Crop disease and pest reasoning method and system based on time sequence cross-modal knowledge graph

Abstract

The present application relates to the field of intelligent agriculture and crop disease and pest intelligent diagnosis technology, and provides a crop disease and pest reasoning method and system based on a time sequence cross-modal knowledge graph, comprising: acquiring multi-modal fusion features in a crop planting scene to generate candidate time sequence facts; performing boundary perception incremental updating on the time sequence cross-modal knowledge graph based on the candidate time sequence facts allowed to be written, to obtain an updated time sequence cross-modal knowledge graph; reasoning, through a symbolic reasoning path based on logical rules and a representation reasoning path based on graph representation learning, about a candidate disease and pest state observed at present to obtain an explicit logical reasoning result and an implicit semantic reasoning result; calculating a fusion weight based on a cognitive uncertainty evaluation quantity in a reasoning process of current input data to perform weighted fusion, and outputting a crop disease and pest mixed reasoning result. The method improves the accuracy and robustness of crop disease and pest reasoning.

Description

Technical Field

[0001] This invention relates to the field of smart agriculture and intelligent diagnosis of crop diseases and pests, specifically to a crop disease and pest reasoning method and system based on temporal cross-modal knowledge graphs. Background Technology

[0002] High-value crops grown in protected environments are susceptible to various pests and diseases induced by complex microclimate changes during their growth, leading to severe economic losses. Traditional pest and disease control relies heavily on the experience and judgment of plant protection personnel and post-treatment application of pesticides, which has inherent drawbacks such as strong subjectivity, difficulty in knowledge transfer, and difficulty in large-scale application.

[0003] With the evolution of agricultural digitalization technologies, knowledge graph-based decision support has become an important research direction in smart agriculture. Existing solutions mostly focus on constructing static agricultural knowledge graphs, storing crop and pest entities and their relationships in a structured form of basic triples. For example, the Chinese invention patent application CN121052349A, "Method and System for Identifying Causal Relationships of Pests and Diseases and a Method and System for Constructing a Causal Graph of Pests and Diseases," acquires multimodal data and constructs event candidate triples; generates personalized question prompt templates based on the contextual information of the event candidate triples; calls a large language model for causal relationship determination to perform causal relationship determination based on the question prompt templates, obtaining the causal determination result; and constructs a causal graph of pests and diseases based on the event candidate triples and the causal determination result. However, this method lacks the ability to explicitly express temporal change information, making it difficult to characterize the nonlinear evolution of diseases during their incubation, occurrence, and spread. Although some studies have introduced multimodal data, they mostly use feature splicing or shallow association methods, failing to establish effective alignment and noise resistance mechanisms for cross-modal semantics. Meanwhile, existing knowledge graphs often employ deterministic evolution rules when processing the time dimension, failing to fully consider the impact of crop growth stage characteristics and facility microclimate disturbances on disease evolution trajectories. This static structure leads to relatively fixed graph relationship boundaries, making it prone to semantic representation biases when facing sudden microclimate changes in real agricultural scenarios, and making it difficult to effectively constrain the nonlinear transitions of entity states.

[0004] Existing technologies still suffer from insufficient adaptability when dealing with complex inference tasks related to pests and diseases that involve strong temporal dependencies and multimodal coupling. Current hybrid inference models, when integrating symbolic logic and deep learning features, typically employ a fixed linear weight allocation method, lacking the ability to quantitatively assess the uncertainty of multimodal sensing data. In actual agricultural facility environments, sensing data often suffers from quality degradation. For example, the clarity of disease phenotypic images may decrease due to factors such as occlusion and changes in lighting, or environmental sensing time-series data may contain high-frequency noise and missing data due to equipment errors. When faced with such low-quality observation data, or when inference rules corresponding to different modal features conflict, existing models tend to use ambiguous, missing, or noisy data features as reliable inference bases, leading to diagnostic conclusions with significant confidence biases and making it difficult to guarantee the reliability of hybrid inference results under complex observation conditions.

[0005] In summary, it is necessary to provide a hybrid inference method for crop diseases and pests, which can still infer the process of disease and pest status changing over time by combining crop growth stages and microclimate changes such as temperature and humidity in the facility, even when disease phenotypic images are blurred or environmental sensor data are noisy or missing. Furthermore, it can adaptively adjust the inference results according to the reliability of different data sources. Summary of the Invention

[0006] The technical problem to be solved by this invention is how to characterize the dynamic evolution law under the influence of time and environmental disturbances, so as to improve the accuracy and robustness of crop disease and pest inference.

[0007] The present invention solves the above-mentioned technical problems through the following technical means:

[0008] This invention provides a crop disease and pest reasoning method based on temporal cross-modal knowledge graphs, comprising the following steps: S1. Acquire multimodal perception data in crop planting scenarios, and perform cross-modal semantic alignment and fusion on the multimodal perception data to obtain multimodal fusion features; S2. Construct candidate fact units based on multimodal fusion features; construct stage perturbation time series labels based on crop growth stage identifiers and microclimate perturbation identifiers; associate and encapsulate time intervals, candidate fact units, and stage perturbation time series labels to generate candidate time series facts; S3. Calculate the boundary penetration energy of each relation type in the candidate time-series facts, and introduce the observation confidence coefficient to jointly construct the update admission index of the candidate time-series facts; perform boundary-aware incremental updates on the time-series cross-modal knowledge graph based on the candidate time-series facts that have been approved for writing, and obtain the updated time-series cross-modal knowledge graph. S4. Based on the updated temporal cross-modal knowledge graph, the current observed candidate pest and disease status is inferred through the symbolic reasoning path based on logical rules and the representational reasoning path based on graph representation learning, respectively, to obtain explicit logical reasoning results and implicit semantic reasoning results. S5. Calculate the fusion weight based on the cognitive uncertainty assessment of the current input data during the reasoning process, and then perform weighted fusion of the explicit logical reasoning results and the implicit semantic reasoning results according to the fusion weight to output the hybrid reasoning results of crop diseases and pests.

[0009] Furthermore, the multimodal sensing data mentioned in step S1 includes at least pest and disease phenotypic images, environmental sensing time-series data, and plant protection text data.

[0010] Furthermore, the cross-modal semantic alignment and fusion described in step S1 includes the following steps: S11. Use the Vision Transformer visual encoder to extract features from the preprocessed phenotypic images of pests and diseases, output high-dimensional visual feature vectors, and generate pest and disease symptom representation information. S12. A one-dimensional convolutional neural network with a temporal encoder is used to extract the temporal dependencies of the environmental sensing temporal data after sliding window processing, output temporal feature vectors, and generate environmental driving representation information. S13. Use a pre-trained large language model to perform natural language understanding and word embedding calculation on plant protection text data to generate semantic representation information of pests and diseases and semantic representation information of prevention and control. S14. Introduce a shared projection head to map the representation information of each modality to a shared semantic space with consistent dimensions. Within this shared semantic space, adopt a self-supervised contrastive learning strategy to introduce the InfoNCE loss function for joint alignment training to obtain multimodal fusion features.

[0011] Further, step S2 includes the following steps: S21. Perform entity recognition and relationship classification on the pest and disease symptom representation information, environmental driving representation information, pest and disease semantic representation information and prevention and control semantic representation information to generate candidate fact units; wherein, the candidate fact unit includes head entity, relationship type and tail entity, and the head entity and tail entity are selected from pest and disease entity, crop state entity, environmental factor entity and prevention and control measure entity; S22. Determine crop growth stage identifiers based on crop phenotypic change characteristics, generate microclimate disturbance identifiers based on microclimate fluctuation characteristics, and construct stage disturbance time series labels based on the coupling relationship between the crop growth stage identifiers and the microclimate disturbance identifiers. S23. Associate and encapsulate the candidate fact unit, stage perturbation time series label and time interval to generate candidate time series facts containing head entity, relation type, tail entity, time interval and stage perturbation time series label.

[0012] Further, step S3 includes the following steps: S31. Based on the crop growth stage identifier, the crop growth stage identifier is mapped into a stage-sensitive vector through a preset or trained stage embedding mapping. Based on the microclimate disturbance identifier, the microclimate disturbance identifier is mapped into a disturbance response vector through a preset or trained disturbance embedding mapping. and in combination with relation types The corresponding perturbation resonance index during the construction stage of the relation mapping matrix This leads to the construction of anisotropic dynamic evolution boundary matrix. As shown in the following formula:

[0013]

[0014] in, This represents the Sigmoid function. Representing relation type The corresponding relational mapping matrix, Representing relation type The corresponding basic direction control matrix is ​​used to characterize the basic evolution direction constraints of each state dimension under this relationship type; Representing relation type The corresponding boundary scale mapping matrix is ​​used to characterize the stage perturbation resonance index. The adjustment direction and adjustment magnitude of the anisotropic dynamic evolution boundary scale; Representing relation type The corresponding resonance expansion coefficient is used to adjust the stage sensitivity vector. With disturbance response vector The intensity of the influence of the coupling result on the anisotropic dynamic evolution boundary; This indicates element-coupled operations. This indicates the construction of a diagonal matrix from vectors; S32, Based on the head entity representation Relational representation Time interval representation and stage sensitivity vector With disturbance response vector The coupled displacement is used to construct the predictive evolution center. As shown in the following formula:

[0015] in, Representing relation type The corresponding relational mapping matrix; S33, Tail Entity Representation Predictive Evolution Center and the anisotropic dynamic evolution boundary matrix Construct boundary penetration energy As shown in the following formula:

[0016] in, Representing relation type The corresponding penetration enhancement coefficient; when When the current tail entity state breaks through the anisotropic dynamic evolution boundary under the corresponding relationship, the state transition between the head entity, the relationship type and the tail entity is constrained according to the boundary penetration energy. S34. The quality of the multimodal sensing data corresponding to the candidate time-series facts is assessed to obtain the degree of observation degradation, and then the observation reliability coefficient is constructed. An evolutionary consistency coefficient is constructed based on the boundary penetration energy of the current candidate time series facts, the boundary penetration energy in the historical evolutionary states, and the degree of state evolution deviation. As shown in the following formula:

[0017] Based on observation confidence coefficient Evolutionary Consistency Coefficient To further construct and update the access index As shown in the following formula:

[0018] in, Used to characterize the reliability of observations for current candidate time-series facts. It is used to characterize the degree of consistency between the current candidate temporal facts and the evolutionary patterns of existing temporal cross-modal knowledge graphs. For relation type The corresponding trusted consistency imbalance penalty coefficient; When the updated access index When the value is greater than a preset admission threshold, the candidate temporal fact is determined to meet the graph writing condition; when the updated admission index... When the value is not greater than the preset admission threshold, the candidate time series fact is written into the unverified cache area, and the updated admission index is recalculated after obtaining supplementary observation information; S35. For candidate time series facts that meet the map writing conditions, update the admission index accordingly. With the anisotropic dynamic evolution boundary matrix Construct boundary-aware incremental updates to obtain updated graph representation objects. As shown in the following formula:

[0019] in, This represents the graph representation object before the update. This represents the graph representation object generated from the current candidate time-series facts. This represents the Sigmoid function. For relation type The corresponding update sensitivity coefficient is used to adjust the update access index. The strength of the impact on the update magnitude of the spectral representation object; For relation type The corresponding boundary correction gain coefficient is used to adjust the anisotropic dynamic evolution boundary matrix. The correction strength of the diagonal boundary scale to the update magnitude of the map representation object; This represents the vector formed by extracting the diagonal elements of the anisotropic dynamic evolution boundary matrix. Indicates and All one vectors of the same dimension This indicates element-coupled operations.

[0020] Furthermore, the symbolic reasoning path described in step S4 specifically includes: Based on pest and disease phenotypic rules, environmental induction rules, and stage evolution rules, an explicit conditional judgment structure is constructed. The candidate pest and disease states corresponding to the current observation are input into this explicit conditional judgment structure to obtain explicit logical reasoning results under a unified probability measure space. .

[0021] Furthermore, the representation of the reasoning path described in step S4 specifically includes: Head entity representation based on updated temporal cross-modal knowledge graph Relation type representation Tail entity representation and time interval representation The validity of candidate temporal facts is scored using representational reasoning in the latent semantic space, and representational reasoning energy is constructed. As shown in the following formula:

[0022] in, Represents the L2 norm distance between vectors; Based on the stated inference energy The negative exponential or inverse mapping is used to generate initial implicit semantic inference probabilities, and then a normalized activation function is used to map these initial implicit semantic inference probabilities to the unified probability measure space to obtain the aligned implicit semantic inference result. .

[0023] Further, step S5 includes the following steps: S51. For the joint output of the symbolic reasoning path and the representational reasoning path, construct a cognitive uncertainty assessment quantity that characterizes the degree of knowledge blind spot accumulation under the current observation conditions. As shown in the following formula:

[0024]

[0025] in, This indicates the boundary penetration strength of the current candidate temporal facts on the dynamic evolution of the temporal cross-modal knowledge graph. This represents the boundary penetration energy. The non-negative part; when At that time, the boundary penetration strength A value of zero indicates that the current candidate state has not broken through the anisotropic dynamic evolution boundary under the corresponding relationship; Indicates time Multimodal input quality degradation. The symbolic rule conflict quantity is used to characterize the degree of conflict in rule support. The spectral evolution mismatch represents the degree of deviation from the state evolution. The boundary penetration enhancement coefficient is the preset or trained model parameter; S52, Assessment based on cognitive uncertainty Boundary penetration strength and path divergence Constructing the initial cognitive transfer potential As shown in the following formula:

[0026]

[0027] in, To prevent the stability constant from having a denominator of zero; This represents the boundary penetration coupling coefficient, which is a preset or trained model parameter used to adjust the boundary penetration strength. The strength of the impact on the migration trend of dual-path contribution; This represents the path divergence amplification factor, which is a preset or trained model parameter used to adjust the path divergence degree. The amplifying effect on cognitive transfer potential; S53, Initial cognitive transfer potential By performing rule conflict suppression and map mismatch enhancement correction, the corrected cognitive transfer potential is obtained. As shown in the following formula:

[0028] in, This represents the reliability correction coefficient, which is a preset or trained model parameter used to adjust the mismatch in the graph evolution. Boundary penetration strength and symbol rule conflict quantity Cognitive transfer potential The correction strength; S54, Based on the corrected cognitive transfer potential Determine the weights of symbolic reasoning paths The sum represents the inference path weight. As shown in the following formula:

[0029] Results of explicit logical reasoning Results of implicit semantic reasoning Perform weighted fusion to output the mixed inference results for crop diseases and pests. As shown in the following formula:

[0030] in, This is the final result of the hybrid reasoning.

[0031] This invention provides a crop disease and pest reasoning system based on a time-series cross-modal knowledge graph. The system executes the aforementioned method during runtime and includes the following modules: The feature extraction module is used to acquire multimodal perception data in crop planting scenarios, and to perform cross-modal semantic alignment and fusion on the multimodal perception data to obtain multimodal fusion features; The candidate time-series fact construction module is used to construct candidate fact units based on multimodal fusion features; construct stage perturbation time-series labels based on crop growth stage identifiers and microclimate perturbation identifiers; and associate and encapsulate time intervals, candidate fact units, and stage perturbation time-series labels to generate candidate time-series facts. The knowledge graph update module is used to calculate the boundary penetration energy of each relation type in the candidate time-series facts, and introduces the observation confidence coefficient to jointly construct the update admission index of the candidate time-series facts; based on the candidate time-series facts that have been approved for writing, the time-series cross-modal knowledge graph is updated with boundary awareness to obtain the updated time-series cross-modal knowledge graph. The dual-path reasoning module is used to reason about the current observed candidate pest and disease status based on the updated temporal cross-modal knowledge graph, through a symbolic reasoning path based on logical rules and a representational reasoning path based on graph representation learning, respectively, to obtain explicit logical reasoning results and implicit semantic reasoning results. The fusion output module is used to calculate the fusion weight based on the cognitive uncertainty assessment of the current input data during the reasoning process, and then to perform weighted fusion of the explicit logical reasoning results and the implicit semantic reasoning results according to the fusion weight, and output the hybrid reasoning results of crop diseases and pests.

[0032] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method.

[0033] The advantages of this invention are: (1) This invention generates anisotropic dynamic evolutionary boundaries by introducing a coupling mechanism between crop growth stage sensitivity vectors and microclimate disturbance response vectors. Through this boundary, the graph relationships are no longer limited to fixed triples, but can form dynamic constraints according to time and environmental changes, thereby characterizing the occurrence and development process of pests and diseases, providing a temporal basis for subsequent reasoning, and solving the problem that traditional static knowledge graphs cannot accurately describe the nonlinear temporal evolution of pests and diseases.

[0034] (2) This invention designs a dual-path hybrid reasoning mechanism consisting of a symbolic reasoning path and a graph representation learning path, and dynamically adjusts the reasoning weights through cognitive blind spot potential and cognitive transfer potential. This mechanism can cope with low-quality data, such as blurred images, missing sensor data, or high noise levels. Simultaneously, the rule conflict suppression and graph mismatch enhancement correction mechanism can suppress the excessive influence of the symbolic reasoning path when symbolic rule conflicts are strong, and enhance logical constraints when graph evolution is abnormal. Through these methods, this invention improves the reliability, stability, and interpretability of pest and disease diagnosis results under low-quality perceptual data.

[0035] (3) This invention overcomes the limitations of traditional static map update methods by using boundary penetration energy as the basis for anomaly identification. By combining the predicted evolution center and the dynamic boundary matrix, the system can identify abnormal changes in the state of pests and diseases when the penetration energy is greater than zero. In addition, by jointly constraining the observation confidence coefficient and the evolution consistency coefficient, low-quality or non-evolutionary candidate facts are filtered out, avoiding noise data from affecting map updates, and enabling entity and relation representations to maintain good accuracy and stability.

[0036] (4) When outputting the final hybrid inference result for pests and diseases, this invention retains the explicit logical link of symbolic reasoning and aligns the energy representing the reasoning to a unified probability space. When the two reasoning paths diverge, the system can adjust the weights based on the decline in data quality or rule mismatch. As a result, farmers and plant protection experts can understand the main reasoning basis of the system, reducing the difficulty of understanding the diagnostic results and facilitating the application of this invention in actual agricultural scenarios. Attached Figure Description

[0037] Figure 1 This is a schematic diagram illustrating the steps of the crop disease and pest reasoning method based on temporal cross-modal knowledge graph according to an embodiment of the present invention; Figure 2 This is a flowchart of a crop disease and pest reasoning method based on a temporal cross-modal knowledge graph, according to an embodiment of the present invention. Figure 3 This is a schematic diagram illustrating the construction and updating of a temporal cross-modal knowledge graph according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the temporal fact structure and dynamic evolution boundary in an embodiment of the present invention; Figure 5 This is a schematic diagram of dual-path hybrid reasoning based on cognitive uncertainty perception, as described in an embodiment of the present invention. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] Example 1 This embodiment provides a crop disease and pest inference method based on a temporal cross-modal knowledge graph. It aims to extract multi-source perceptual features through cross-modal semantic alignment and encapsulate spatiotemporal information into graph-based temporal facts by combining crop growth stages and microclimate perturbation markers. Furthermore, it constructs anisotropic dynamic evolutionary boundaries for graph relationships, constrained by observational reliability and evolutionary consistency. Based on this, a dual-path hybrid mechanism is designed, consisting of symbolic logic inference based on expert knowledge and implicit semantic inference based on graph representation learning. Then, by quantifying the quality degradation and rule conflict degree of the input data, a cognitive uncertainty assessment quantity is generated to dynamically adjust the fusion weight of the dual-path inference. Ultimately, it achieves accurate dynamic inference and adaptive diagnosis of the occurrence and development patterns of diseases and pests, effectively overcoming the defects of static graph semantic representation fixation and significantly improving the robustness and decision reliability of the hybrid inference system under low-quality perceptual data constraints. The specific implementation steps and flowchart of the method are as follows: Figure 1 and Figure 2 As shown, this embodiment uses tomato crop and its common diseases as an example to illustrate the following steps: S1. Acquire multimodal perception data in crop planting scenarios, and perform cross-modal semantic alignment and fusion on the multimodal perception data to obtain multimodal fusion features; the multimodal perception data includes at least pest and disease phenotypic images, environmental sensing time-series data, and plant protection text data; The cross-modal semantic alignment and fusion includes the following steps: S11. The Vision Transformer is used to extract features from the preprocessed phenotypic images of pests and diseases, outputting high-dimensional visual feature vectors to generate symptom representation information of pests and diseases. In this embodiment, the phenotypic images of pests and diseases are collected at regular intervals by a binocular multispectral camera or an agricultural inspection robot deployed on the sliding rail between rows in the greenhouse to capture early lesion texture features on the surface of leaves, stems or fruits, such as water-soaked dark green lesions in the early stage of late blight of tomatoes.

[0040] S12. A one-dimensional convolutional neural network with a temporal encoder is used to extract the temporal dependencies from the environmental sensing time-series data after sliding window processing, outputting a temporal feature vector to generate environmental driving characterization information. In this embodiment, the environmental sensing time-series data is collected in real time using IoT sensor nodes based on low-power local area network protocols such as LoRaWAN, reflecting macroscopic environmental parameters reflecting greenhouse microclimate changes. The collected indicators include air temperature and humidity, photosynthetically active radiation intensity, carbon dioxide concentration, soil moisture, etc., and the sampling frequency is preferably set to once every 10 to 30 minutes.

[0041] S13. A pre-trained large language model is used to perform natural language understanding and word embedding calculation on plant protection text data to generate semantic representation information of pests and diseases and semantic representation information of prevention and control. In this embodiment, unstructured text data is extracted from the local agricultural expert system database, standard plant protection manual and historical prevention and control logs by means of web crawling or interface calling. S14. A shared projection head is introduced to map the representation information of each modality to a shared semantic space with consistent dimensions, such as a 256-dimensional space. Within this shared semantic space, a self-supervised contrastive learning strategy is adopted to introduce the InfoNCE loss function for joint alignment training, obtaining multimodal fusion features. Its core physical logic lies in forcing the cosine distance between different modal features describing the same disease state in the shared vector space to narrow, while excluding features unrelated to the disease, thereby achieving low-level semantic alignment of heterogeneous features. Specifically, in actual greenhouse environments, sensor data is easily degraded due to physical interference, such as camera lenses fogging due to high humidity causing image blurring, or sensors being obstructed and generating noise. To avoid inferior modalities dominating the fusion result, this embodiment calculates modality confidence factors based on the aligned features. For example, the Laplacian variance algorithm is used to evaluate the sharpness of the input image, generating confidence weights for the visual modality. Furthermore, by utilizing adaptive weighted concatenation or gated attention mechanisms, the aligned modal representations are dynamically weighted and fused to generate the final multimodal fusion feature. When the image is severely degraded due to fogging ( When the weight of visual features is reduced, the fusion mechanism will automatically reduce the fusion weight of visual features and correspondingly increase the proportion of highly reliable environmental-driven representation information and textual semantic representation information, thereby ensuring that the output multimodal fusion features can always represent the overall state of the current tomato plant with high fidelity.

[0042] S2. Construct candidate fact units based on multimodal fusion features; construct stage perturbation time series labels based on crop growth stage identifiers and microclimate perturbation identifiers; associate and encapsulate time intervals, candidate fact units, and stage perturbation time series labels to generate candidate time series facts; this step aims to instantiate the multimodal fusion features extracted from the front end into dynamic graph nodes and edges with timestamps and state labels. By introducing the coupling perturbation between environmental microclimate and tomato growth period, the homogeneous assumption of graph relationship evolution is broken, and anisotropic dynamic evolutionary boundaries are given to the relationships in the graph, thereby accurately depicting the nonlinear evolution logic of tomato diseases under complex microclimates. The specific implementation includes the following steps: S21. Perform entity recognition and relationship classification on the pest and disease symptom representation information, environmental driving representation information, pest and disease semantic representation information and prevention and control semantic representation information to generate candidate fact units; wherein, the candidate fact unit includes head entity, relationship type and tail entity, and the head entity and tail entity are selected from pest and disease entity, crop state entity, environmental factor entity and prevention and control measure entity; S22. Determine crop growth stage identifiers based on crop phenotypic change characteristics, generate microclimate disturbance identifiers based on microclimate fluctuation characteristics, and construct stage disturbance time series labels based on the coupling relationship between the crop growth stage identifiers and the microclimate disturbance identifiers; in this embodiment, crop growth stage identifiers are determined based on the phenotypic change characteristics of tomato crops (e.g., determining the current flowering and fruit setting stage through visual features); microclimate disturbance identifiers are generated based on environmental characteristics (e.g., extracting time series fluctuation characteristics of sudden temperature drop and high humidity); the crop growth stage identifiers and microclimate disturbance identifiers are spliced ​​across spatial tensors to construct stage disturbance time series labels; S23. Associate and encapsulate candidate fact units, stage perturbation time series labels, and time intervals to generate candidate time series facts (such as tomato plant, suspected infection, early stage of late blight, [timestamp], flowering period - high humidity and sudden temperature drop perturbation) containing head entity, relation, tail entity, time interval, and stage perturbation time series labels, and use them as data to be admitted for subsequent map update steps.

[0043] S3. In actual greenhouse environments, the evolution speed and direction of diseases are highly susceptible to interference from microclimate and crop immunity. Therefore, combining crop growth characteristics with environmental disturbance characteristics, anisotropic dynamic evolution boundaries are constructed for relationships in the temporal cross-modal knowledge graph. The boundary penetration energy of each relationship type in candidate temporal facts is calculated, and an observation confidence coefficient is introduced to jointly construct the update admission index for candidate temporal facts. Based on the approved candidate temporal facts, the temporal cross-modal knowledge graph is incrementally updated with boundary awareness to obtain the updated temporal cross-modal knowledge graph. Specific implementation methods are as follows: Figure 3 and Figure 4 As shown, it includes the following steps: S31. Based on the crop growth stage identifier, the crop growth stage identifier is mapped into a stage-sensitive vector through a preset or trained stage embedding mapping. Based on the microclimate disturbance identifier, the microclimate disturbance identifier is mapped into a disturbance response vector through a preset or trained disturbance embedding mapping. and in combination with relation types The corresponding perturbation resonance index during the construction stage of the relation mapping matrix This leads to the construction of anisotropic dynamic evolution boundary matrix. As shown in the following formula:

[0044]

[0045] in, This represents the Sigmoid function. Representing relation type The corresponding relational mapping matrix, Representing relation type The corresponding basic direction control matrix is ​​used to characterize the basic evolution direction constraints of each state dimension under this relationship type; Representing relation type The corresponding boundary scale mapping matrix is ​​used to characterize the stage perturbation resonance index. The adjustment direction and adjustment magnitude of the anisotropic dynamic evolution boundary scale; Representing relation type The corresponding resonance expansion coefficient is used to adjust the stage sensitivity vector. With disturbance response vector The intensity of the influence of the coupling result on the anisotropic dynamic evolution boundary; This indicates element-coupled operations. This represents constructing a diagonal matrix from vectors; specifically, in this embodiment, All adopt standard neural network initialization strategies. The optimal value is set within the range of 0.5 to 1.0. The physical meaning of this formula is that when tomatoes are subjected to strong environmental disturbances and are in a susceptible growth period, the boundary space representing the transfer of disease status will undergo asymmetrical stretching and expansion in specific susceptible characteristic dimensions.

[0046] S32. Based on the above evolutionary boundaries, calculate the deviation between the theoretical target point and the actual observation point of the map entity evolution to identify anomalous signals. (Based on head entity representation) Relational representation Time interval representation and stage sensitivity vector With disturbance response vector The coupled displacement is used to construct the predictive evolution center. As shown in the following formula:

[0047] in, Representing relation type The corresponding relational mapping matrix.

[0048] S33, Tail Entity Representation Predictive Evolution Center and the anisotropic dynamic evolution boundary matrix Construct boundary penetration energy As shown in the following formula:

[0049] in, Representing relation type The corresponding penetration enhancement coefficient; when When the current tail entity state breaks through the anisotropic dynamic evolution boundary under the corresponding relationship, the state transition between the head entity, the relationship, and the tail entity is constrained based on the boundary penetration energy. In this embodiment, when the calculated... At this time, it indicates that the tomato disease status observed by the current multimodal sensing data has broken through the constraints of the conventional dynamic evolution boundary, that is, the system has keenly captured the signal of abnormal outbreak or unexpected deterioration of the disease.

[0050] S34. To prevent noisy and distorted multimodal sensing data from writing incorrect knowledge into the underlying graph, this embodiment introduces a graph noise-resistant purification update mechanism. For newly generated candidate time-series facts, the observation degradation degree of the multimodal sensing data is extracted to construct an observation reliability coefficient. An evolutionary consistency coefficient is constructed based on the boundary penetration energy of the current candidate time series facts, the boundary penetration energy in the historical evolutionary states, and the degree of state evolution deviation. As shown in the following formula:

[0051] Based on observation confidence coefficient Evolutionary Consistency Coefficient To further construct and update the access index As shown in the following formula:

[0052] in, Used to characterize the reliability of observations for current candidate time-series facts. It is used to characterize the degree of consistency between the current candidate temporal facts and the evolutionary patterns of existing temporal cross-modal knowledge graphs. For relation type The corresponding trusted consistency imbalance penalty coefficient; S35. For candidate time series facts that meet the map writing conditions, update the admission index accordingly. With the anisotropic dynamic evolution boundary matrix Construct boundary-aware incremental updates to obtain updated graph representation objects. As shown in the following formula:

[0053] in, This represents the graph representation object before the update. This represents the graph representation object generated from the current candidate time-series facts. This represents the Sigmoid function. For relation type The corresponding update sensitivity coefficient is used to adjust the update access index. The strength of the impact on the update magnitude of the spectral representation object; For relation type The corresponding boundary correction gain coefficient is used to adjust the anisotropic dynamic evolution boundary matrix. The correction strength of the diagonal boundary scale to the update magnitude of the map representation object; This represents the vector formed by extracting the diagonal elements of the anisotropic dynamic evolution boundary matrix. Indicates and All one vectors of the same dimension This indicates element-coupled operations.

[0054] Specifically, in this embodiment, The preferred value is between 0.2 and 0.5. When the value exceeds the preset admission threshold of 0.65, the system is authorized to execute the above incremental update formula. The absolute value penalty term... This ensures that the system will question the authenticity of data when extremely high-resolution images severely violate the objective laws of plant pathology. Simultaneously, through spatial modulation, the model allows for larger parameter update amplitudes in the feature dimension of tomato disease evolution boundary expansion, while strictly limiting the update step size in the boundary convergence dimension, thus maximizing the high fidelity and robustness of the knowledge graph.

[0055] S4, such as Figure 5 As shown, based on the updated temporal cross-modal knowledge graph, the current observed candidate pest and disease states are inferred through symbolic reasoning paths based on logical rules and representational reasoning paths based on graph representation learning, respectively, yielding explicit logical reasoning results and implicit semantic reasoning results. In this embodiment, in actual agricultural facility environments, a single data-driven model is highly susceptible to extreme observational noise, resulting in distorted conclusions. To address this, this step innovatively constructs a dual-path reasoning engine that deeply integrates expert logical rules and graph representation learning. This mechanism organically decouples and parallelizes prior knowledge of plant pathology (possessing strong logical guidance and white-box interpretability) with the implicit mining capabilities of graph neural networks for high-dimensional complex modal features (possessing deep generalization mining capabilities), thereby providing a solid logical baseline guarantee for crop pest and disease diagnosis under extremely harsh observation conditions.

[0056] The symbolic reasoning path specifically includes: Based on the phenotypic rules of pests and diseases, the environmental inducement rules, and the stage evolution rules, explicit logical reasoning is performed on the candidate pest and disease states corresponding to the current observation, and the explicit logical reasoning results under a unified probability measure space are obtained. Specifically, in this embodiment, the symbolic reasoning path acts like an experienced plant protection expert. For example, regarding tomato late blight, when the environmental sensing rule matches "high humidity and suitable temperature" and the phenotypic rule matches "water-soaked lesions appear on leaves," a deterministic high probability of late blight is strictly triggered. Even if the underlying image data is partially degraded due to fogging, this branch can still provide a judgment basis that conforms to objective physiological evolution laws based on environmental induction rules. The aforementioned inference path specifically includes: Based on the head entity representation in the aforementioned temporal cross-modal knowledge graph Relation type representation Tail entity representation and time interval representation The validity of candidate temporal facts is scored using representational reasoning in the latent semantic space, and representational reasoning energy is constructed. As shown in the following formula:

[0057] in, This represents the L2 norm distance of a vector.

[0058] Based on the stated inference energy An implicit semantic reasoning result is generated, and the implicit semantic reasoning result is mapped to the unified probability measure space to obtain an aligned implicit semantic reasoning result. Specifically, in this embodiment, the physical meaning of the formula is as follows: The head entity state, which integrates the time dimension and relational features, is transferred to the prediction space using a relational mapping matrix, and the Euclidean spatial distance difference between it and the actually observed tail entity state is calculated. The smaller the distance, the higher the potential validity of characterizing the evolution of the disease state.

[0059] S5, such as Figure 5 As shown, a fusion weight is calculated based on the cognitive uncertainty of the current input data during the reasoning process. Then, the explicit logical reasoning result and the implicit semantic reasoning result are weighted and fused according to the fusion weight to output a hybrid reasoning result for crop diseases and pests. The specific implementation includes the following steps: S51. For the joint output of the symbolic reasoning path and the representational reasoning path, construct a cognitive uncertainty assessment quantity that characterizes the degree of knowledge blind spot accumulation under the current observation conditions. As shown in the following formula:

[0060]

[0061] in, This indicates the boundary penetration strength of the current candidate temporal facts on the dynamic evolution of the temporal cross-modal knowledge graph. This represents the boundary penetration energy. The non-negative part; when At that time, the boundary penetration strength A value of zero indicates that the current candidate state has not broken through the anisotropic dynamic evolution boundary under the corresponding relationship; Indicates time Multimodal input quality degradation. The symbolic rule conflict quantity is used to characterize the degree of conflict in rule support. The spectral evolution mismatch represents the degree of deviation from the state evolution. The boundary penetration enhancement coefficient is the preset or trained model parameter; S52, Assessment based on cognitive uncertainty Boundary penetration strength and path divergence Constructing the initial cognitive transfer potential As shown in the following formula:

[0062]

[0063] in, To prevent the stability constant from having a denominator of zero; This represents the boundary penetration coupling coefficient, which is a preset or trained model parameter used to adjust the boundary penetration strength. The strength of the impact on the migration trend of dual-path contribution; This represents the path divergence amplification factor, which is a preset or trained model parameter used to adjust the path divergence degree. The amplifying effect on cognitive transfer potential; S53, Initial cognitive transfer potential By performing rule conflict suppression and map mismatch enhancement correction, the corrected cognitive transfer potential is obtained. As shown in the following formula:

[0064] in, This represents the reliability correction coefficient, which is a preset or trained model parameter used to adjust the mismatch in the graph evolution. Boundary penetration strength and symbol rule conflict quantity Cognitive transfer potential The correction strength; S54, Based on the corrected cognitive transfer potential Determine the weights of symbolic reasoning paths The sum represents the inference path weight. As shown in the following formula:

[0065] Results of explicit logical reasoning Results of implicit semantic reasoning Perform weighted fusion to output the mixed inference results for crop diseases and pests. As shown in the following formula:

[0066] in, This is the final result of the hybrid reasoning.

[0067] Among them, when the input data quality deteriorates, leading to the cognitive uncertainty assessment quantity Increase, and the path divergence degree When it increases, the cognitive transfer potential Increase, thereby adaptively increasing the weight of the symbolic reasoning path with strong logical interpretability. This reduces the weights of the inference paths that are susceptible to noise interference. .

[0068] In this embodiment, to address the common problem of multimodal sensor data degradation in agricultural greenhouse scenarios (such as lens fogging due to high humidity, sensor signal loss, etc.), a rigorous cognitive uncertainty assessment mechanism is established. By quantifying the divergence between data quality degradation and the dual-path inference model, a transfer potential is dynamically generated. This enables adaptive deprivation of the decision-making power of the noise-sensitive "graph representation learning network" under extreme observation conditions, forcibly increasing the minimum contribution of the "expert logic rules," thereby maximizing the reliability and robustness of the hybrid inference results for pests and diseases. The physical significance and innovative logic of the above fusion mechanism lie in: when the quality of agricultural environmental input data severely degrades, leading to the aforementioned cognitive uncertainty assessment quantity... The degree of divergence increases, and the two reasoning paths become severely conflicting (i.e., the path divergence degree). When significantly increased, the cognitive transfer potential driven by the exponential amplification effect... This will increase dramatically. At this point, the system automatically breaks the weight balance, greatly increasing the weight of symbolic reasoning paths that have strong logical interpretability and a strong basis in plant pathology. This corresponds to a sharp reduction in the weights of representation inference paths that are susceptible to high-noise data interference. This provides a solid logical basis for the diagnosis of crops such as tomatoes under extreme working conditions.

[0069] Example 2 It should be further explained that, based on the same inventive concept, this embodiment provides a crop disease and pest reasoning system based on a time-series cross-modal knowledge graph. When the system runs, it executes the method described in Embodiment 1, including the following modules: The feature extraction module is used to acquire multimodal perception data in crop planting scenarios, and to perform cross-modal semantic alignment and fusion on the multimodal perception data to obtain multimodal fusion features; The candidate time-series fact construction module is used to construct candidate fact units based on multimodal fusion features; construct stage perturbation time-series labels based on crop growth stage identifiers and microclimate perturbation identifiers; and associate and encapsulate time intervals, candidate fact units, and stage perturbation time-series labels to generate candidate time-series facts. The knowledge graph update module is used to calculate the boundary penetration energy of each relation type in the candidate time-series facts, and introduces the observation confidence coefficient to jointly construct the update admission index of the candidate time-series facts; based on the candidate time-series facts that have been approved for writing, the time-series cross-modal knowledge graph is updated with boundary awareness to obtain the updated time-series cross-modal knowledge graph. The dual-path reasoning module is used to reason about the current observed candidate pest and disease status based on the updated temporal cross-modal knowledge graph, through a symbolic reasoning path based on logical rules and a representational reasoning path based on graph representation learning, respectively, to obtain explicit logical reasoning results and implicit semantic reasoning results. The fusion output module is used to calculate the fusion weight based on the cognitive uncertainty assessment of the current input data during the reasoning process, and then to perform weighted fusion of the explicit logical reasoning results and the implicit semantic reasoning results according to the fusion weight, and output the hybrid reasoning results of crop diseases and pests.

[0070] This embodiment also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in Embodiment 1.

[0071] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A crop disease and pest reasoning method based on temporal cross-modal knowledge graph, characterized in that, Includes the following steps: S1. Acquire multimodal perception data in crop planting scenarios, and perform cross-modal semantic alignment and fusion on the multimodal perception data to obtain multimodal fusion features; the multimodal perception data includes at least pest and disease phenotypic images, environmental sensing time-series data, and plant protection text data; S2. Construct candidate fact units based on multimodal fusion features. Candidate fact units include head entities, relation types, and tail entities. Head entities and tail entities are selected from pest and disease entities, crop status entities, environmental factor entities, and control measure entities. Construct stage perturbation time series labels based on crop growth stage identifiers and microclimate perturbation identifiers. Associate and encapsulate time intervals, candidate fact units, and stage perturbation time series labels to generate candidate time series facts. S3. Calculate the boundary penetration energy of each relation type in the candidate time-series facts, and introduce the observation confidence coefficient to jointly construct the update admission index of the candidate time-series facts. When the boundary penetration energy is greater than 0, it is used to characterize that the crop disease state observed by the current multimodal sensing data has broken through the constraints of the conventional dynamic evolution boundary, that is, the system has keenly captured the abnormal outbreak or unexpected deterioration signal of the disease. The observation confidence coefficient is used to characterize the observation confidence level of the current candidate time-series facts. The update admission index is used to determine whether the candidate time-series facts meet the map writing conditions. The temporal cross-modal knowledge graph is updated by using candidate temporal facts whose update admission index is greater than a preset threshold to perform boundary-aware incremental updates, resulting in an updated temporal cross-modal knowledge graph. S4. Based on the updated temporal cross-modal knowledge graph, the current observed candidate pest and disease status is inferred through the symbolic reasoning path based on logical rules and the representational reasoning path based on graph representation learning, respectively, to obtain explicit logical reasoning results and implicit semantic reasoning results. S5. Calculate the fusion weight based on the cognitive uncertainty assessment of the current input data during the reasoning process, and then perform weighted fusion of the explicit logical reasoning results and the implicit semantic reasoning results according to the fusion weight to output the hybrid reasoning results of crop diseases and pests.

2. The crop disease and pest reasoning method based on temporal cross-modal knowledge graph as described in claim 1, characterized in that, The cross-modal semantic alignment and fusion described in step S1 includes the following steps: S11. Use the Vision Transformer visual encoder to extract features from the preprocessed phenotypic images of pests and diseases, output high-dimensional visual feature vectors, and generate pest and disease symptom representation information. S12. A one-dimensional convolutional neural network with a temporal encoder is used to extract the temporal dependencies of the environmental sensing temporal data after sliding window processing, output temporal feature vectors, and generate environmental driving representation information. S13. Use a pre-trained large language model to perform natural language understanding and word embedding calculation on plant protection text data to generate semantic representation information of pests and diseases and semantic representation information of prevention and control. S14. Introduce a shared projection head to map the representation information of each modality to a shared semantic space with consistent dimensions. Within this shared semantic space, adopt a self-supervised contrastive learning strategy to introduce the InfoNCE loss function for joint alignment training to obtain multimodal fusion features.

3. The crop disease and pest reasoning method based on temporal cross-modal knowledge graph as described in claim 2, characterized in that, Step S2 includes the following steps: S21. Perform entity recognition and relationship classification on the pest and disease symptom representation information, environmental driving representation information, pest and disease semantic representation information and prevention and control semantic representation information to generate candidate fact units. S22. Determine crop growth stage identifiers based on crop phenotypic change characteristics, generate microclimate disturbance identifiers based on microclimate fluctuation characteristics, and construct stage disturbance time series labels based on the coupling relationship between the crop growth stage identifiers and the microclimate disturbance identifiers. S23. Associate and encapsulate the candidate fact unit, stage perturbation time series label and time interval to generate candidate time series facts containing head entity, relation type, tail entity, time interval and stage perturbation time series label.

4. The crop disease and pest reasoning method based on temporal cross-modal knowledge graph according to claim 3, characterized in that, Step S3 includes the following steps: S31. Based on the crop growth stage identifier, the crop growth stage identifier is mapped into a stage-sensitive vector through a preset or trained stage embedding mapping. Based on the microclimate disturbance identifier, the microclimate disturbance identifier is mapped into a disturbance response vector through a preset or trained disturbance embedding mapping. and in combination with relation types The corresponding perturbation resonance index during the construction stage of the relation mapping matrix This leads to the construction of anisotropic dynamic evolution boundary matrix. As shown in the following formula: in, This represents the Sigmoid function. Representing relation type The corresponding relational mapping matrix, Representing relation type The corresponding basic direction control matrix is ​​used to characterize the basic evolution direction constraints of each state dimension under this relationship type; Representing relation type The corresponding boundary scale mapping matrix is ​​used to characterize the stage perturbation resonance index. The adjustment direction and adjustment magnitude of the anisotropic dynamic evolution boundary scale; Representing relation type The corresponding resonance expansion coefficient is used to adjust the stage sensitivity vector. With disturbance response vector The intensity of the influence of the coupling result on the anisotropic dynamic evolution boundary; This indicates element-coupled operations. This indicates the construction of a diagonal matrix from vectors; S32, Based on the head entity representation Relational representation Time interval representation and stage sensitivity vector With disturbance response vector The coupled displacement is used to construct the predictive evolution center. As shown in the following formula: in, Representing relation type The corresponding relational mapping matrix; S33, Tail Entity Representation Predictive Evolution Center and the anisotropic dynamic evolution boundary matrix Construct boundary penetration energy As shown in the following formula: in, Representing relation type The corresponding penetration enhancement coefficient; S34. The quality of the multimodal sensing data corresponding to the candidate time-series facts is assessed to obtain the degree of observation degradation, and then the observation reliability coefficient is constructed. An evolutionary consistency coefficient is constructed based on the boundary penetration energy of the current candidate time series facts, the boundary penetration energy in the historical evolutionary states, and the degree of state evolution deviation. As shown in the following formula: Based on observation confidence coefficient Evolutionary Consistency Coefficient To further construct and update the access index As shown in the following formula: in, Used to characterize the reliability of observations for current candidate time-series facts. It is used to characterize the degree of consistency between the current candidate temporal facts and the evolutionary patterns of existing temporal cross-modal knowledge graphs. For relation type The corresponding trusted consistency imbalance penalty coefficient; When the updated access index When the value is greater than a preset admission threshold, the candidate temporal fact is determined to meet the graph writing condition; when the updated admission index... When the value is not greater than the preset admission threshold, the candidate time series fact is written into the unverified cache area, and the updated admission index is recalculated after obtaining supplementary observation information; S35. For candidate time series facts that meet the map writing conditions, update the admission index accordingly. With the anisotropic dynamic evolution boundary matrix Construct boundary-aware incremental updates to obtain updated graph representation objects. As shown in the following formula: in, This represents the graph representation object before the update. This represents the graph representation object generated from the current candidate time-series facts. This represents the Sigmoid function. For relation type The corresponding update sensitivity coefficient is used to adjust the update access index. The strength of the impact on the update magnitude of the spectral representation object; For relation type The corresponding boundary correction gain coefficient is used to adjust the anisotropic dynamic evolution boundary matrix. The correction strength of the diagonal boundary scale to the update magnitude of the map representation object; This represents the vector formed by extracting the diagonal elements of the anisotropic dynamic evolution boundary matrix. Indicates and All one vectors of the same dimension This indicates element-coupled operations.

5. The crop disease and pest reasoning method based on temporal cross-modal knowledge graph according to claim 4, characterized in that, The symbolic reasoning path described in step S4 specifically includes: Based on pest and disease phenotypic rules, environmental induction rules, and stage evolution rules, an explicit conditional judgment structure is constructed. The candidate pest and disease states corresponding to the current observation are input into this explicit conditional judgment structure to obtain explicit logical reasoning results under a unified probability measure space. .

6. The crop disease and pest reasoning method based on temporal cross-modal knowledge graph according to claim 4, characterized in that, The representation of the reasoning path described in step S4 specifically includes: Head entity representation based on updated temporal cross-modal knowledge graph Relation type representation Tail entity representation and time interval representation The validity of candidate temporal facts is scored using representational reasoning in the latent semantic space, and representational reasoning energy is constructed. As shown in the following formula: in, Represents the L2 norm distance between vectors; Based on the stated inference energy The negative exponential or inverse mapping is used to generate initial implicit semantic inference probabilities, and then a normalized activation function is used to map these initial implicit semantic inference probabilities to a unified probability measure space to obtain aligned implicit semantic inference results. .

7. The crop disease and pest reasoning method based on temporal cross-modal knowledge graph according to claim 6, characterized in that, Step S5 includes the following steps: S51. For the joint output of the symbolic reasoning path and the representational reasoning path, construct a cognitive uncertainty assessment quantity that characterizes the degree of knowledge blind spot accumulation under the current observation conditions. As shown in the following formula: in, This indicates the boundary penetration strength of the current candidate temporal facts on the dynamic evolution of the temporal cross-modal knowledge graph. This represents the boundary penetration energy. The non-negative part; when At that time, the boundary penetration strength A value of zero indicates that the current candidate state has not broken through the anisotropic dynamic evolution boundary under the corresponding relationship; Indicates time Multimodal input quality degradation. The symbolic rule conflict quantity is used to characterize the degree of conflict in rule support. The spectral evolution mismatch represents the degree of deviation from the state evolution. The boundary penetration enhancement coefficient is the preset or trained model parameter; S52, Assessment based on cognitive uncertainty Boundary penetration strength and path divergence Constructing the initial cognitive transfer potential As shown in the following formula: in, To prevent the stability constant from having a denominator of zero; This represents the boundary penetration coupling coefficient, which is a preset or trained model parameter used to adjust the boundary penetration strength. The strength of the impact on the migration trend of dual-path contribution; This represents the path divergence amplification factor, which is a preset or trained model parameter used to adjust the path divergence degree. The amplifying effect on cognitive transfer potential; S53, Initial cognitive transfer potential By performing rule conflict suppression and map mismatch enhancement correction, the corrected cognitive transfer potential is obtained. As shown in the following formula: in, This represents the reliability correction coefficient, which is a preset or trained model parameter used to adjust the mismatch in the graph evolution. Boundary penetration strength and symbol rule conflict quantity Cognitive transfer potential The correction strength; S54, Based on the corrected cognitive transfer potential Determine the weights of symbolic reasoning paths The sum represents the inference path weight. As shown in the following formula: Results of explicit logical reasoning Results of implicit semantic reasoning Perform weighted fusion to output the mixed inference results for crop diseases and pests. As shown in the following formula: in, This is the final result of the hybrid reasoning.

8. A crop disease and pest reasoning system based on a temporal cross-modal knowledge graph, characterized in that, Includes the following modules: The feature extraction module is used to acquire multimodal perception data in crop planting scenarios, and to perform cross-modal semantic alignment and fusion on the multimodal perception data to obtain multimodal fusion features; the multimodal perception data includes at least pest and disease phenotypic images, environmental sensing time-series data, and plant protection text data; The candidate time-series fact construction module is used to construct candidate fact units based on multimodal fusion features. The candidate fact unit includes a head entity, a relation type, and a tail entity. The head entity and tail entity are selected from pest and disease entities, crop status entities, environmental factor entities, and control measure entities. The module constructs stage perturbation time-series labels based on crop growth stage identifiers and microclimate perturbation identifiers. The module associates and encapsulates the time interval, candidate fact units, and stage perturbation time-series labels to generate candidate time-series facts. The knowledge graph update module is used to calculate the boundary penetration energy of each relation type in the candidate time-series facts, and introduces the observation credibility coefficient to jointly construct the update admission index of the candidate time-series facts. When the boundary penetration energy is greater than 0, it is used to characterize that the tomato disease status observed by the current multimodal sensing data has broken through the constraints of the conventional dynamic evolution boundary, that is, the system has keenly captured the abnormal outbreak or unexpected deterioration signal of the disease. The observation credibility coefficient is used to characterize the observation credibility of the current candidate time-series facts. The update admission index is used to determine whether the candidate time-series facts meet the graph writing conditions. The temporal cross-modal knowledge graph is updated by using candidate temporal facts whose update admission index is greater than a preset threshold to perform boundary-aware incremental updates, resulting in an updated temporal cross-modal knowledge graph. The dual-path reasoning module is used to reason about the current observed candidate pest and disease status based on the updated temporal cross-modal knowledge graph, through a symbolic reasoning path based on logical rules and a representational reasoning path based on graph representation learning, respectively, to obtain explicit logical reasoning results and implicit semantic reasoning results. The fusion output module is used to calculate the fusion weight based on the cognitive uncertainty assessment of the current input data during the reasoning process, and then to perform weighted fusion of the explicit logical reasoning results and the implicit semantic reasoning results according to the fusion weight, and output the hybrid reasoning results of crop diseases and pests.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-7.

Images