Intelligent operation and maintenance method and system based on multi-modal large model data platform

Abstract

The application provides a multi-modal large model-based data center intelligent operation and maintenance method and system, relating to the technical field of data processing, comprising acquiring multi-modal data to construct a knowledge graph, expanding the knowledge graph based on node features and connection weights, matching similar scenarios to identify historical decision information, generating and evaluating a decision scheme set, selecting the optimal scheme and generating a new decision scheme through reinforcement learning optimization, and finally executing operation and maintenance operations and recording the results. The method realizes intelligent decision-making and self-optimization of data center operation and maintenance, improving operation and maintenance efficiency and accuracy.

Description

Technical Field

[0001] This invention relates to data processing technology, and more particularly to a method and system for intelligent operation and maintenance of a data platform based on a multimodal large model. Background Technology

[0002] With the deepening of enterprise digital transformation, data platforms, as a unified platform for managing and processing various enterprise data resources, have become an important component of enterprise informatization. Data platforms integrate data generated by various business systems, providing enterprises with one-stop data analysis and application services. However, with the surge in data volume and the increase in business complexity, the operation and maintenance management of data platforms faces unprecedented challenges. Traditional manual operation and maintenance methods are no longer sufficient to meet the demands for efficient, accurate, and timely response in large-scale data processing environments.

[0003] In recent years, with the development of artificial intelligence technology, especially the breakthroughs in large-scale model and multimodal data processing technology, new technical paths have been provided for the intelligent operation and maintenance of data platforms. Multimodal large models can simultaneously process multiple data types such as text, images, logs, and time-series data, and through knowledge graphs and reasoning technology, achieve a deep understanding of operation and maintenance scenarios and intelligent decision-making.

[0004] In existing technologies, data platform operation and maintenance methods are difficult to effectively integrate and utilize multimodal data. A large amount of unstructured data, such as system logs and monitoring images, is not fully utilized, resulting in a lack of comprehensive data support for operation and maintenance decisions and affecting the accuracy of fault diagnosis and prediction.

[0005] Existing operation and maintenance systems lack effective knowledge accumulation and experience transfer mechanisms, making it difficult to transform historical operation and maintenance experience into reusable knowledge. Every time a similar problem is encountered, it needs to be re-analyzed and solved, which greatly reduces operation and maintenance efficiency and increases operation and maintenance costs.

[0006] Traditional operation and maintenance methods lack adaptive learning capabilities, cannot automatically adjust decision-making strategies according to different scenarios, and are difficult to continuously optimize from past operation and maintenance practices, resulting in unstable execution effects of operation and maintenance solutions and difficulty in coping with complex and ever-changing data platform operating environments. Summary of the Invention

[0007] This invention provides a data platform intelligent operation and maintenance method and system based on a multimodal large model, which can solve the problems in the prior art.

[0008] A first aspect of this invention provides an intelligent operation and maintenance method for a data platform based on a multimodal large model, comprising:

[0009] Acquire multimodal data in the data platform operation and maintenance scenario, construct node features of a knowledge graph based on the multimodal data, extract scenario features of the multimodal data, construct node relationships of the knowledge graph based on the scenario features, analyze the association strength between the nodes, and construct connection weights between nodes.

[0010] The knowledge graph is expanded using the node features and the connection weights, and similar scenarios in the knowledge graph are matched based on the scene features to identify historical decision information related to the current scene. The scene features and the historical decision information are input into the reasoning module of the knowledge graph to generate a set of decision schemes, and the execution effect score of the set of decision schemes is calculated.

[0011] The optimal decision scheme is selected based on the performance score, and the execution result of the optimal decision scheme is used as the reward signal of the reinforcement learning algorithm to dynamically adjust the inference module. Based on the adjusted inference module, the optimal decision scheme is optimized using the reinforcement learning algorithm to generate a new decision scheme.

[0012] The new decision-making scheme is used to generate operation and maintenance instructions, which are then executed according to a preset execution flow. The execution process and results of the decision-making scheme are recorded.

[0013] Based on the multimodal data, node features of the knowledge graph are constructed, and scene features of the multimodal data are extracted. Based on the scene features, node relationships of the knowledge graph are constructed, and the association strength between the nodes is analyzed. The connection weights between nodes are constructed, including:

[0014] The scene features of the multimodal data are extracted and mapped to the knowledge graph space to construct the node features of the knowledge graph. Based on the scene features, semantic association analysis is performed on the node features of the knowledge graph to construct the connection relationship between the nodes. The association strength between the nodes is calculated according to the connection relationship, and the association strength is used as the connection weight between the nodes.

[0015] The knowledge graph is expanded using the node features and connection weights, and similar scenes in the knowledge graph are matched based on the scene features to identify historical decision information related to the current scene, including:

[0016] Introduce scene feature vectors into node feature vectors to calculate inter-node connection weights, and use the node feature vectors and inter-node connection weights to establish a knowledge antibody feature matrix containing multiple feature patterns.

[0017] Based on the knowledge antibody feature matrix, calculate the matching coefficient between the node feature vector and the feature pattern in the knowledge antibody feature matrix; calculate the expansion weight based on the node connection weight and the matching coefficient; dynamically expand the initial knowledge graph based on the expansion weight to obtain the expanded knowledge graph; extract the current scene features from the expanded knowledge graph based on the expansion weight, and combine the current scene features with the historical features and relationship features in the expanded knowledge graph to construct a scene feature vector;

[0018] The similarity between the scene feature vector and the historical scene feature vector is calculated using the extended weights. Similar historical scenes are identified from the extended knowledge graph based on the similarity, and corresponding decision features are extracted from the similar historical scenes based on the matching coefficient.

[0019] The decision features are fused with the scene feature vector to obtain the decision weight. The weighted function value of the decision weight and the matching coefficient is input into the knowledge antibody feature matrix for feature matching. Based on the feature matching result, historical decision information related to the current scene is obtained.

[0020] The scene features and historical decision information are input into the reasoning module of the knowledge graph to generate a set of decision solutions. The execution performance score of the set of decision solutions is then calculated, including:

[0021] Scene features and historical decisions are used as inputs to the reasoning module of the knowledge graph, and a set of nonlinear state equations is constructed in the reasoning module of the knowledge graph. The system evolution trajectory is obtained by solving the set of nonlinear state equations, and the dynamic evolution features of the knowledge graph are extracted based on the state sequence in the system evolution trajectory.

[0022] The reasoning module of the knowledge graph is used to reason about the scene features and the historical decision information based on the temporal change pattern of the dynamic evolution features, generating multiple candidate decision schemes and forming a decision scheme set.

[0023] Calculate the gradient vector of the system evolution trajectory, identify multiple control points from the gradient vector, use the control points as the benchmark points for scoring, and calculate the corresponding evaluation value based on the state information of each candidate decision scheme in the decision scheme set at the control point.

[0024] Based on the corresponding evaluation value, the corresponding neuron activation intensity is calculated, and a plasticity connection matrix representing the dynamic correlation strength between the control points is constructed. The neuron activation intensity and the plasticity connection matrix are iteratively calculated to obtain a stable network resonance state. Based on the network resonance state, the execution effect score of the candidate decision scheme is calculated.

[0025] The reasoning module of the knowledge graph uses the temporal change pattern of the dynamic evolution features to reason about the scene features and the historical decision information, generating multiple candidate decision schemes and forming a decision scheme set, including:

[0026] Scene features and historical decision information are used to construct feature vectors that represent temporal characteristics. Temporal evolution analysis is performed on the feature vectors to obtain their evolutionary rules, and dynamic change features are extracted from the evolutionary rules.

[0027] The feature vector is input into the reasoning module of the knowledge graph and transformed into a feature space to obtain a transformed feature expression. The stable region of the transformed feature expression is determined according to the evolution law of the feature vector. The stable region is used to constrain the transformed feature expression to obtain multiple feature parameters characterizing the stability of the system.

[0028] The reasoning module of the knowledge graph constructs a feature weight distribution based on the feature parameters, and calculates the changing trend of the transformed feature expression based on the feature weight distribution;

[0029] The dynamic change features are input into the reasoning module of the knowledge graph, and the transformed feature expression is reasoned based on the change trend to obtain the reasoning result. Under the constraint of the stable region, multiple candidate decision schemes are generated according to the reasoning result.

[0030] For the candidate decision schemes, an evaluation index including scheme reliability, scheme diversity and information gain is calculated. Based on the evolution law of the feature vector, a dynamically changing screening threshold is determined. The evaluation index is compared with the screening threshold, and candidate decision schemes that meet the screening threshold are selected to form a decision scheme set.

[0031] Based on the performance score, the optimal decision scheme is selected, and the execution result of the optimal decision scheme is used as the reward signal for the reinforcement learning algorithm. The inference module is dynamically adjusted as follows:

[0032] A decision ranking value is constructed based on the results of the execution effect score of the decision scheme. A phase synchronization matrix is ​​constructed by calculating the phase difference of multiple neurons based on the decision ranking value. The phase synchronization matrix is ​​used to perform co-mapping on the features of different oscillation modes to obtain the neural oscillation mode prediction value corresponding to the decision scheme. The decision ranking value is co-mapped with the neural oscillation mode prediction value through the phase synchronization matrix to obtain the comprehensive decision evaluation value. The optimal decision scheme is selected based on the comprehensive decision evaluation value.

[0033] The actual execution result of the optimal decision scheme is compared with the comprehensive evaluation value of the decision to construct an instant reward value. The neural oscillation synchronicity change value is calculated based on the phase synchronization matrix. The instant reward value is co-mapped with the neural oscillation synchronicity change value through the phase synchronization matrix to obtain a compensation reward value. The compensation reward value is accumulated over time to obtain a cumulative reward value. The cumulative reward value is input into the reinforcement learning algorithm as a reward signal.

[0034] The gradient values ​​of the inference module parameters are calculated based on the reinforcement learning algorithm. The constraint difference between the inference module parameters and the target neural oscillation mode is calculated based on the phase synchronization matrix. The inference module is dynamically adjusted by co-mapping the gradient values ​​and the historical update values ​​of the inference module parameters.

[0035] Based on the adjusted inference module, the reinforcement learning algorithm is used to optimize the optimal decision scheme and generate a new decision scheme, including:

[0036] A chaotic feature matrix is ​​constructed based on the adjusted inference module output. A reward value is generated based on the chaotic feature matrix. The state distribution of the decision scheme is determined based on the reward value. A decision optimization sequence is generated based on the state distribution.

[0037] The decision optimization sequence is input into the inference module, and the parameters of the inference module are iteratively updated according to the reward value to obtain the updated inference module parameters;

[0038] The decision scheme is expanded and optimized using the updated inference module parameters to obtain a multidimensional decision set. The optimization index of each decision scheme in the multidimensional decision set is calculated based on the chaotic feature matrix. A new decision scheme is selected from the multidimensional decision set according to the optimization index.

[0039] A second aspect of the present invention provides an intelligent operation and maintenance system for a data platform based on a multimodal large model, comprising:

[0040] The first unit is used to acquire multimodal data in the data platform operation and maintenance scenario, construct node features of the knowledge graph based on the multimodal data, extract scenario features of the multimodal data, construct node relationships of the knowledge graph based on the scenario features, analyze the association strength between the nodes, and construct connection weights between the nodes.

[0041] The second unit is used to expand the knowledge graph using the node features and the connection weights, and to match similar scenes in the knowledge graph based on the scene features to identify historical decision information related to the current scene; the scene features and the historical decision information are input into the reasoning module of the knowledge graph to generate a set of decision schemes, and the execution effect score is calculated for the set of decision schemes.

[0042] The third unit is used to select the optimal decision scheme based on the execution effect score, use the execution result of the optimal decision scheme as the reward signal of the reinforcement learning algorithm, and dynamically adjust the inference module; based on the adjusted inference module, the reinforcement learning algorithm is used to optimize the optimal decision scheme and generate a new decision scheme.

[0043] The fourth unit is used to generate operation and maintenance instructions using the new decision-making scheme, execute the operation and maintenance instructions according to the preset execution process, and record the execution process and results of the decision-making scheme.

[0044] A third aspect of the present invention provides an electronic device, comprising:

[0045] processor;

[0046] Memory used to store processor-executable instructions;

[0047] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0048] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0049] The beneficial effects of this application are as follows:

[0050] This invention achieves efficient and automated operation and maintenance of a data platform through an intelligent operation and maintenance method based on a multimodal large model, thereby improving operation and maintenance efficiency and quality. This method utilizes multimodal data to construct a knowledge graph and performs similar scenario matching based on scenario features, effectively identifying historical decision-making information related to the current scenario and providing reliable knowledge support for operation and maintenance decisions.

[0051] By inputting scene features and historical decision-making information into the reasoning module of the knowledge graph, this method can generate a diverse set of decision-making schemes and select the optimal scheme based on the execution performance score, greatly improving the accuracy and intelligence of operation and maintenance decisions. Simultaneously, this method introduces a reinforcement learning algorithm, using the execution results of the decision scheme as a reward signal to dynamically adjust the reasoning module, achieving continuous optimization and self-improvement of the decision schemes.

[0052] This invention records the execution process and results of decision-making schemes, establishing a complete operational knowledge loop. This enables the system to continuously learn from historical experience, improving its adaptability and processing efficiency in complex operational scenarios. Overall, this method organically combines multimodal data analysis, knowledge graph reasoning, and reinforcement learning to construct an intelligent and adaptive data platform operational system, significantly reducing operational costs and enhancing the intelligence level of operational processes. Attached Figure Description

[0053] Figure 1 This is a flowchart illustrating the intelligent operation and maintenance method for a data platform based on a multimodal large model, as described in an embodiment of the present invention.

[0054] Figure 2 This is a flowchart illustrating the network behavior evaluation process based on dynamic feature reasoning and neuron activation intensity, as described in an embodiment of the present invention.

[0055] Figure 3 This is a flowchart illustrating the decision-making scheme generation process based on dual-path feature analysis in an embodiment of the present invention. Detailed Implementation

[0056] 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 with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and 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.

[0057] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0058] Figure 1 This is a flowchart illustrating the intelligent operation and maintenance method for a data platform based on a multimodal large model, as described in an embodiment of the present invention. Figure 1 As shown, the method includes:

[0059] Acquire multimodal data in the data platform operation and maintenance scenario, construct node features of a knowledge graph based on the multimodal data, extract scenario features of the multimodal data, construct node relationships of the knowledge graph based on the scenario features, analyze the association strength between the nodes, and construct connection weights between nodes.

[0060] The knowledge graph is expanded using the node features and the connection weights, and similar scenarios in the knowledge graph are matched based on the scene features to identify historical decision information related to the current scene. The scene features and the historical decision information are input into the reasoning module of the knowledge graph to generate a set of decision schemes, and the execution effect score of the set of decision schemes is calculated.

[0061] The optimal decision scheme is selected based on the performance score, and the execution result of the optimal decision scheme is used as the reward signal of the reinforcement learning algorithm to dynamically adjust the inference module. Based on the adjusted inference module, the optimal decision scheme is optimized using the reinforcement learning algorithm to generate a new decision scheme.

[0062] The new decision-making scheme is used to generate operation and maintenance instructions, which are then executed according to a preset execution flow. The execution process and results of the decision-making scheme are recorded.

[0063] In one optional implementation, node features of a knowledge graph are constructed based on the multimodal data, and scene features of the multimodal data are extracted; node relationships of the knowledge graph are constructed based on the scene features, and the association strength between the nodes is analyzed. The construction of connection weights between nodes includes:

[0064] The scene features of the multimodal data are extracted and mapped to the knowledge graph space to construct the node features of the knowledge graph. Based on the scene features, semantic association analysis is performed on the node features of the knowledge graph to construct the connection relationship between the nodes. The association strength between the nodes is calculated according to the connection relationship, and the association strength is used as the connection weight between the nodes.

[0065] Multimodal data includes, but is not limited to, text, image, audio, and video data. Different feature extraction methods are used for different types of multimodal data. For text data, a pre-trained language model is used to extract semantic features; for image data, a convolutional neural network is used to extract visual features; for audio data, acoustic feature extraction techniques are used to obtain audio features; and for video data, temporal and spatial information are combined to extract video features.

[0066] Taking an e-commerce scenario as an example, multimodal data includes product description text, product images, user review audio, and product demonstration videos. Keywords and semantic information, such as "high-definition monitor," "27 inches," and "4K resolution," are extracted from the product description text; visual features such as color, shape, and category are extracted from the product images; sentiment and key evaluation points are extracted from the user review audio; and dynamic usage scenarios and functional demonstration information are extracted from the product demonstration videos.

[0067] When extracting scene features from multimodal data, cross-modal fusion technology is employed to fuse features from different modalities. Specifically, an attention mechanism is used to weight and fuse features from different modalities, automatically adjusting the importance weights of each feature based on the needs of the current task. For example, when judging product quality, more attention is paid to user reviews and product display videos; while when classifying products, more attention is paid to product images and descriptive text. By calculating the correlation between features from different modalities, the weight coefficients for feature fusion are determined, achieving intelligent feature fusion.

[0068] After scene features are extracted, they are mapped to a knowledge graph space to construct node features. The knowledge graph space is a high-dimensional feature space, where each node represents an entity, and the node feature is the representation of that entity in the feature space. The mapping process is implemented using a deep neural network. The extracted scene features are input into the neural network, undergo multiple layers of nonlinear transformations, and output node feature representations that conform to the distribution of the knowledge graph space.

[0069] In e-commerce scenarios, nodes can be entities such as products, categories, attributes, and users. For a product node, its features include information such as product name, brand, price range, and main functions; for a user node, its features include information such as user preferences, purchase history, and browsing behavior. These features are represented in the knowledge graph space in vector form, with each vector dimension corresponding to a specific semantic meaning.

[0070] Semantic association analysis is performed on the node features of the knowledge graph based on scene characteristics to construct connections between nodes. The semantic association analysis employs graph neural network technology, using a message passing mechanism to transmit information between nodes and uncover potential relationships. Specifically, for any two nodes, their semantic similarity in the feature space is calculated; if the similarity exceeds a preset threshold, a connection is established.

[0071] Taking e-commerce scenarios as an example, there is an "have" relationship between the product node "high-definition display" and the attribute node "4K resolution"; there is a "belong to" relationship between the product node "high-definition display" and the category node "electronic devices"; and there are relationships such as "purchase", "browse", and "favorite" between user nodes and product nodes. These relationships are automatically discovered and established through semantic association analysis.

[0072] After establishing node connections, it is necessary to calculate the association strength between nodes, using this strength as the connection weight. Association strength calculation is based on factors such as node feature similarity, interaction frequency, and temporal information. For node feature similarity, cosine similarity is used to calculate the cosine of the angle between the feature vectors of two nodes; the smaller the angle, the higher the similarity. For interaction frequency, the number of historical interactions between nodes is counted; the more frequent the interactions, the stronger the association. For temporal information, the temporal distribution of interactions is considered, with recent interactions having a higher weight than older interactions.

[0073] In practical applications, such as product recommendation systems, the association strength between user nodes and product nodes can be calculated based on factors such as the user's purchase history, browsing time, and review content. For example, suppose user A has recently browsed high-definition monitor-related products multiple times, spent a significant amount of time on a particular 27-inch 4K monitor, and viewed its detailed specifications and user reviews. In this case, the association strength between user A and that monitor node would be relatively high. Specifically, if the user browsed the monitor 5 times, spent an average of 3 minutes on each visit, and viewed its details 2 times, a comprehensive association strength value can be calculated as the connection weight between these two nodes.

[0074] The calculation method for association strength differs depending on the type of relationship. For example, for the "have" relationship, the association strength can be calculated based on the frequency and importance of the attribute in the description text; for the "belong to" relationship, the association strength can be calculated based on the typicality of the product and category; and for the "purchase" relationship, the association strength can be calculated based on purchase frequency, rating score, etc.

[0075] Once the knowledge graph is constructed, various intelligent applications can be used to leverage the graph structure. For example, implicit relationships between entities can be discovered through graph traversal algorithms; groups of entities with similar characteristics can be identified through community detection algorithms; and structural information in the knowledge graph can be compressed into low-dimensional vector representations through graph embedding techniques for downstream tasks such as recommendation and classification.

[0076] Taking a recommendation system as an example, based on a constructed knowledge graph, products related to a user's interests can be found through graph traversal. If user A shows interest in high-definition monitors, the system can follow the relational paths in the knowledge graph to find other products related to high-definition monitors, such as graphics cards, HDMI cables, or other display devices with similar characteristics. By calculating the cumulative value of the association strength along the path, the recommendation results can be ranked, with products with higher association strength ranked higher.

[0077] In practical applications, the knowledge graph can also be dynamically updated according to business needs. When new multimodal data is input, the above process is repeated to extract scene features, update node features and relationships, and adjust connection weights. This ensures that the knowledge graph reflects the latest data status and business changes.

[0078] Furthermore, to improve the quality of knowledge graphs, a feedback mechanism can be introduced to adjust parameters during the knowledge graph construction process based on the effectiveness of knowledge graph applications, such as recommendation accuracy and user satisfaction metrics. For example, if the prediction accuracy for a certain type of relationship is found to be low, the feature extraction and association strength calculation methods for that type of relationship can be adjusted.

[0079] Using the methods described above, knowledge graphs constructed from multimodal data can comprehensively capture the complex relationships between entities, providing powerful knowledge support for intelligent applications. This method is applicable to various scenarios such as e-commerce, content recommendation, and intelligent customer service, effectively improving the system's intelligence level and user experience.

[0080] In one optional implementation, the knowledge graph is expanded using the node features and the connection weights, and similar scenes in the knowledge graph are matched based on the scene features to identify historical decision information related to the current scene, including:

[0081] Introduce scene feature vectors into node feature vectors to calculate inter-node connection weights, and use the node feature vectors and inter-node connection weights to establish a knowledge antibody feature matrix containing multiple feature patterns.

[0082] Based on the knowledge antibody feature matrix, calculate the matching coefficient between the node feature vector and the feature pattern in the knowledge antibody feature matrix; calculate the expansion weight based on the node connection weight and the matching coefficient; dynamically expand the initial knowledge graph based on the expansion weight to obtain the expanded knowledge graph; extract the current scene features from the expanded knowledge graph based on the expansion weight, and combine the current scene features with the historical features and relationship features in the expanded knowledge graph to construct a scene feature vector;

[0083] The similarity between the scene feature vector and the historical scene feature vector is calculated using the extended weights. Similar historical scenes are identified from the extended knowledge graph based on the similarity, and corresponding decision features are extracted from the similar historical scenes based on the matching coefficient.

[0084] The decision features are fused with the scene feature vector to obtain the decision weight. The weighted function value of the decision weight and the matching coefficient is input into the knowledge antibody feature matrix for feature matching. Based on the feature matching result, historical decision information related to the current scene is obtained.

[0085] We acquire the node feature vectors and extracted scene feature vectors from the constructed knowledge graph. We then perform dimensionality normalization on both types of vectors to ensure consistent numerical ranges for subsequent calculations. We concatenate the scene feature vectors to the end of the node feature vectors, forming enhanced node feature vectors. Finally, we calculate the element-wise product of the node feature vectors and the scene feature vectors, capturing the interaction information between them. The result of this element-wise product is added to the enhanced node feature vectors to form a complete fused feature vector.

[0086] In intelligent healthcare scenarios, node feature vectors contain patient basic information, symptom characteristics, examination results, etc., with a dimension of 120; scenario feature vectors contain the medical environment, seasonal factors, disease prevalence trends, etc., with a dimension of 80. After normalizing the two types of vectors respectively, they are concatenated to obtain a 200-dimensional vector; the element-wise product of the two vectors is calculated to obtain the interaction features; finally, a 280-dimensional fused feature vector containing the original features and interaction features is formed, comprehensively describing the patient's comprehensive characteristics in a specific scenario.

[0087] When calculating the connection weights between nodes, for any two nodes i and j, their fused feature vectors are taken, and the cosine similarity between the vectors is calculated. Direct relationships between nodes are considered, such as the diagnostic relationship between a patient and a disease, and a correlation coefficient is assigned. The importance of nodes in the scene is considered, and the correlation between node features and scene features is calculated. The similarity, correlation coefficient, and correlation are weighted and summed to obtain the connection weights between nodes. In practical applications, the similarity weight can be set to 0.4, the correlation coefficient weight to 0.3, and the correlation weight to 0.3, achieving a comprehensive consideration of multiple factors.

[0088] A knowledge antibody feature matrix containing multiple feature patterns is established by using node feature vectors and inter-node connection weights. Historical scene data is collected, with each scene containing multiple nodes and their features and relationships. Cluster analysis is performed on historical scenes to identify typical scene patterns. For each scene, representative features are extracted to form feature patterns. All feature patterns are organized into a matrix form, with rows representing different patterns and columns representing feature dimensions. The matrix element values ​​represent the importance of a specific feature in a specific pattern.

[0089] In intelligent healthcare scenarios, by analyzing 10,000 historical cases, 20 typical disease patterns are identified. Each pattern is represented by a 300-dimensional feature vector, forming a 20×300 knowledge antibody feature matrix. For example, the first row of the matrix represents respiratory system disease patterns, with columns having higher values ​​corresponding to respiratory symptom features; the second row represents digestive system disease patterns, with columns having higher values ​​corresponding to abdominal symptom features.

[0090] The matching coefficient between the node feature vector and the feature pattern is calculated based on the knowledge antibody feature matrix, and the fused feature vector of the node is taken. The feature vector is then multiplied by each row of the knowledge antibody feature matrix to obtain the matching score. The matching score is normalized by applying the softmax function to obtain the matching coefficient between the node and each feature pattern.

[0091] In practical applications, assuming the dot product of the fused feature vector of a certain patient node and the 20 feature patterns of the knowledge antibody feature matrix is ​​[0.65, 0.12, 0.08, 0.05, 0.03, ...], after normalization, the matching coefficients are [0.73, 0.09, 0.06, 0.04, 0.02, ...], indicating that the patient feature is the best match for the first disease pattern.

[0092] The extended weight is calculated based on the connection weights and matching coefficients between nodes. For each edge in the knowledge graph, the matching coefficients of the two connected nodes are obtained. The weighted average of the matching coefficients of the two nodes is calculated, and the weights can be set according to the importance of the nodes. The extended weight of the edge is calculated by combining the connection weights of the edge and the weighted average of the matching coefficients of the nodes. For each node in the knowledge graph, the extended weights of the edges connected to it are aggregated to obtain the extended weight of the node.

[0093] In a practical implementation, the connection weight can be set to 60%, the matching coefficient to 40%, and the extended weight of the edge can be calculated. For example, for an edge connecting a patient node and a disease node, the connection weight is 0.85, and the weighted average of the matching coefficients of the two nodes is 0.7. Then the extended weight of the edge is calculated as 0.85×0.6+0.7×0.4=0.79.

[0094] The initial knowledge graph is dynamically expanded based on the expansion weight. An expansion threshold is set, and nodes and edges with expansion weights higher than the threshold are selected. Additional information related to the selected nodes is retrieved from an external knowledge base. New nodes are created to represent the additional information, and connections are established between the new nodes and the original nodes. The matching coefficient of the new nodes and the expansion weight of the new edges are calculated. The new nodes and edges that meet the conditions are added to the knowledge graph to form the expanded knowledge graph.

[0095] In the intelligent healthcare scenario, the initial knowledge graph contains 200 nodes and 500 edges. An expansion threshold of 0.7 is set, selecting nodes with an expansion weight greater than 0.7, including nodes representing specific patients, diseases, and symptoms. Relevant information for these nodes, such as treatment plans, medication information, and complications, is retrieved from the medical knowledge base. New nodes are created to represent this information and connected to the original nodes. After calculation and selection, 100 new nodes and 300 new edges are finally added to the knowledge graph, forming the expanded knowledge graph.

[0096] Based on the extended weights, the current scene features are extracted from the extended knowledge graph to determine the central nodes related to the current scene, such as the current patient node. A graph attention network is applied to aggregate the neighbor node information within a certain number of hops, starting from the central node. During the aggregation process, the weight of the neighbor node is determined by the extended weight of the edge. The weighted average of the node representations obtained by aggregation is calculated, with the weight being the extended weight of the node. The weighted average result is used as the current scene feature.

[0097] In practical applications, taking a patient node as the center, information from neighboring nodes within two hops is aggregated, including nodes related to symptoms, examination results, diseases, and medications. Assuming there are 30 neighboring nodes, each represented by a 150-dimensional vector, these are aggregated using a graph attention network to obtain the contextual representations of the 30 nodes. A weighted average is then performed based on the node expansion weights to obtain a 150-dimensional feature vector for the current scene.

[0098] The current scene features are combined with historical and relational features from the expanded knowledge graph to construct a scene feature vector. Historical scenes related to the current scene are retrieved from the knowledge graph. Feature representations of historical scenes are extracted. Relationship features between the current scene and historical scenes, such as time differences and environmental differences, are extracted. The current scene features, historical features, and relational features are concatenated to form a complete scene feature vector.

[0099] In the intelligent healthcare scenario, the current patient scenario features are 150-dimensional vectors, the relevant historical patient scenario features are also 150-dimensional vectors, and the relationship features are 50-dimensional vectors. After concatenation, a complete 350-dimensional scenario feature vector is obtained, which comprehensively describes the current patient scenario and its relationship with historical scenarios.

[0100] The similarity between the scene feature vector and the historical scene feature vector is calculated using extended weights to obtain the complete scene feature vector; all historical scene feature vectors are retrieved from the knowledge graph; the weighted cosine similarity between the scene feature vector and each historical scene feature vector is calculated, with the weights being the extended weights of the feature dimensions; the similarity is sorted to determine the set of most similar historical scenes.

[0101] In practical applications, suppose the knowledge graph stores 5000 historical medical scenarios, each represented by a 350-dimensional feature vector. The similarity between the current scenario and these 5000 historical scenarios is calculated, with results ranging from 0 to 1. After sorting, the 10 historical scenarios with the highest similarity are [0.95, 0.92, 0.88, 0.86, 0.83, 0.81, 0.79, 0.77, 0.75, 0.72], indicating a high degree of similarity between these historical scenarios and the current scenario.

[0102] Similar historical scenes are identified from the expanded knowledge graph based on similarity. A similarity threshold is set, and historical scenes with similarity higher than the threshold are filtered out. For the filtered historical scenes, their complete information in the knowledge graph is extracted, including node features, relationship features, and decision information. The similar historical scenes are sorted by similarity to form a set of similar scenes.

[0103] In the context of intelligent healthcare, a similarity threshold of 0.8 is set, and five historical scenarios with a similarity greater than 0.8 are selected from 5000 historical scenarios. Complete information about these five scenarios is extracted, including patient characteristics, disease characteristics, and treatment plans. These scenarios are then sorted by similarity to form a set of similar healthcare scenarios, providing a reference for subsequent decision-making.

[0104] Based on the matching coefficient, the corresponding decision features are extracted from similar historical scenarios. For each similar historical scenario, its decision features, such as treatment plan, drug selection, and follow-up plan, are extracted. The weight of the decision features is calculated by combining the similarity and matching coefficient of the similar historical scenarios. The decision features of each historical scenario are weighted and averaged according to the weight to obtain the comprehensive decision features.

[0105] In practical applications, the decision features of five similar historical scenarios are each 100-dimensional vectors, with similarity scores of [0.95, 0.92, 0.88, 0.86, 0.83] and matching coefficients of [0.85, 0.82, 0.79, 0.76, 0.72]. The decision weight for each scenario is calculated as the normalized product of the similarity score and the matching coefficient [0.25, 0.23, 0.21, 0.20, 0.11]. A weighted average of the five decision features based on their decision weights yields a 100-dimensional comprehensive decision feature.

[0106] The decision features and scene feature vectors are fused to obtain the decision weights. The scene feature vectors are then input into an attention network to generate attention weights. The decision features are weighted according to the attention weights to highlight the important decision factors in the current scene. The weighted decision features are then normalized to obtain the decision weight vector, which represents the priority of each decision option.

[0107] In intelligent healthcare scenarios, a 350-dimensional scene feature vector is input into an attention network to generate a 100-dimensional attention weight vector, representing the degree of attention given to each dimension of the decision features. The attention weights are then element-wise multiplied with the 100-dimensional decision features to obtain weighted decision features. After normalization, the resulting decision weight vector indicates the priority of each treatment plan.

[0108] The weighted function value of the decision weight and the matching coefficient is input into the knowledge antibody feature matrix for feature matching. The weighted function value of the decision weight and the matching coefficient can be calculated using a linear combination, and the weights are set according to the specific application. The weighted function value is used as a query vector, and the similarity is calculated with each row of the knowledge antibody feature matrix. The most matching feature pattern is identified based on the similarity. The corresponding historical decision information is extracted from the feature pattern.

[0109] In practical applications, the decision weight is set to 70%, and the matching coefficient to 30%, and a weighted function value is calculated. This value is then compared with the 20 feature patterns of the knowledge antibody feature matrix to obtain a similarity array. Assuming the highest similarity is 0.91, corresponding to the 3rd feature pattern, it indicates that the historical decision information under this pattern is most suitable for the current scenario.

[0110] Based on the feature matching results, historical decision information related to the current scenario is obtained. From the most matching feature pattern, the corresponding historical decision records are retrieved. Information such as decision content, decision reasons, execution methods, and effect evaluation are extracted. The historical decision information is adaptively adjusted according to the characteristics of the current scenario. A decision suggestion report is generated, which includes recommended decisions, explanations of reasons, and expected effects.

[0111] In a smart healthcare scenario, 10 historical decision records are retrieved from the most matching feature patterns. Each record includes the diagnosis, treatment plan, medication recommendations, and prognosis. Core information from these records is extracted, such as drug A dosage of 10mg three times daily; treatment duration of 7 days; and physiotherapy twice weekly. The drug dosage and treatment frequency are adjusted based on the patient's current specific circumstances (e.g., age, weight, liver and kidney function). Finally, a decision recommendation report is generated, recommending the most suitable personalized treatment plan for the current patient.

[0112] In a specific application example in the field of intelligent manufacturing, the initial knowledge graph of a factory's equipment management system contains 300 nodes (equipment, components, fault types, etc.) and 800 edges (compositional relationships, fault associations, etc.). When abnormal vibration of a critical piece of equipment is detected, the system executes the following process: Introducing current scene features (vibration frequency, equipment load, ambient temperature, etc.) to calculate the connection weights between nodes; constructing a knowledge antibody feature matrix containing 30 fault modes; calculating the matching coefficient between equipment nodes and feature modes, finding a high matching degree of 0.87 with the bearing wear mode; expanding the knowledge graph to 500 nodes and 1200 edges based on the expanded weights, adding historical maintenance records, spare parts information, etc.; extracting current scene features and comparing them with historical scenes, identifying 5 historical cases with a similarity exceeding 0.85; extracting decision features from these cases, fusing them to obtain decision weights; and finally matching the optimal maintenance solution: replacing a specific model of bearing while inspecting the lubrication system, with an estimated maintenance time of 4 hours, avoiding a major failure after 72 hours. This solution is highly consistent with historical successful cases, and the expected results are reliable.

[0113] In one optional implementation, the scene features and the historical decision information are jointly input into the reasoning module of the knowledge graph to generate a set of decision schemes, and the execution effect score of the set of decision schemes is calculated, including:

[0114] Scene features and historical decisions are used as inputs to the reasoning module of the knowledge graph, and a set of nonlinear state equations is constructed in the reasoning module of the knowledge graph. The system evolution trajectory is obtained by solving the set of nonlinear state equations, and the dynamic evolution features of the knowledge graph are extracted based on the state sequence in the system evolution trajectory.

[0115] The reasoning module of the knowledge graph is used to reason about the scene features and the historical decision information based on the temporal change pattern of the dynamic evolution features, generating multiple candidate decision schemes and forming a decision scheme set.

[0116] Calculate the gradient vector of the system evolution trajectory, identify multiple control points from the gradient vector, use the control points as the benchmark points for scoring, and calculate the corresponding evaluation value based on the state information of each candidate decision scheme in the decision scheme set at the control point.

[0117] Based on the corresponding evaluation value, the corresponding neuron activation intensity is calculated, and a plasticity connection matrix representing the dynamic correlation strength between the control points is constructed. The neuron activation intensity and the plasticity connection matrix are iteratively calculated to obtain a stable network resonance state. Based on the network resonance state, the execution effect score of the candidate decision scheme is calculated.

[0118] like Figure 2As shown, the method includes:

[0119] Using scene features and historical decisions as inputs to the knowledge graph's reasoning module, a nonlinear state equation system is constructed within the module. Scene features are extracted as an n-dimensional vector F, with each component representing a specific scene feature value. Historical decision information is encoded as an m-dimensional vector D, with each component representing a specific decision parameter value. The two vectors are concatenated to form the complete input vector I. In the knowledge graph's reasoning module, a state variable vector S is defined, representing the state value of key nodes. A nonlinear state equation system is constructed to describe the relationship between the rate of change of the state variable over time and the current state, the input vector, and time.

[0120] In intelligent transportation scenarios, the scenario feature vector contains 20 components, such as current traffic flow (250 vehicles / hour), average vehicle speed (45 km / h), weather condition coding (sunny day = 1, current value = 1), and time period coding (evening peak = 3, current value = 3), etc.; the historical decision vector contains 15 components, such as traffic light timing scheme parameters (red light duration = 60 seconds, green light duration = 40 seconds), lane allocation strategy (number of straight lanes = 2, number of left-turn lanes = 1), etc. The state variable vector contains 25 components, representing the traffic status of key nodes in the road network, such as intersection queue length (35 meters), traffic efficiency (0.75), and safety risk index (0.15), etc.

[0121] The system evolution trajectory is obtained by solving a set of nonlinear state equations. Based on the state sequence in the system evolution trajectory, dynamic evolution features of the knowledge graph are extracted. An initial state is set, corresponding to the currently observed system state. An appropriate numerical integration method, such as the fourth-order Runge-Kutta method, is selected to solve the set of nonlinear state equations. The integration step size and the total simulation time are set, and the state values ​​at each time point are calculated to obtain the state sequence. The state sequence is organized into a matrix form. The first-order and second-order differences of the state variables are calculated. The statistical and temporal features of the state sequence are extracted. The above features are combined to form a dynamic evolution feature vector.

[0122] In the intelligent transportation scenario, the initial state is set to the current measured traffic state. A step size h = 10 seconds is chosen, and the total simulation duration T = 3600 seconds (1 hour). The fourth-order Runge-Kutta method is used to solve the state equations, resulting in a state sequence matrix containing 360 rows (time points) and 25 columns (state variables). The calculated queue length has a mean of 28 meters, a standard deviation of 12 meters, a maximum value of 52 meters, and a minimum value of 10 meters. The calculated traffic efficiency has a mean of 0.82, a standard deviation of 0.08, a maximum value of 0.95, and a minimum value of 0.65. These features together form a dynamic evolution feature vector with a dimension of 250.

[0123] This paper utilizes the reasoning module of a knowledge graph to infer scene features and historical decision information based on the temporal changes of dynamic evolutionary characteristics, generating multiple candidate decision schemes and forming a decision scheme set. A target state is defined to represent the desired ideal state. A state prediction model is constructed based on dynamic evolutionary feature vectors. A decision parameter space is defined, containing all possible combinations of decision parameters. A series of candidate decision parameter combinations are generated in the decision parameter space. For each candidate decision parameter combination, the final state of the system under that decision is predicted. The deviation between the final state and the target state is calculated. The top N candidate decision parameter combinations with the smallest deviations are selected to form a decision scheme set.

[0124] In the intelligent transportation scenario, the target state is set as follows: queue length not exceeding 20 meters, traffic efficiency not less than 0.9, and safety risk index not exceeding 0.1. A predictive model is constructed based on extracted dynamic evolution features. The decision parameter space includes all combinations of traffic light timing schemes and lane allocation strategies. By searching the parameter space, 200 candidate decision parameter combinations are generated, and the 8 decision schemes with the smallest deviations are selected to form a decision scheme set. For example, Scheme 1 has a 65-second red light, a 45-second green light, 2 straight lanes, and 2 left-turn lanes; Scheme 2 has a 55-second red light, a 55-second green light, 3 straight lanes, and 1 left-turn lane.

[0125] The gradient vector of the system's evolution trajectory is calculated, and multiple control points are identified from the gradient vector. These control points are used as the benchmark points for scoring calculation. For each state variable in the system's evolution trajectory, its derivative (gradient) with respect to time is calculated to form a gradient vector. The norm of the gradient vector is calculated. Local maxima and local minima of the gradient vector norm are found. The moments when the signs of the gradient vector components change are found. Based on the above feature points, M key moments are selected as control points.

[0126] In an intelligent transportation scenario, the gradient vector of the 1-hour evolution trajectory is calculated, identifying five local maxima of the norm, located at minutes 28, 35, 42, 49, and 56. Three local minima and the moments of change in the gradient sign of key state variables are also identified. Considering these features, six control points are selected, located at minutes 15, 28, 35, 42, 49, and 56.

[0127] The evaluation value is calculated based on the state information of each candidate decision scheme at the control point in the decision scheme set. For each decision scheme in the decision scheme set, the state of the system at the control point is calculated. Multiple evaluation indicators are defined, such as state deviation, resource consumption, and operational complexity. Weights are assigned to each evaluation indicator. The comprehensive evaluation value at each control point is calculated.

[0128] In intelligent transportation scenarios, evaluation indicators include: the deviation of queue length from the target value (20 meters) (weight 0.3), the deviation of traffic efficiency from the target value (0.9) (weight 0.3), the deviation of safety risk index from the target value (0.1) (weight 0.2), and the frequency of traffic light adjustments (weight 0.1) and lane change frequency (weight 0.1). Taking Scheme 1 as an example, at the 15-minute control point, the comprehensive evaluation value is calculated to be 0.563. Similarly, the evaluation values ​​of Scheme 1 at other control points are calculated, resulting in the evaluation value vector [0.563, 0.482, 0.415, 0.387, 0.402, 0.425].

[0129] The corresponding neuron activation intensity is calculated based on the corresponding evaluation value. A nonlinear activation function, such as the Sigmoid function, is defined. The evaluation value is normalized. The corresponding neuron activation intensity is calculated. The neuron activation intensity vector of the decision scheme at all control points is obtained.

[0130] In the intelligent transportation scenario, the Sigmoid function is chosen as the activation function, with parameters α=5 and β=0.5. For Scheme 1, the evaluation value at the 15-minute control point is 0.563, and the calculated neuron activation intensity is 0.73. Similarly, the neuron activation intensity of Scheme 1 at other control points is calculated, resulting in the activation intensity vector [0.73, 0.79, 0.83, 0.85, 0.84, 0.82]. The activation intensity vector for Scheme 3 is [0.82, 0.85, 0.88, 0.90, 0.89, 0.87].

[0131] Construct a plastic connection matrix to characterize the dynamic correlation strength between control points. Create an M×M matrix to represent the connection strength between M control points. For any two control points, calculate the time interval between them. Calculate the correlation coefficient of the state vector at the control point. Considering the decay effect of the time interval, define the formula for calculating the connection strength. Normalize the matrix.

[0132] In an intelligent transportation scenario, six control points form a 6×6 plasticity connectivity matrix. The Pearson correlation coefficient between the state vectors of two control points is calculated; for example, the correlation coefficient between the first and second control points is 0.85. Assuming an attenuation coefficient γ = 0.02, the connectivity strength W12 = 0.85 × exp(-0.02 × 13) = 0.65. The connectivity strength of all control point pairs is calculated, resulting in the complete plasticity connectivity matrix. The first row is [0, 0.65, 0.48, 0.32, 0.25, 0.18]. The matrix is ​​then normalized to ensure that the sum of each row is 1.

[0133] A stable network resonance state is obtained by iteratively calculating the neuron activation intensity and the plasticity connection matrix. The initial state is set as the neuron activation intensity vector of the decision scheme. An iterative update rule is defined. An iterative termination condition is set. Iterative calculation is performed until the termination condition is reached to obtain the final activation intensity vector.

[0134] In the intelligent transportation scenario, taking Scheme 1 as an example, the initial activation intensity vector is [0.73, 0.79, 0.83, 0.85, 0.84, 0.82]. Let the learning rate λ = 0.3 and the iteration threshold ε = 0.001. After 12 iterations, the activation intensity vector converges to [0.95, 0.96, 0.96, 0.95, 0.94, 0.93], reaching a stable network resonance state. The state of Scheme 3 after iteration is [0.98, 0.99, 0.99, 0.98, 0.97, 0.96].

[0135] The execution effect scores of candidate decision schemes are calculated based on the network resonance state, and importance weights are assigned to each control point. The execution effect score of the decision scheme is calculated as the weighted sum of the activation intensity of each control point in the resonance state. The scores of all candidate decision schemes are ranked.

[0136] In the intelligent transportation scenario, the importance weights of the six control points are set to [0.1, 0.15, 0.2, 0.25, 0.2, 0.1]. For Scheme 1, its network resonance state is [0.95, 0.96, 0.96, 0.95, 0.94, 0.93], and its execution performance score is 0.950. The scores of other schemes are: Scheme 2 is 0.925, Scheme 3 is 0.978, Scheme 4 is 0.912, Scheme 5 is 0.936, Scheme 6 is 0.945, Scheme 7 is 0.958, and Scheme 8 is 0.970. Scheme 3 has the best execution performance, with a score of 0.978.

[0137] In smart manufacturing scenarios, this method is applied to optimize production line scheduling, selecting 10 options from 200 candidate decision-making schemes for final evaluation. By calculating neuron activation intensity, constructing a plasticity connection matrix, and iterating to the network resonance state, the execution effectiveness score of each scheme is obtained. The highest-scoring scheme is 0.985, and its core strategy is: first, conduct a small-batch trial production for 2 hours, adjusting parameters based on quality feedback; in the middle stage, maintain maximum capacity for 4 hours; and in the final 2 hours, gradually reduce capacity while strengthening quality inspection. This scheme is expected to improve production efficiency by 15%, reduce the defect rate by 8%, and reduce overall costs by 12%.

[0138] In intelligent logistics scenarios, this method was applied to optimize delivery routes. Through network resonance calculation, the solution with the highest score of 0.967 was obtained. Its core strategies are: adopting a regional segmented delivery strategy, using dynamic path planning within each region; avoiding congested road sections during peak hours; and using time window clustering for delivery points. This solution is expected to reduce delivery time by 18%, reduce fuel consumption by 15%, and increase customer satisfaction by 12%.

[0139] The above method enables a complete technical solution that integrates scene features and historical decision information into a knowledge graph reasoning module, generating a set of decision schemes and calculating execution performance scores. This method organically combines nonlinear system theory with neural network dynamics, establishing an end-to-end solution from scene features to decision scoring, providing a scientific basis and technical support for intelligent decision-making.

[0140] In one optional implementation, the reasoning module of the knowledge graph infers the scene features and the historical decision information based on the temporal change patterns of the dynamic evolution features, generating multiple candidate decision schemes and forming a decision scheme set, including:

[0141] Scene features and historical decision information are used to construct feature vectors that represent temporal characteristics. Temporal evolution analysis is performed on the feature vectors to obtain their evolutionary rules, and dynamic change features are extracted from the evolutionary rules.

[0142] The feature vector is input into the reasoning module of the knowledge graph and transformed into a feature space to obtain a transformed feature expression. The stable region of the transformed feature expression is determined according to the evolution law of the feature vector. The stable region is used to constrain the transformed feature expression to obtain multiple feature parameters characterizing the stability of the system.

[0143] The reasoning module of the knowledge graph constructs a feature weight distribution based on the feature parameters, and calculates the changing trend of the transformed feature expression based on the feature weight distribution;

[0144] The dynamic change features are input into the reasoning module of the knowledge graph, and the transformed feature expression is reasoned based on the change trend to obtain the reasoning result. Under the constraint of the stable region, multiple candidate decision schemes are generated according to the reasoning result.

[0145] For the candidate decision schemes, an evaluation index including scheme reliability, scheme diversity and information gain is calculated. Based on the evolution law of the feature vector, a dynamically changing screening threshold is determined. The evaluation index is compared with the screening threshold, and candidate decision schemes that meet the screening threshold are selected to form a decision scheme set.

[0146] like Figure 3As shown, the method includes:

[0147] The scenario features and historical decision-making information are used to construct feature vectors representing temporal characteristics. Static and dynamic attributes are extracted from the scenario features. Static attributes include environmental parameters, resource status, and constraints, while dynamic attributes include time series data, rate of state change, and evolutionary trends. Key data such as decision parameters, execution time series, and effect evaluation are extracted from the historical decision-making information. Static attributes are normalized to ensure comparability of features with different dimensions. Dynamic attributes are temporally sampled to retain data from key time points. The processed static and dynamic attributes are then organized into feature vectors according to a predefined structure.

[0148] In the smart grid load dispatch scenario, scenario features include static attributes such as grid topology, substation capacity, and line load, as well as dynamic attributes such as load curves, voltage fluctuations, and power factor. Historical decision information includes generation dispatch strategies, load transfer operations, and emergency response measures over the past 24 hours. Static attributes such as substation capacity (original value 500 MVA) are normalized to 0.75, and line load factor (original value 65%) is normalized to 0.65. Dynamic attributes such as load curves are sampled at 30-minute intervals over the past 24 hours, resulting in data at 48 time points. These processed features are organized into a 256-dimensional feature vector, where the first 64 dimensions represent static attributes and the last 192 dimensions represent dynamic attributes.

[0149] Evolutionary analysis of eigenvectors over time yields their evolutionary patterns. Dynamic characteristics are extracted from these patterns, and eigenvectors from multiple time points are collected to form a time-series dataset. The difference between eigenvectors at adjacent time points is calculated to obtain the rate of change. A sliding window averaging of the rate of change is applied to reduce noise. Abrupt change points in the rate of change are identified, typically corresponding to transitions in the system state. The autocorrelation coefficients of each dimension of the eigenvectors are calculated to identify periodic change patterns. An evolutionary pattern model is constructed based on the rate of change, abrupt change points, periodicity, and frequency characteristics. Key features characterizing the system's dynamic behavior are extracted from this model to form dynamic change characteristics.

[0150] In a smart grid load dispatch scenario, feature vectors from the past 7 days were collected, with 48 time points per day, totaling 336 samples. The differences in feature vectors between adjacent time points were calculated, and the rate of change was averaged using a sliding window of length 5. Significant abrupt changes in grid load were identified around 07:30, 12:00, and 18:30 each day, corresponding to the morning peak, midday peak, and evening peak. The autocorrelation coefficients of each dimension of the feature vectors were calculated, revealing a clear 24-hour periodicity in the load curve, with the correlation coefficient reaching 0.92 at the 24-hour delay. Based on these analyses, a load evolution model was constructed, and eight key dynamic change features were extracted, including the daily load growth rate (maximum 0.08 / hour), peak duration (average 3.5 hours), and peak-to-valley ratio (average 1.8).

[0151] The feature vectors are input into the reasoning module of the knowledge graph and transformed into a feature space to obtain the transformed feature representation. A feature mapping layer of the knowledge graph is constructed, containing multiple nonlinear transformation units. The feature vectors are input into the feature mapping layer and a high-dimensional feature representation is obtained through nonlinear transformation. Dimensionality reduction techniques are applied to project the high-dimensional feature representation onto a low-dimensional semantic space. In the semantic space, similar feature vectors are mapped to nearby positions to form a cluster structure. Topological relationships and distance information in the semantic space are extracted to construct the transformed feature representation.

[0152] In the smart grid load dispatching scenario, the feature mapping layer of the knowledge graph contains three nonlinear transformation units with 512, 256, and 128 nodes, respectively. The 256-dimensional feature vectors are input into the feature mapping layer and undergo nonlinear transformation to obtain a 128-dimensional high-dimensional feature representation. The t-SNE dimensionality reduction technique is applied to project the 128-dimensional feature representation onto a 3-dimensional semantic space. In the semantic space, the feature vectors for weekdays form one cluster with coordinate centers approximately (0.4, 0.6, 0.5); the feature vectors for weekends form another cluster with coordinate centers approximately (-0.3, -0.4, 0.2); and the feature vectors under special weather conditions form a third cluster with coordinate centers approximately (0.1, -0.2, -0.6). The topological relationships and internal structures of these clusters are extracted to construct a 64-dimensional transformed feature representation.

[0153] Based on the evolution law of eigenvectors, the stable region of the transformed feature expression is determined. This stable region is then used to constrain the transformed feature expression, yielding multiple feature parameters characterizing system stability. The range and pattern of change of the eigenvectors during the time-series evolution are analyzed. In the transformed feature space, the distribution area of ​​historical data is marked. Regions where the system operates stably within the historical data are identified and defined as stable regions. The geometric characteristics of the stable regions, such as center point, boundary, and shape parameters, are calculated. Based on the geometric characteristics of the stable regions, constraints are constructed to limit the reasonable range of the feature expression. The transformed feature expression is projected onto the constraints to obtain a feature expression that meets the stability requirements. Key parameters are extracted from the constrained feature expression as feature parameters characterizing system stability.

[0154] In a smart grid load dispatch scenario, the variation range of feature vectors over the past 7 days was analyzed, with load fluctuations ranging from 45% to 75% of total capacity. In the transformed 3D semantic space, the distribution of 336 historical data points was marked. A region of stable system operation was identified, approximating an ellipsoid with its center point coordinates (0.2, 0.1, 0.0) and semi-axis lengths (0.5, 0.4, 0.3). Constraints were constructed based on this ellipsoid, requiring new feature expressions to fall within the ellipsoid or be no more than 0.1 units from the surface. Projecting the current feature expression (0.3, 0.2, -0.1) onto the constraints yielded a feature expression (0.3, 0.2, -0.1) that met the stability requirements, indicating that this point was within the stable region. Five key parameters were extracted from the constrained feature representation: distance to the center point (0.14), coordinates projected onto the principal axis (0.25, 0.15, -0.10), and stability margin (0.36). These parameters characterize the stability state of the system.

[0155] The reasoning module of the knowledge graph constructs a feature weight distribution based on feature parameters. The changing trend of the transformed feature expression is calculated according to the feature weight distribution. The feature parameters are then input into the weight generation network of the knowledge graph. In the weight generation network, importance scores for each dimension are calculated based on the values ​​of the feature parameters. The importance scores are normalized to obtain the feature weight distribution. The transformed feature expression is weighted according to the feature weight distribution. The gradient direction of the weighted feature expression is calculated, representing the most likely direction of change in the system. Combining the observed change patterns in historical data, the position of the feature expression at future time points is predicted. The changing trend, including the direction, rate, and acceleration of change, is extracted from the predicted position sequence.

[0156] In a smart grid load dispatching scenario, five feature parameters are input into a weighted generation network, which consists of two layers with 32 and 64 nodes respectively. Based on the values ​​of the feature parameters, importance scores for each dimension of the 64-dimensional feature representation are calculated. Softmax normalization is applied to the scores to obtain the feature weight distribution, with weight values ​​ranging from 0.005 to 0.042. Based on this weight distribution, the feature representation is weighted to highlight the influence of important dimensions. The gradient direction of the weighted feature representation in the 3D semantic space is calculated, resulting in a vector (0.05, 0.03, -0.02), representing the most likely direction of change in the system. Combining this with historical data patterns, the position sequence of the feature representation over the next 6 hours is predicted. The trend of change is extracted from these positions: the direction remains relatively stable, while the rate gradually increases, from an initial 0.017 / hour to a final 0.057 / hour, indicating that the system load is accelerating.

[0157] The dynamic change features are input into the reasoning module of the knowledge graph. Reasoning is performed on the transformed feature representation based on the change trend to obtain the reasoning result. Under the constraint of the stable region, multiple candidate decision schemes are generated based on the reasoning result. The extracted dynamic change features are combined with the transformed feature representation to form complete input information. In the reasoning module of the knowledge graph, knowledge nodes related to the current scene are activated. Based on the activated knowledge nodes and input information, reasoning rules are applied for forward derivation. Considering the system's change trend, the possible system state changes caused by different decisions are predicted. The predicted state changes are evaluated, and state transition paths that meet the objectives are selected. Based on each valid state transition path, a corresponding decision scheme is generated. It is checked whether the decision scheme will cause the system to leave the stable region, and schemes that do not meet the stability constraints are eliminated. Decision schemes that meet all constraints are retained to form a candidate decision scheme set.

[0158] In the smart grid load dispatching scenario, eight dynamic change features are combined with 64-dimensional feature representations and input into the reasoning module of the knowledge graph. 150 knowledge nodes related to power load dispatching, energy distribution, and stability control are activated in the knowledge graph. Based on these knowledge nodes and the input information, expert rules for power system dispatching are applied for reasoning. Considering the accelerating trend of load growth, the system's state changes over the next three hours are predicted under different dispatching strategies. Fifteen possible dispatching strategies are generated, such as increasing the output of conventional generator units, starting standby power supplies, and reducing non-critical loads. The effectiveness of each strategy is evaluated, and ten effective state transition paths are selected, corresponding to ten candidate decision schemes. These schemes are checked to see if they would cause the system to leave the stable region. Two schemes were found to potentially cause line overload under high load conditions, pushing the system out of the stable region; therefore, these two schemes were eliminated. Finally, eight candidate decision schemes that satisfy all constraints are retained, such as Scheme 1 "Increase the output of generator unit No. 3 by 30MW while reducing non-critical loads in area B by 15MW", and Scheme 2 "Start the standby gas turbine to provide 40MW of emergency capacity", etc.

[0159] Evaluation metrics, including reliability, diversity, and information gain, are calculated for candidate decision-making schemes. A dynamically changing screening threshold is determined based on the evolution of eigenvectors. The evaluation metrics are compared with the screening threshold, and candidate decision-making schemes that meet the threshold are selected to form a decision-making scheme set. For each candidate decision-making scheme, its reliability is evaluated based on historical data and expert knowledge, and a reliability score is calculated. The similarity between decision-making schemes is analyzed, and a diversity index is calculated to ensure sufficient coverage of the scheme set. The information gain of each decision-making scheme is calculated to evaluate its contribution to the current decision problem. A comprehensive evaluation metric is calculated by combining reliability, diversity, and information gain. A dynamically adjusted screening threshold is set according to the characteristics of state changes, allowing the threshold to adapt to the system state. The evaluation metrics of each decision-making scheme are compared with the current screening threshold. Decision-making schemes whose evaluation metrics exceed the screening threshold are selected to form the final decision-making scheme set.

[0160] In a smart grid load dispatch scenario, eight candidate decision-making schemes are evaluated. Based on historical data and power system expert knowledge, the reliability score of each scheme is assessed; for example, Scheme 1 scores 0.85, and Scheme 2 scores 0.78. The cosine similarity matrix between the schemes is calculated. Scheme 1 and Scheme 3 have a similarity of 0.72, indicating they are relatively similar; Scheme 2 and Scheme 4 have a similarity of 0.35, indicating they are significantly different. The information gain of each scheme is calculated: Scheme 1 has a gain of 0.42, and Scheme 2 has a gain of 0.56. Combining these three indicators, the comprehensive evaluation index for Scheme 1 is 0.67, and for Scheme 2 it is 0.73. Based on the system state, the reliability threshold is set to 0.75, the diversity threshold to 0.4, the information gain threshold to 0.45, and the comprehensive evaluation threshold to 0.65. By comparing the evaluation indicators and thresholds of each scheme, Scheme 1, Scheme 2 and Scheme 7 were finally selected to form the decision scheme set, which are "increase the output of generator unit No. 3 by 30MW and reduce the non-critical load in area B by 15MW", "start the standby gas turbine to provide 40MW of emergency capacity" and "adjust the grid topology to transfer the load from the high load area to the low load area".

[0161] The above method achieves the technical objective of using a knowledge graph's reasoning module to infer scenario features and historical decision information based on the temporal changes of dynamic evolutionary characteristics, generating multiple candidate decision schemes and forming a decision scheme set. This method has broad application prospects in complex system decision-making scenarios such as smart grid load scheduling, and can effectively improve the scientific nature and adaptability of decision-making.

[0162] In one optional implementation, the optimal decision scheme is selected based on the execution performance score, and the execution result of the optimal decision scheme is used as the reward signal for the reinforcement learning algorithm. Dynamically adjusting the inference module includes:

[0163] A decision ranking value is constructed based on the results of the execution effect score of the decision scheme. A phase synchronization matrix is ​​constructed by calculating the phase difference of multiple neurons based on the decision ranking value. The phase synchronization matrix is ​​used to perform co-mapping on the features of different oscillation modes to obtain the neural oscillation mode prediction value corresponding to the decision scheme. The decision ranking value is co-mapped with the neural oscillation mode prediction value through the phase synchronization matrix to obtain the comprehensive decision evaluation value. The optimal decision scheme is selected based on the comprehensive decision evaluation value.

[0164] The actual execution result of the optimal decision scheme is compared with the comprehensive evaluation value of the decision to construct an instant reward value. The neural oscillation synchronicity change value is calculated based on the phase synchronization matrix. The instant reward value is co-mapped with the neural oscillation synchronicity change value through the phase synchronization matrix to obtain a compensation reward value. The compensation reward value is accumulated over time to obtain a cumulative reward value. The cumulative reward value is input into the reinforcement learning algorithm as a reward signal.

[0165] The gradient values ​​of the inference module parameters are calculated based on the reinforcement learning algorithm. The constraint difference between the inference module parameters and the target neural oscillation mode is calculated based on the phase synchronization matrix. The inference module is dynamically adjusted by co-mapping the gradient values ​​and the historical update values ​​of the inference module parameters.

[0166] A decision ranking value is constructed based on the performance evaluation scores of the decision schemes. The performance evaluation scores of all candidate decision schemes are collected and ranked from highest to lowest. The ranked scores are then normalized so that the highest score corresponds to a normalized value of 1, and the lowest score corresponds to a normalized value of 0. A ranking weight is assigned to each decision scheme, with higher-ranked schemes receiving higher weights. The normalized score is multiplied by the ranking weight to obtain a weighted score. This weighted score is then converted into a decision ranking value, ensuring that the ranking value reflects the relative merits of the decision schemes.

[0167] In an intelligent medical diagnosis scenario, suppose there are five candidate decision-making schemes with performance scores of 0.92, 0.87, 0.83, 0.76, and 0.65, respectively. Normalizing these scores yields normalized values ​​of 1.00, 0.81, 0.67, 0.41, and 0.00. Assigning ranking weights of [0.40, 0.30, 0.15, 0.10, 0.05] results in weighted scores of 0.40, 0.24, 0.10, 0.04, and 0.00. Converting these to decision ranking values ​​yields [1.00, 0.60, 0.25, 0.10, 0.00], clearly reflecting the relative merits of the five decision-making schemes.

[0168] A phase synchronization matrix is ​​constructed by calculating the phase difference of multiple neurons based on the decision ranking value. A set of neural oscillators is defined, with each oscillator corresponding to a key variable in the decision-making process. The phase value of each oscillator is initialized, and the initial phase can be set based on historical data or domain knowledge. The phase change rate of each oscillator is adjusted according to the decision ranking value, with the key variable corresponding to the decision scheme with the higher ranking value receiving a larger phase change. The phase difference between any two oscillators is calculated to form a phase difference matrix. The phase difference is transformed by applying a phase synchronization function to obtain the phase synchronization strength. The phase synchronization strength is organized into a matrix form to form the phase synchronization matrix.

[0169] In the intelligent medical diagnostic scenario, eight neural oscillators are defined, corresponding to eight key variables: patient symptoms, vital signs, test results, imaging features, disease probability, treatment plan, drug selection, and prognosis prediction. The initial phase values ​​of each oscillator are [0.2, 0.5, 0.3, 0.6, 0.4, 0.7, 0.1, 0.8]. The phase change rate is adjusted based on the sorting value [1.00, 0.60, 0.25, 0.10, 0.00], and the new phase values ​​of the eight oscillators are calculated as [0.8, 0.7, 0.9, 0.6, 0.5, 0.3, 0.4, 0.2]. The phase difference between any two oscillators is calculated; for example, the phase difference between the first and second oscillators is 0.1. The synchronization strength is calculated using a phase synchronization function; for example, the synchronization strength between the first and second oscillators is 0.9. Construct an 8×8 phase synchronization matrix, with the first row being [1.0, 0.9, 0.7, 0.6, 0.5, 0.3, 0.4, 0.2].

[0170] This paper utilizes a phase synchronization matrix to collaboratively map the features of different oscillation modes to obtain the predicted values ​​of neural oscillation modes corresponding to the decision scheme, and collects the neural oscillation mode features observed in the historical decision-making process. The features of the current decision problem are input into a pattern recognition network to predict possible oscillation modes. The similarity between the predicted mode and the historical mode is calculated, and the best matching historical mode is identified. The feature parameters of the best matching historical mode are extracted. The historical mode features are combined with the current phase synchronization matrix, and feature mapping is performed through matrix multiplication or convolution operations. Key features are extracted from the mapping results to form the predicted values ​​of neural oscillation modes.

[0171] In intelligent medical diagnosis scenarios, 100 typical neural oscillation patterns were collected from a historical database, each described by 20 feature parameters. The features of the current diagnostic problem were input into a pattern recognition network, predicting three possible oscillation patterns with similarities of 0.88, 0.72, and 0.65 to historical patterns, respectively. Feature parameters of the historical pattern with the highest similarity were extracted, including alpha wave frequency (10Hz), beta wave frequency (20Hz), alpha wave amplitude (15μV), beta wave amplitude (8μV), and alpha-beta phase coupling strength (0.6). These feature parameters were combined with the current 8×8 phase synchronization matrix and mapped using matrix multiplication to obtain a 64-dimensional feature vector. Key components are extracted from this vector to form a 12-dimensional neural oscillation pattern prediction value [0.82, 0.75, 0.68, 0.63, 0.57, 0.52, 0.48, 0.43, 0.39, 0.35, 0.32, 0.28].

[0172] The decision ranking value is obtained by co-mapping the decision ranking value with the neural oscillation mode prediction value through a phase synchronization matrix, and the decision ranking value is represented as a vector. The decision ranking value is transformed by the phase synchronization matrix to obtain the ranking mapping value. The dot product of the ranking mapping value and the neural oscillation mode prediction value is calculated to obtain the preliminary evaluation score. A nonlinear activation function is applied to the preliminary evaluation score to adjust the score distribution. According to the characteristics of the decision domain, weight coefficients are set to balance the influence of ranking factors and oscillation mode factors. The weighted evaluation score is normalized to obtain the final decision comprehensive evaluation value. All candidate decision schemes are ranked according to the decision comprehensive evaluation value, and the scheme with the highest evaluation value is selected as the optimal decision scheme.

[0173] In the intelligent medical diagnosis scenario, the decision ranking value [1.00, 0.60, 0.25, 0.10, 0.00] is represented as a 5-dimensional vector. The ranking value is transformed using an 8×8 phase synchronization matrix to obtain an 8-dimensional ranking mapping value [0.86, 0.72, 0.65, 0.59, 0.53, 0.47, 0.42, 0.38]. The generalized dot product of this mapping value and the 12-dimensional neural oscillation pattern prediction value is calculated, yielding a preliminary evaluation score of 3.25. The score is adjusted using the Sigmoid activation function, resulting in 0.96. With the ranking factor weight set to 0.6 and the oscillation pattern factor weight set to 0.4, the weighted evaluation score is calculated as 0.6×1.0 + 0.4×0.96 = 0.984. The above calculations were repeated for all candidate schemes, resulting in comprehensive decision evaluation values ​​of [0.984, 0.731, 0.412, 0.286, 0.153] for the five schemes. Scheme 1 had the highest evaluation value and was selected as the optimal decision scheme.

[0174] The process involves comparing the actual execution result of the optimal decision-making scheme with the comprehensive decision evaluation value to construct an immediate reward value, executing the optimal decision-making scheme, and obtaining the actual execution result. A set of evaluation indicators is defined to measure the quality of the execution result. The actual score of the execution result is calculated based on the evaluation indicators. The actual score is compared with the expected comprehensive decision evaluation value, and the difference between the two is calculated. Positive or negative immediate rewards are given based on the difference. The absolute magnitude of the difference is considered; the larger the difference, the greater the reward or penalty. The calculated reward value is limited to a predetermined range to prevent the influence of extreme values.

[0175] In the intelligent medical diagnosis scenario, the optimal decision-making plan (Plan 1) is executed, and the actual execution results are obtained. Evaluation indicators include diagnostic accuracy (0.95), treatment effectiveness (0.88), side effect severity (0.15), medical resource consumption (0.25), and patient satisfaction (0.90). Based on these indicators, the actual score is calculated to be 0.92. Comparing the actual score (0.92) with the expected comprehensive decision evaluation value (0.984), the difference is calculated to be -0.064. Since the actual score is slightly lower than expected, a small negative immediate reward is given, with the immediate reward value set at -0.05.

[0176] The specific implementation method is as follows: calculating the neural oscillation synchronicity change value based on the phase synchronization matrix, co-mapping the instantaneous reward value with the neural oscillation synchronicity change value through the phase synchronization matrix to obtain the compensation reward value, accumulating the compensation reward value over time to obtain the cumulative reward value, inputting the cumulative reward value into the reinforcement learning algorithm as the reward signal, and calculating the gradient value of the inference module parameters based on the reinforcement learning algorithm.

[0177] This example demonstrates the dynamic adjustment inference module through an intelligent logistics delivery scenario. The logistics delivery decision system presents six candidate delivery plans with performance scores of [0.88, 0.85, 0.79, 0.76, 0.72, 0.68]. The constructed decision ranking value is [1.00, 0.85, 0.55, 0.40, 0.20, 0.00]. Ten neural oscillators are defined, corresponding to key variables such as delivery time, route length, number of vehicles, loading rate, fuel consumption, traffic conditions, weather impact, customer time window, delivery priority, and warehouse inventory. A 10×10 phase synchronization matrix is ​​constructed, and the predicted values ​​of the neural oscillation patterns are extracted. The decision ranking values ​​are mapped through the phase synchronization matrix, and the comprehensive decision evaluation value for each plan is calculated as [0.925, 0.842, 0.631, 0.483, 0.296, 0.142]. Plan 1 is selected as the optimal plan.

[0178] After implementing the optimal solution, the actual score was 0.90, a difference of -0.025 from the expected 0.925, with an immediate reward value of -0.02. The calculated neural oscillation synchronization change value was 0.65, indicating an improvement in delivery network coordination. A collaborative mapping between the immediate reward value of -0.02 and the synchronization change value of 0.65 yielded a compensation reward value of -0.008. This compensation reward value was accumulated over time and combined with historical rewards [0.015, 0.020, -0.005, 0.010, 0.012] to calculate a cumulative reward value of 0.009. This cumulative reward value was input into the policy gradient algorithm. The state space includes delivery area features, vehicle status, and order information, while the action space includes route planning strategies, vehicle allocation strategies, and time window adjustment strategies. The gradient values ​​of the inference module parameters were calculated to optimize the delivery decision model. After 10 rounds of iterative optimization, the weight of the system's merging strategy for small orders during peak hours increased from 0.35 to 0.48, improving delivery efficiency by 12%.

[0179] In one optional implementation, based on the adjusted inference module, the reinforcement learning algorithm is used to optimize the optimal decision scheme and generate a new decision scheme, including:

[0180] A chaotic feature matrix is ​​constructed based on the adjusted inference module output. A reward value is generated based on the chaotic feature matrix. The state distribution of the decision scheme is determined based on the reward value. A decision optimization sequence is generated based on the state distribution.

[0181] The decision optimization sequence is input into the inference module, and the parameters of the inference module are iteratively updated according to the reward value to obtain the updated inference module parameters;

[0182] The decision scheme is expanded and optimized using the updated inference module parameters to obtain a multidimensional decision set. The optimization index of each decision scheme in the multidimensional decision set is calculated based on the chaotic feature matrix. A new decision scheme is selected from the multidimensional decision set according to the optimization index.

[0183] The adjusted inference module utilizes reinforcement learning algorithms to optimize the optimal decision-making scheme, generating new schemes. During this process, a chaotic feature matrix is ​​constructed based on the output of the adjusted inference module. Specifically, features are extracted from the vector data output by the inference module to obtain key dimensional information, such as decision information, environmental state, and expected results. These multidimensional features are then converted into state points in a chaotic feature space using a nonlinear mapping function, forming a feature matrix with chaotic dynamics characteristics. For example, for a decision-making problem in an intelligent transportation system, data such as traffic flow, vehicle density, and weather conditions output by the inference module can be transformed into an 80×80 chaotic feature matrix, where each element's value is in the interval [-1, 1], representing the mapping value of that feature in the chaotic space.

[0184] When generating reward values ​​based on the chaotic feature matrix, a feature sensitivity evaluation method is used. For each element in the chaotic feature matrix, its influence on the decision result is calculated, and features with high influence are assigned higher weights. The comprehensive reward value is calculated by weighted summation. In the specific implementation, a reward function can be defined. This function takes the chaotic feature matrix as input and outputs a scalar value as the reward for the current decision. For example, in a resource scheduling scenario, if the decision scheme can reduce resource consumption while meeting the task completion time, a higher reward value will be obtained. If the resource utilization rate increases by 5%, the reward value increases by 0.2; if the task completion time is shortened by 10%, the reward value increases by 0.3.

[0185] The state distribution for determining decision schemes based on reward values ​​is achieved through probability mapping. The softmax function is used to convert reward values ​​into a probability distribution, ensuring that states with high reward values ​​have a higher selection probability. In practice, the probability value of each state is calculated, forming a state probability vector. For example, for five candidate states, the normalized probability distribution is [0.15, 0.25, 0.35, 0.10, 0.15], indicating that the third state has the highest selection probability.

[0186] When generating decision optimization sequences based on state distributions, a Monte Carlo tree search method is employed. Starting from the initial state, multiple samples are taken based on the state distribution to explore decision paths. Each sample generates a decision sequence, and the cumulative reward of that sequence is recorded. After multiple rounds of sampling, the decision sequence with the highest cumulative reward is selected as the optimization result. In a specific implementation, the number of samples can be set to 1000, with each sample generating a decision sequence of length 20. The top 5 sequences with the highest cumulative reward values ​​are then selected to form the final decision optimization sequence.

[0187] After the decision optimization sequence is input into the inference module, the parameters of the inference module are iteratively updated based on the reward value. The policy gradient method is used to calculate the gradient of the reward value relative to the model parameters, and the gradient ascent method is used to update the parameters, adjusting them towards directions that yield higher rewards. In the actual update process, a parameter update rule with a learning rate of 0.01 is used, and the learning rate decays every 100 iterations with a decay coefficient of 0.95. In this way, the inference module can gradually adapt to the decision-making environment and generate better decision solutions.

[0188] When expanding and optimizing decision schemes using the updated inference module parameters to obtain a multidimensional decision set, parameter perturbation techniques are employed to add varying degrees of random perturbation to the updated inference module parameters, generating multiple variant models. Each variant model generates different decision schemes for the same decision problem, thus forming a multidimensional decision set. In the specific implementation, up to 30 variant models can be generated, with the parameter perturbation range of each model controlled within ±5% of the original parameters, ensuring that the generated decision schemes have a certain degree of diversity while maintaining basic effectiveness.

[0189] When calculating the optimization index of each decision scheme in a multidimensional decision set based on the chaotic feature matrix, its performance is evaluated by combining multiple dimensions of the decision. Evaluation dimensions include resource consumption, time efficiency, stability, and adaptability. For each dimension, a corresponding evaluation function is designed to map the decision scheme to the chaotic feature matrix and calculate its score on each dimension. A comprehensive optimization index is formed by weighted combination of the scores from each dimension. For example, in smart grid load balancing decision-making, a weight of 0.4 can be set for resource utilization, 0.3 for system stability, and 0.3 for load balancing, to calculate the comprehensive score for each decision scheme.

[0190] When selecting new decision schemes from a multidimensional decision set based on optimization indicators, not only a single optimal solution is considered, but also the diversity of decision schemes. Using the Pareto optimality principle, a set of non-dominated solutions is selected from the multidimensional decision set, forming the Pareto front. From the Pareto front, the most suitable scheme is selected as the new decision scheme based on the current system state and decision requirements. In specific implementation, the top three decision schemes with the highest comprehensive scores can be selected to form a candidate set, and then the final decision is made in conjunction with the specific needs of the current system. For example, under different load conditions, decision schemes with different emphases are selected: under light load, a scheme prioritizing resource utilization is selected; under high load, a scheme prioritizing system stability is selected. This dynamic selection mechanism enables the decision system to better adapt to complex and changing environments.

[0191] The method further includes:

[0192] The knowledge acquisition module enables multi-source data collection through a visually configurable interface. Operations personnel can set API interface parameters, including interface address, authentication information, request method, request headers, and request body, in the configuration interface. Uploaded documents support formats such as PDF, Word, Excel, and Markdown; video materials support mainstream formats such as MP4 and AVI. When setting up scheduled collection tasks, the collection frequency can be specified, such as hourly, daily, weekly, or a custom time interval. For example, a data platform operations team configured the Git repository API of the data platform source code, setting it to automatically pull the latest code at 2 AM every day; simultaneously, they uploaded 500 historical fault solution documents and 50 fault handling video tutorials, establishing a comprehensive knowledge acquisition channel.

[0193] The knowledge content extraction module employs corresponding large-scale model processing techniques for different types of data. For text data, a natural language processing model is used to extract key information, segmenting long texts into knowledge fragments containing complete semantics. Each fragment includes a problem description and a solution. For image data, an image recognition model is used to extract textual information, chart data, and visual elements from images, converting them into structured descriptions. For audio data, a speech recognition model is used to convert audio into text, annotating the speaker, key content, and timestamps. For video data, video frame analysis and speech recognition are combined to extract key scenes and explanatory content, generating a text summary with a time index. In practical applications, for a fault handling video containing a failed ETL task in a data platform, the system extracted the operation steps, error log screenshots, and the engineer's verbal explanation from the video, forming a complete knowledge entry.

[0194] The knowledge construction and annotation module performs structured processing on the extracted knowledge. Based on semantic similarity algorithms, the system clusters knowledge items with a similarity exceeding 0.85 to form knowledge clusters. For each knowledge item, key entities such as component names, error codes, and operation steps are extracted to construct an entity relationship network. Deep learning models are used to analyze the correlations between knowledge items, identifying causal relationships, inclusion relationships, and similarity relationships. Key data is annotated, including dimensions such as problem type, severity, involved components, difficulty of resolution, and applicable scenarios. For example, for the knowledge content of data access components, the system annotates attributes such as "data source type = relational database," "error type = connection timeout," "severity = high," and "solution = check network connection and database parameter configuration," facilitating subsequent accurate retrieval.

[0195] The knowledge system storage module employs a dual-track storage strategy. The unstructured knowledge system storage portion uses a distributed document database to store source code, technical documents, failure cases, and internet solutions in their original formats, while creating full-text and vector indexes for each document. The structured knowledge system storage portion uses a graph database to construct a table relationship graph and a problem-solution semantic graph. The table relationship graph records the relationships between various tables in the data platform, including field mappings, foreign key constraints, and business dependencies. The semantic graph connects nodes such as problem descriptions, failure causes, solution steps, and verification methods, forming a complete problem-solving knowledge network. In a certain data platform environment, the system stores 5000 source code files, 2000 table structure definitions, and 1500 historical problem and solution records. Through the graph database, it has established over 10,000 entity nodes and 30,000 relationship edges, achieving efficient knowledge storage and rapid retrieval.

[0196] The problem description module provides a visual interface that supports multiple input methods. Operations personnel can describe problem details through text input boxes; it also supports uploading auxiliary information such as error screenshots and log files; it offers voice input, which is automatically converted to text; the interface integrates form guidance to help users fill in structured information such as problem type, occurrence time, and scope of impact; the system records users' historical queries for easy problem tracking and correlation analysis. For example, if an operations personnel enters "Data integration task A1023 continues to fail, error log shows HDFS connection timeout" on the interface and uploads an error screenshot, the system integrates this information to form a complete problem description.

[0197] The knowledge retrieval module searches for relevant knowledge based on the problem description, converts the problem description into a vector representation, and calculates the similarity with each entry in the knowledge base. It employs a hybrid retrieval strategy, combining keyword matching and semantic understanding to improve retrieval accuracy. For problems containing images, it extracts image content for multimodal retrieval. It optimizes the search ranking by incorporating user interaction history and problem-solving feedback. It also supplements the search results by querying entities and relationships related to the problem from the graph database. Regarding the aforementioned HDFS connection timeout issue, the system retrieved 10 relevant knowledge entries, including 3 similar cases, 2 HDFS configuration documents, and 5 related troubleshooting methods, with similarities of 0.92, 0.87, 0.85, 0.79, 0.78, 0.75, 0.72, 0.70, 0.68, and 0.65, respectively.

[0198] The prompt word encapsulation module transforms search results into prompt words that the large model can process efficiently. The module analyzes the retrieved knowledge content, extracts key information points, organizes information into structured prompts based on preset prompt word templates, adds contextual information such as system environment and related component configurations, adds task instructions explicitly requiring the large model to provide specific solution steps, and sets output format requirements to standardize the presentation of the solution. For example, the system encapsulates retrieved HDFS connection timeout related knowledge into prompt words: "Background: Data platform data integration task A1023 reports an HDFS connection timeout error. Related knowledge: 1. Common causes of HDFS connection timeout include network problems, configuration errors, and permission issues; 2. Integration task logs show connection failure after 3 attempts; 3. 80% of similar cases are caused by firewall rules. Task: Please analyze possible causes based on the above information and provide detailed troubleshooting steps, in the format of 'Cause Analysis' and 'Solution Steps'."

[0199] The large model management module enables efficient management and application of model resources. It supports new model configuration, including selecting the basic model type, setting parameters, and configuring the deployment environment; provides model fine-tuning capabilities, adapting to specific domain knowledge using a small number of data samples; supports model training, leveraging large amounts of labeled data to enhance the model's professional capabilities in operational fields; implements unified API management for models, providing standardized interfaces for models with different capabilities; and monitors model performance metrics, including response time, accuracy, and resource consumption, enabling dynamic model scheduling. In actual deployment, the system is configured with a general-purpose large language model for text understanding and generation, a dedicated image recognition model for screenshot analysis, and domain expert models to solve complex technical problems, all coordinated and invoked through a unified API interface.

[0200] The multi-stage reasoning module performs in-depth analysis based on a large model and generates solutions. First, it generates a preliminary solution based on retrieved knowledge and problem descriptions. Then, it evaluates the completeness and applicability of the preliminary solution, identifying any necessary supplementary information. Through multiple rounds of interaction, it requests supplementary information from the knowledge base to optimize the solution. Finally, it generates a final solution with detailed steps and clear operations. The reasoning process and evidence chain are recorded to support the interpretability of the solution. For the HDFS connection timeout issue, the system first generates a preliminary solution including network checks and configuration verification. Evaluation reveals a lack of firewall configuration information specific to the environment. A secondary search supplements the relevant knowledge, ultimately providing a complete solution with "checking whether firewall rules are blocking communication on port 8020" as the primary troubleshooting step, accurately pinpointing the root cause of the problem.

[0201] A data engineer discovered that the daily incremental synchronization task for the data warehouse was failing repeatedly, with system logs showing "data source connection error." The engineer entered a problem description and uploaded screenshots of the error logs through the intelligent operations and maintenance assistant interface. The system vectorized the problem and retrieved 7 similar cases and 3 related documents from the knowledge base, with similarity ranging from 0.95 to 0.75. The knowledge retrieval results showed that 80% of similar problems were related to data source network connectivity, 15% to account permissions, and 5% to changes in the source system. The system encapsulated prompts, called a large model for multi-round reasoning, and after generating a preliminary solution, discovered that specific data source configuration information was missing. It automatically supplemented the relevant content from the knowledge base, ultimately generating a complete solution. The engineer followed the suggestions and found that the problem was caused by a change in the source system's IP address. Updating the configuration successfully resolved the issue. The entire process, from problem submission to resolution, took only 15 minutes, reducing the time by 80% compared to traditional manual troubleshooting.

[0202] A second aspect of the present invention provides an intelligent operation and maintenance system for a data platform based on a multimodal large model, comprising:

[0203] The first unit is used to acquire multimodal data in the data platform operation and maintenance scenario, construct node features of the knowledge graph based on the multimodal data, extract scenario features of the multimodal data, construct node relationships of the knowledge graph based on the scenario features, analyze the association strength between the nodes, and construct connection weights between the nodes.

[0204] The second unit is used to expand the knowledge graph using the node features and the connection weights, and to match similar scenes in the knowledge graph based on the scene features to identify historical decision information related to the current scene; the scene features and the historical decision information are input into the reasoning module of the knowledge graph to generate a set of decision schemes, and the execution effect score is calculated for the set of decision schemes.

[0205] The third unit is used to select the optimal decision scheme based on the execution effect score, use the execution result of the optimal decision scheme as the reward signal of the reinforcement learning algorithm, and dynamically adjust the inference module; based on the adjusted inference module, the reinforcement learning algorithm is used to optimize the optimal decision scheme and generate a new decision scheme.

[0206] The fourth unit is used to generate operation and maintenance instructions using the new decision-making scheme, execute the operation and maintenance instructions according to the preset execution process, and record the execution process and results of the decision-making scheme.

[0207] A third aspect of the present invention provides an electronic device, comprising:

[0208] processor;

[0209] Memory used to store processor-executable instructions;

[0210] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0211] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0212] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0213] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. 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 or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A data platform intelligent operation and maintenance method based on a multimodal large model, characterized in that, include: Acquire multimodal data in the data platform operation and maintenance scenario, construct node features of a knowledge graph based on the multimodal data, and extract scenario features from the multimodal data; Based on the scene features, the node relationships of the knowledge graph are constructed, and the association strength between the nodes is analyzed to construct the connection weights between the nodes. The knowledge graph is expanded using the node features and the connection weights, and similar scenarios in the knowledge graph are matched based on the scene features to identify historical decision information related to the current scene. The scene features and the historical decision information are input into the reasoning module of the knowledge graph to generate a set of decision schemes, and the execution effect score of the set of decision schemes is calculated. Based on the performance score, the optimal decision scheme is selected, and the execution result of the optimal decision scheme is used as the reward signal for the reinforcement learning algorithm to dynamically adjust the inference module, including: A decision ranking value is constructed based on the results of the execution effect score of the decision scheme. A phase synchronization matrix is ​​constructed by calculating the phase difference of multiple neurons based on the decision ranking value. The phase synchronization matrix is ​​used to perform co-mapping on the features of different oscillation modes to obtain the neural oscillation mode prediction value corresponding to the decision scheme. The decision ranking value is co-mapped with the neural oscillation mode prediction value through the phase synchronization matrix to obtain the comprehensive decision evaluation value. The optimal decision scheme is selected based on the comprehensive decision evaluation value. The actual execution result of the optimal decision scheme is compared with the comprehensive evaluation value of the decision to construct an instant reward value. The neural oscillation synchronicity change value is calculated based on the phase synchronization matrix. The instant reward value is co-mapped with the neural oscillation synchronicity change value through the phase synchronization matrix to obtain a compensation reward value. The compensation reward value is accumulated over time to obtain a cumulative reward value. The cumulative reward value is input into the reinforcement learning algorithm as a reward signal. The gradient values ​​of the inference module parameters are calculated based on the reinforcement learning algorithm. The constraint difference between the inference module parameters and the target neural oscillation mode is calculated based on the phase synchronization matrix. The inference module is dynamically adjusted by co-mapping the gradient values ​​and the historical update values ​​of the inference module parameters. Based on the adjusted inference module, the optimal decision scheme is optimized using the reinforcement learning algorithm to generate a new decision scheme. The new decision-making scheme is used to generate operation and maintenance instructions, which are then executed according to a preset execution flow. The execution process and results of the decision-making scheme are recorded.

2. The method according to claim 1, characterized in that, Node features of a knowledge graph are constructed based on the multimodal data, and scene features of the multimodal data are extracted. Based on the scene features, a knowledge graph node relationship is constructed, and the association strength between the nodes is analyzed. The connection weights between nodes are constructed, including: Extract scene features from the multimodal data, and map the scene features to the knowledge graph space to construct node features of the knowledge graph; Based on the scene features, semantic association analysis is performed on the node features of the knowledge graph to construct the connection relationship between the nodes; the association strength between the nodes is calculated according to the connection relationship, and the association strength is used as the connection weight between the nodes.

3. The method according to claim 1, characterized in that, The knowledge graph is expanded using the node features and connection weights, and similar scenes in the knowledge graph are matched based on the scene features to identify historical decision information related to the current scene, including: Introduce scene feature vectors into node feature vectors to calculate inter-node connection weights, and use the node feature vectors and inter-node connection weights to establish a knowledge antibody feature matrix containing multiple feature patterns. Based on the knowledge antibody feature matrix, calculate the matching coefficient between the node feature vector and the feature pattern in the knowledge antibody feature matrix; calculate the expansion weight based on the node connection weight and the matching coefficient; dynamically expand the initial knowledge graph based on the expansion weight to obtain the expanded knowledge graph; extract the current scene features from the expanded knowledge graph based on the expansion weight, and combine the current scene features with the historical features and relationship features in the expanded knowledge graph to construct a scene feature vector; The similarity between the scene feature vector and the historical scene feature vector is calculated using the extended weights. Similar historical scenes are identified from the extended knowledge graph based on the similarity, and corresponding decision features are extracted from the similar historical scenes based on the matching coefficient. The decision features are fused with the scene feature vector to obtain the decision weight. The weighted function value of the decision weight and the matching coefficient is input into the knowledge antibody feature matrix for feature matching. Based on the feature matching result, historical decision information related to the current scene is obtained.

4. The method according to claim 1, characterized in that, The scene features and historical decision information are input into the reasoning module of the knowledge graph to generate a set of decision solutions. The execution performance score of the set of decision solutions is then calculated, including: Scene features and historical decisions are used as inputs to the reasoning module of the knowledge graph, and a set of nonlinear state equations is constructed in the reasoning module of the knowledge graph. The system evolution trajectory is obtained by solving the set of nonlinear state equations, and the dynamic evolution features of the knowledge graph are extracted based on the state sequence in the system evolution trajectory. The reasoning module of the knowledge graph is used to reason about the scene features and the historical decision information based on the temporal change pattern of the dynamic evolution features, generating multiple candidate decision schemes and forming a decision scheme set. Calculate the gradient vector of the system evolution trajectory, identify multiple control points from the gradient vector, use the control points as the benchmark points for scoring, and calculate the corresponding evaluation value based on the state information of each candidate decision scheme in the decision scheme set at the control point. Based on the corresponding evaluation value, the corresponding neuron activation intensity is calculated, and a plasticity connection matrix representing the dynamic correlation strength between the control points is constructed. The neuron activation intensity and the plasticity connection matrix are iteratively calculated to obtain a stable network resonance state. Based on the network resonance state, the execution effect score of the candidate decision scheme is calculated.

5. The method according to claim 4, characterized in that, The reasoning module of the knowledge graph uses the temporal change pattern of the dynamic evolution features to reason about the scene features and the historical decision information, generating multiple candidate decision schemes and forming a decision scheme set, including: Scene features and historical decision information are used to construct feature vectors that represent temporal characteristics. Temporal evolution analysis is performed on the feature vectors to obtain their evolutionary rules, and dynamic change features are extracted from the evolutionary rules. The feature vector is input into the reasoning module of the knowledge graph and transformed into a feature space to obtain a transformed feature expression. The stable region of the transformed feature expression is determined according to the evolution law of the feature vector. The stable region is used to constrain the transformed feature expression to obtain multiple feature parameters characterizing the stability of the system. The reasoning module of the knowledge graph constructs a feature weight distribution based on the feature parameters, and calculates the changing trend of the transformed feature expression based on the feature weight distribution; The dynamic change features are input into the reasoning module of the knowledge graph, and the transformed feature expression is reasoned based on the change trend to obtain the reasoning result. Under the constraint of the stable region, multiple candidate decision schemes are generated according to the reasoning result. For the candidate decision schemes, an evaluation index including scheme reliability, scheme diversity and information gain is calculated. Based on the evolution law of the feature vector, a dynamically changing screening threshold is determined. The evaluation index is compared with the screening threshold, and candidate decision schemes that meet the screening threshold are selected to form a decision scheme set.

6. The method according to claim 1, characterized in that, Based on the adjusted inference module, the reinforcement learning algorithm is used to optimize the optimal decision scheme and generate a new decision scheme, including: A chaotic feature matrix is ​​constructed based on the adjusted inference module output. A reward value is generated based on the chaotic feature matrix. The state distribution of the decision scheme is determined based on the reward value. A decision optimization sequence is generated based on the state distribution. The decision optimization sequence is input into the inference module, and the parameters of the inference module are iteratively updated according to the reward value to obtain the updated inference module parameters; The decision scheme is expanded and optimized using the updated inference module parameters to obtain a multidimensional decision set. The optimization index of each decision scheme in the multidimensional decision set is calculated based on the chaotic feature matrix. A new decision scheme is selected from the multidimensional decision set according to the optimization index.

7. A data platform intelligent operation and maintenance system based on a multimodal large model, used to implement the method of any one of claims 1-6, characterized in that, include: The first unit is used to acquire multimodal data in the data platform operation and maintenance scenario, construct node features of the knowledge graph based on the multimodal data, and extract scenario features of the multimodal data; Based on the scene features, the node relationships of the knowledge graph are constructed, and the association strength between the nodes is analyzed to construct the connection weights between the nodes. The second unit is used to expand the knowledge graph using the node features and the connection weights, and to match similar scenes in the knowledge graph based on the scene features to identify historical decision information related to the current scene; the scene features and the historical decision information are input into the reasoning module of the knowledge graph to generate a set of decision schemes, and the execution effect score is calculated for the set of decision schemes. The third unit is used to select the optimal decision scheme based on the execution effect score, use the execution result of the optimal decision scheme as the reward signal of the reinforcement learning algorithm, and dynamically adjust the inference module; based on the adjusted inference module, the reinforcement learning algorithm is used to optimize the optimal decision scheme and generate a new decision scheme. The fourth unit is used to generate operation and maintenance instructions using the new decision-making scheme, execute the operation and maintenance instructions according to the preset execution process, and record the execution process and results of the decision-making scheme.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images