Method for constructing knowledge graph of fruit and vegetable preservation strategy and intelligent decision system

Abstract

The application discloses a fruit and vegetable preservation strategy knowledge graph construction method and an intelligent decision system, and the construction method comprises the following steps: constructing a fruit and vegetable preservation multi-physical field coupling field ontology, generating a standardized vector by using a structure-semantic dual representation strategy and a gate fusion network; converting acquired multi-sensing data into natural language corpus conforming to the ontology structure; guiding a large language model to perform small sample knowledge extraction through a dependency syntax structure, and performing dynamic optimization by using reinforcement learning; adopting multi-source entity alignment driven by self-supervised fusion and ontology constraint; absorbing new triplets after alignment and fusion into a knowledge graph, and completing real-time updating and multi-dimensional verification not dependent on a cloud at an edge end; selecting a RotatE model in a complex space as a basic embedding model for training, performing knowledge completion, and finally outputting an interpretable preservation strategy suggestion. The application improves the coverage of characteristic fruits and vegetables and emerging technologies.

Description

Technical Field

[0001] This invention relates to the field of agricultural artificial intelligence and decision support technology, specifically to a method for constructing a knowledge graph of fruit and vegetable preservation strategies and an intelligent decision-making system. Background Technology

[0002] The performance boundary of an artificial intelligence model inherently depends on the quality of its training data. Regardless of how optimized the algorithm architecture is, without high-quality data, the model will be unable to learn the core knowledge and rules of the domain. While existing general-purpose language models perform exceptionally well on massive amounts of general data, in knowledge graph-driven intelligent decision-making for fruit and vegetable preservation, if the accuracy and coverage of the specialized corpus used in knowledge graph construction cannot be guaranteed, the AI ​​will fail to meet practical needs in vertical applications and become merely a decorative element. Furthermore, intelligent decision-making for fruit and vegetable preservation involves a large number of dynamic environmental parameters, varietal physiological characteristics, and implicit process experience. Traditional extraction techniques struggle to accurately identify complex semantic relationships such as temperature and humidity coupling, gas composition regulation, and ethylene release, thus hindering deep knowledge mining.

[0003] Existing NLP models struggle with terminology recognition in the fruit and vegetable preservation field. This field involves numerous specialized terms (such as "controlled atmosphere storage" and "physiological chilling injury"), exhibiting strong semantic heterogeneity. When the same concept (such as "pre-cooling") has multiple expressions in different scenarios, and a unified ontology standard is lacking, significant difficulties arise in entity alignment and high error rates in entity and relation extraction. Furthermore, existing knowledge graphs primarily focus on common fruits and vegetables (such as apples and cabbage), lacking comprehensive coverage of specialty vegetables (such as lilies and arrowhead) and emerging preservation technologies (such as spectral non-destructive testing and heat shock preservation). They rely heavily on static literature and expert experience, lacking automatic access to real-time monitoring data and incremental learning mechanisms, thus limiting the timeliness of the knowledge graphs. Moreover, knowledge completion algorithms have limited reasoning capabilities in vertical domains, with insufficient prediction accuracy for missing entities or relations, making it difficult to support high-precision, interpretable fruit and vegetable preservation decisions. Summary of the Invention

[0004] The purpose of this invention is to provide a method for constructing a knowledge graph of fruit and vegetable preservation strategies and an intelligent decision-making system to solve the problems mentioned in the background art.

[0005] The specific technical solution provided by this invention is as follows: a method for constructing a knowledge graph of fruit and vegetable preservation strategies, comprising the following operational steps: A multi-physics coupled ontology for fruit and vegetable preservation is constructed, and standardized vectors are generated by using a structure-semantic dual representation strategy and a gated fusion network. The acquired multi-sensor data is transformed into natural language corpus that conforms to the ontology structure; The large language model is guided to extract knowledge from small samples by dependency syntax structure, and dynamic optimization is carried out by reinforcement learning. Multi-source entity alignment is driven by a combination of self-supervised fusion and ontology constraints; The newly aligned and merged triples are absorbed into the knowledge graph, and real-time updates and multi-dimensional verification are completed at the edge without relying on the cloud. The RotatE model in complex space is selected as the basic embedding model for training, knowledge completion is performed, and finally interpretable preservation strategy suggestions are output.

[0006] Preferably, constructing a multi-physics coupling domain ontology for fruit and vegetable preservation includes: representing the structure of the multi-physics coupling domain ontology for fruit and vegetable preservation as: DOM={concept, attribute, relation, instance, axiom}, wherein the attribute dimension includes the ternary coupling relationship of "environmental parameters-gas composition-physical response", and using the OWL-DL description language to encode the core preservation mechanism into computable axioms, so that the ternary coupling relationship has logical reasoning ability.

[0007] Preferably, the process of converting the acquired multi-sensor data into natural language corpus that conforms to the ontology structure includes: constructing a predefined template for each coupling relationship based on the ontology of multi-physics coupling in fruit and vegetable preservation; inserting the filtered data into the predefined template according to the ontology concept-attribute mapping function; and converting the multi-dimensional sensor numerical stream into structured descriptive sentences to complete the generation of natural language corpus.

[0008] Preferably, the small-sample knowledge extraction includes: constructing a small-sample support set using the generated dynamic corpus combined with a small amount of manually annotated literature corpus; setting a prompt template under the constraints of the multi-physics coupling domain of fruit and vegetable preservation ontology; performing preliminary extraction based on the prompt template using a two-stage hybrid extraction method to generate candidate triples; using a dependency parser to parse all statements in the support set to obtain the dependency tree of each sentence; summarizing several typical dependency paths for each ontology relation to construct a dependency syntax pattern library; and simultaneously verifying and correcting relation types based on the predefined path pattern library.

[0009] Preferably, the two-stage hybrid extraction includes: the first stage uses a fine-tuned large language model for initial screening; the second stage performs dependency syntax verification and relation correction.

[0010] Preferably, dynamic optimization using reinforcement learning includes: optimizing the extractor using a reinforcement learning mechanism based on proximal policy optimization, and using dependency parsing results and ontology consistency as reward signals for self-optimization.

[0011] Preferably, the multi-source entity alignment driven by a combination of self-supervised fusion and ontology constraints includes: unified modeling and self-supervised pre-alignment of multi-source data; constructing a lightweight graph attention network for iterative entity alignment; evidence fusion when multiple data sources contribute conflicting values ​​to the same entity-attribute; and logical consistency testing of the fused triples.

[0012] Preferably, the multidimensional verification includes: First verification: the original sensor data is subjected to Kalman filtering residual analysis again, and the data is marked when the covariance of the measurement residual exceeds a preset threshold; Second verification: the ontology constraints are verified to check whether the new triplet conflicts with the ontology axiom library; Third verification: the consistency of historical trends is detected, and the expected value of the current parameter is given based on the embedded spatiotemporal prediction model. When the actual observation deviation is greater than the set deviation, manual review is triggered.

[0013] Preferably, the final output of an interpretable preservation strategy suggestion includes: selecting a RotatE model in complex space as the basic embedding model for training; incorporating ontology axiom constraints into the loss function; directly using the trained RotatE model to answer specific queries in the preservation decision; constructing the query into a link prediction task to predict tail entities, calculating scores for all possible candidate tail entities, outputting the top-10 candidates with the highest scores, and providing a normalized probability distribution; simultaneously starting an attention-based graph neural network to reason along multi-hop paths in the graph, highlighting causal chains, and outputting an interpretable preservation strategy suggestion.

[0014] On the other hand, the present invention also provides a fruit and vegetable preservation intelligent decision-making system driven by a knowledge graph of fruit and vegetable preservation strategies, including the following functional modules: The intelligent filing module for fruit and vegetable information is used to input multimodal information to file fruit and vegetable information. After the filing is completed, it automatically associates and matches basic physiological knowledge in the atlas. Multi-scenario preservation strategy intelligent recommendation module: used to semantically match the optimal combination of preservation solutions from the knowledge graph based on the multimodal information of fruits and vegetables, and output them in order of priority; Intelligent control and decision-making module for storage environment parameters: It is used to convert the optimal parameter knowledge in the map into set values ​​that can be directly sent to the intelligent controller for intelligent control of storage environment parameters; Disease prediction and control decision-making module: It is used to predict high-risk diseases based on environmental and physiological conditions by utilizing the causal reasoning chain of knowledge graphs, and recommend targeted control measures. The intelligent shelf-life prediction module is used to dynamically predict the remaining shelf life of fruits and vegetables based on the standard shelf life of the product category in the knowledge graph and the calculated real-time environmental correction coefficient. Loss warning and dynamic strategy optimization module: used to monitor environmental deviations in real time, adjust preservation combination strategies, and achieve dynamic adaptive decision-making; Knowledge Base Management Module: This module provides knowledge graph operation and maintenance tools for different user roles to manage and maintain knowledge.

[0015] Compared with the prior art, the beneficial effects achieved by the present invention are: (1) This invention solves the problems of conceptual ambiguity and terminology barriers in the field of fruit and vegetable preservation by constructing an ontology containing the axiom of multi-physics coupling and generating standardized vectors by using structure-semantic dual embedding and gating fusion. It provides a unified and computable semantic benchmark and reduces the initial alignment error rate.

[0016] (2) This invention reduces the high error rate and lack of interpretability in the extraction of complex coupling relationships in vertical domains by injecting the dual constraints of dependency syntax rules and ontology axioms into the small sample extraction and reinforcement learning optimization of large language models.

[0017] (3) This invention systematically eliminates entity ambiguity and attribute conflict problems when multiple sources of knowledge converge by self-supervised pre-aligned embedding, graph attention network iterative seeding, reputation-time adaptive conflict fusion and inviolable axiom verification, so that the constructed fusion knowledge graph has high consistency and high credibility. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the dynamic knowledge graph self-evolution construction method provided in the embodiments of the present invention; Figure 2 This is a schematic diagram of the architecture of a fruit and vegetable preservation intelligent decision-making system driven by a knowledge graph of fruit and vegetable preservation strategies, provided in an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention are within the scope of protection of the present invention.

[0020] Example 1: In this embodiment, as Figure 1As shown, this invention employs a dynamic knowledge graph self-evolution construction method driven by the coupling of three domains: physics, information, and ontology. Unlike traditional static document dependence and general extraction techniques, this invention constructs a multi-physics coupled domain ontology, introduces edge-based real-time sensing corpus generation, integrates dependency syntax and large language models for interpretable small-sample extraction, and achieves incremental learning through gated semantic alignment and lightweight graph neural networks. Finally, it completes knowledge reasoning under the constraints of first principles in coupled fruit and vegetable physiology.

[0021] For example, the construction of a knowledge graph method for fruit and vegetable preservation strategies includes: Step S1: Construct a multi-physics coupled ontology for fruit and vegetable preservation, and generate standardized vectors using a structure-semantic dual representation strategy and a gated fusion network.

[0022] In this embodiment, the present invention addresses the problems of insufficient coverage of specialty fruits and vegetables and emerging preservation technologies, as well as the difficulties in entity alignment caused by the heterogeneity of terms such as "pre-cooling" and "controlled atmosphere." It establishes a standardized domain ontology covering over 200 types of specialty fruits and vegetables and explicitly expressing multi-physics coupling mechanisms, providing a semantic foundation with strong logical constraints for all subsequent knowledge processing. Its implementation process includes ontology construction, OWL-DL axiom formalization, and structure-semantic joint embedding generation.

[0023] For example, in constructing an ontology for the multiphysics coupling of fruit and vegetable preservation, this invention uses current guidelines, norms, and standards for the storage and transportation of fresh agricultural products, as well as the first principles of "Postharvest Physiology of Fruits and Vegetables," as knowledge sources to construct an ontology DOM consisting of a five-tuple of concepts, attributes, relations, instances, and axioms. Conceptual layer: This invention utilizes the respiratory metabolism type of fruits and vegetables after harvest to establish two mutually exclusive base classes: "respiratory climacteric fruits and vegetables" and "non-respiratory climacteric fruits and vegetables". Based on these two mutually exclusive base classes, more than 200 kinds of characteristic fruits and vegetables, such as lily and arrowhead, are further subdivided.

[0024] Attribute Groups: The following attribute groups are defined in this invention: Environmental physical properties: temperature (unit: °C, accuracy ±0.3 °C), relative humidity (unit: %RH, accuracy ±2%), ethylene concentration (unit: ppm, accuracy ±0.05 ppm). partial pressure; Physiological response attribute: respiratory intensity ( ), Ethylene release rate ( ), cell membrane permeability, epidermal browning index, and incidence of cold damage; Process control attributes: precooling method, modified atmosphere ratio, thermal shock treatment temperature-time combination, and characteristic wavelength of non-destructive spectral testing.

[0025] Relationship Definition: In addition to the conventional subclass (is_a) and the quantitative descriptive property bound to that class (hasProperty), the relationship part of this invention also defines a coupling relationship template of "temperature and humidity-ethylene-physiology", including but not limited to synergizes_with, inhibits, causes, and modulates_response. Coupling relationships are established with other ontology classes through OWL object properties, for example: `isCausedBy` is associated with `StorageEnvironment`: indicating that chilling injury is caused by improper storage conditions.

[0026] The "affects" clause is related to "FruitOrVegetable": it indicates the specific affected plant individual.

[0027] isObservedAs is linked to PhysiologicalResponse: it links chilling injury to changes in physiological indicators such as abrupt changes in respiration intensity and abnormal ethylene production.

[0028] The relational template Temperature×RelativeHumidity→EthyleneProductionRate indicates the combined effect of temperature and humidity on the ethylene release rate.

[0029] In this invention, the attribute dimensions primarily revolve around a ternary coupling of physical, physiological, and technological aspects. To enable logical reasoning within this coupling, the core preservation mechanism is encoded into computable axioms using the OWL-DL description language. This embodiment illustrates the axioms with specific examples. Regarding the synergistic effect of temperature, humidity, and ethylene, a general implied axiom is defined: for any stored individual fruit or vegetable... When it is at temperature and relative humidity Under these conditions, the ethylene release rate exhibits a variety-specific nonlinear function. It also defines mutual exclusion and triggering axioms, such as the sufficient condition for "physiological chilling injury" to occur: when the storage temperature is lower than the critical temperature of the variety. And duration When the safety threshold is exceeded, the individual fruit or vegetable must belong to the ChillingInjury class. This means that the OWL-DL axioms are used to force logical triggering and reasoning within the ChillingInjury class, forcing the instance to establish an association with the ChillingInjury class. Once the axioms in this invention are written into the ontology, the subsequent knowledge graph can achieve automatic reasoning. Similarly, when the edge sensor detects that the lily is in a state of... After 10 hours in the environment, even if the input data does not contain the word "chilling injury", the knowledge graph can deduce through axioms that the lily individual has undergone chilling injury and mark it.

[0030] Finally, to make the ontology computable in the vector space, this invention employs a dual representation strategy of structural embedding and semantic embedding: Structural embedding generation: Using the constructed OWL ontology file as input, the Word2Vec model is run, treating the ontology as a directed graph where nodes include classes, individuals, and attributes, and edges represent axioms and relations. Multi-strategy random walk generation of corpus sequences: A depth-first approach is used for concept hierarchy (walk parameters...). To capture fine-grained parent-child relationships; to prioritize breadth over depth in relational chains. To capture lateral interactions, in this embodiment, the length of each sequence is controlled within 40 elements, and 200 sequences are generated around each node. Then, the node embeddings are trained using a Skip-gram model with a window size c=5, a negative sampling number of 10, and an embedding dimension of 128. The training objective is to maximize the log probability, ultimately obtaining a 128-dimensional structural embedding vector for entity e. ; Semantic embedding generation: Standard labels, a list of synonyms (e.g., "controlled atmosphere storage," "CA storage"), and textual description annotations are extracted from each entity in the ontology and concatenated into a natural language text. The pre-trained model Sentence-BERT is then used to encode this text into a 384-dimensional semantic vector. The model has been optimized on large-scale semantic similarity tasks and can effectively capture contextual information of synonyms and definition texts.

[0031] Gated Fusion Network: In this invention, to adaptively integrate structural and semantic information, a lightweight gated fusion layer is constructed as follows: first, the structural embedding vector is... and semantic vectors By mapping independent linear layers to a unified 256-dimensional space and calculating gating coefficients, the final entity anchoring vector is obtained. In this invention, the gated fusion network utilizes equivalent alignment pairs from existing general knowledge bases (such as fruit and vegetable entities in DBpedia) for supervised training, employing cosine similarity marginal loss as the loss function. After training convergence, the final entity anchoring vector... That is, as the unique semantic identifier of an entity in subsequent fusion and reasoning, this invention eliminates terminological heterogeneity from the root.

[0032] Step S2: Deploy edge intelligent nodes to transform real-time IoT sensor data into natural language corpus that conforms to the ontology structure, thereby generating dynamic knowledge.

[0033] In this embodiment, the problem of timeliness loss caused by the reliance on static documents and inability to absorb dynamic environmental data in traditional knowledge graphs is addressed. This invention deploys edge intelligent nodes, utilizes Kalman filtering to suppress noise, and transforms the purified data into high-fidelity natural language corpus using ontology templates.

[0034] For example: This invention collects corresponding data at a set frequency using multi-parameter sensors. All sensors aggregate the raw data stream to an edge computing platform via the RS-485 / Modbus protocol. However, raw sensor signals often contain electronic noise and environmental disturbances, and directly generating corpus would introduce a large number of spurious fluctuations. Therefore, this invention runs Kalman filtering on edge nodes to denoise and smooth the data, eliminating measurement noise. The filtered values ​​must be converted into natural language fragments containing semantics to be effectively utilized by large language models and extraction models. Based on the multi-physics coupling domain ontology for fruit and vegetable preservation defined in step S1, this invention constructs predefined templates for each coupling relationship, including ordinary description templates, contrast change templates, and threshold trigger templates. The filtered data is inserted into the predefined templates according to the ontology concept-attribute mapping function, converting the multi-dimensional sensor numerical stream into structured descriptive sentences, thus completing corpus generation. Simultaneously, to ensure the timeliness of the knowledge graph and prevent the loss of key transient events, this invention establishes a dual-condition trigger mechanism: Periodic triggering: Every n hours, the edge nodes summarize the filtered data from the past m hours and generate no less than 20 descriptive statements by variety batch, covering all active storage units to ensure real-time corpus increment; Mutation trigger: When the rate of change of any parameter exceeds a preset threshold (temperature change) within 15 minutes. Humidity changes Ethylene changes When a sensor data point is generated, a highly timely description is immediately created for that part, and metadata with timestamps down to the second level is assigned. Each generated corpus entry is appended with contextual information, including the sensor ID, timestamp, and confidence score (initial values ​​based on the filter covariance). This invention combines edge computing and Kalman filtering to transform raw sensor data into high-fidelity, ontology-compliant natural language corpus in real time. This solves the problem that existing dynamic environment data cannot be automatically integrated into the knowledge graph construction pipeline, providing a continuous and accurate real-time knowledge source for the graph and compressing update latency to the hourly level.

[0035] Step S3: Guide the large language model to extract knowledge from small samples through dependency syntax structure, and use reinforcement learning for dynamic optimization.

[0036] In this embodiment, to address the problem of scarce labeled data and high error rate in general model extraction for complex coupling relationships (such as temperature and humidity coupling and gas composition regulation) in vertical domains, the present invention constructs a hybrid enhanced few-sample knowledge extraction method, the specific operation process of which is as follows: Constructing a small-sample support set and ontology-constrained prompts: Utilizing the generated corpus combined with a small amount of manually annotated high-quality literature corpus, several "tuple-text" pairs are stratified and sampled according to relation type to construct a small-sample support set. To enable the model to distinguish between easily confused cooperative and inhibitory relations, the support set explicitly includes positive and negative samples. Designing prompt templates under ontology constraints: The prompts input to the large language model in this invention are not open-ended questions, but rather embed ontology patterns, such as "In the above text, based on the ontology relation set {cause, inhibit, cooperative}, extract the (head entity, relation, tail entity) triples"; binding the extraction task to a preset ontology relation space avoids the model generalizing to meaningless or illegal relations.

[0037] To construct a dependency syntax pattern library and define rules, a dependency syntax pattern library for the fruit and vegetable preservation domain was pre-constructed to provide linguistic interpretability and filter noise in the extraction results. Specifically, the Stanford CoreNLP dependency parser was used to parse all statements in the support set, obtaining the dependency tree for each sentence; for each ontology relation, several typical dependency path templates were summarized.

[0038] The two-stage hybrid extraction process involves two phases: The first phase is fine-tuning the large language model for initial screening. This invention uses ChatGLM-6B as the base, performing LoRa (Locally-Oriented Relationship) parameter fine-tuning on the support set. The adapter is only injected into the query and value weight matrix of the attention layer. The fine-tuning goal is to maximize the conditional probability of generating triples. After inputting a given text, the fine-tuned large language model (such as ChatGLM) performs initial extraction based on prompts, generating candidate triples. The second phase involves dependency parsing and relation correction. For each candidate triple output by the LLM, entity linking is performed in the original text sentence to determine the word fragment range of the head and tail entities in the original sentence. Dependency parsing is then performed on the sentence to extract the shortest dependency path between the head and tail entities, and this path is matched with the path templates allowed by the relation in the pattern library for similarity. If the path pattern is completely matched, it is directly retained and the confidence is increased; if the path matches a pattern of another relation in the library, the candidate relation is corrected to the other relation; if the path does not match any pattern, the candidate triple is pruned and discarded. This stage improves accuracy, and at the same time, the relation types are corrected to linguistically reasonable ontological relations, thus achieving interpretability.

[0039] End-to-end reinforcement learning self-optimization: To enable the extraction model to continuously evolve, this invention introduces the Proximal Policy Optimization (PPO) algorithm. Dependency parsing and ontology constraints are used as reward signals to fine-tune the LoRA adapter parameters through reinforcement learning. The results of dependency parsing and ontology consistency are used as reward signals for self-optimization. The training process of the Proximal Policy Optimization (PPO) algorithm includes: sampling a batch of text from a dynamic corpus in each iteration, generating triples using the current policy, calculating the reward, and updating the policy parameters based on clip loss. The advantage function uses generalized advantage estimation to reduce variance. During training, the model gradually learns to generate triples that are both validated by syntactic rules and strictly follow the physical axioms of fruit and vegetable preservation, maintaining high accuracy even on entirely new dynamic corpora lacking annotations.

[0040] Step S4: Employ multi-source entity alignment driven by a combination of self-supervised fusion and ontology constraints.

[0041] In this embodiment, when a large number of generated triples are aggregated with the existing knowledge base, it is necessary to resolve the heterogeneity of entities referring to the same process as "precooling," "precooling," and "PC," as well as the conflicting values ​​given by different sources for the optimal storage temperature of the same variety. This invention employs a strategy combining self-supervised fusion and ontology constraints to resolve entity ambiguity and attribute value conflicts in multi-source knowledge fusion, control the entity alignment error rate, eliminate noisy knowledge that violates the basic axioms of fruit and vegetable preservation, and ensure the high consistency and reliability of the knowledge graph.

[0042] For example, the operational process combining self-supervised fusion with ontology constraints includes: Step S41: Perform unified modeling and self-supervised pre-alignment on multi-source data.

[0043] In this embodiment, the present invention uniformly converts all knowledge to be fused into an RDF graph, regardless of whether the source is RDF triples extracted from literature, tables entered by experts, or existing standard knowledge bases. During this process, a preliminary type constraint is applied using the constructed multiphysics coupling domain ontology for fruit and vegetable preservation: all entities must belong to classes already defined in the multiphysics coupling domain ontology for fruit and vegetable preservation (such as variety, environmental parameters, physiological state), attribute predicates must correspond to ontology attributes, and invalid triples with incompatible types are filtered out. Then, multi-dimensional normalization processing is performed.

[0044] Finally, the heterogeneous graphs from different sources are treated as multiple views, and a cross-graph self-supervised method is used for initial embedding alignment. Specifically, for any entity... The anchoring vector is calculated using the trained gating fusion network. Simultaneously, lightweight knowledge graph embedding models (such as TransE) are used to pre-train the triplet structure to obtain its structural embedding. Concatenating vectors Input a shared multilayer perceptron encoder to generate the final aligned representation. The training objective is to maximize the final alignment representation of two entities in different graphs, given that they are known to be the same entity (based on a small number of alignment seeds constructed from a thesaurus). The cosine similarity is used to minimize the similarity of negative samples. The loss function follows the marginal loss, which makes synonymous entities from different sources closer to each other in the vector space.

[0045] Step S42: Construct a lightweight graph attention network for iterative entity alignment.

[0046] In this embodiment, a multi-source heterogeneous information graph is constructed using all entities to be fused as nodes and various relationships between them as edges (including is_a, hasProperty, and domain relationships). A two-layer graph attention network is used to iteratively refine the node representations. The embeddings of initial seed pairs (such as those known to be equivalent to "CA storage" and "controlled atmosphere storage") are forced to be close during training. The GNN propagates this equivalence information to its neighboring nodes through message passing, thereby discovering more implicit alignments, such as discovering those closely related to "CA storage". "concentration" and " The "concentration" attribute automatically aligns the corresponding entities in another graph. After each iteration, entity pairs with a cosine similarity higher than a set threshold and the closest mutuality are automatically added to the seed set, gradually expanding the alignment range. This process continues until no new seed pairs are generated.

[0047] Step S43: Perform evidence fusion when multiple data sources contribute conflicting values ​​to the same entity-attribute.

[0048] In this embodiment, a dynamic reputation value is maintained for each data source, initially set based on expert evaluation (e.g., 0.95 for relevant standards, 0.85 for expert experience, and 0.70 for unverified dynamic corpora). Each time a conflict is manually verified, if the data source is confirmed correct, its reputation value increases slightly using an exponential moving average; if incorrect, it decreases significantly. An information freshness factor is also introduced simultaneously. ,in, For the present time, The time when this knowledge was generated or collected. To control the decay rate. For rapidly changing properties such as environmental parameters, Setting it to a relatively large value (e.g., 0.01, with a half-life of approximately 70 hours) is beneficial for stable attributes such as the physiological threshold of the variety. Approaching 0. Finally, for the entity... The fusion value is calculated from multiple candidate values ​​of an attribute. For numerical attributes, the values ​​are directly weighted and summed; for categorical attributes, the category with the highest weighted probability is selected. The certainty of the fusion result is measured by the weighted entropy. If the entropy value is too high (too dispersed), multiple values ​​are retained and marked as uncertain, and then handed over to experts or downstream decision-making modules for processing according to risk preferences.

[0049] Step S44: Before the merged triples enter the final knowledge graph, a logical consistency check is performed.

[0050] In this embodiment, the fused RDF dataset is loaded into an ontology inference engine (such as HermiT or Pellet) and consistency checks are performed. The inference engine automatically applies previously defined axioms to examine fact triples. The conflict handling mechanism includes: if a logical contradiction is found during the examination, the data is not simply discarded, but a source analysis is initiated: checking whether the sensor is faulty (via NIS indicators), checking whether the entity alignment is incorrect, and checking whether the ontology axioms apply to the specific subspecies. If a knowledge error is confirmed, the conflicting triple is rejected, logged, and the data source reputation is reduced; if physical anomalies cannot be ruled out (such as the discovery of a new cold-resistant species), the conflict is temporarily suspended, and the ontology extension module is triggered to prompt human expert intervention. This mechanism ensures that the atlas absolutely does not contain noisy knowledge that violates known physiological laws.

[0051] Step S5: Rapidly absorb and verify new knowledge at the edge.

[0052] In this embodiment, to maintain the real-time performance of the knowledge graph, the present invention needs to incorporate the aligned and fused new triples into the knowledge graph and complete real-time updates and multi-dimensional verification at the edge, independent of the cloud. Specifically: when the alignment and fusion result outputs a new triple, an incremental update is triggered on the edge node. By deploying a lightweight GraphSAGE model with ≤3 layers and <5M parameters, neighborhood aggregation is performed based on the local subgraphs related to the new triple, and entity embeddings are incrementally updated. Furthermore, the present invention constructs a triple verification rule: First verification: The raw sensor data is subjected to Kalman filtering and residual analysis again. When the covariance of the measurement residual exceeds a preset threshold, it is marked. The second verification step involves validating the ontology constraints and checking whether the new triplet conflicts with the ontology axiom library. The third verification step involves checking the consistency of historical trends. Based on an embedded spatiotemporal prediction model, the expected values ​​of the current parameters are provided. If the actual observed deviation exceeds the set deviation, manual review is triggered. Finally, only triples that pass triple verification are written into the ontology. To address the issue of missing features of specialty fruits and vegetables or emerging technologies in the initial ontology, this invention also includes an automatic expansion mechanism based on statistical frequency. For example, if an unidentified entity name appears more than 50 times within 7 days, or a new preservation process (such as "pulse intense light treatment") is continuously extracted more than 30 times and cannot be mapped to an existing ontology class, all context triples of that entity are automatically collected, candidate definition descriptions are generated, and these are added to the ontology as temporary concept nodes, marked "pending review." Then, an ontology expansion suggestion package is output at the edge, containing suggested parent classes, attributes, synonyms, etc., and pushed to the domain expert client. After confirmation by the expert client, the expansion axiom is wirelessly distributed to the edge node, incorporating the new term and its relationships into the ontology, thus achieving self-growth of the knowledge system.

[0053] Step S6: Based on the completed high-quality dynamic graph, provide users with high-precision prediction of missing knowledge and interpretable preservation decision suggestions.

[0054] In this embodiment, the RotatE model in complex space is selected as the basic embedding model for training because it can effectively model symmetric, antisymmetric, inverted, and combinatorial relationships, making it very suitable for the transitivity of coupling effects in the fruit and vegetable preservation atlas (such as "high humidity synergistic with ethylene", "ethylene causes chilling injury", "high humidity synergistic with chilling injury"). Simultaneously, ontology axiom constraints are incorporated into the loss function, forcing the embedding space to respect the encoded physiological rules. The trained RotatE model can be directly used to answer specific queries in preservation decisions, constructing the query as a link prediction task to predict tail entities. For all possible candidate tail entities (all GasMixture class instances in the atlas), a score is calculated; the higher the score (i.e., the larger the absolute value of the negative number), the more likely the candidate is to be the correct answer. The top-10 candidates with the highest scores are output, along with a normalized probability distribution. Finally, to achieve interpretable reasoning, an attention-based graph neural network is activated to reason along multi-hop paths in the atlas, highlighting causal chains such as "high humidity synergistic - ethylene accumulation - increased risk of physiological chilling injury," and directly outputting preservation strategy suggestions.

[0055] Example 2 In this embodiment, as shown in Figure 2, the present invention further constructs a fruit and vegetable preservation intelligent decision-making system driven by the fruit and vegetable preservation strategy knowledge graph method constructed in Embodiment 1, including: The intelligent fruit and vegetable information filing module is used to input fruit and vegetable categories, quality status, harvesting information, and circulation scenarios through manual input / visual recognition, and automatically matches them with basic physiological knowledge within the ontology. The manual input channel includes a structured form containing fields such as variety name (supporting fuzzy matching of the ontology thesaurus), harvest maturity (e.g., fruit firmness, soluble solids content), harvest time, geographical coordinates of the origin, and pre-cooling method. During input, the system uses the constructed ontology thesaurus in real time for entity disambiguation, automatically mapping to standard concepts. The visual recognition channel integrates a lightweight image classification model (such as MobileNetV3), deployed on edge nodes. After photographing the appearance of fruits and vegetables, the model identifies the category (supporting 200+ varieties) and estimates the maturity level based on the "color-maturity" mapping attribute defined in the ontology. The recognition result is linked to the ontology instance, automatically filling in the filing information. After filing, the system uses the fruit and vegetable batch as a new node in the ontology, automatically establishing an instance_of relationship with the variety in the ontology. By querying the SPARQL map, information such as the variety's respiration type (climacteric / non-climacteric), critical chilling injury temperature, optimal storage temperature and humidity range, list of known susceptible diseases, and baseline value for expected shelf life are extracted. This information is presented in card format, forming a "digital twin" file for the batch. Simultaneously, as the batch circulates in the cold chain, edge sensors continuously generate monitoring data, and through an incremental update mechanism, dynamic attributes are constantly added to the batch's nodes, maintaining real-time synchronization between the digital file and the physical entity.

[0056] The multi-scenario preservation strategy intelligent recommendation module is used to semantically match the optimal combination of preservation solutions from the knowledge graph based on the fruit and vegetable category, circulation scenario, and expected storage duration, and output them in order of priority to avoid the bias of empirical solution selection.

[0057] Specifically, in this module, three typical scenarios are defined as different instances in the ontology: Short-term circulation scenarios: circulation time <72 hours, environmental conditions are room temperature or simple insulation; Medium- to long-term storage scenarios: storage time > 1 week, equipped with controlled atmosphere and cold chain facilities; Perishable berry category: varieties such as strawberries and bayberries have high respiration rates and thin skins.

[0058] Each scenario instance is bound to a set of constraints through axioms. For example, the axiom for the "perishable berries scenario" requires that the respiration type of fruits and vegetables be catatonic and that the peel thickness attribute be "thin". The ontology class, maturity, and expected storage and transportation days of the current batch are then converted into SPARQL queries. All class instances semantically related to these conditions are retrieved from the graph, and a comprehensive score is calculated for candidate strategies based on a multidimensional ranking function. The highest-scoring strategy is marked as the top-priority option, and the next 1-2 are considered alternatives. Furthermore, if a strategy conflicts with a mutually exclusive axiom trained on the current batch (e.g., "heat-induced treatment" and "cold-sensitive fruits and vegetables" are mutually exclusive), it is automatically classified as a disabled strategy with an explanation of the conflict.

[0059] Intelligent control and decision-making module for storage environment parameters: Based on the optimal parameter library of the spectrum, combined with real-time environmental data, it automatically outputs: precise temperature control range, humidity adjustment value, and controlled atmosphere. / The system integrates proportions, ventilation frequency, and ethylene removal methods, enabling automatic control via intelligent cold storage. The core of this invention module relies on the "variety-scenario-optimal parameter" triplet in a knowledge graph. This knowledge originates from fused multi-source literature and standard data, missing parameters supplemented by RotatE, and incremental absorption of real-time sensor data feedback optimization results. The specific query logic includes: using the current batch variety, maturity, and expected storage duration as conditions, performing graph traversal to obtain recommended values ​​and tolerance ranges for the corresponding attributes. Then, it continuously subscribes to current cold storage environment readings from the edge node MQTT message queue, compares them in real-time with the optimal parameters, and outputs the corresponding temperature, humidity, and gas deviations. The output deviations are input into a preset PID control logic (or a reinforcement learning-based control strategy) to generate specific control quantities: refrigeration unit start / stop or frequency converter adjustment to the target temperature; humidification / dehumidification device adjustment to the target humidity; nitrogen generator / The injection valve adjusts the gas regulation ratio; the ethylene removal unit starts and stops on a cycle, etc. Commands are sent to the cold storage PLC via the Modbus / OPC UA protocol to complete closed-loop control. Continuous monitoring is implemented; if the system fails to converge to the optimal range after adjustment, the strategy optimization module is triggered.

[0060] Disease prediction and control decision-making module: By utilizing the causal reasoning chain of knowledge graph, high-risk diseases can be predicted in advance based on environmental and physiological conditions before disease symptoms appear, and targeted control measures can be recommended.

[0061] The intelligent shelf-life prediction module dynamically predicts the remaining shelf life of fruits and vegetables based on the standard shelf life of the product category within the knowledge graph and the calculated real-time environmental correction coefficients, providing accurate data for outbound distribution and sales prioritization. Specifically, the standard shelf-life baseline is obtained by querying the attribute value defined in the ontology instance for that variety from the knowledge graph (e.g., the expected shelf life of lilies under optimal conditions is 30 days). This value is derived from a fusion of multi-source standards and literature. The calculated real-time environmental correction coefficients include: constructing a multi-factor correction model, extracting correction factors such as temperature correction factor, ethylene correction factor, and chilling injury correction factor from the graph relationships, calculating the comprehensive deterioration rate, and then predicting the remaining shelf life based on the number of days stored and the consumed quality life. An automatic warning is issued when the remaining shelf life is lower than a preset near-expiration threshold (e.g., 3 days).

[0062] Loss warning and strategy dynamic optimization module: Real-time monitoring of environmental deviations. When the environment deviates from the optimal threshold or the quality indicators of fruits and vegetables deteriorate, the module automatically triggers graph traversal reasoning to quickly generate an adjusted preservation combination strategy and achieve dynamic adaptive decision-making.

[0063] Knowledge Base Management Module: Provides a visual, low-code knowledge graph operation and maintenance tool for different user roles, supporting the creation, deletion, modification, and query of knowledge, adjustment of strategy adaptation relationships, and batch import of industry standards to ensure the sustainable growth of the knowledge graph.

[0064] It should be noted that, in this invention, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0065] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for constructing a knowledge graph of fruit and vegetable preservation strategies, characterized by: The following steps are included: A multi-physics coupled ontology for fruit and vegetable preservation is constructed. A structure-semantic dual representation strategy is used to generate structure embedding vectors and semantic embedding vectors for the ontology. The structure embedding vectors and semantic embedding vectors are fused through a gated fusion network to generate standardized vectors for the entities. The acquired multi-sensor data is transformed into natural language corpus that conforms to the ontology structure; The large language model is guided to extract knowledge from small samples by dependency syntax structure, and dynamic optimization is carried out by reinforcement learning. The small sample knowledge extraction includes the following process: Constructing a small sample support set and ontology-constrained prompts: Using the generated corpus combined with a small amount of manually annotated high-quality literature corpus, several "tuple-text" pairs are stratified and sampled according to relation type to construct a small sample support set and design prompt templates under ontology constraints. A dependency syntax pattern library for the field of fruit and vegetable preservation is constructed. A dependency parser is used to parse all statements in the support set to obtain the dependency tree of each sentence. For each ontology relation, several typical dependency path templates are summarized. Two-stage mixed extraction: The first stage is to fine-tune the large language model for initial screening. After inputting the given text, the fine-tuned large language model performs preliminary extraction based on the prompts to generate candidate triples. The second stage is to perform dependency syntax verification and relation correction on candidate triples based on the dependency syntax pattern library; Multi-source entity alignment is driven by a combination of self-supervised fusion and ontology constraints; The newly aligned and merged triples are absorbed into the knowledge graph, and real-time updates and multi-dimensional verification are completed at the edge without relying on the cloud. The RotatE model in complex space is selected as the basic embedding model for training, knowledge completion is performed, and finally interpretable preservation strategy suggestions are output.

2. The method for constructing a knowledge graph of fruit and vegetable preservation strategies according to claim 1, characterized in that: The construction of the multiphysics coupling domain ontology for fruit and vegetable preservation includes: representing the structure of the multiphysics coupling domain ontology for fruit and vegetable preservation as follows: DOM = {concepts, attributes, relationships, instances, axioms} Among them, the attribute dimension includes the three-element coupling relationship of "environmental parameters-gas composition-physiological response". The core preservation mechanism is encoded into computable axioms using the OWL-DL description language, so that the three-element coupling relationship has logical reasoning ability.

3. The method for constructing a knowledge graph of fruit and vegetable preservation strategies according to claim 2, characterized in that: The process of converting the acquired multi-sensor data into natural language corpus that conforms to the ontology structure includes: Based on the multiphysics coupling domain ontology for fruit and vegetable preservation, a predefined template is constructed for each coupling relationship; The filtered data is inserted into a predefined template based on the ontology concept-attribute mapping function; The multidimensional sensor numerical stream is converted into structured descriptive sentences to complete the generation of natural language corpus.

4. The method for constructing a knowledge graph of fruit and vegetable preservation strategies according to claim 3, characterized in that: The dynamic optimization using reinforcement learning includes: using a reinforcement learning mechanism based on proximal policy optimization to optimize the extractor, and using dependency parsing results and ontology consistency as reward signals for self-optimization.

5. The method for constructing a knowledge graph of fruit and vegetable preservation strategies according to claim 4, characterized in that: The multi-source entity alignment driven by a combination of self-supervised fusion and ontology constraints includes: Unified modeling and self-supervised pre-alignment of multi-source data; Construct a lightweight graph attention network for iterative entity alignment; Evidence fusion is performed when multiple data sources contribute conflicting values ​​to the same entity-attribute. Perform a logical consistency check on the merged triples.

6. The method for constructing a knowledge graph of fruit and vegetable preservation strategies according to claim 5, characterized in that: The multidimensional verification includes: First verification: The raw sensor data is subjected to Kalman filtering and residual analysis again. When the covariance of the measurement residual exceeds a preset threshold, it is marked. The second verification step involves validating the ontology constraints and checking whether the new triplet conflicts with the ontology axiom library. The third verification step involves checking the consistency of historical trends and providing expected values ​​for current parameters based on an embedded spatiotemporal prediction model. If the actual observation deviation exceeds the set deviation, a manual review is triggered.

7. The method for constructing a knowledge graph of fruit and vegetable preservation strategies according to claim 6, characterized in that: The final output, which can be interpreted as preservation strategy recommendations, includes: The RotatE model in complex space was selected as the basic embedding model for training. Incorporate ontology axiom constraints into the loss function; The trained RotatE model is directly used to answer specific queries in the food preservation decision-making process. The query is constructed as a link prediction task to predict tail entities. Scores are calculated for all possible candidate tail entities, the top-10 candidates with the highest scores are output, and a normalized probability distribution is provided. Simultaneously launch an attention-based graph neural network, reason along multi-hop paths in the graph, highlight causal chains, and output interpretable preservation strategy suggestions.

8. A smart decision-making system for fruit and vegetable preservation, driven by a knowledge graph of fruit and vegetable preservation strategies generated by the method for constructing a knowledge graph of fruit and vegetable preservation strategies according to any one of claims 1 to 7, characterized in that: Includes the following functional modules: The intelligent filing module for fruit and vegetable information is used to input multimodal information to file fruit and vegetable information. After the filing is completed, it automatically associates and matches basic physiological knowledge in the atlas. Multi-scenario preservation strategy intelligent recommendation module: used to semantically match the optimal combination of preservation solutions from the knowledge graph based on the multimodal information of fruits and vegetables, and output them in order of priority; Intelligent control and decision-making module for storage environment parameters: It is used to convert the optimal parameter knowledge in the map into set values ​​that can be directly sent to the intelligent controller for intelligent control of storage environment parameters; Disease prediction and control decision-making module: It is used to predict high-risk diseases based on environmental and physiological conditions by utilizing the causal reasoning chain of knowledge graphs, and recommend targeted control measures. The intelligent shelf-life prediction module is used to dynamically predict the remaining shelf life of fruits and vegetables based on the standard shelf life of the product category in the knowledge graph and the calculated real-time environmental correction coefficient. Loss warning and dynamic strategy optimization module: used to monitor environmental deviations in real time, adjust preservation combination strategies, and achieve dynamic adaptive decision-making; Knowledge Base Management Module: This module provides knowledge graph operation and maintenance tools for different user roles to manage and maintain knowledge.

Images