Crop disease and pest knowledge base self-learning and incremental updating method and system

Abstract

The application discloses a crop disease and pest knowledge base self-learning and incremental updating method and system, belongs to the technical field of crop disease and pest diagnosis and knowledge base management, and has the technical scheme as follows: a first knowledge network is constructed through three-level credibility weighted fusion and graph node matching; the quality score and the compensated use frequency of a knowledge item are obtained through frequency statistics and multi-condition threshold value; a second knowledge network containing a new node and an associated edge is obtained through attention heat map extraction and attribute node generation; a third knowledge network and a rollback event record are obtained through semantic version control and execution rollback condition judgment; a fourth knowledge network and a migration record are obtained through subgraph clone adaptation for cross-crop knowledge migration and graph network fusion; and the application is based on knowledge network evolution and feedback driving, realizes knowledge base self-learning and incremental updating through exposure compensation, version control, cross-crop migration and other technologies, and improves knowledge coverage, updating efficiency and application reliability.

Description

Technical Field

[0001] This invention relates to the field of crop disease and pest diagnosis and knowledge base management technology, and more specifically to a self-learning and incremental update method and system for a crop disease and pest knowledge base. Background Technology

[0002] With the deep integration of agricultural intelligence and big data technologies, large-scale agricultural models are increasingly being applied in areas such as intelligent question answering, agricultural technology decision-making, and disease diagnosis. Current research mainly focuses on model training optimization and the construction of evaluation systems. For example, techniques such as dynamic word embedding disambiguation and regional language pattern enhancement coding are used to train large-scale Transformer architecture models, or agricultural knowledge graphs are constructed for factual verification. However, these methods often treat the knowledge base as a static resource, lacking a closed-loop linkage with actual application scenarios, and making it difficult to leverage multi-source feedback data from users, experts, and other sources to drive dynamic iteration of knowledge. Agricultural knowledge has significant timeliness and regionality; pests and diseases are constantly mutating, and crop varieties are continuously updated. Existing technologies have failed to design adaptive update mechanisms for these characteristics, resulting in a decline in the accuracy of the knowledge base as the environment changes. In addition, the lack of technical design for quality assessment, version control, and cross-scenario migration in the knowledge management process makes it difficult to continuously optimize and reuse the knowledge base, failing to meet the requirements of precision agriculture for real-time, reliable, and generalizable knowledge. Therefore, existing technologies have shortcomings. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention aims to provide a self-learning and incremental update method and system for a crop pest and disease knowledge base. The method involves collecting raw feedback data through a multi-source feedback interface, constructing a first knowledge network based on three-level credibility weighted fusion and graph node matching, and extracting consensus knowledge. Based on system records and the first knowledge network, a quality score and compensated usage frequency are obtained through frequency statistics and multi-condition threshold judgment. Based on the first knowledge network and the compensated usage frequency, a second knowledge network is obtained through dynamic weight adjustment and node fusion via attention heatmap extraction and attribute node generation. Based on the second knowledge network and update records, a third knowledge network is obtained through semantic version control, quality indicator differential calculation, and rollback condition judgment. Based on the third knowledge network and actual application feedback, a fourth knowledge network is obtained through cross-crop knowledge transfer and graph network fusion via multi-dimensional semantic similarity calculation and subgraph cloning adaptation. This achieves closed-loop self-learning and incremental update of the knowledge base, solving the problems of static knowledge bases, lack of closed-loop feedback mechanisms, difficulty in dynamically evaluating knowledge quality, low update efficiency, and poor cross-scenario reusability in existing technologies. This improves the knowledge base's adaptability, knowledge accuracy, update security, and cross-crop knowledge reusability.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] This invention provides a method for self-learning and incremental updating of a crop disease and pest knowledge base, including:

[0006] Raw feedback data is collected based on a multi-source feedback interface. The first knowledge network is constructed and consensus knowledge is extracted through three-level credibility weighted fusion and graph node matching.

[0007] Based on system records and the first knowledge network, the quality score of each knowledge item is obtained through frequency statistics and multi-condition threshold judgment, and the usage frequency after compensation is calculated.

[0008] Based on the first knowledge network and the usage frequency after compensation, the first knowledge network is dynamically weighted and fused by extracting attention heatmaps and generating attribute nodes to obtain a second knowledge network containing newly added nodes and associated edges.

[0009] Based on the second knowledge network and update records, the third knowledge network and rollback event records are obtained by performing semantic version control and quality indicator differential calculation, and then executing rollback condition judgment.

[0010] Based on the third knowledge network and feedback from practical applications, cross-crop knowledge transfer and graph network fusion are performed through multi-dimensional semantic similarity calculation and subgraph cloning adaptation to obtain the fourth knowledge network and transfer records.

[0011] As a further improvement of the present invention, the step of collecting raw feedback data based on a multi-source feedback interface, constructing a first knowledge network and extracting consensus knowledge through three-level credibility weighted fusion and graph node matching includes:

[0012] The original feedback data is collected based on the multi-source feedback interface. The final credibility of each piece of original feedback data is obtained by three-level credibility weighted fusion, which includes three-level credibility assignment and weighted average fusion.

[0013] Based on the attribute dimension division in agricultural knowledge and agricultural ontology, an initial list of entity nodes and a set of concept nodes are generated to construct an initial knowledge network skeleton.

[0014] Based on the original feedback data and the final credibility, entities and attribute phrases are extracted through natural language processing. The attribute phrases are then semantically matched with the concept nodes in the knowledge network skeleton. The weights of the associated edges are dynamically updated according to the co-occurrence relationship of node pairs in the same feedback content, thus obtaining the first knowledge network and consensus knowledge.

[0015] As a further improvement of the present invention, the step of obtaining a quality score for each knowledge item and calculating the compensated usage frequency based on system records and the first knowledge network through frequency statistics and multi-condition threshold judgment includes:

[0016] Based on system records, the usage frequency score of each knowledge item is obtained through frequency statistics and ranking normalization.

[0017] Based on the correctness indicators of the diagnostic results in the original feedback data, the diagnostic accuracy score of each knowledge item is obtained by counting the number of positive and negative feedbacks.

[0018] A timeliness score is obtained based on the difference between the last update time of each knowledge item and the system record time, using a segmented scoring standard.

[0019] Based on the usage frequency score, diagnostic accuracy score, and timeliness score, a quality score for each knowledge item is obtained through weighted averaging and multi-condition threshold judgment.

[0020] Based on the usage frequency recorded in the system, the compensated usage frequency of each knowledge item is calculated through underexposure coefficient compensation.

[0021] As a further improvement of the present invention, the calculation of the compensated usage frequency of each knowledge item based on the usage frequency recorded in the system through underexposure coefficient compensation includes:

[0022] Based on system records, the exposure underexposure coefficient of each knowledge item is obtained by calculating the ratio of the number of times a knowledge item is judged as relevant but not included in the return list to the total number of times it is judged as relevant.

[0023] Based on the original usage frequency of all knowledge items in the knowledge base, the baseline frequency is obtained by statistically analyzing the median.

[0024] Based on the original usage frequency, underexposure coefficient, and reference frequency of each knowledge entry, the product of the underexposure coefficient and the reference frequency is weighted and summed using a preset compensation coefficient to obtain the compensated usage frequency.

[0025] As a further improvement of the present invention, the step of dynamically adjusting the weights and fusing nodes of the first knowledge network based on the first knowledge network and the compensated usage frequency, through attention heatmap extraction and attribute node generation, to obtain a second knowledge network containing newly added nodes and associated edges, includes:

[0026] Based on the first knowledge network and the original feedback data, the set of salient regions and the average attention value of each salient region are obtained through similarity comparison, attention heatmap extraction and connected component analysis.

[0027] Based on the set of salient regions, the first knowledge network after adding attribute nodes is obtained through multimodal description model, semantic similarity matching, initial association edge establishment and one-hop neighborhood diffusion;

[0028] Based on the original feedback data, the first knowledge network after adding attribute nodes, and the usage frequency after compensation, the first knowledge network after dynamic weight adjustment is obtained through negative feedback positioning and comprehensive weight adjustment.

[0029] Based on the first knowledge network after dynamic weight adjustment, a second knowledge network and newly added nodes and associated edge records are obtained through weighted calculation of vector similarity and text similarity and node merging operation.

[0030] As a further improvement of the present invention, the step of obtaining a set of salient regions and the average attention value of each salient region based on the first knowledge network and the original feedback data through similarity comparison, attention heatmap extraction and connected component analysis includes:

[0031] Based on the first knowledge network, a new knowledge extraction trigger signal is obtained by similarity comparison and by combining the user confirmation identifier and final credibility in the original feedback data.

[0032] Based on the new knowledge, trigger signals and their corresponding multimodal diagnostic results are extracted. The average weight of attention is extracted through a multimodal large model to obtain an attention heatmap.

[0033] Based on the attention heatmap, a set of salient regions and the average attention value of each salient region are obtained through threshold segmentation and connected component analysis.

[0034] As a further improvement of the present invention, the first knowledge network obtained based on the set of salient regions, through a multimodal description model, semantic similarity matching, initial association edge establishment, and one-hop neighborhood diffusion, after adding attribute nodes, includes:

[0035] Based on the set of salient regions, a text description is generated and attribute phrases are extracted using a multimodal description model to obtain a set of candidate attribute phrases;

[0036] Based on the candidate attribute phrase set and the concept node set in the first knowledge network, the attribute phrases are classified into the corresponding concept nodes by semantic similarity matching to obtain the new attribute node set and its corresponding attention mapping weights.

[0037] Based on the newly added attribute node set and its corresponding attention mapping weights, the first knowledge network after adding attribute nodes is obtained by establishing initial associated edges and one-hop neighborhood diffusion.

[0038] As a further improvement of the present invention, the step of obtaining a third knowledge network and rollback event record based on the second knowledge network and update record, through semantic version control and quality indicator differential calculation, and then performing rollback condition judgment, includes:

[0039] Based on the second knowledge network and update records, semantic version numbers are generated through semantic version control and changes in newly added nodes, deleted nodes and associated edges are extracted. Incremental snapshot data is obtained through serialization.

[0040] Based on the second knowledge network and update records, regional information is extracted through regular expression matching of geographic keywords to obtain regional branch version records;

[0041] Based on system logs, the changes in quality indicators are obtained through time-series database queries and quality indicator difference calculations. The changes in quality indicators include changes in accuracy, changes in complaint rate, and changes in error rate.

[0042] Based on the changes in the quality indicators and the preset change threshold, a rollback operation is performed through Boolean logic to obtain the third knowledge network and rollback event records.

[0043] As a further improvement of the present invention, based on the third knowledge network and feedback from actual applications, cross-crop knowledge transfer and graph network fusion are performed through multi-dimensional semantic similarity calculation and subgraph cloning adaptation to obtain a fourth knowledge network and transfer records, including:

[0044] Based on the names of pests and diseases, symptom descriptions, and control methods of the source and target crops in the third knowledge network, a multi-dimensional semantic similarity calculation is performed using a bidirectional encoder representation method to obtain a comprehensive similarity.

[0045] Based on the comprehensive similarity, expert verification records and actual application feedback records in the original feedback data, a migration credibility score is obtained through ratio statistics and weighted evaluation.

[0046] Based on the source crop subgraph in the third knowledge network, subgraph cloning and adaptation are performed by querying the threshold of associated edges, replacing entity node identifiers, and scaling the weight of associated edges to obtain the adapted subgraph to be accessed.

[0047] Based on the subgraph to be accessed and the target crop subgraph in the third knowledge network, cross-crop knowledge transfer and graph network fusion are performed through semantic similarity threshold matching, associated edge merging and node insertion to obtain the fourth knowledge network and transfer records.

[0048] This invention provides a self-learning and incremental update system for a crop disease and pest knowledge base, comprising:

[0049] The feedback fusion module collects raw feedback data based on multi-source feedback interfaces, constructs the first knowledge network and extracts consensus knowledge through three-level credibility weighted fusion and graph node matching;

[0050] The quality assessment module, based on system records and the first knowledge network, obtains the quality score of each knowledge item and calculates the usage frequency after compensation through frequency statistics and multi-condition threshold judgment.

[0051] The incremental update module, based on the first knowledge network and the compensated usage frequency, performs dynamic weight adjustment and node fusion on the first knowledge network through attention heatmap extraction and attribute node generation to obtain a second knowledge network containing newly added nodes and associated edges.

[0052] The version migration module, based on the second knowledge network and update records, performs semantic version control and quality index differential calculation, and then executes rollback condition judgment to obtain the third knowledge network and rollback event records; based on the third knowledge network and actual application feedback, it performs cross-crop knowledge migration and graph network fusion through multi-dimensional semantic similarity calculation and subgraph cloning adaptation to obtain the fourth knowledge network and migration records.

[0053] This invention is based on a knowledge network construction and evolution mechanism driven by multi-source feedback data. Through core technologies such as three-level credibility weighted fusion, attention heatmap extraction, semantic version control, and cross-crop subgraph cloning adaptation, and based on quantitative criteria such as feedback credibility, knowledge usage frequency, quality assessment indicators, and multi-dimensional semantic similarity, it realizes a closed-loop process of automated construction, incremental update, version rollback, and cross-crop knowledge migration from multi-source feedback data to a structured knowledge network. It solves the problems of lagging updates, underestimation of long-tail knowledge, difficulty in version management, and low efficiency of cross-crop knowledge reuse in traditional knowledge bases. It improves the self-learning ability, incremental update efficiency, accuracy of knowledge quality assessment, and reliability of cross-crop knowledge migration of the knowledge base, and enhances the accuracy and applicability of crop disease and pest diagnosis. Attached Figure Description

[0054] Figure 1 This is a flowchart illustrating the steps of the self-learning and incremental update method for the crop disease and pest knowledge base of this invention.

[0055] Figure 2 A flowchart illustrating the steps of obtaining a quality score for each knowledge item and calculating the usage frequency after compensation in this invention.

[0056] Figure 3 This is a flowchart illustrating the steps of new knowledge injection and network update based on attention heatmaps in this invention.

[0057] Figure 4 This is a flowchart illustrating the steps of cross-crop knowledge transfer and graph network fusion in this invention.

[0058] Figure 5 This is a schematic diagram of the structure of the crop disease and pest knowledge base self-learning and incremental update system of the present invention. Detailed Implementation

[0059] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof.

[0060] Identical parts are indicated by the same reference numerals. It should be noted that the terms "front," "rear," "left," "right," "up," and "down" used in the following description refer to directions in the accompanying drawings, while the terms "bottom surface," "top surface," "inner," and "outer" refer to directions toward or away from the geometric center of a specific part, respectively.

[0061] The term "and / or" in the following text is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0062] like Figure 1 As shown, this invention provides a self-learning and incremental update method for a crop disease and pest knowledge base, including:

[0063] Raw feedback data is collected based on a multi-source feedback interface. The first knowledge network is constructed and consensus knowledge is extracted through three-level credibility weighted fusion and graph node matching.

[0064] Based on system records and the first knowledge network, the quality score of each knowledge item is obtained through frequency statistics and multi-condition threshold judgment, and the usage frequency after compensation is calculated.

[0065] Based on the first knowledge network and the usage frequency after compensation, the first knowledge network is dynamically weighted and fused by extracting attention heatmaps and generating attribute nodes to obtain a second knowledge network containing newly added nodes and associated edges.

[0066] Based on the second knowledge network and update records, the third knowledge network and rollback event records are obtained by performing semantic version control and quality indicator differential calculation, and then executing rollback condition judgment.

[0067] Based on the feedback from the third knowledge network and practical applications, cross-crop knowledge transfer and graph network fusion are carried out through multi-dimensional semantic similarity calculation and subgraph cloning adaptation to obtain the fourth knowledge network and transfer records.

[0068] The multi-source feedback interface is an application programming interface endpoint used to receive feedback data related to the diagnosis of different types of pests and diseases. It includes three API endpoints: a user feedback interface, an expert feedback interface, and an actual effect feedback interface. Each interface receives feedback data from different sources and dimensions. The raw feedback data consists of feedback information related to the entire process of crop pest and disease diagnosis collected through the multi-source feedback interface. This includes user-side diagnostic result verification, control effect evaluation, and improvement suggestions; expert-side diagnostic correction opinions, knowledge base review opinions, and professional knowledge supplements; and the actual application side's control measures usage, effectiveness evaluation, and actual control results. This data is collected through preset parameters of each API endpoint and uniformly stored in the database. A three-level credibility weighted fusion is used to process the raw feedback data from different sources. The system employs a trusted quantification and fusion computing approach to achieve credible stratification and comprehensive evaluation of multi-source feedback data, providing a reliable data foundation for subsequent knowledge network construction. Graph node matching involves node retrieval and matching operations based on a fine-tuned BERT-Agr model and semantic similarity calculation using agricultural corpora. This maps feedback data to the node system of the knowledge network, enabling the connection between feedback data and the knowledge network. The first knowledge network is a heterogeneous graph network in the field of crop diseases and pests, constructed using a graph database as its storage medium. It consists of three types of nodes: entity nodes, attribute nodes, and concept nodes, as well as membership edges connecting attribute nodes and concept nodes, and association edges connecting various types of nodes. Consensus knowledge comprises disease and pest diagnosis knowledge extracted from the topology of the first knowledge network that has been verified by multi-source feedback and has high correlation strength.

[0069] The system records various operation and feedback-related log data generated during the operation of the knowledge base, including retrieval logs, feedback data tables, and knowledge item update logs. This data is obtained through real-time recording and storage of operations such as querying, calling, and receiving feedback from the knowledge base. Frequency statistics are a method of statistically calculating the retrieval and usage of each knowledge item within a specified time range, used to quantify the actual application frequency of each knowledge item. Multi-condition thresholds are multiple quantitative judgment thresholds set for knowledge item quality assessment, including preset quality score thresholds, diagnostic accuracy thresholds, retrieval time thresholds, and negative feedback rate thresholds, used as criteria for identifying low-quality knowledge content and achieving tiered screening of knowledge item quality. The quality score for each knowledge item is a quantitative evaluation value that comprehensively considers three dimensions: usage frequency, diagnostic accuracy, and timeliness, used to comprehensively quantify the overall quality level of the knowledge item. Compensated usage frequency is the frequency value corrected by introducing an exposure compensation mechanism to adjust the actual usage frequency of the knowledge item, used to correct the insufficient exposure of long-tail knowledge caused by sorting truncation, and more realistically reflect the actual usage needs of the knowledge item.

[0070] The attention heatmap is a heatmap of the same size as the input image, generated from the last layer of self-attention average weights extracted from the multimodal large model. Each pixel corresponds to an attention value of 0-1, used to locate salient feature regions of pests and diseases in the image. Attribute nodes are node types in the knowledge network that represent specific descriptive values ​​such as symptoms and prevention measures. Dynamic weight adjustment is a processing method that updates the weights of associated edges in the first knowledge network in batches based on the knowledge item quality assessment index, used to achieve dynamic matching between the weights of associated edges and the quality of knowledge items and optimize the knowledge network topology. Node fusion is a processing operation that merges highly similar attribute nodes under the same concept node in the first knowledge network, used to eliminate node redundancy in the knowledge network and improve the simplicity and consistency of the knowledge network. The second knowledge network is an updated heterogeneous graph network obtained by extracting attention heatmaps, generating new attribute nodes, dynamic weight adjustment and node fusion based on the first knowledge network. It includes nodes and associated edges added due to the extraction of new knowledge, as well as the network topology optimized by weight adjustment and node fusion.

[0071] The update log is a detailed record of all update operations from the first knowledge network to the second knowledge network, including new node information, changes in associated edge weights, node fusion operations, and knowledge correction content. Semantic version control is a version management method that expands the version control object from knowledge items to the node set and edge weight matrix of the second knowledge network. It is used to accurately record the evolution process of the second knowledge network and supports independent version control for multiple regional branches. Quality index differential calculation is a method of calculating the difference between quality monitoring indicators corresponding to different versions of the knowledge network. It is used to quantitatively evaluate the impact of version updates on the application effect of the knowledge network. Rollback condition judgment is a version rollback trigger judgment operation based on the quality index differential calculation results. It is used to promptly identify version updates that do not meet quality standards to ensure the application stability of the knowledge base. The third knowledge network is a stable heterogeneous graph network obtained after performing a version rollback operation on the second knowledge network. The rollback event log is a detailed record of the version rollback operation, including the rollback trigger reason, version information before and after the rollback, the restored content of node and edge weights, and the rollback execution time.

[0072] Practical application feedback refers to the feedback data collected by the third knowledge network during the practical application of cross-crop pest and disease diagnosis, including verification of diagnostic results on different crops, evaluation of the effectiveness of control measures, and feedback on knowledge transfer adaptability. Multi-dimensional semantic similarity calculation is a method for quantifying the semantic similarity of pest and disease knowledge across different crops from three dimensions: name, symptom description, and control method. This is used to identify transferable similar pest and disease knowledge between different crops. Subgraph cloning and adaptation is a process of cloning and adapting similar knowledge subgraphs from the source crop knowledge network to the target crop knowledge network, enabling rapid reuse and adaptation of similar knowledge across different crops. Cross-crop knowledge transfer is the process of reusing mature pest and disease diagnosis knowledge from the source crop to the target crop knowledge base, used to compensate for deficiencies in the content of the target crop knowledge base. The system achieves cross-crop knowledge reuse; graph network fusion is the operation of merging the cross-crop knowledge subgraphs after subgraph cloning and adaptation with the original knowledge network of the target crop, which is used to achieve the organic integration of cross-crop knowledge and the original knowledge of the target crop to form a unified knowledge network; the fourth knowledge network is a cross-crop fused heterogeneous graph network obtained by cross-crop knowledge transfer and graph network fusion based on the third knowledge network, which integrates the disease and pest diagnosis knowledge of different crops and has the ability to reuse cross-crop knowledge; the migration record is a detailed record of the entire process of cross-crop knowledge transfer, including source crop and target crop information, similarity calculation results, migration credibility assessment value, subgraph cloning and adaptation content, network fusion operation and migration knowledge annotation, which is obtained by the system recording each step of the knowledge transfer operation in real time.

[0073] This embodiment assesses the credibility of multi-source feedback data by establishing a three-level credibility weighted fusion mechanism. It constructs a knowledge relationship network composed of entity nodes, attribute nodes, and concept nodes, and uses incremental calculation of associated edge weights to achieve feedback-driven knowledge evolution. It introduces attention heatmap extraction and a networked cold start mechanism to automatically inject new knowledge and assign initial weights. It corrects the underestimation of long-tail knowledge through exposure underestimation coefficient calculation and compensation formulas. It achieves regionally sensitive knowledge network version management and automatic rollback through semantic version control and quality indicator differential calculation. It enables cross-crop knowledge transfer and reuse and graph network fusion through multi-dimensional semantic similarity calculation and subgraph cloning adaptation. This realizes a leap from passive entry storage to active conceptual evolution in the crop pest and disease knowledge base, improving its self-learning ability, incremental update efficiency, and cross-crop knowledge reuse capability, while ensuring the accuracy, stability, and applicability of the knowledge base in practical applications.

[0074] Furthermore, this embodiment provides a method for collecting raw feedback data based on a multi-source feedback interface, constructing a first knowledge network and extracting consensus knowledge through three-level credibility weighted fusion and graph node matching. The steps include:

[0075] The original feedback data is collected based on the multi-source feedback interface. The final credibility of each piece of original feedback data is obtained by three-level credibility weighted fusion, which includes three-level credibility assignment and weighted average fusion.

[0076] Based on the attribute dimension division in agricultural knowledge and agricultural ontology, an initial list of entity nodes and a set of concept nodes are generated to construct an initial knowledge network skeleton.

[0077] Based on the original feedback data and the final credibility, entities and attribute phrases are extracted through natural language processing. The attribute phrases are semantically matched with the concept nodes in the knowledge network skeleton. The weights of the associated edges are dynamically updated according to the co-occurrence relationship of node pairs in the same feedback content, thus obtaining the first knowledge network and consensus knowledge.

[0078] The three-level credibility weighted fusion is a comprehensive processing method that performs hierarchical assignment and fusion calculation of the credibility of multi-source feedback data to achieve a quantitative assessment of the credibility of feedback data from different sources. It includes two core steps: three-level credibility assignment and weighted average fusion. The three-level credibility assignment is a method of assigning different basic credibility values ​​according to the source type of feedback data. This is achieved by assigning differentiated basic credibility values ​​to three types of sources: expert feedback, user feedback, and system inference feedback. The weighted average fusion is a fusion algorithm that combines the basic credibility of each feedback and its corresponding weight to calculate the final credibility of multiple feedback data for the same diagnostic task. This is achieved by comprehensively calculating the credibility of multi-source feedback for the same diagnostic task through a formula of weighted summation and then taking the average. The final credibility of each feedback data is the quantified credibility value obtained after the three-level credibility weighted fusion of a single feedback data.

[0079] The agricultural knowledge system comprises a professional knowledge system related to the diagnosis of crop diseases and pests, including basic professional information on various crops, diseases and pests, symptoms, and control measures. This information is compiled and organized through agricultural professional literature, industry standards, and expert experience. The attribute dimension division in agricultural ontology is a dimensional division method for the attribute characteristics of crop disease and pest diagnosis knowledge, obtained through summarizing the core characteristic dimensions of disease and pest diagnosis and expert verification in the agricultural field. The initial list of entity nodes is the initial basic dataset of entity nodes in the knowledge network, containing information on common crop and disease / pest types. The set of concept nodes is the basic dataset of concept nodes in the knowledge network, containing various attribute dimension information of disease and pest diagnosis knowledge. The initialized knowledge network skeleton is the basic structure of the knowledge network without feedback data, possessing only basic nodes and preset edge relationships.

[0080] Natural Language Processing (NLP) parsing is the process of extracting entity and attribute information from feedback text. It is used to extract structured knowledge network node information from unstructured feedback text. Entities are specific objects related to crops and pests extracted from the original feedback data text, obtained through the entity extraction step of NLP parsing. Attribute phrases are descriptive text fragments related to symptoms and prevention measures for pest diagnosis extracted from the original feedback data text, obtained through the attribute information extraction step of NLP parsing. Concept nodes are node types that represent abstract attribute dimensions in the knowledge network. The co-occurrence relationship of node pairs in the same feedback content is the association relationship between two nodes that appear simultaneously in a single feedback text. The association edge weight is the weight value of the association edge connecting two nodes in the knowledge network, which is dynamically updated based on the co-occurrence relationship of node pairs in the same feedback content and the final credibility of the feedback data.

[0081] Specifically, the raw feedback data collected from the multi-source feedback interface is first subjected to a three-level credibility weighted fusion. A basic credibility score is assigned to the feedback data from different sources using the three-level credibility assignment, where the basic credibility score for expert feedback is assigned as follows: The basic credibility of user feedback is assigned a value. The basic credibility assignment of the system inference feedback is... ,and Then, by weighted averaging, the results of the same diagnostic task are fused. Feedback data, based on the formula The final credibility of each piece of feedback data is calculated using a formula derived from a weighted arithmetic mean formula combined with the actual needs of feedback data credibility assessment. For the first The basic credibility of the feedback. For the first The source weight of each feedback item , The total number of feedback data is assigned a one-to-one correspondence between the source weight and the basic credibility assignment rules, and the final credibility calculation and storage of all original feedback data are completed.

[0082] Then, the results of the attribute dimension division in agricultural knowledge and agricultural ontology are sorted out. Common crops and pests are extracted to construct an initial list of entity nodes. Attribute dimensions for pest diagnosis, such as lesion characteristics, location of occurrence, and type of control agent, are extracted to construct a set of concept nodes. At the same time, the membership edge relationship between attribute nodes and concept nodes is predefined. The membership edge is set according to the attribute dimension association rules of agricultural ontology. In this way, an initial knowledge network skeleton containing only basic nodes and pre-set membership edges is constructed. This network skeleton uses a graph database as the storage carrier and contains two types of basic nodes, entity nodes and concept nodes, and one type of basic edge relationship, membership edges.

[0083] Next, the raw feedback data used for the final credibility calculation was parsed using natural language processing. The BERT-Agr model, finely tuned on a corpus in the field of crop diseases and pests, was used to extract entities and attribute phrases. This model uses the basic BERT model as its backbone and adds an agricultural domain vocabulary embedding layer to the model's encoding layer. The input layer uses dual features of word segmentation and part-of-speech tagging. The hidden layer has 12 encoder layers and 16 attention heads. The training corpus consists of 100,000 texts related to the diagnosis of crop diseases and pests. The training batch size is 32, the learning rate is 0.00001, and the number of training epochs is 20. The model input is the text string of the raw feedback data, and the output is a text sequence with entity types and attribute phrase boundaries labeled. Entities are divided into two categories: crop entities and disease entities. Attribute phrases are divided into two categories: symptom description phrases and prevention and control measure phrases.

[0084] The extracted attribute phrases are semantically matched with the concept nodes in the initialized knowledge network skeleton, using an improved cosine similarity formula. Computational attribute phrases With concept nodes The semantic similarity formula is based on the fundamental cosine similarity formula and incorporates a semantic correction coefficient for the agricultural domain. It is derived that, For the BERT-Agr model fine-tuned for the agricultural field, attribute phrases The semantic feature vector obtained by encoding. For the BERT-Agr model fine-tuned for the agricultural domain, the concept nodes The semantic feature vector is obtained by encoding the standard descriptive text; among which the semantic correction coefficient for the agricultural field is... The value ranges from 0.9 to 1.05, with a preferred value of 0.95. This value is used to correct semantic matching deviations in the context of crop pests and diseases using general cosine similarity, and to compensate for the insufficient fit of general semantic encoding with agricultural terminology and pest-specific expressions. If the maximum semantic similarity between an attribute phrase and a concept node is greater than or equal to a preset semantic similarity threshold... If the semantic similarity is less than the preset semantic similarity threshold, then the attribute phrase is assigned to the corresponding concept node. If no corresponding attribute node exists, a new attribute node is created and a membership edge is established between it and the concept node. If the attribute phrase is positive, it is marked as a candidate for a new concept. The semantic similarity threshold is the core threshold for distinguishing whether an attribute phrase can be categorized into the existing concept system or marked as a candidate for a new concept; its basic value range is [value range missing]. The basic values ​​are determined by testing the BERT-Agr model on the labeled test set in the field of agricultural pests and diseases. The similarity value is based on the peak value corresponding to the harmonic average of the model's semantic matching accuracy and recall. It can be dynamically adjusted according to the knowledge base iteration stage and concept granularity requirements.

[0085] Finally, based on the co-occurrence relationship of node pairs in the same feedback content, and combined with the final credibility of the feedback data, the weights of the association edges in the knowledge network are dynamically updated. First, a set of node pairs extracted from a single feedback data point is constructed, where a node pair is any combination of two different nodes in the set. Then, the increment of the association edge weights is calculated according to the positive or negative type of the feedback. The formula for the weight increment of positive feedback is... This formula is derived based on the basic exponential decay formula combined with feedback confidence and node path distance, where For learning rate, To determine the final credibility of the feedback data, Distance attenuation coefficient, For the first Nodes in round iteration and Shortest path distance For the node co-occurrence frequency correction coefficient, This represents the current iteration number, indicating the current weight update's position in the iteration. Each feedback data processing cycle corresponds to a time-series identifier of the network state before the weight update. The distance decay coefficient is a hyperparameter controlling the influence of the shortest path distance of a node on the weight increment. It is obtained through experimental parameter tuning based on the knowledge network topology scale and feedback propagation depth requirements, and its value range is typically [value range missing]. In this embodiment, it is preferably set to 0.5; the node co-occurrence frequency correction coefficient is a correction factor that adjusts the weight increment based on the co-occurrence frequency of node pairs in a single feedback data. It is obtained by statistically analyzing the co-occurrence density of node pairs in the same feedback content, and its value range is usually [value missing]. In this embodiment, it is preferably set to 1.1.

[0086] The shortest path distance is calculated using an improved Dijkstra's algorithm. This algorithm adds node type weight constraints to the path search stage, giving higher priority to path searches between entity nodes. For positive feedback, the formula for updating the weight of associated edges is... For negative feedback, weight decay is applied, and the formula for updating the weight of associated edges is as follows: ,in For penalty coefficient, and To limit the maximum and minimum values ​​of the associated edge weights, the first knowledge network is obtained after updating the associated edge weights driven by all feedback data. Then, the associated attribute nodes of specific crop and disease pairs are processed according to the formula. Calculate scores and sort them, then select the top-ranked. The attribute nodes, combined with the preset concept node importance weight threshold. Consensus knowledge validated by multi-source feedback was filtered and marked, among which... , where is the weighting coefficient of the edge weights associated with crop nodes, disease nodes, and attribute nodes, respectively, and ; For attribute nodes The consensus score is used to quantify the correlation strength between the attribute node and the corresponding crop node and disease node, as well as the degree of multi-source feedback verification. This represents the current weight value of the edge connecting the current crop node and the attribute node. The current weight value of the edge connecting the current disease node and the attribute node; the threshold for the importance weight of the concept node. The importance of each concept node (i.e., attribute dimension) in agricultural ontology is preset based on its basic importance. Its value range is usually 0.5 to 0.8, and preferably 0.6.

[0087] For example, this embodiment assumes that the raw feedback data to be processed is for crops. Diseases Diagnostic feedback was collected in total. Expert feedback, User feedback, The system infers feedback, assuming the basic credibility of expert feedback in the three-level credibility assignment. Basic credibility of user feedback The reliability of the system's inference feedback The corresponding source weights are consistent with the basic credibility values; knowledge extracted from the agricultural field... Crops The initial list of entity nodes is constructed based on disease-related structures, and the attribute dimensions in agricultural ontology are used to divide them. A set of concept nodes is constructed based on each attribute dimension to build an initial knowledge network skeleton; an improved BERT-Agr model is then used to extract entities from this batch of feedback data. Extract symptom description phrases Phrases related to prevention and control measures The semantic similarity threshold for semantic matching between attribute phrases and concept nodes. Semantic correction coefficient in the agricultural field Calculated With concept nodes The semantic similarity between (lesion shape) and other parameters was 0.85 and 0.82, respectively. With concept nodes The semantic similarity (for the type of pesticide) was 0.88, which is greater than the preset semantic similarity threshold. Therefore, Create new attribute nodes and establish membership edges with their corresponding concept nodes; the final credibility of a single expert feedback is calculated using a weighted average. This feedback is positive, and the learning rate... Distance attenuation coefficient Node co-occurrence frequency correction coefficient ,entity Shortest path distance to each attribute node After calculating the weight increment Maximum weight of associated edges Minimum value Penalty coefficient After updating the weights of the associated edges for this feedback, and processing all feedback data sequentially, the first knowledge network is obtained; in the consensus knowledge filtering, the weighting coefficients of the associated edge weights of crop nodes, disease nodes, and attribute nodes are used. Select the top rankings The attribute nodes, and the preset concept node importance weight threshold. Concept node The importance weights were 0.65 and 0.72, respectively, both greater than the threshold. Therefore, the top rankings are as follows. The attribute nodes are marked as consensus knowledge. In this embodiment, the basic credibility value, semantic matching threshold, learning rate, distance decay coefficient, and concept node importance weight threshold are merely examples; those skilled in the art can set them according to actual circumstances, and this embodiment does not impose any limitations on them.

[0088] This embodiment achieves accurate credibility assessment of multi-source feedback data through three-level credibility weighted fusion, constructs a standardized knowledge network framework using agricultural domain knowledge and agricultural ontology, efficiently extracts entity and attribute phrases from feedback text through an improved natural language processing model, and realizes the evolution of the knowledge network and accurate screening of consensus knowledge through dynamic updating of feedback-driven association edge weights. It realizes the transformation from multi-source unstructured feedback data to a structured concept-driven knowledge network, improves the credibility, standardization, and accuracy of knowledge extraction in knowledge network construction, and makes the extraction of consensus knowledge more in line with the real needs of actual diagnostic feedback.

[0089] Furthermore, this embodiment provides a method for obtaining a quality score for each knowledge item and calculating the compensated usage frequency based on system records and a first knowledge network through frequency statistics and multi-condition threshold judgment. The steps include:

[0090] Based on system records, the usage frequency score of each knowledge item is obtained through frequency statistics and ranking normalization.

[0091] Based on the correctness indicators of the diagnostic results in the original feedback data, the diagnostic accuracy score of each knowledge item is obtained by counting the number of positive and negative feedbacks.

[0092] A timeliness score is obtained based on the difference between the last update time of each knowledge item and the system record time, using a segmented scoring standard.

[0093] Based on usage frequency score, diagnostic accuracy score and timeliness score, a quality score for each knowledge item is obtained by weighted average and multi-condition threshold judgment.

[0094] Based on the usage frequency recorded in the system, the compensated usage frequency of each knowledge item is calculated through underexposure coefficient compensation.

[0095] Among them, ranking normalization is a calculation method that relativizes the frequency of knowledge item usage to eliminate the impact of changes in the size of the knowledge base on the frequency score; the usage frequency score of each knowledge item is a quantitative value obtained after ranking normalization, used to characterize the relative usage popularity of the knowledge item in the knowledge base; the diagnostic result correctness identifier is a field in the original feedback data that records the user's judgment on the correctness of the diagnostic result, with values ​​of correct, incorrect, or uncertain; the number of positive and negative feedbacks is the number of correct diagnoses, incorrect diagnoses, and uncertain diagnoses associated with each knowledge item, statistically obtained from the original feedback data; the diagnostic accuracy score of each knowledge item is a quantitative value calculated by combining the positive feedback rate and the diagnostic accuracy rate, used to characterize the diagnostic reliability of the knowledge item; the last update time of each knowledge item is the timestamp of the most recent modification of the knowledge item content, recorded in the knowledge base metadata; the system recording time is the current moment of quality assessment; and the segmented scoring standard is a rule that assigns different timeliness scores to different intervals divided according to time differences, used to quantify... The timeliness of knowledge entries is assessed by dividing the time difference into five intervals: 0-30 days, 31-90 days, 91-180 days, 181-365 days, and over 366 days, corresponding to score values ​​of 1.0, 0.8, 0.5, 0.3, and 0.1, respectively. The timeliness score is a quantitative value calculated based on the segmented scoring criteria, used to characterize the novelty of the knowledge entry. The quality score for each knowledge entry is a quantitative value calculated using a weighted average formula, combining three dimensions: usage frequency score, diagnostic accuracy score, and timeliness score, used to comprehensively quantify the overall quality level of the knowledge entry. The usage frequency recorded in the system is the number of times each knowledge entry is retrieved per unit time, statistically obtained from the retrieval logs. The underexposure coefficient compensation calculation is a calculation method to compensate for the underexposure of knowledge entries caused by sorting truncation, used to correct the underestimation of long-tail knowledge. The compensated usage frequency for each knowledge entry is the corrected frequency value obtained after the underexposure coefficient compensation calculation, used to more accurately reflect the actual usage needs of the knowledge entry.

[0096] Specifically, such as Figure 2 As shown, firstly, usage frequency statistics and ranking normalization are performed based on the retrieval logs recorded in the system. The retrieval logs record detailed information for each knowledge base query, including the query identifier, timestamp, query content, and a list of returned knowledge item identifiers. For each knowledge item... Statistics on its performance within a preset time window Search count within Calculate the length of the time window. To obtain the frequency of use Then, all knowledge items are sorted in descending order to obtain the ranking of each knowledge item. Let the total number of entries in the knowledge base be... The ranking normalization formula is adopted. The frequency of use score for each knowledge item is calculated using a formula derived based on the principle of linear normalization. hour To avoid division by zero errors.

[0097] Then, a diagnostic accuracy score is calculated based on the original feedback data, where each feedback record in the original feedback data contains an indicator of the correctness of the diagnostic result. The value can be correct, incorrect, or uncertain; for each knowledge entry The number of positive feedbacks is obtained by statistically analyzing all related feedback records from the feedback data. Number of negative feedback Uncertain number of times Total number of diagnoses First, calculate the positive feedback rate. ,like but Then calculate the diagnostic accuracy. ,like but Finally, the diagnostic accuracy score is obtained by combining the two. ,in , The corresponding weighting coefficients are used in this comprehensive formula, which is derived based on the commonly used weighting method for accuracy assessment and the different levels of emphasis placed on positive feedback rate and diagnostic accuracy in this scheme.

[0098] Next, a timeliness score is calculated based on the last update time of each knowledge item. The metadata records the last update time. The system records the current evaluation time. Calculate the time difference The unit is days; the timeliness score is determined according to the preset segmented scoring criteria. :like but ;like but ;like but ;like but ;like but The segmented scoring criteria are based on the law of knowledge obsolescence, in which the smaller the time difference, the higher the timeliness score.

[0099] Then, a quality score is calculated based on the frequency of use score, diagnostic accuracy score, and timeliness score, using a weighted average formula. The weightings for the frequency score, diagnostic accuracy score, and timeliness score are as follows: , , And satisfy This weighting allocation highlights the core importance of diagnostic accuracy while also considering usage frequency and timeliness; after calculating the quality score, low-quality content is identified based on multi-condition thresholds. For the quality score: if the quality score... The score is less than the preset quality score threshold or diagnostic accuracy score. Less than the preset diagnostic accuracy scoring threshold or time difference Greater than the preset time difference threshold or positive feedback rate If the positive feedback rate is less than the preset threshold, the knowledge item is marked as low-quality content. These thresholds are set based on experience and are used to filter out knowledge items of poor quality.

[0100] Finally, based on the usage frequency recorded in the system, the compensated usage frequency is calculated using an underexposure factor. Specifically, the underexposure factor calculation is performed based on the usage frequency recorded in the system by first counting the number of times each knowledge item was deemed relevant but not included in the return list. Total number of times it was determined to be relevant Through formula The underexposure coefficient was calculated, and then the usage frequency of all knowledge items in the knowledge base in the system records was statistically analyzed. The baseline frequency was obtained by using the median statistical method. Finally, through the improved compensation formula The formula calculates the post-compensation usage frequency for each knowledge item, introducing a frequency adaptation coefficient into the basic compensation formula. This is used to adjust the compensation amount based on the type of knowledge item, where For the preset compensation coefficient, For the first The frequency adaptation coefficient of each knowledge item is set in the range of 0.95-1.05 and is dynamically adjusted according to the crop type to which the knowledge item belongs.

[0101] For example, this embodiment assumes that a certain knowledge base has 10 knowledge entries, for one of the knowledge entries The number of searches was calculated from the search logs over the past 30 days. Next, use frequency. Times / day; assuming the knowledge item is ranked in descending order of usage frequency. Then use frequency scoring. Knowledge items were obtained from the raw feedback data. Number of positive feedback links Number of negative feedback Uncertain number of times Total number of diagnoses Then the positive feedback rate Diagnostic accuracy Diagnostic accuracy score ; Assuming knowledge items Last updated time Distance from current time If it is 15 days, then the time difference The timeliness score falls within the 0-30 day range. A quality score is calculated based on usage frequency, diagnostic accuracy, and timeliness. Due to quality rating The quality score is greater than the preset quality score threshold of 0.6, the diagnostic accuracy score is greater than the preset diagnostic accuracy score threshold of 0.6, the time difference of 15 days is less than the preset time difference threshold of 180 days, and the positive feedback rate is greater than the preset positive feedback rate threshold of 0.7. Therefore, the knowledge item... Not marked as low-quality content; ultimately, statistics on insufficient exposure data and knowledge entries are compiled. of , The underexposure coefficient was calculated. The baseline frequency of all knowledge entries in the knowledge base Preset compensation coefficient Frequency adaptation coefficient Frequency of use after calculation and compensation In this embodiment, the weight allocation and threshold settings are merely examples, and those skilled in the art can adjust them according to actual circumstances. This embodiment does not impose any limitations on these aspects.

[0102] This embodiment achieves standardized evaluation of knowledge item usage frequency through improved frequency statistics and ranking normalization, precise quantification of diagnostic accuracy through multi-dimensional weighted calculation, scientific evaluation of knowledge item timeliness through a combination of segmented scoring and time decay, and comprehensive judgment of overall knowledge item quality through a combination of weighted averaging and multi-condition threshold judgment. Furthermore, it introduces an exposure compensation mechanism to correct the usage frequency bias of long-tail knowledge, realizing multi-dimensional and refined quality evaluation of knowledge items in the crop pest and disease knowledge base. This improves the accuracy and objectivity of quality scoring, solves the problem of underestimated usage frequency of long-tail knowledge due to insufficient exposure, more realistically reflects the actual application value and demand of knowledge items, and enhances the scientificity and effectiveness of knowledge base management.

[0103] Furthermore, this embodiment provides a method for calculating the compensated usage frequency of each knowledge item based on the usage frequency recorded in the system through an underexposure coefficient compensation, the steps of which include:

[0104] Based on system records, the exposure underexposure coefficient of each knowledge item is obtained by calculating the ratio of the number of times a knowledge item is judged as relevant but not included in the return list to the total number of times it is judged as relevant.

[0105] Based on the original usage frequency of all knowledge items in the knowledge base, the baseline frequency is obtained by statistically analyzing the median.

[0106] Based on the original usage frequency, underexposure coefficient, and reference frequency of each knowledge entry, the product of the underexposure coefficient and the reference frequency is weighted and summed using a preset compensation coefficient to obtain the compensated usage frequency.

[0107] The number of knowledge items judged as relevant but not included in the returned list during each retrieval is the sum of the number of times a specific knowledge item was not displayed due to ranking truncation in multiple retrievals, obtained from the full ranking log. This is achieved by counting knowledge items in each retrieval record whose theoretical similarity exceeds the relevance threshold but whose actual returned position exceeds the truncation range. The relevance threshold is used to determine whether a knowledge item is relevant to the query request; it is set based on historical click-through rate statistics and semantic similarity distribution of the retrieval system, typically ranging from 0.5 to 0.9, with 0.7 being preferred. The total number of times a knowledge item is judged as relevant during each retrieval is the sum of the number of times a knowledge item is judged as relevant during each retrieval, obtained from the full ranking log. The total number of times a specific knowledge item is determined to be relevant to the query in multiple searches is obtained by counting the knowledge items that meet the relevance criteria in each search record; the underexposure coefficient is the ratio of the above two counts, ranging from 0 to 1, used to quantify the degree of exposure loss of knowledge items due to sorting truncation; the baseline frequency is the median of the original usage frequency of all knowledge items in the knowledge base, used to characterize the typical usage level of the knowledge base as a whole, obtained by taking the median value after sorting the usage frequencies of all knowledge items; the preset compensation coefficient is a hyperparameter that controls the intensity of exposure compensation, used to adjust the correction magnitude of underexposure on usage frequency, obtained through experimental verification or empirical setting.

[0108] This embodiment uses the full sorting log recorded by the system to accurately quantify the degree of insufficient exposure of knowledge items. By obtaining the baseline frequency through median statistics and combining it with the preset compensation coefficient, a scientific exposure compensation calculation formula is constructed. This achieves accurate correction of the usage frequency of long-tail knowledge, improves the accuracy and objectivity of the usage frequency in reflecting the real demand for knowledge items, and effectively solves the problem of insufficient exposure and underestimation of value of long-tail knowledge caused by sorting truncation. It also provides frequency data that is more in line with the actual application needs for subsequent adjustment of knowledge network weights.

[0109] Furthermore, this embodiment provides a method for dynamically adjusting the weights and fusing nodes of a first knowledge network based on a first knowledge network and its compensated usage frequency, through attention heatmap extraction and attribute node generation, to obtain a second knowledge network containing newly added nodes and associated edges. The steps include:

[0110] Based on the first knowledge network and the original feedback data, the set of salient regions and the average attention value of each salient region are obtained through similarity comparison, attention heatmap extraction and connected component analysis.

[0111] Based on the set of salient regions, the first knowledge network after adding attribute nodes is obtained through multimodal description model, semantic similarity matching, initial association edge establishment and one-hop neighborhood diffusion;

[0112] Based on the original feedback data, the first knowledge network after adding attribute nodes, and the usage frequency after compensation, the first knowledge network after dynamic weight adjustment is obtained through negative feedback positioning and comprehensive weight adjustment.

[0113] Based on the first knowledge network with dynamically adjusted weights, the second knowledge network and newly added nodes and associated edge records are obtained through weighted calculation of vector similarity and text similarity and node merging operation.

[0114] The similarity comparison method is a semantic similarity calculation and matching process that compares the diagnosed pest and disease types with the existing pest and disease knowledge in the first knowledge network. This is used to determine whether the diagnosis results match the knowledge base content and to provide a trigger for new knowledge extraction. The connected component analysis method is an image processing method that performs connectivity detection and region segmentation on pixel regions in the attention heatmap that meet the attention value threshold. This method is used to extract complete salient feature regions of pests and diseases from the heatmap. The salient region set is a set of pixel regions obtained by connected component analysis whose attention values ​​are greater than the preset attention value threshold. Each region contains key information such as bounding box, area, and center coordinates. The average attention value of each salient region is the arithmetic mean of the attention values ​​of all pixels in a single salient region. The attention value threshold is obtained based on the global mean and global standard deviation of the attention heatmap to achieve dynamic adaptive segmentation for different image feature distributions. Under the usual image attention distribution, the attention value threshold ranges from 0.3 to 0.7, and is usually about 0.45.

[0115] The multimodal description model is a deep learning model capable of generating text descriptions of image regions, used to transform the visual features of salient regions into natural language descriptions of symptoms or prevention measures. Semantic similarity matching is a processing method that calculates the semantic similarity between candidate attribute phrases and concept nodes in the first knowledge network and classifies them, used to map attribute phrases to corresponding concept nodes to achieve the concept attribution determination of attribute nodes. Initial association edge establishment is a processing operation that constructs initial association edges and assigns initial weights to newly added attribute nodes and their corresponding crop and disease entity nodes, used to achieve the initial connection between newly added nodes and the knowledge network. One-hop neighborhood diffusion is a processing method that establishes weak association edges between newly added attribute nodes and the directly associated nodes of crop and disease entity nodes, i.e., one-hop neighborhood nodes, used to accelerate the cold start of newly added nodes in the knowledge network. The knowledge network after adding attribute nodes is an intermediate knowledge network formed by adding newly created attribute nodes and their initial association edges and diffusion association edges on the basis of the first knowledge network, retaining the node and edge relationships of the original network, while adding the network structure corresponding to the new knowledge.

[0116] Negative feedback positioning involves filtering out negative feedback such as diagnostic errors and poor prevention and control effects from the original feedback data and locating the corresponding related edges in the knowledge network. This method is used to identify erroneous or low-reliability knowledge associations in the knowledge network and to provide a basis for weight adjustment. Comprehensive weight adjustment is a processing method that combines knowledge item quality assessment indicators and feedback types to batch update the weights of related edges in the knowledge network. This method is used to achieve dynamic matching between the weights of related edges and knowledge quality and practical application feedback. The knowledge network after dynamic weight adjustment is a knowledge network with optimized related edge weights after negative feedback positioning and comprehensive weight adjustment. This eliminates the impact of erroneous associations and strengthens the association strength of high-reliability knowledge.

[0117] The weighted calculation of vector similarity and text similarity involves simultaneously calculating the vector semantic similarity and text keyword similarity between attribute nodes, and then summing them using preset weights to obtain a comprehensive similarity. This method is used to comprehensively evaluate the similarity between attribute nodes and provide a basis for determining node fusion. The node merging operation involves merging the content, transferring weights, and deleting redundant nodes of attribute nodes under the same concept node that meet the fusion conditions. This is used to eliminate node redundancy in the knowledge network to improve its simplicity and consistency. The record of newly added nodes and associated edges is a detailed ledger of newly added attribute nodes, initial associated edges, diffused associated edges, and weight adjustment records in the second knowledge network. This is used to trace the update process of the knowledge network and is obtained through real-time recording when the system performs knowledge network update operations.

[0118] Specifically, such as Figure 3As shown, firstly, a similarity comparison is performed between the first knowledge network and the original feedback data. The text information of the diagnosed pest and disease types is transformed into vectors using an improved Sentence-BERT model. This model adds a vocabulary embedding layer for agricultural pests and diseases to the encoding layer. The training corpus consists of 80,000 text data points on agricultural pests and diseases, with a batch size of 16, a learning rate of 0.0001, and 15 training epochs. The semantic matching similarity between these vectors and all pest and disease knowledge vectors in the first knowledge network is then calculated. ,like Less than the preset similarity matching threshold Furthermore, the user confirmation flag in the original feedback data is correct, and the final credibility of the feedback data corresponds to... Greater than or equal to the preset confidence threshold Then, a new knowledge extraction trigger signal is generated. The similarity matching threshold is used to determine whether there is a significant difference between the diagnostic result and the existing knowledge in the knowledge base. Its setting method is based on semantic similarity distribution statistics and the system's sensitivity requirements for new knowledge, and the value range is usually 0.7 to 0.9, preferably 0.8. The credibility threshold is used to filter feedback data with sufficient reliability to trigger new knowledge extraction. Its setting method is determined based on the basic credibility level division in the three-level credibility weighted fusion, and the value range is usually 0.6 to 0.9, preferably 0.7.

[0119] Based on the new knowledge, the multimodal diagnostic results corresponding to the trigger signals are extracted. The average weight of the last layer of self-attention is extracted through the improved VisionTransformer multimodal large model, and an attention heatmap of the same size as the input image is generated. This model introduces region attention constraints in the self-attention layer. The input is a 3-channel RGB image, and the output is a single-channel attention heatmap with attention values ​​for each pixel. Calculate the global mean of the attention heatmap. and standard deviation Set a preset attention threshold. The attention value extracted is greater than For pixel regions, an improved connected component analysis algorithm is used, which adds a region area filtering constraint and removes regions with areas smaller than [a certain value]. From the tiny regions, we obtain the set of salient regions. And calculate the average attention value for each region. ,in For the region The number of pixels. Among them, The area of ​​the smallest salient region; The index is the sequence number of the significant region, and its value range is... , This represents the total number of salient regions obtained after connected component analysis. These are the pixel coordinates in the attention heatmap.

[0120] Then, based on the set of salient regions, each region... The corresponding image patch input is used to improve the multimodal description model and generate text descriptions. This model, based on BLIP-2, adds a layer of agricultural symptom keyword reinforcement to the decoder, and uses an improved BERT-Agr model to extract text descriptions. Extract candidate attribute phrase set from For each candidate attribute phrase Calculate its relationship with all concept nodes in the first knowledge network. semantic similarity ,in This is the visual semantic mapping correction coefficient, typically taken as 1.02. If... Greater than or equal to the preset attribute matching threshold If the candidate attribute phrase is selected, it is assigned to the corresponding concept node; further, the semantic similarity between the candidate attribute phrase and all existing attribute nodes under the concept node is calculated. ,like Less than the preset attribute matching threshold That is, there is no semantic similarity exceeding If there is an existing attribute node, then create a new attribute node. And establish membership edges with the corresponding concept nodes. The attribute matching threshold is set in a similar way to the semantic similarity threshold mentioned above, with a value range of 0.75 to 0.95, preferably 0.90. Attention mapping weights are also calculated. ,in The weight values ​​are used to characterize the visual salience of new attribute nodes. This is the set of salient regions corresponding to the new attribute node. This represents the total number of salient regions in the set of salient regions corresponding to the new attribute node. For the first The average attention value for each salient region is used to calculate the initial associated edge weights. Perform initial edge establishment, new attribute node and crop node Disease nodes The initial associated edge weight formulas are as follows: ,in As the benchmark weight, , These are the crop fit coefficient and the disease fit coefficient, both ranging from 0.5 to 1.0, with the benchmark weight limited to [0.1, 1.0].

[0121] Then, perform one-hop neighborhood diffusion to establish weak associations between the new attribute nodes and the one-hop neighbor nodes of the crop nodes and disease nodes, respectively, to obtain the knowledge network after the addition of the attribute nodes. For the one-hop neighbor nodes of the crop nodes... The weighting formula is For the one-hop neighbor nodes of the disease node The weighting formula is ;in The diffusion coefficient controls the strength of the association propagated from a new attribute node to its one-hop neighbor nodes. It is obtained through experiments based on the size of the knowledge network and the cold start speed requirements for new knowledge access, and its value typically ranges from [value missing]. In this embodiment, it is preferably set to ; These are the initial edge weights between the new attribute node and the crop node and the disease node, respectively. These are the weights of existing associated edges in the knowledge network between crop nodes, disease nodes, and their respective one-hop neighbor nodes, obtained by reading the corresponding attribute fields of the associated edges from the graph database. This indicates taking the smaller of the two values.

[0122] Then, based on the original feedback data, negative feedback localization is performed to filter out negative feedback where the diagnostic results are marked as incorrect or the prevention and control effect evaluation is invalid. The corresponding node pairs are located, and combined with the knowledge network after adding attribute nodes and the usage frequency after compensation, a comprehensive weight adjustment is performed. The current accuracy rate is obtained from the quality assessment module. Current usage frequency and corresponding benchmark values Calculate the accuracy adjustment factor Frequency adjustment coefficient ,in , These are correction factors for the accuracy adjustment coefficient and the frequency adjustment coefficient, respectively. They are used to adjust the sensitivity of the quality indicators to changes in the adjustment coefficients. These factors are obtained through experimental parameter tuning based on the knowledge base's tolerance for fluctuations in diagnostic accuracy and usage frequency. The typical value range is [value range missing]. In this embodiment, it is preferably set to 0.95. Let the node pair to be adjusted be... In the The current associated edge weight in the round of iteration is ,in To update the iteration sequence number, the formula for adjusting the overall weight is: Overall weight Limited to Within the scope. For nodes involved in negative feedback, weight decay is applied, and the formula for updating the weight of associated edges is: ,in To perform weight decay on the node With nodes New weight values ​​for the edges connecting them; This is the original weight value of the associated edge in the current iteration round; This is the penalty coefficient, used to amplify or reduce the impact of negative feedback on weight decay. It is obtained through experimental parameter tuning based on the system's required penalty for erroneous knowledge, and its value range is typically [range missing]. In this embodiment, it is preferably set to ; This is the weight decay, calculated in the same way as the positive feedback weight increment, i.e. ,in For learning rate, To ensure the final credibility of negative feedback data, Distance attenuation coefficient, For the first Nodes in round iteration and Shortest path distance This is a correction coefficient for the co-occurrence frequency of nodes; The minimum value limit for the weights of associated edges is defined as a lower threshold for the weights of associated edges in the knowledge network. This limit is used to prevent the weights from dropping too low due to continuous decay, which could lead to the failure of knowledge associations. The minimum value is set experimentally based on the stability requirements of the knowledge network and the control of weight decay boundaries. The typical value range is [insert range here]. In this embodiment, it is preferably set to 0. After completing the above comprehensive weight adjustment and negative feedback weight decay, the knowledge network with dynamically adjusted weights is obtained.

[0123] Finally, based on the dynamically weighted knowledge network, all attribute nodes under the same concept node are periodically scanned, and the vector similarity between each pair is calculated using an improved Sentence-BERT model. The Jaccard keyword similarity algorithm was used to extract keywords, and the text similarity was obtained by combining jieba word segmentation with an agricultural dictionary. The formula for comprehensive similarity is: ;in, To achieve comprehensive similarity, it is used to comprehensively measure the overall similarity between two attribute nodes at the semantic vector level and the text keyword level; These are the vector similarity weight coefficients. Let be the text similarity weight coefficient, and both satisfy . These parameters are used to adjust the contribution ratio of vector similarity and text similarity to the overall similarity score. They are obtained through experimental parameter tuning based on the difference in importance between vector semantic matching and keyword matching in attribute node fusion. In this embodiment, they are preferably set to... , ; This is a comprehensive similarity correction coefficient used to adjust the contribution weight of Jaccard keyword similarity in the overall similarity score, compensating for the measurement differences between text keyword matching and deep semantic matching. It is determined based on the distribution statistics of vector similarity and text similarity, and its value typically ranges from [value missing]. In this embodiment, it is preferably set to If we consider the overall similarity Greater than the preset fusion threshold If the nodes meet the conditions of having the same concept and complementary content, then a node merging operation is performed: the node with more complete fields is retained, symptom descriptions and prevention methods are merged and duplicate content is removed, the weights of the associated edges of redundant nodes are transferred to the retained nodes by taking the maximum value, redundant nodes are deleted, and information on newly added nodes, changes in associated edge weights, and node fusion status are recorded to obtain the second knowledge network and records of newly added nodes and associated edges. The fusion threshold is used to determine whether two attribute nodes meet the fusion conditions. It is determined based on the similarity distribution of attribute nodes and the requirement for the simplicity of the knowledge network, and its value range is usually from 0.75 to 0.90, preferably set to 0.80.

[0124] For example, this embodiment assumes that the first knowledge network contains Disease and pest knowledge, diagnosis of crops from raw feedback data Diseases The similarity score was calculated by comparing its match with the existing pest and disease knowledge in the first knowledge network. Set a preset similarity matching threshold Preset credibility threshold The feedback was marked as "correct" by the user, indicating a final credibility level. The new knowledge extraction trigger condition is met, and a trigger signal is generated; the corresponding multimodal diagnostic result includes a lesion image, and an attention heatmap is generated through the improved VisionTransformer model, and the global mean of the heatmap is calculated. Global standard deviation of heatmap Preset attention value threshold The set of salient regions was obtained through connected component analysis. Their average attention scores are respectively Input the image patches corresponding to the three regions into the multimodal description model to generate text descriptions. , , Extract the candidate attribute phrase set Calculate the semantic similarity between each phrase and the concept node. The similarity to the "lesion shape" was 0.88. The similarity to "lesion characteristics" is 0.86. The similarity with "prevention agent type" is 0.91. A preset attribute matching threshold is set. , The maximum semantic similarity between the corresponding candidate phrase and the existing attribute nodes under the concept node is less than the attribute matching threshold. Since no matching existing attribute node was found, a new attribute node was created. , Reuse existing attribute nodes; calculate attention map weights. Set benchmark weights , New attribute node With crop nodes Initial associated edge weights New attribute node With disease nodes Initial associated edge weights Similarly, we can obtain , Set the diffusion coefficient. , with crop nodes One-hop neighborhood node (Assuming it's an existing attribute node) Establish weak association weights, and assign weights to the existing association edges between this neighboring node and the crop node. Then the weak association weight For disease nodes One-hop neighborhood node (Assuming it's another existing attribute node), set the weight of the existing association edge between this neighboring node and the disease node. Then the weak association weight Similarly, new attribute nodes. Calculate the corresponding weak association weights to obtain the knowledge network after adding attribute nodes.

[0125] Two node pairs were selected from the original feedback data. Negative feedback, obtain the current accuracy from the quality assessment module. Benchmark accuracy Current usage frequency Reference frequency Set adjustment coefficient correction factor Calculate the accuracy adjustment factor Frequency adjustment coefficient Let node pairs In the The current associated edge weight in the round of iteration is Then the weight after comprehensive weight adjustment To reduce the execution weight of this node, a penalty coefficient is set. Learning rate Distance attenuation coefficient Node co-occurrence frequency correction coefficient The ultimate credibility of negative feedback Shortest path distance between nodes Calculate the weight decay Minimum weight limit for associated edges Then the weight after decay This yields a knowledge network with dynamically adjusted weights.

[0126] Scan the attribute nodes under the "lesion shape" concept node and calculate With existing nodes vector similarity Text similarity Set vector similarity weight coefficients Text similarity weight coefficient Comprehensive similarity correction coefficient The overall similarity was calculated. Preset fusion threshold Based on comprehensive similarity Less than the preset fusion threshold Therefore, the fusion conditions are not met, and the independent node is retained, ultimately resulting in a result that includes the newly added node. And a second knowledge network with corresponding associated edges. In this embodiment, the thresholds, weight coefficients, model training parameters, etc., are merely examples; those skilled in the art can set them according to actual circumstances, and this embodiment does not impose any limitations on them.

[0127] This embodiment achieves precise localization of salient features of pests and diseases in images through attention heatmap extraction and connected component analysis. It automatically generates and classifies visual features into structured attribute nodes through a multimodal description model and semantic similarity matching. It achieves networked cold start of new knowledge through initial association edge establishment and one-hop neighborhood diffusion. It dynamically optimizes the weights of the knowledge network through negative feedback localization and quality-driven comprehensive weight adjustment. It simplifies the structure and eliminates redundancy of the knowledge network through weighted calculation of vector similarity and text similarity and node merging operations. This fully automates the process from multimodal feedback data to incremental updates of the knowledge network, enabling incremental updates and conceptual evolution of the crop pest and disease knowledge base. It improves the knowledge base's ability to capture new knowledge, the efficiency of network topology optimization, and the consistency of knowledge content, while ensuring the dynamic adaptability and accuracy of the knowledge network.

[0128] Furthermore, this embodiment provides a method for obtaining a set of salient regions and the average attention value of each salient region based on a first knowledge network and original feedback data, through similarity comparison, attention heatmap extraction, and connected component analysis. The steps include:

[0129] Based on the first knowledge network, the new knowledge extraction trigger signal is obtained by similarity comparison and by combining the user confirmation mark and final credibility in the original feedback data.

[0130] Based on the new knowledge, trigger signals and their corresponding multimodal diagnostic results are extracted. The average weight of attention is extracted through a multimodal large model to obtain an attention heatmap.

[0131] Based on the attention heatmap, the set of salient regions and the average attention value of each salient region are obtained through threshold segmentation and connected component analysis.

[0132] Among them, the user confirmation identifier in the original feedback data is a Boolean identifier used to mark whether the user or expert agrees with the correctness of the diagnosis result; the new knowledge extraction trigger signal is a control signal used to start the new knowledge extraction process, which is automatically generated when the diagnosis result does not match the knowledge base, the user confirms the diagnosis is correct, and the final credibility of the feedback meets the standard; the multimodal diagnosis result is a set of diagnosis results containing various types of data such as pest and disease images and text descriptions, which is automatically generated and stored by the diagnosis system after completing the pest and disease diagnosis; the multimodal large model is a deep learning model with image feature extraction and self-attention calculation capabilities, used to extract attention heatmaps from pest and disease images; the average weight of self-attention is the row average or column average of the self-attention weight matrix of the last layer of the multimodal large model. The average calculation results are as follows: Attention heatmap is a heatmap of the same size as the input image, which maps the average weight of self-attention. The brightness of each pixel corresponds to the attention weight value, which is used to visualize and locate the salient feature regions of pests and diseases; Connected component analysis is a method of merging adjacent foreground pixels into independent regions by performing connectivity detection on the foreground region after threshold segmentation, which is used to extract the complete feature regions of pests and diseases. It is achieved by scanning image pixels and marking the boundaries and ranges of connected regions; The salient region set and the average attention value of each salient region are the set of independent foreground regions obtained by connected component analysis, and the arithmetic mean of the attention values ​​of all pixels in each region. The salient region set is a concentrated representation of the features of pests and diseases, and the average attention value is a quantitative representation of the importance of the region.

[0133] This embodiment achieves precise triggering of new knowledge extraction based on similarity comparison and credibility assessment of multi-source feedback data. It realizes end-to-end generation from diagnostic images to attention heatmaps based on the self-attention mechanism of a multimodal large model. It achieves automated localization and quantification of salient features of pests and diseases in images based on adaptive threshold segmentation and connected component analysis. Finally, it obtains a set of salient regions that can accurately reflect the typical symptoms of pests and diseases and their attention weights. This realizes precise triggering of new knowledge extraction and efficient localization of salient features of pests and diseases, improves the knowledge network's ability to capture and utilize visual information, improves the pertinence and accuracy of new attribute node generation, and provides a high-quality feature data source for incremental updates of the knowledge base.

[0134] Furthermore, this embodiment provides a method for obtaining a knowledge network with added attribute nodes based on a set of salient regions, through a multimodal description model, semantic similarity matching, initial association edge establishment, and one-hop neighborhood diffusion. The steps include:

[0135] Based on the set of salient regions, a multimodal description model is used to generate text descriptions and extract attribute phrases to obtain a set of candidate attribute phrases.

[0136] Based on the candidate attribute phrase set and the concept node set in the first knowledge network, the attribute phrases are classified into corresponding concept nodes through semantic similarity matching, resulting in a new set of attribute nodes and their corresponding attention mapping weights;

[0137] Based on the newly added set of attribute nodes and their corresponding attention mapping weights, the first knowledge network after adding attribute nodes is obtained by establishing initial associated edges and one-hop neighborhood diffusion.

[0138] The text description is a descriptive text obtained by a multimodal description model through natural language conversion of the visual features of salient regions, containing key information such as symptoms, characteristics, or control methods of pests and diseases; the candidate attribute phrase set is a set of phrases such as symptom description fragments and control measure fragments related to pest and disease diagnosis extracted from the text description; the new attribute node set is a set of attribute nodes newly created after the candidate attribute phrases are semantically similar and no existing attribute nodes are matched, and each new attribute node is associated with a corresponding concept node; its corresponding attention mapping weight is a quantitative value representing the strength of association between the new attribute node and the salient region.

[0139] This embodiment achieves cross-modal conversion from visual salient regions to natural language descriptions based on a multimodal description model. It achieves accurate classification of new attribute phrases into existing concept systems and the creation or reuse of attribute nodes based on semantic similarity matching. It achieves scientific assignment of initial associated edges of new nodes based on attention mapping weights. It achieves rapid cold start and information propagation of new nodes in the knowledge network based on one-hop neighborhood diffusion. Finally, it obtains the first knowledge network after adding attribute nodes with integrated visual knowledge. It realizes full-process automation of new knowledge from feature extraction to network entry, improves the knowledge network's capture efficiency and representation ability of visual pest and disease knowledge, improves the efficiency of incremental updates of the knowledge base and the network adaptability of new knowledge, and ensures the rationality of new knowledge access and network consistency through attention weight mapping and neighborhood diffusion mechanisms.

[0140] Furthermore, this embodiment provides a method for obtaining a third knowledge network and rollback event records based on a second knowledge network and update records, through semantic version control and quality indicator differential calculation, and then by performing rollback condition judgment. The steps include:

[0141] Based on the second knowledge network and update records, semantic version numbers are generated through semantic version control and changes in newly added nodes, deleted nodes and associated edges are extracted. Incremental snapshot data is obtained through serialization.

[0142] Based on the second knowledge network and update records, regional information is extracted through regular expression matching of geographic keywords to obtain regional branch version records;

[0143] Based on system logs, the changes in quality indicators are obtained through time-series database queries and quality indicator difference calculations. The changes in quality indicators include changes in accuracy, complaint rate, and error rate.

[0144] Based on the change in quality indicators and the preset change threshold, a rollback operation is performed through Boolean logic to obtain the third knowledge network and rollback event records.

[0145] The semantic version number is a string identifier used to uniquely identify the version status of the knowledge network. The major version number corresponds to major changes in the node structure, the minor version number corresponds to batch adjustments of the weights of associated edges, and the revision number corresponds to minor corrections of local nodes or edges. New nodes are entity nodes, attribute nodes, or concept nodes added to the second knowledge network compared to the previous version. Deleted nodes are redundant or erroneous nodes removed from the second knowledge network compared to the previous version. Changes in associated edges refer to changes in the weights of associated edges in the second knowledge network, such as adjustments, additions, or deletions, including information such as the changed edge ID, old and new weight values, and change type. Incremental snapshot data includes data on newly added nodes, deleted nodes, and associated edges. The dataset obtained after JSON serialization of the changes is used to accurately record incremental changes to the knowledge network; the geographic keyword regular expression matching method is a text processing method that uses regular expressions built based on a preset geographic keyword dictionary to extract region-related information from update records or knowledge content, used to identify the applicable regions corresponding to the updated content of the knowledge network; the regional information is descriptive information representing the applicable geographic scope of the updated content of the knowledge network, including geographic identifiers at the provincial, municipal, and regional levels; the regional branch version record is version management data maintained for different regions, containing information such as semantic version numbers, regional tags, and version pointers, used to realize the regionalization of the knowledge network. This control; the system log records events such as diagnostic requests, feedback reception, version updates, and system errors during the operation of the knowledge base; the time-series database query performs time-dimensional retrieval and data aggregation operations on the system log, used to extract quality indicator-related data according to a specified time interval; the quality indicator change is the difference between the quality monitoring indicators of the current version and the previous version, including the change in accuracy, the change in complaint rate, and the change in error rate; the accuracy change is the difference between the diagnostic accuracy of the current version and the diagnostic accuracy of the previous version, and the diagnostic accuracy is calculated based on the user feedback data of the corresponding version; the complaint rate change is the difference between the user complaint rate of the current version and the user complaint rate of the previous version. The difference in user complaint rates is the ratio of the number of complaints to the total number of diagnoses. The error rate change is the difference between the current version's system error rate and the previous version's system error rate. The system error rate is the ratio of the number of internal system errors to the total number of diagnoses. The preset change thresholds are critical values ​​for quality indicator changes used to determine whether to trigger a rollback operation. These include accuracy decrease thresholds, complaint rate increase thresholds, and error rate increase thresholds, which are set through historical operational data statistics and stability requirement analysis of the agricultural knowledge base. The Boolean logic judgment is a judgment method based on logical comparison between the change in quality indicators and the preset change thresholds, used to determine whether the current version meets the rollback trigger conditions.

[0146] Specifically, semantic version control is first performed based on the second knowledge network and update records. The version number change rules are determined according to the type of change in the update records: if there is a major change in the node structure, the major version number is incremented by 1; if there is a batch adjustment of the weights of related edges, the minor version number is incremented by 1; if it is only a minor correction to a local node or edge, the revision number is incremented by 1, thus generating a new semantic version number. ,in Major version number, This is the minor version number. The revision number is used; then, the list of newly added nodes in the second knowledge network is extracted. List of deleted nodes and list of changes in associated edges The three types of data are converted into standard JSON format using a JSON serialization tool to obtain incremental snapshot data. The serialization process uses UTF-8 encoding to ensure cross-platform data compatibility. Then, based on the second knowledge network and update records, regional information is extracted through improved geographic keyword regular expression matching. A dictionary containing provincial, municipal, and regional geographic keywords is constructed, and multi-pattern regular expressions are written. ,in This is a regular expression pattern for provincial-level geographical keywords. This is a regular expression pattern for city-level geographical keywords. This involves using regular expression patterns for regional geographic keywords; matching the knowledge description text and feedback source region annotations in the update records to extract a set of regional information. If the updated content applies to all regions, then mark it as such. Otherwise As a regional label, combined with semantic version number and version pointer, a regional branch version record is generated. ,in This is a pointer to the applicable network status for the current region. This is a Boolean-type globally applicable flag. A value of true indicates that the current knowledge network update is not geographically restricted and is applicable to full deployment. A value of false indicates that branch management is required based on the regional tags in the regional information set. Subsequently, based on system logs, quality indicator data is extracted through time-series database queries, and a time window is set. ,in This refers to the effective date of the previous version. The total number of diagnoses within the current time window is retrieved from the time-series database, based on the effective date of the current version. Number of positive feedback Number of complaints Number of system errors Calculate the diagnostic accuracy of the previous version. Current version diagnostic accuracy The previous version's complaint rate Current version complaint rate Previous version error rate Current version error rate The change in quality indicators, specifically the change in accuracy, is obtained through differential calculation. Changes in complaint rate Error rate change value This difference calculation is directly derived based on the basic difference formula and the statistical characteristics of the quality indicators. and These represent the number of positive feedback entries in the previous version and the current version, respectively. and The numbers represent the number of user feedback reports from the previous version and the current version, respectively, with an uncertain number of responses. and These represent the total number of diagnoses for the previous version and the current version, respectively. and These are the number of user complaints for the previous version and the current version, respectively. and These represent the number of internal system errors in the previous and current versions, respectively. Finally, a Boolean logic check is performed based on the change in quality metrics and a preset change threshold to set the preset change threshold, including an accuracy decrease threshold. The threshold for an increase in the complaint rate Error rate rise threshold Construct a Boolean logic judgment rule: if the accuracy changes by a certain value... Less than the accuracy drop threshold Or change in complaint rate Greater than the threshold for rising complaint rate Or error rate change value greater than the error rate rise threshold This will trigger a rollback operation, changing the version pointer of the corresponding regional branch. Revert to the previous stable version Based on the incremental snapshot data from the previous version Perform node and edge restoration operations in the graph database: mark newly added nodes. To delete the node and reactivate it. 1. Restore the weights of associated edges to The third knowledge network is obtained, and rollback event logs are recorded simultaneously. ,in For the rollback execution timestamp; the For the previous stable version The weight values ​​of the corresponding associated edges are determined based on incremental snapshot data. The information on changes in associated edges recorded in the database is obtained. The preset change threshold and weight coefficient values ​​are merely examples; those skilled in the art can set them according to actual circumstances, and this embodiment does not impose any limitations on them.

[0147] This embodiment achieves refined recording and version identification of knowledge network changes through semantic version control, accurate extraction of regional information and regional branch version management through geographic keyword regular expression matching, quantitative evaluation of quality indicator changes through time-series database query and differential calculation, and rapid rollback of substandard versions through Boolean logic judgment and rollback operation. It realizes regionally sensitive version management and stability assurance of the knowledge network, improves the refinement, regional adaptability and operational stability of knowledge base version control, and effectively avoids the impact of low-quality versions on diagnostic results.

[0148] Furthermore, this embodiment provides a step-by-step approach based on a third knowledge network and practical application feedback, using multi-dimensional semantic similarity calculation and subgraph cloning adaptation to perform cross-crop knowledge transfer and graph network fusion, resulting in a fourth knowledge network and transfer records. The steps include:

[0149] Based on the names of pests and diseases, symptom descriptions, and control methods of the source and target crops in the third knowledge network, a comprehensive similarity is obtained by multi-dimensional semantic similarity calculation through a bidirectional encoder representation method.

[0150] Based on comprehensive similarity, expert verification records and actual application feedback records in the original feedback data, a migration credibility score is obtained through ratio statistics and weighted evaluation.

[0151] Based on the source crop subgraph in the third knowledge network, the subgraph is cloned and adapted by querying the threshold of associated edges, replacing entity node identifiers, and scaling the weight of associated edges to obtain the adapted subgraph to be accessed.

[0152] Based on the subgraph to be accessed and the target crop subgraph in the third knowledge network, cross-crop knowledge transfer and graph network fusion are carried out through semantic similarity threshold matching, associated edge merging and node insertion to obtain the fourth knowledge network and transfer records.

[0153] In the third knowledge network, the source crops are crop types in the knowledge base with mature disease and pest diagnosis knowledge that can be used as a source of transfer; the target crops are crop types in the knowledge base with insufficient knowledge of corresponding diseases and pests that need to be supplemented; the corresponding disease and pest names are the specific names of diseases and pests that are common to the source and target crops or have similar characteristics; the symptom description is a textual description of the symptoms exhibited by the disease or pest on the source crop; and the control method text is a textual description of effective control measures for the disease or pest on the source crop. The bidirectional encoder representation method is a text semantic vector encoding method based on an improved Sentence-BERT model, used to convert disease and pest-related text into high-dimensional semantic vectors and calculate similarity; the comprehensive similarity... The overall similarity quantification, obtained by weighted summation after calculating semantic similarity from three dimensions—pest / disease name, symptom description, and control method—is used to determine the similarity between source and target crop pest / disease knowledge. The expert verification records in the original feedback data are expert opinions and scores on the feasibility of cross-crop knowledge transfer, collected through an expert feedback interface. The practical application feedback records are feedback data on the effects of small-scale trials of source crop knowledge on the target crop, including evaluations of control effectiveness and symptom matching. The transfer credibility score is a quantitative assessment value of transfer feasibility obtained by combining the comprehensive similarity, expert verification records, and practical application feedback records. The source crop subgraph is extracted from the third knowledge network and... The system consists of a local subnetwork composed of attribute nodes, entity nodes, and their associated edges that are strongly associated with the source crop and specific pests and diseases; the associated edge threshold query is a query operation that filters strongly associated edges in the source crop subgraph based on a preset weight threshold; entity node identifier replacement is a processing operation that replaces the source crop entity node ID in the source crop subgraph with the target crop entity node ID, used to adapt the subgraph to the target crop network; associated edge weight scaling is a processing method that adjusts the weight of associated edges in the source crop subgraph by multiplying it by the migration reliability score, used to optimize the initial weights based on migration reliability; the adapted subgraph to be accessed is a standardized subnetwork that can be accessed from the target crop subgraph after entity node replacement and weight scaling; Third Knowledge Network The target crop subgraph in the network is a local subnetwork composed of all nodes and associated edges related to the target crop in the third knowledge network; semantic similarity threshold matching is a processing method that calculates the semantic similarity between the attribute nodes in the subgraph to be accessed and the existing attribute nodes in the target crop subgraph, and compares it with the preset semantic similarity threshold of attribute nodes, which is used to determine whether attribute nodes need to be created repeatedly; associated edge merging is an operation that merges the weights of associated edges of the same node pair in the subgraph to be accessed and the target crop subgraph by taking the maximum value, which is used to retain the optimal knowledge association strength; node insertion is an operation that adds attribute nodes that do not exist in the target crop subgraph in the subgraph to be accessed to the target crop subgraph, which is used to fill the knowledge gaps of the target crop.

[0154] Specifically, such as Figure 4As shown, the source crop is first extracted from the third knowledge network. With target crop Text related to pests and diseases, including pest and disease names. Symptom description Prevention and control methods This model employs an improved bidirectional encoder representation method for multi-dimensional semantic similarity calculation. It adds an agricultural vocabulary enhancement module to the encoding layer of the basic Sentence-BERT model. This module includes an agricultural domain vocabulary embedding layer and a gated fusion unit, which are used to extract enhanced word vectors of agricultural-specific vocabulary and perform weighted fusion with the original word vectors. The training corpus consists of 120,000 textual data on agricultural pests and diseases, with a batch size of 24, a learning rate of 0.0001, and 18 training epochs. The similarity of pest and disease names is calculated for each epoch. Symptom description similarity Similarity of prevention and control methods , among which The superscript represents the semantic vector function extracted by the improved bidirectional encoder representation method. Indicates the text corresponding to the source crop, superscript The text representing the target crop. and The names of pests and diseases in the source crop and the target crop are displayed separately. and These are descriptions of symptoms in the source crop and the target crop, respectively. and The texts describe the control methods for the source crop and the target crop, respectively. Then, a weighted formula is used... Calculate the overall similarity, with weights assigned as follows: The formula is derived based on the degree of influence of each dimension on knowledge transfer; then, expert verification records and actual application feedback records are collected from the original feedback data, and the number of times expert verification passed is counted. Total number of verifications Calculate expert verification score Statistical analysis of the number of successful applications in actual use Total number of applications Calculate the actual application score Through weighted formula The transfer confidence score is calculated. If the transfer confidence score is... Greater than or equal to the migration confidence threshold If the target crop knowledge base is insufficient, subsequent migration operations will be triggered.

[0155] Then, the source crop subgraph is extracted from the third knowledge network. The system filters edges whose weights are greater than or equal to a preset threshold by querying the associated edge weight. The associated edges retain the source crop entity nodes. Disease entity nodes and strongly related attribute nodes Perform entity node identifier replacement, replacing the source crop entity node. Replace the node ID with the target crop The node ID is then used to scale the weights of all associated edges, updating the weights to the specified values. The adapted subgraph to be accessed is obtained. ,in Represents nodes in the source crop subgraph With nodes The original weight values ​​of the edges connecting them. This represents the initial weight value of the corresponding associated edge in the subgraph to be connected after weight scaling. A migration reliability score is used to proportionally scale the weights of associated edges based on migration reliability; ultimately, the subgraph to be integrated is... With the target crop subgraph in the third knowledge network Perform graph network fusion, for each attribute node to be connected. Calculate its relationship with Semantic similarity of all attribute nodes under the same concept node The semantic similarity is calculated using the same improved cosine similarity formula as mentioned above, i.e. , For semantic correction coefficients in the agricultural field For the new attribute nodes to be merged in the subgraph to be integrated, For the target crop subgraph and Existing attribute nodes belonging to the same concept node; if Greater than or equal to the preset attribute matching threshold Then, perform edge merging. For entity nodes that are jointly associated, the edge weight is taken as... Add associated edges directly to new entity nodes, where The existing weight values ​​of the corresponding associated edges in the target crop subgraph. The weight value of the associated edge in the subgraph to be integrated is the weighted value after weight scaling; if the similarity is lower than the threshold, node insertion is performed. Add to The process involves establishing related edges and, after fusion, obtaining a fourth knowledge network. Simultaneously, migration records are recorded, including source and target crop information, overall similarity, migration credibility score, and information on newly added nodes and related edges. The model training parameters, weights for each dimension, and preset thresholds in this embodiment are merely examples; those skilled in the art can set them according to actual circumstances, and this embodiment does not impose any limitations on them.

[0156] This embodiment achieves accurate similarity matching of cross-crop knowledge through an improved bidirectional encoder representation method, obtains reliable transfer credibility scores through comprehensive multi-dimensional feedback, realizes standardized transformation of source crop knowledge through subgraph cloning adaptation, and achieves seamless knowledge access through semantic similarity threshold matching and graph network fusion. It realizes efficient transfer and reuse of mature crop pest and disease knowledge to crops with knowledge gaps, improves the knowledge coverage, incremental update efficiency and cross-crop adaptation capability of the knowledge base, and ensures the accuracy and reliability of transferred knowledge.

[0157] Furthermore, such as Figure 5 As shown in the embodiments of this application, a self-learning and incremental update system for a crop disease and pest knowledge base is provided. This system includes a feedback fusion module, a quality assessment module, an incremental update module, and a version migration module.

[0158] Feedback fusion module: Based on the multi-source feedback interface, it collects raw feedback data, and constructs the first knowledge network and extracts consensus knowledge through three-level credibility weighted fusion and graph node matching.

[0159] Quality assessment module: Based on system records and the first knowledge network, it obtains the quality score of each knowledge item and calculates the usage frequency after compensation through frequency statistics and multi-condition threshold judgment;

[0160] Incremental update module: Based on the first knowledge network and the usage frequency after compensation, the module dynamically adjusts the weights and merges the nodes of the first knowledge network through attention heatmap extraction and attribute node generation to obtain a second knowledge network containing newly added nodes and associated edges.

[0161] Version migration module: Based on the second knowledge network and update records, the module performs semantic version control and quality index differential calculation, and then executes rollback condition judgment to obtain the third knowledge network and rollback event records; based on the third knowledge network and actual application feedback, the module performs cross-crop knowledge migration and graph network fusion through multi-dimensional semantic similarity calculation and subgraph cloning adaptation to obtain the fourth knowledge network and migration records.

[0162] The feedback fusion module, quality assessment module, incremental update module, and version migration module are all located on the server. The server receives data transmitted from the acquisition devices and performs further analysis. The feedback fusion module performs credibility fusion and node matching on multi-source feedback data to construct the first knowledge network and extract consensus knowledge. The quality assessment module performs frequency statistics, accuracy evaluation, and timeliness scoring on knowledge items to obtain a quality score and calculate the usage frequency after compensation. The incremental update module extracts new attribute nodes through attention heatmaps and performs dynamic weight adjustment and node fusion on the first knowledge network to obtain the second knowledge network. The version migration module performs semantic version control and automatic rollback on the second knowledge network to obtain the third knowledge network, and obtains the fourth knowledge network through cross-crop knowledge migration. These modules are connected sequentially and work collaboratively to form a complete closed loop from data acquisition to knowledge evolution, realizing the self-learning and incremental update of the knowledge base.

[0163] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0164] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0165] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0166] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A self-learning and incremental update method for a crop disease and pest knowledge base, characterized in that, include: Raw feedback data is collected based on a multi-source feedback interface. The first knowledge network is constructed and consensus knowledge is extracted through three-level credibility weighted fusion and graph node matching. The multi-source feedback interface is an application programming interface endpoint used to receive feedback data related to the diagnosis of different types of pests and diseases, including three API endpoints: user feedback interface, expert feedback interface, and actual effect feedback interface. The three-level credibility weighted fusion is a comprehensive processing method for assigning and fusing credibility to multi-source feedback data in a hierarchical manner, including two core steps: three-level credibility assignment and weighted average fusion. The three-level credibility assignment is an assignment method that assigns different basic credibility values ​​according to the source type of the feedback data, which is achieved by assigning differentiated basic credibility to the three types of sources: expert feedback, user feedback, and system inference feedback. Based on system records and the first knowledge network, the quality score of each knowledge item is obtained through frequency statistics and multi-condition threshold judgment, and the usage frequency after compensation is calculated. Based on the first knowledge network and the usage frequency after compensation, the first knowledge network is dynamically weighted and fused by extracting attention heatmaps and generating attribute nodes to obtain a second knowledge network containing newly added nodes and associated edges. Based on the second knowledge network and update records, the third knowledge network and rollback event records are obtained by performing semantic version control and quality indicator differential calculation, and then executing rollback condition judgment. Based on the third knowledge network and feedback from practical applications, cross-crop knowledge transfer and graph network fusion are performed through multi-dimensional semantic similarity calculation and subgraph cloning adaptation to obtain the fourth knowledge network and transfer records. The process of collecting raw feedback data based on a multi-source feedback interface, constructing a first knowledge network and extracting consensus knowledge through three-level credibility weighted fusion and graph node matching includes: The original feedback data is collected based on the multi-source feedback interface. The final credibility of each piece of original feedback data is obtained by three-level credibility weighted fusion, which includes three-level credibility assignment and weighted average fusion. Based on the attribute dimension division in agricultural knowledge and agricultural ontology, an initial list of entity nodes and a set of concept nodes are generated to construct an initial knowledge network skeleton. Based on the original feedback data and the final credibility, entities and attribute phrases are extracted through natural language processing. The attribute phrases are then semantically matched with the concept nodes in the knowledge network skeleton. The weights of the associated edges are dynamically updated according to the co-occurrence relationship of node pairs in the same feedback content, thus obtaining the first knowledge network and consensus knowledge.

2. The method for self-learning and incremental updating of a crop disease and pest knowledge base according to claim 1, characterized in that, The process of obtaining a quality score for each knowledge item and calculating its compensated usage frequency based on system records and the first knowledge network, through frequency statistics and multi-condition threshold judgment, includes: Based on system records, the usage frequency score of each knowledge item is obtained through frequency statistics and ranking normalization. Based on the correctness indicators of the diagnostic results in the original feedback data, the diagnostic accuracy score of each knowledge item is obtained by counting the number of positive and negative feedbacks. A timeliness score is obtained based on the difference between the last update time of each knowledge item and the system record time, using a segmented scoring standard. Based on the usage frequency score, diagnostic accuracy score, and timeliness score, a quality score for each knowledge item is obtained through weighted averaging and multi-condition threshold judgment. Based on the usage frequency recorded in the system, the compensated usage frequency of each knowledge item is calculated through underexposure coefficient compensation.

3. The method for self-learning and incremental updating of a crop disease and pest knowledge base according to claim 2, characterized in that, The calculation of the compensated usage frequency for each knowledge item, based on the usage frequency recorded in the system and through underexposure coefficient compensation, includes: Based on system records, the exposure underexposure coefficient of each knowledge item is obtained by calculating the ratio of the number of times a knowledge item is judged as relevant but not included in the return list to the total number of times it is judged as relevant. Based on the original usage frequency of all knowledge items in the knowledge base, the baseline frequency is obtained by statistically analyzing the median. Based on the original usage frequency, underexposure coefficient, and reference frequency of each knowledge entry, the product of the underexposure coefficient and the reference frequency is weighted and summed using a preset compensation coefficient to obtain the compensated usage frequency.

4. The method for self-learning and incremental updating of a crop disease and pest knowledge base according to claim 1, characterized in that, The process involves dynamically adjusting the weights and fusing nodes in the first knowledge network based on the first knowledge network and the compensated usage frequency, through attention heatmap extraction and attribute node generation, to obtain a second knowledge network containing newly added nodes and associated edges. This includes: Based on the first knowledge network and the original feedback data, the set of salient regions and the average attention value of each salient region are obtained through similarity comparison, attention heatmap extraction and connected component analysis. Based on the set of salient regions, the first knowledge network after adding attribute nodes is obtained through multimodal description model, semantic similarity matching, initial association edge establishment and one-hop neighborhood diffusion; Based on the original feedback data, the first knowledge network after adding attribute nodes, and the usage frequency after compensation, the first knowledge network after dynamic weight adjustment is obtained through negative feedback positioning and comprehensive weight adjustment. Based on the first knowledge network after dynamic weight adjustment, a second knowledge network and newly added nodes and associated edge records are obtained through weighted calculation of vector similarity and text similarity and node merging operation.

5. The method for self-learning and incremental updating of a crop disease and pest knowledge base according to claim 4, characterized in that, Based on the first knowledge network and the original feedback data, the process of obtaining a set of salient regions and the average attention value of each salient region through similarity comparison, attention heatmap extraction, and connected component analysis includes: Based on the first knowledge network, a new knowledge extraction trigger signal is obtained by similarity comparison and by combining the user confirmation identifier and final credibility in the original feedback data. Based on the new knowledge, trigger signals and their corresponding multimodal diagnostic results are extracted. The average weight of attention is extracted through a multimodal large model to obtain an attention heatmap. Based on the attention heatmap, a set of salient regions and the average attention value of each salient region are obtained through threshold segmentation and connected component analysis.

6. The method for self-learning and incremental updating of a crop disease and pest knowledge base according to claim 4, characterized in that, The first knowledge network, obtained based on the set of salient regions and through a multimodal description model, semantic similarity matching, initial association edge establishment, and one-hop neighborhood diffusion, with newly added attribute nodes, includes: Based on the set of salient regions, a text description is generated and attribute phrases are extracted using a multimodal description model to obtain a set of candidate attribute phrases; Based on the candidate attribute phrase set and the concept node set in the first knowledge network, the attribute phrases are classified into the corresponding concept nodes by semantic similarity matching to obtain the new attribute node set and its corresponding attention mapping weights. Based on the newly added attribute node set and its corresponding attention mapping weights, the first knowledge network after adding attribute nodes is obtained by establishing initial associated edges and one-hop neighborhood diffusion.

7. The method for self-learning and incremental updating of a crop disease and pest knowledge base according to claim 1, characterized in that, Based on the second knowledge network and update records, the third knowledge network and rollback event records are obtained through semantic version control and quality indicator differential calculation, followed by rollback condition judgment, including: Based on the second knowledge network and update records, semantic version numbers are generated through semantic version control and changes in newly added nodes, deleted nodes and associated edges are extracted. Incremental snapshot data is obtained through serialization. Based on the second knowledge network and update records, regional information is extracted through regular expression matching of geographic keywords to obtain regional branch version records; Based on system logs, the changes in quality indicators are obtained through time-series database queries and quality indicator difference calculations. The changes in quality indicators include changes in accuracy, changes in complaint rate, and changes in error rate. Based on the changes in the quality indicators and the preset change threshold, a rollback operation is performed through Boolean logic to obtain the third knowledge network and rollback event records.

8. The method for self-learning and incremental updating of a crop disease and pest knowledge base according to claim 1, characterized in that, Based on the third knowledge network and feedback from practical applications, cross-crop knowledge transfer and graph network fusion are performed through multi-dimensional semantic similarity calculation and subgraph cloning adaptation to obtain a fourth knowledge network and transfer records, including: Based on the names of pests and diseases, symptom descriptions, and control methods of the source and target crops in the third knowledge network, a multi-dimensional semantic similarity calculation is performed using a bidirectional encoder representation method to obtain a comprehensive similarity. Based on the comprehensive similarity, expert verification records and actual application feedback records in the original feedback data, a migration credibility score is obtained through ratio statistics and weighted evaluation. Based on the source crop subgraph in the third knowledge network, subgraph cloning and adaptation are performed by querying the threshold of associated edges, replacing entity node identifiers, and scaling the weight of associated edges to obtain the adapted subgraph to be accessed. Based on the subgraph to be accessed and the target crop subgraph in the third knowledge network, cross-crop knowledge transfer and graph network fusion are performed through semantic similarity threshold matching, associated edge merging and node insertion to obtain the fourth knowledge network and transfer records.

9. A self-learning and incremental update system for a crop disease and pest knowledge base, used to implement the self-learning and incremental update method for a crop disease and pest knowledge base as described in any one of claims 1-8, characterized in that, The self-learning and incremental update system for a crop disease and pest knowledge base includes: The feedback fusion module collects raw feedback data based on multi-source feedback interfaces, constructs the first knowledge network and extracts consensus knowledge through three-level credibility weighted fusion and graph node matching; The quality assessment module, based on system records and the first knowledge network, obtains the quality score of each knowledge item and calculates the usage frequency after compensation through frequency statistics and multi-condition threshold judgment. The incremental update module, based on the first knowledge network and the compensated usage frequency, performs dynamic weight adjustment and node fusion on the first knowledge network through attention heatmap extraction and attribute node generation to obtain a second knowledge network containing newly added nodes and associated edges. The version migration module, based on the second knowledge network and update records, performs semantic version control and quality index differential calculation, and then executes rollback condition judgment to obtain the third knowledge network and rollback event records; based on the third knowledge network and actual application feedback, it performs cross-crop knowledge migration and graph network fusion through multi-dimensional semantic similarity calculation and subgraph cloning adaptation to obtain the fourth knowledge network and migration records.

Images