W p is the predicate parameter matrix, and e0 is the object vector. For the middle section sub-region: e t After optimization, the predicate matrix matching degree with the middle section attribute is higher, and the rationality score is closer to 1; for the non-middle section sub-region: e f The matching degree with the predicate matrix of the non-middle section attribute is higher, and the rationality score is closer to 0.

[0159] It should be noted that the rationality scores of the M sub-regions are normalized to obtain the normalized score, which converts the rationality of a single triple into the relative importance proportion of each sub-region in the global.

[0160] It should be noted that the middle section feature constraint loss is used to quantify the deviation loss between the sub-region actually concerned by the knowledge graph and the sub-region theoretically concerned, and to force the knowledge graph to pay attention to the middle section optimization.

[0161] The formula is In the formula, the square ensures that the deviation is amplified regardless of the sign, and the sum value reflects the global deviation. For the middle section, if the normalized score < the region importance distribution, the loss value will increase; for the non-middle section, because the region importance distribution is 0, if the normalized score is too large, the loss will also increase. The middle section feature constraint loss directly aims at the middle section knowledge weakening problem, so as to focus the optimization target on improving the normalized score of the middle section to the region importance distribution level through mathematical constraint, and avoid the natural deviation of the knowledge graph to the data sufficient non-middle section.

[0162] It should be noted that the embedding vector, the parameter matrix and the object vector are iteratively optimized based on the middle section feature constraint loss by gradient descent method, so as to minimize the loss and gradually approach the normalized score to the regional importance distribution.

[0163] It should be noted that the optimized target embedding vector, target parameter matrix and attribute triple are solidified to form a target knowledge graph. The target knowledge graph has the following beneficial effects:

[0164] 1. The embedding vector of the middle section sub-region is more strongly associated with the high importance attribute (such as high weight, oxygen partial pressure fluctuation).

[0165] 2. The triple of the middle section is more easily identified as reasonable and important during retrieval, thereby solving the problem of weak semantics and loose association in the middle section of the traditional knowledge graph and providing a structured knowledge base for the preferential calling of middle section information for LLM reasoning.

[0166] In an embodiment of the present application, an updated LLM question and answer model is obtained according to the target knowledge graph, a reasoning answer of the updated LLM question and answer model is obtained, and a middle section reference rate of the reasoning answer is determined, comprising:

[0167] S61, loading the target knowledge graph into the LLM question and answer model to obtain an updated LLM question and answer model;

[0168] S62, inputting a preset question to the updated LLM question and answer model; wherein the preset question includes a prompt word, and the prompt word requires the updated LLM question and answer model to give the cause of the question and the cause area;

[0169] S63, repeating S62 Q times to obtain Q sets of reasoning answers;

[0170] S64, determining the number q of cause areas belonging to the middle section sub-region set in the Q sets of reasoning answers to obtain a middle section reference rate U, and the middle section reference rate is

[0171] It should be noted that the target knowledge graph has strengthened the semantic expression of the middle section information through the middle section feature constraint loss. Loading it into the LLM question and answer model enables the LLM question and answer model to preferentially retrieve and call the middle section key evidence in the graph during reasoning, thereby distinguishing from the model without loading the target knowledge graph. The updated LLM question and answer model can directly obtain the middle section oxygen partial pressure fluctuation, high update position weight and other core information from the optimized graph, thereby providing a basic model for verifying whether the middle section knowledge is effectively utilized.

[0172] It should be noted that the preset question needs to focus on the core scene related to the coloration of blue and white porcelain. For example, what is the reason for the bluish glaze of blue and white porcelain? Where is the corresponding kiln area? The answer to the preset question is directly related to the kiln environment. By prompting the word, the cause and the cause area are required to be given, so as to force the model to output specific area information, that is, to avoid the model from only answering that the temperature is improper and avoiding the area, and to ensure that the middle section can be accurately counted whether it is cited.

[0173] It should be noted that a single inference may be affected by random factors, resulting in occasional high or low citation of the middle section, which cannot reflect the true effect. Repeatedly executing Q times (preferably 30 times) can reduce accidental errors through multiple sampling: if the target knowledge graph indeed strengthens the middle section information, the proportion of citing the middle section in multiple answers will tend to be stable; if it is not effectively strengthened, the proportion will continue to be low or fluctuate, so as to ensure that the calculation of the middle section citation rate can truly reflect the attention ability of the model to the middle section knowledge.

[0174] It should be noted that the middle section citation rate is used to quantify the actual calling proportion of the model to the middle section knowledge. If the target knowledge graph is effective, the model will frequently cite the middle section sub-area set when answering the coloration question. For example, the fluctuation of the oxygen partial pressure causes the cobalt material to be insufficiently reduced. If the target knowledge graph does not effectively strengthen the middle section, the model will still tend to cite the front section (such as too high temperature) or the rear section (such as insufficient suction).

[0175] In an embodiment of the present application, the quality evaluation result of the target knowledge graph is generated according to the middle section citation rate of the inference answer, comprising:

[0176] S71, taking the sum of the update position weights of the sub-areas in the middle section sub-area set as the expected citation rate;

[0177] S72, calculating the citation rate difference between the middle section citation rate of the inference answer and the expected citation rate;

[0178] S73, if the absolute value of the citation rate difference is greater than a preset quality threshold, the target knowledge graph is unqualified; if the absolute value of the citation rate difference is less than or equal to the preset quality threshold, the target knowledge graph is qualified.

[0179] It should be noted that the expected citation rate is the theoretical middle section citation proportion that the target knowledge graph should reach, and its value is equal to the sum of the update position weights of all sub-areas in the middle section sub-area set. The update position weight directly reflects the importance proportion of the sub-area in the target knowledge graph. If the weight sum of the middle section sub-area is higher, it means that the proportion of the target knowledge graph theoretically should take this area as the core evidence is higher. For example, the weight sum of the middle section is 0.4, which means that the proportion of the model inference citing the middle section should be close to 40%. The weight sum of the middle section is defined as the expected citation rate.

[0180] The reference rate difference is the difference between the actual middle section reference rate and the expected reference rate. Thus, the deviation between the proportion of the model actually referencing the middle section and the proportion expected by the target knowledge graph is quantified:

[0181] If the difference is positive and large, it means that the model excessively references the middle section;

[0182] If the difference is negative and large, it means that the model still does not sufficiently reference the middle section;

[0183] If the difference is close to 0, it means that the actual reference is consistent with the theoretical expectation.

[0184] The preset quality threshold is an acceptable deviation range set according to actual needs, used to judge whether the deviation is within a reasonable range:

[0185] If the absolute value of the reference rate difference is greater than the threshold, it means that the reference of the model to the middle section deviates too much from the design goal of the target knowledge graph, i.e., the target knowledge graph may not sufficiently strengthen the semantics of the middle section. For example, the embedding vector does not effectively highlight the middle section features, resulting in the model still ignoring the middle section, which is determined to be unqualified;

[0186] If the absolute value of the difference is less than or equal to the threshold, it means that the proportion of the model referencing the middle section is basically consistent with the expectation of the target knowledge graph, i.e., the middle section information is effectively called in reasoning, and the target knowledge graph achieves the expected design goal, which is determined to be qualified.

[0187] In an embodiment of the present application, if the target knowledge graph is unqualified, a reference rate loss is constructed, and S56 is returned; the reference rate loss and the middle section feature constraint loss are combined to form a comprehensive loss, and S57 is executed with the comprehensive loss until the target knowledge graph is qualified.

[0188] It should be noted that the core role of the reference rate loss is to quantify the reference rate difference between the actual middle section reference rate and the expected reference rate, and to convert the reference rate difference into the reference rate loss, including: the reference rate loss is the square of the reference rate difference, because: the square of the reference rate difference ensures that the deviation is equally punished regardless of the sign, avoiding optimization bias to one side.

[0189] The comprehensive loss is a weighted combination of the reference rate loss and the middle section feature constraint loss. For example, the reference rate loss is weighted by a first loss weight, and the middle section feature constraint loss is weighted by a second loss weight. The first loss weight and the second loss weight are both not 0, and the sum is 1.

[0190] Embodiment two:

[0191] A knowledge graph construction system based on traditional ceramic techniques, applied to any one of the knowledge graph construction methods based on traditional ceramic techniques, comprising:

[0192] The data acquisition module is configured to divide the kiln along a longitudinal direction into M sub-regions at fixed intervals, and to obtain observation data of each sub-region, wherein the observation data includes normalized temperature values and oxygen partial pressure values.

[0193] The zero point extraction module is configured to determine a zero point region in the M sub-regions and a mid-section oxygen partial pressure fluctuation value corresponding to the zero point region according to the observation data.

[0194] The weight compensation module is configured to compensate a preset position weight of each sub-region according to the mid-section oxygen partial pressure fluctuation value, to obtain an updated position weight of each sub-region, with the zero point region as the center.

[0195] The data construction module is configured to construct a knowledge graph data set based on the zero point region, the mid-section oxygen partial pressure fluctuation value, and the updated position weight.

[0196] The graph construction module is configured to calculate a mid-section feature constraint loss, and to construct a target knowledge graph by combining the knowledge graph data set.

[0197] The graph test module is configured to obtain an updated LLM question and answer model according to the target knowledge graph, to obtain an inference answer of the updated LLM question and answer model, and to determine a mid-section reference rate of the inference answer.

[0198] The graph evaluation module is configured to generate a quality evaluation result of the target knowledge graph according to the mid-section reference rate of the inference answer.

[0199] The above describes an embodiment of the present embodiment, but the present embodiment is not limited to the specific implementation described above, which is only illustrative and not limiting, and a person of ordinary skill in the art can make many forms under the inspiration of the present embodiment, which are all within the protection of the present embodiment.

Claims

1. A knowledge graph construction method based on traditional ceramic techniques, characterized by, The method comprises the following steps: S1, dividing the kiln into M sub-regions along the longitudinal direction at fixed intervals, and obtaining observation data of each sub-region; wherein the observation data comprises normalized temperature value and oxygen partial pressure value; S2, determining the zero point region in the M sub-regions according to the observation data, and the corresponding middle oxygen partial pressure fluctuation value of the zero point region, comprising: S21, determining the observation time period sampling time points; S22, at the tth moment of the observation time period, calculate the temperature value difference of the mth sub-region and the m+1th sub-region and the oxygen partial pressure value difference ; S23, for the mth sub-region and the m+1th sub-region, the statistical quantity is calculated by integrating the temperature value difference and the oxygen partial pressure value difference at each time point: ； wherein, is a statistical quantity of the mth sub-region in the observation time period, denotes a fixed interval; S24, along the front to the rear of the kiln, the product of the statistical quantities of adjacent sub-regions is calculated in the order of the sub-regions to extract the first sub-region with the product ≤0, and the sub-region is taken as the zero point region; S25, calculating the oxygen partial pressure value in the zero-point region. The variance at each sampling time is used to obtain the fluctuation value of oxygen partial pressure in the middle section; S3, taking the zero point region as the center, compensating the preset position weight of each sub-region according to the middle oxygen partial pressure fluctuation value to obtain the updated position weight of each sub-region; S4, based on the zero point region, the middle oxygen partial pressure fluctuation value and the updated position weight, a knowledge graph data set is constructed; S5, the middle feature constraint loss is calculated, and the target knowledge graph is constructed by combining the knowledge graph data set; S6, the updated LLM question and answer model is obtained according to the target knowledge graph, the inference answer of the updated LLM question and answer model is obtained, and the middle reference rate of the inference answer is determined; S7, the quality evaluation result of the target knowledge graph is generated according to the middle reference rate of the inference answer. 2.The knowledge graph construction method based on traditional ceramic techniques according to claim 1, wherein, Taking the zero point region as the center, compensating the preset position weight of each sub-region according to the middle oxygen partial pressure fluctuation value to obtain the updated position weight of each sub-region, comprising: S31, based on the zero point region, the middle sub-region set is extracted from the M sub-regions; S32, determine the interval distance from the kth sub-region in the middle sub-region set to the zero point region ; S33, depending on the interval distance The spatial proximity of the kth sub-region to the zero point region is calculated as follows: ； wherein, represents the spatial proximity of the kth sub-region to the zero-point region, represents a preset bandwidth parameter, is greater than 0; S34, the middle oxygen partial pressure fluctuation value is normalized to obtain the compensation weight; S35, multiplying the product of the compensation weight and the spatial proximity as a compensation factor of the kth sub-region ; S36, the preset position weight of the kth sub-region is weighted and compensated by the compensation factor to obtain the initial updated position weight of the kth sub-region; S37, the position weight of the sub-region in the non-middle sub-region set is kept unchanged, the initial updated position weight of each sub-region in the middle sub-region set is determined, and the position weight and the initial updated position weight are normalized to obtain the updated position weight of each sub-region. 3.The knowledge graph construction method based on traditional ceramic techniques according to claim 2, characterized in that, Based on the zero point region, the middle oxygen partial pressure fluctuation value and the updated position weight, a knowledge graph data set is constructed, comprising: S41, defining the kiln entity corresponding to the observation time period as KilnRun, and the mth sub-region entity as KilnSection(m); and constructing the attribute triple of the kiln entity, comprising: ； ； ； wherein, represents a zero point region, represents a middle sub-region set, represents a middle segment oxygen partial pressure fluctuation value, represents a first predicate, represents a second predicate, represents a third predicate; S42, constructing the attribute triple of the mth sub-region entity, comprising: ； wherein, denotes an update position weight of the mth sub-region, denotes a fourth predicate; S43, combining the attribute triple of the kiln entity and the attribute triple of the sub-region entity to obtain the knowledge graph data set. 4.The knowledge graph construction method based on traditional ceramic techniques according to claim 3, characterized in that, The middle feature constraint loss is calculated, and the target knowledge graph is constructed by combining the knowledge graph data set, comprising: S51, generating the region importance distribution of the mth sub-region, as follows: ； wherein, represents the region importance distribution of the mth sub-region, if the mth sub-region does not belong to the middle sub-region set, then is 0. S52, define the corresponding embedding vector for the mth sub-region entity ; S53, define a corresponding parameter matrix for each predicate ; S54, evaluating the mth attribute triple, comprising: ； wherein, represents a rationality score of the mth attribute triple, represents an activation function, represents a transposed vector of , if the mth sub-region belongs to the set of middle sub-regions, is valued as , is a first object vector, if the mth sub-region does not belong to the set of middle sub-regions, is valued as , is a second object vector; S55, obtaining the rationality score of the M sub-regions, and performing normalization processing to obtain the normalized score of each sub-region; wherein the sum of the normalized scores of the M sub-regions is 1; S56, calculate the middle section feature constraint loss, including: ； wherein, denotes the mid-section feature constraint loss, denotes the normalized score of the mth sub-region; S57, updating the embedding vector by gradient descent method based on the middle section feature constraint loss , a parameter matrix , a first object vector , and a second object vector , until a preset number of iterations is reached, to obtain a target embedding vector and a target parameter matrix; S58, solidify the target embedding vector, the target parameter matrix and the attribute triple of each sub-region to form a full target knowledge graph. 5.The knowledge graph construction method based on traditional ceramic techniques according to claim 4, characterized in that, According to the target knowledge graph, an updated LLM question and answer model is obtained, the inference answer of the updated LLM question and answer model is acquired, and the middle section reference rate of the inference answer is determined, including: S61, load the target knowledge graph into the LLM question and answer model to obtain an updated LLM question and answer model; S62, input a preset question into the updated LLM question and answer model; wherein the preset question includes a prompt word, which requires the updated LLM question and answer model to give the cause of the question and the cause area of the question; S63, repeat S62 Q times to obtain Q groups of inference answers; S64, determine the number q of the cause area in the Q-group reasoning answer belonging to the middle sub-region set, get the middle reference rate U, the middle reference rate is . 6.The knowledge graph construction method based on traditional ceramic techniques according to claim 5, characterized in that, According to the middle section reference rate of the inference answer, a quality evaluation result of the target knowledge graph is generated, including: S71, take the sum of the updated position weights of the sub-regions in the middle section sub-region set as the expected reference rate; S72, calculate the reference rate difference between the middle section reference rate of the inference answer and the expected reference rate; S73, if the absolute value of the reference rate difference is greater than a preset quality threshold, the target knowledge graph is unqualified; if the absolute value of the reference rate difference is less than or equal to the preset quality threshold, the target knowledge graph is qualified. 7.The knowledge graph construction method based on traditional ceramic techniques according to claim 6, characterized in that, If the target knowledge graph is unqualified, a reference rate loss is constructed, and S56 is returned; the reference rate loss and the middle section feature constraint loss are combined to form a comprehensive loss, and S57 is executed with the comprehensive loss until the target knowledge graph is qualified.

8. A knowledge graph construction system based on traditional ceramic technology, applied to the knowledge graph construction method based on traditional ceramic technology in any one of claims 1-7, characterized in that, Including: The data acquisition module is configured to divide the kiln along the longitudinal direction into M sub-regions at fixed intervals, and acquire observation data of each sub-region; wherein the observation data includes normalized temperature values and oxygen partial pressure values; The zero point extraction module is configured to determine a zero point region in the M sub-regions and a middle section oxygen partial pressure fluctuation value corresponding to the zero point region according to the observation data; The weight compensation module is configured to compensate a preset position weight of each sub-region according to the middle section oxygen partial pressure fluctuation value, taking the zero point region as the center, to obtain an updated position weight of each sub-region; The data construction module is configured to construct a knowledge graph data set based on the zero point region, the middle section oxygen partial pressure fluctuation value and the updated position weight; The graph construction module is configured to calculate a middle section feature constraint loss, and construct a target knowledge graph by combining the knowledge graph data set; The graph test module is configured to obtain an updated LLM question and answer model according to the target knowledge graph, acquire an inference answer of the updated LLM question and answer model, and determine a middle section reference rate of the inference answer; The graph evaluation module is configured to generate a quality evaluation result of the target knowledge graph according to the middle section reference rate of the inference answer.

Images