Multi-modal large model-based intelligent operation and maintenance method and system for data medium station

Abstract

The invention provides an intelligent operation and maintenance method and system for a data medium table based on a multi-modal large model, and relates to the technical field of data processing, and the method comprises the steps: obtaining multi-modal data to construct a knowledge graph, expanding the knowledge graph based on node features and connection weights, matching similar scenes to recognize historical decision information, and generating and evaluating a decision scheme set. An optimal scheme is selected, a new decision scheme is generated through reinforcement learning optimization, and finally operation and maintenance operation is executed and a result is recorded. According to the method, intelligent decision-making and self-optimization of operation and maintenance of the data medium station are realized, and the operation and maintenance efficiency and accuracy are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a data center intelligent operation and maintenance method and system based on a multi-modal large model. BACKGROUND

[0002] With the deepening of enterprise digital transformation, data center as a platform for unified management and processing of various data resources of enterprises has become an important part of enterprise informatization construction. Data center integrates data generated by various business systems to provide one-stop data analysis and application services for enterprises. However, with the rapid increase of data volume and the increase of business complexity, the operation and maintenance management of data center faces unprecedented challenges. Traditional manual operation and maintenance methods have been difficult to meet the efficient, accurate and timely response requirements in large-scale data processing environment.

[0003] In recent years, with the development of artificial intelligence technology, especially the breakthrough of large model and multi-modal data processing technology, it provides a new technical path for intelligent operation and maintenance of data center. Multi-modal large model can process various data types such as text, image, log, time series data, etc., and realize deep understanding and intelligent decision of operation and maintenance scene through knowledge graph and reasoning technology.

[0004] In the prior art, the data center operation and maintenance method cannot effectively integrate and utilize multi-modal data, and a large amount of unstructured data such as system logs and monitoring images cannot be fully utilized, resulting in lack of comprehensive data support for operation and maintenance decision, affecting the accuracy of fault diagnosis and prediction.

[0005] The existing operation and maintenance system lacks effective knowledge accumulation and experience inheritance mechanism, and it is difficult to transform historical operation and maintenance experience into reusable knowledge. Each time a similar problem occurs, it needs to be analyzed and solved again, which greatly reduces the operation and maintenance efficiency and increases the operation and maintenance cost.

[0006] Traditional operation and maintenance methods lack self-adaptive learning ability, cannot automatically adjust decision strategies according to different scenes, and are difficult to continuously optimize from past operation and maintenance practices, resulting in unstable execution effect of operation and maintenance scheme and difficult to cope with complex and changeable data center operation environment. SUMMARY

[0007] The embodiment of the present application provides a data center intelligent operation and maintenance method and system based on a multi-modal large model, which can solve the problems in the prior art.

[0008] The first aspect of the embodiment of the present application provides a data center intelligent operation and maintenance method based on a multi-modal large model, comprising: The multi-modal data in the data acquisition platform operation scene is acquired, node features of a knowledge graph are constructed based on the multi-modal data, and scene features of the multi-modal data are extracted; node relationships of the knowledge graph are constructed based on the scene features, and the correlation strength between the nodes is analyzed to construct connection weights between the nodes; The knowledge graph is expanded using the node features and the connection weights, similar scenes in the knowledge graph are matched based on the scene features, and historical decision information related to the current scene is identified; the scene features and the historical decision information are jointly input into an inference module of the knowledge graph, a decision scheme set is generated, and an execution effect score of the decision scheme set is calculated; The optimal decision scheme is selected based on the execution effect score, the execution result of the optimal decision scheme is taken as a reward signal of a reinforcement learning algorithm, and the inference module is dynamically adjusted; the optimal decision scheme is optimized using the reinforcement learning algorithm based on the adjusted inference module to generate a new decision scheme; The new decision scheme is used to generate an operation instruction, the operation instruction is executed according to a preset execution process, and the execution process and the execution result of the decision scheme are recorded.

[0009] The node features of the knowledge graph are constructed based on the multi-modal data, and the scene features of the multi-modal data are extracted; the node relationships of the knowledge graph are constructed based on the scene features, and the correlation strength between the nodes is analyzed to construct connection weights between the nodes, including: The scene features of the multi-modal data are extracted, the scene features are mapped to the knowledge graph space to construct the node features of the knowledge graph, the node features of the knowledge graph are analyzed based on the scene features to construct the connection relationships between the nodes, and the correlation strength between the nodes is calculated as the connection weights between the nodes.

[0010] The knowledge graph is expanded using the node features and the connection weights, similar scenes in the knowledge graph are matched based on the scene features, and historical decision information related to the current scene is identified, including: The scene feature vector is introduced into the node feature vector to calculate the connection weights between the nodes, and the node feature vector and the connection weights between the nodes are used to establish a knowledge antibody feature matrix containing multiple feature modes; calculate a matching coefficient between the node feature vector and a feature pattern in the knowledge antibody feature matrix based on the knowledge antibody feature matrix; calculate an expansion weight according to the inter-node connection weight and the matching coefficient, and perform dynamic expansion on the initial knowledge graph based on the expansion weight to obtain an expanded knowledge graph; extract a current scene feature from the expanded knowledge graph based on the expansion weight, and combine the current scene feature with historical features and relationship features in the expanded knowledge graph to construct a scene feature vector; calculate a similarity between the scene feature vector and a historical scene feature vector using the expansion weight, identify a similar historical scene from the expanded knowledge graph according to the similarity, and extract corresponding decision features from the similar historical scene based on the matching coefficient; perform feature fusion on the decision features and the scene feature vector to obtain a decision weight, input the decision weight and a weighted function value of the matching coefficient into the knowledge antibody feature matrix for feature matching, and obtain historical decision information related to the current scene based on a result of the feature matching.

[0011] input the scene feature and the historical decision information into an inference module of a knowledge graph, generate a decision scheme set, and calculate an execution effect score of the decision scheme set, including: input the scene feature and the historical decision information into an inference module of a knowledge graph, generate a decision scheme set, and calculate an execution effect score of the decision scheme set, including: use the inference module of the knowledge graph to reason the scene feature and the historical decision information based on a time sequence change rule of the dynamic evolution feature, generate a plurality of candidate decision schemes and form a decision scheme set; calculate a gradient vector of the system evolution trajectory, identify a plurality of control points from the gradient vector, use the control points as reference points for score calculation, and calculate a corresponding evaluation value based on state information of each candidate decision scheme in the decision scheme set at the control points; calculate a corresponding neuron activation intensity based on the corresponding evaluation value, construct a plasticity connection matrix representing dynamic correlation strength between the control points, perform iterative calculation on the neuron activation intensity and the plasticity connection matrix to obtain a stable network resonance state, and calculate an execution effect score of the candidate decision scheme based on the network resonance state.

[0012] The inference module of the knowledge graph reasons the scene features and the historical decision information based on the time sequence variation law of the dynamic evolution characteristics, generates a plurality of candidate decision schemes and constitutes a decision scheme set, which comprises: The scene features and the historical decision information are constructed into a feature vector representing time sequence characteristics, the evolution law of the feature vector is obtained by time sequence evolution analysis of the feature vector, and the dynamic change characteristics are extracted from the evolution law; The feature vector is input into the inference module of the knowledge graph for feature space conversion to obtain a converted feature expression, the stable region of the converted feature expression is determined according to the evolution law of the feature vector, the converted feature expression is constrained by using the stable region, and a plurality of feature parameters representing system stability are obtained; The inference module of the knowledge graph is used to construct a feature weight distribution based on the feature parameters, and the change trend of the converted feature expression is calculated according to the feature weight distribution; The dynamic change characteristics are input into the inference module of the knowledge graph, the converted feature expression is reasoned based on the change trend to obtain a reasoning result, and a plurality of candidate decision schemes are generated according to the reasoning result under the constraint of the stable region; The evaluation indexes including scheme reliability, scheme diversity and information gain are calculated for the candidate decision schemes, the dynamic change filtering threshold is determined based on the evolution law of the feature vector, the evaluation indexes are compared with the filtering threshold, and the candidate decision schemes satisfying the filtering threshold are selected to constitute a decision scheme set.

[0013] Based on the execution effect score, the optimal decision scheme is selected, the execution result of the optimal decision scheme is taken as the reward signal of the reinforcement learning algorithm, and the inference module is dynamically adjusted, which comprises: A decision ranking value is constructed based on the result of the execution effect score of the decision scheme, a phase difference of a plurality of neurons is calculated according to the decision ranking value to construct a phase synchronization matrix, different oscillation mode characteristics are cooperatively mapped by using the phase synchronization matrix to obtain a neural oscillation mode prediction value corresponding to the decision scheme, the decision ranking value is cooperatively mapped with the neural oscillation mode prediction value through the phase synchronization matrix to obtain a decision comprehensive evaluation value, and the optimal decision scheme is selected based on the decision comprehensive evaluation value; The actual execution result of the optimal decision scheme is compared with the decision comprehensive evaluation value to construct an instant reward value, a neural oscillation synchronization change value is calculated based on the phase synchronization matrix, the instant reward value is cooperatively mapped with the neural oscillation synchronization change value through the phase synchronization matrix to obtain a compensation reward value, the compensation reward value is time sequence accumulated to obtain an accumulated reward value, and the accumulated reward value is input into the reinforcement learning algorithm as a reward signal. calculate gradient values of inference module parameters based on the reinforcement learning algorithm, calculate constraint difference values between the inference module parameters and target neural oscillation patterns according to the phase synchronization matrix, and dynamically adjust the inference module by synergistically mapping the gradient values and historical update values of the inference module parameters.

[0014] based on the adjusted inference module, optimize the optimal decision scheme using the reinforcement learning algorithm, and generate a new decision scheme including: based on the adjusted inference module output, construct a chaotic feature matrix, generate a reward value according to the chaotic feature matrix, determine a state distribution of the decision scheme based on the reward value, and generate a decision optimization sequence based on the state distribution; input the decision optimization sequence into the inference module, iteratively update parameters of the inference module according to the reward value, and obtain updated inference module parameters; use the updated inference module parameters to expand and optimize the decision scheme to obtain a multi-dimensional decision set, calculate optimization indicators of each decision scheme in the multi-dimensional decision set based on the chaotic feature matrix, and select a new decision scheme from the multi-dimensional decision set according to the optimization indicators.

[0015] In a second aspect of the embodiment of the present application, a data center intelligent operation and maintenance system based on a multi-modal large model is provided, including: A first unit is configured to acquire multi-modal data in a data center operation and maintenance scenario, construct node features of a knowledge graph based on the multi-modal data, and extract scene features of the multi-modal data; construct node relationships of the knowledge graph based on the scene features, analyze the association strength between the nodes, and construct connection weights between the nodes; A second unit is configured to expand the knowledge graph using the node features and the connection weights, match similar scenes in the knowledge graph based on the scene features, and identify historical decision information related to the current scene; input the scene features and the historical decision information into an inference module of the knowledge graph together, generate a decision scheme set, and calculate execution effect scores of the decision scheme set; A third unit is configured to select an optimal decision scheme based on the execution effect scores, use the execution result of the optimal decision scheme as a reward signal of a reinforcement learning algorithm, and dynamically adjust the inference module; based on the adjusted inference module, optimize the optimal decision scheme using the reinforcement learning algorithm, and generate a new decision scheme. A fourth unit is configured to generate operation and maintenance operation instructions using the new decision scheme, execute the operation and maintenance operation instructions according to a preset execution process, and record the execution process and execution result of the decision scheme.

[0016] In a third aspect, the present application provides an electronic device, comprising: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0017] In a fourth aspect, the present application provides a computer-readable storage medium having stored thereon computer program instructions, which, when executed by a processor, implement the method described above.

[0018] The present application has the following beneficial effects: The present application realizes efficient and automated operation and maintenance of the data center through the data center intelligent operation and maintenance method based on the multi-modal large model, improves the operation and maintenance efficiency and quality. The method constructs a knowledge graph using multi-modal data and performs similar scene matching based on scene characteristics, effectively identifies historical decision information related to the current scene, and provides reliable knowledge support for operation and maintenance decision-making.

[0019] By inputting the scene characteristics and historical decision information into the inference module of the knowledge graph, the method can generate a diversified set of decision schemes and select the optimal decision scheme based on the execution effect score, greatly improving the accuracy and intelligence level of operation and maintenance decision-making. At the same time, the method introduces a reinforcement learning algorithm, using the execution result of the decision scheme as a reward signal to dynamically adjust the inference module, realizing continuous optimization and self-improvement of the decision scheme.

[0020] The present application records the execution process and result of the decision scheme, establishes a complete operation and maintenance knowledge closed loop, enables the system to continuously learn from historical experience, and improves the system's adaptability and processing efficiency for complex operation and maintenance scenarios. Overall, the method organically combines multi-modal data analysis, knowledge graph reasoning, and reinforcement learning to build an intelligent and adaptive data center operation and maintenance system, significantly reducing operation and maintenance costs and improving the intelligence level of operation and maintenance. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 FIG. 1 is a flowchart of the data center intelligent operation and maintenance method based on the multi-modal large model of the present application; Figure 2 FIG. 2 is a flowchart of the network behavior evaluation method based on dynamic feature reasoning and neuron activation intensity of the present application; Figure 3 FIG. 3 is a flowchart of the decision scheme generation method based on dual-path feature analysis of the present application. DETAILED DESCRIPTION

[0022] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0023] The technical solutions of the present application will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes can not be described in some embodiments.

[0024] Figure 1 The flowchart of the method for intelligent operation and maintenance of a data center based on a multi-modal large model according to an embodiment of the present application is shown in FIG. 1, which comprises the following steps. Figure 1 Obtaining multi-modal data in a data center operation and maintenance scenario, constructing node features of a knowledge graph based on the multi-modal data, and extracting scenario features of the multi-modal data; constructing node relationships of the knowledge graph based on the scenario features, and analyzing the correlation strength between the nodes to construct connection weights between the nodes; Extending the knowledge graph using the node features and the connection weights, and matching similar scenarios in the knowledge graph based on the scenario features to identify historical decision information related to the current scenario; inputting the scenario features and the historical decision information into an inference module of the knowledge graph to generate a set of decision schemes, and calculating the execution effect score of the set of decision schemes; Selecting the optimal decision scheme based on the execution effect score, taking the execution result of the optimal decision scheme as a reward signal of a reinforcement learning algorithm, and dynamically adjusting the inference module; optimizing the optimal decision scheme using the reinforcement learning algorithm based on the adjusted inference module to generate a new decision scheme; Generating an operation and maintenance operation instruction using the new decision scheme, executing the operation and maintenance operation instruction according to a preset execution process, and recording the execution process and execution result of the decision scheme.

[0025] In an optional embodiment, constructing node features of a knowledge graph based on the multi-modal data, and extracting scenario features of the multi-modal data; constructing node relationships of the knowledge graph based on the scenario features, and analyzing the correlation strength between the nodes to construct connection weights between the nodes comprises: ​extracting scene features of the multi-modal data, mapping the scene features to a knowledge graph space to construct node features of the knowledge graph, performing semantic correlation analysis on the node features of the knowledge graph based on the scene features to construct connection relationships between nodes, and calculating correlation strengths between the nodes according to the connection relationships and taking the correlation strengths as connection weights between the nodes.

[0026] The multi-modal data includes but is not limited to text, image, audio and video data, and different feature extraction methods are adopted for different types of multi-modal 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, an acoustic feature extraction technique is used to obtain audio features; and for video data, temporal and spatial information are combined to extract video features.

[0027] Taking an e-commerce scenario as an example, the multi-modal data includes product description text, product pictures, user comment audio and product demonstration video. Key words and semantic information such as "high-definition display", "27 inches" and "4K resolution" are extracted from the product description text; color, shape and category visual features are extracted from the product pictures; sentiment tendency and key evaluation points are extracted from the user comment audio; and dynamic use scenarios and function demonstration information are extracted from the product demonstration video.

[0028] In the process of extracting scene features of multi-modal data, cross-modal fusion technology is used to fuse features of different modalities. In specific implementation, an attention mechanism is used to weight and fuse different modal features, and the importance weight of each modal feature is automatically adjusted according to the requirements of the current task. For example, more attention is paid to user comments and product demonstration videos when judging product quality, and more attention is paid to product pictures and description texts when classifying products. By calculating the correlation between different modal features, the weight coefficient of feature fusion is determined to realize intelligent feature fusion.

[0029] After the 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, and each node represents an entity. The node features are the representation of the entity in the feature space. The mapping process is implemented by a deep neural network. The extracted scene features are input into the neural network, and after multiple layers of nonlinear transformation, the node feature representation conforming to the distribution of the knowledge graph space is output.

[0030] In the e-commerce scenario, the nodes can be product, category, attribute, user and other entities. For product nodes, the features include product name, brand, price range, main functions and other information; for user nodes, the features include user preferences, purchase history, browsing behavior and other information. These features are represented in the knowledge graph space in the form of vectors, and each vector dimension corresponds to a specific semantic meaning.

[0031] The semantic correlation analysis is performed on the node features of the knowledge graph based on scene features to construct the connection relationship between nodes. The semantic correlation analysis adopts a graph neural network technology to pass information between nodes through a message passing mechanism to mine the potential correlation between nodes. In a specific implementation, for any two nodes, the semantic similarity of the nodes in a feature space is calculated, and if the similarity exceeds a preset threshold, a connection relationship is established.

[0032] Taking an e-commerce scenario as an example, the commodity node "high-definition display" has a "has" relationship with the attribute node "4K resolution"; the commodity node "high-definition display" has a "belongs to" relationship with the category node "electronic device"; and the user node has a "purchase", "browse", "collect", and the like relationship with the commodity node. These relationships are automatically discovered and established through semantic correlation analysis.

[0033] After the connection relationship between nodes is established, the correlation strength between nodes needs to be calculated, and the correlation strength is taken as the connection weight between nodes. The correlation strength calculation is based on the similarity of node features, the interaction frequency between nodes, and time sequence information, and the like. For the similarity of node features, the cosine similarity is used to calculate the cosine value of the included angle between two node feature vectors, and the smaller the included angle, the higher the similarity; for the interaction frequency, the historical interaction times between nodes are counted, and the more frequent the interaction, the stronger the correlation; for the time sequence information, the time distribution of the interaction is considered, and the recent interaction weight is higher than the long-term interaction.

[0034] In actual applications, for example, in a commodity recommendation system, the correlation strength between a user node and a commodity node can be calculated based on the user's purchase history, browsing time, evaluation content, and the like. Assuming that user A has recently browsed high-definition display related commodities multiple times, and has a longer stay time for a 27-inch 4K display, and has also viewed detailed parameters and user evaluations, the correlation strength between user A and the display node will be higher. Specifically, the browsing times are 5 times, the average stay time is 3 minutes each time, and the details are viewed 2 times, and then a comprehensive correlation strength value can be calculated as the connection weight between the two nodes.

[0035] For different types of relationships, the calculation method of the correlation strength is different, for example, for the "has" relationship, the correlation strength can be calculated based on the appearance frequency and importance of the attribute in the description text; for the "belongs to" relationship, the correlation strength can be calculated based on the typicality of the commodity and the category; for the "purchase" relationship, the correlation strength can be calculated based on the purchase frequency, evaluation score, and the like.

[0036] After the knowledge graph is constructed, various intelligent applications can be performed using the graph structure, for example, discovering the implicit relationship between entities through a graph traversal algorithm, identifying a group of entities with similar characteristics through a community discovery algorithm, and compressing the structural information in the knowledge graph into a low-dimensional vector representation through a graph embedding technique for downstream tasks such as recommendation, classification, etc.

[0037] Taking a recommendation system as an example, based on the constructed knowledge graph, the user's interest-related goods can be found through graph traversal. If user A shows interest in a high-definition display, the system can find other goods related to the high-definition display along the relationship path in the knowledge graph, such as a graphics card, an HDMI cable, etc., or other display devices with similar characteristics. By calculating the cumulative value of the correlation strength on the path, the recommendation results can be sorted, and the goods with high correlation strength are sorted in the front.

[0038] In actual applications, the knowledge graph can also be dynamically updated according to business needs. When new multi-modal data is input, the above process is repeated to extract scene features, update node features and relationships, and adjust connection weights. This can ensure that the knowledge graph reflects the latest data state and business changes.

[0039] In addition, in order to improve the quality of the knowledge graph, a feedback mechanism can be introduced to adjust the parameters in the knowledge graph construction process according to the effect of the knowledge graph application, such as recommendation accuracy, user satisfaction, etc. For example, if it is found that the prediction accuracy of a certain type of relationship is low, the feature extraction and correlation strength calculation method of this type of relationship can be adjusted.

[0040] Through the above method, the knowledge graph constructed based on multi-modal data can comprehensively capture the complex relationship between entities and provide strong knowledge support for intelligent applications. This method is suitable for e-commerce, content recommendation, intelligent customer service, etc. Various scenarios can effectively improve the intelligent level of the system and user experience.

[0041] In an 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: The scene feature vector is introduced into the node feature vector to calculate the connection weight between nodes, and the knowledge antibody feature matrix containing multiple feature modes is established using the node feature vector and the connection weight between nodes. calculate a matching coefficient between the node feature vector and a feature pattern in the knowledge antibody feature matrix based on the knowledge antibody feature matrix; calculate an expansion weight based on the inter-node connection weight and the matching coefficient, and perform dynamic expansion on the initial knowledge graph based on the expansion weight to obtain an expanded knowledge graph; extract a current scene feature from the expanded knowledge graph based on the expansion weight, and combine the current scene feature with historical features and relationship features in the expanded knowledge graph to construct a scene feature vector; calculate a similarity between the scene feature vector and a historical scene feature vector using the expansion weight, identify a similar historical scene from the expanded knowledge graph according to the similarity, and extract a corresponding decision feature from the similar historical scene based on the matching coefficient; perform feature fusion on the decision feature and the scene feature vector to obtain a decision weight, input the decision weight and a weighted function value of the matching coefficient into the knowledge antibody feature matrix for feature matching, and obtain historical decision information related to the current scene based on a result of the feature matching.

[0042] The constructed knowledge graph node feature vector and the extracted scene feature vector are subjected to dimension normalization processing respectively to ensure that the numerical ranges of the dimensions are consistent, facilitating subsequent calculation. The scene feature vector is added to the tail of the node feature vector in a splicing manner to form an enhanced node feature vector. The element product of the node feature vector and the scene feature vector is calculated to capture the interaction information therebetween, and the element product result is added to the enhanced node feature vector to form a complete fusion feature vector.

[0043] In the intelligent medical scene, the node feature vector contains patient basic information, symptom characteristics, examination results, etc., and has a dimension of 120; the scene feature vector contains a medical environment, seasonal factors, disease epidemic trends, etc., and has a dimension of 80. After normalization of the two types of vectors, a 200-dimensional vector is obtained by splicing; the element product of the two vectors is calculated to obtain interaction features; and finally a 280-dimensional fusion feature vector containing original features and interaction features is formed, which comprehensively describes the comprehensive features of the patient in a specific scene.

[0044] When calculating the inter-node connection weight, for any two nodes i and j, their fusion feature vectors are taken, and the cosine similarity between the vectors is calculated; the direct association relationship between the nodes is considered, such as the diagnosis relationship between the patient and the disease, and an association coefficient is given; the importance of the nodes in the scene is considered, and the correlation between the node features and the scene features is calculated; the similarity, the association coefficient and the correlation are weighted and summed to obtain the inter-node connection weight. In actual application, the similarity weight can be set to 0.4, the association coefficient weight to 0.3, and the correlation weight to 0.3 to realize comprehensive consideration of multiple factors.

[0045] A knowledge antibody feature matrix containing multiple feature patterns is established using node feature vectors and inter-node connection weights, historical scene data is collected, each scene containing multiple nodes and their features and relationships; cluster analysis is performed on the historical scenes to identify typical scene patterns; for each type of scene, representative features are extracted to form feature patterns; all feature patterns are organized in 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.

[0046] In the intelligent medical scene, 20 typical disease patterns are identified by analyzing 10,000 historical cases, each pattern is represented by a 300-dimensional feature vector, forming a 20x300 knowledge antibody feature matrix. For example, the first row in the matrix represents the respiratory system disease pattern, and the columns with higher values correspond to respiratory symptoms features; the second row represents the digestive system disease pattern, and the columns with higher values correspond to abdominal symptoms features.

[0047] Based on the knowledge antibody feature matrix, the matching coefficients between node feature vectors and feature patterns are calculated, and the fused feature vector of the node is obtained; the feature vector is multiplied by each row of the knowledge antibody feature matrix to obtain the matching scores; the matching scores are normalized by applying the softmax function to obtain the matching coefficients of the node and each feature pattern.

[0048] In practical applications, assuming that the dot product result of the fused feature vector of a patient node and the 20 feature patterns of the knowledge antibody feature matrix is [0.65, 0.12, 0.08, 0.05, 0.03,...], and the normalized matching coefficients are [0.73, 0.09, 0.06, 0.04, 0.02,...], it indicates that the patient's features are most matched with the first disease pattern.

[0049] According to the inter-node connection weights and the matching coefficients, the extended weights are calculated, for each edge in the knowledge graph, the matching coefficients of the two connected nodes are obtained; the weighted average of the two node matching coefficients is calculated, the weight can be set according to the node importance; the extended weight of the edge is calculated by combining the connection weight of the edge and the weighted average of the node matching coefficients; 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.

[0050] In specific implementation, the connection weight can be set to 60% and the matching coefficient to 40% to calculate the extended weight of the edge. For example, for the edge connecting the patient node and the disease node, the connection weight is 0.85, and the weighted average of the two node matching coefficients is 0.7, then the extended weight of the edge is calculated as 0.85x0.6+0.7x0.4=0.79.

[0051] The initial knowledge graph is dynamically expanded based on the expansion weight, an expansion threshold is set, and nodes and edges with expansion weight higher than the threshold are screened; additional information related to the screened nodes is retrieved from an external knowledge base; new nodes are created to represent the additional information, and connection relationships between the new nodes and the original nodes are established; the matching coefficients of the new nodes and the expansion weights of the new edges are calculated; and the new nodes and new edges that meet the conditions are added to the knowledge graph to form an expanded knowledge graph.

[0052] In the intelligent medical scenario, the initial knowledge graph contains 200 nodes and 500 edges. Set the expansion threshold to 0.7, and select the nodes with expansion weight greater than 0.7, including specific patient, disease and symptom nodes. Retrieve related information of these nodes from the medical knowledge base, such as treatment plan of disease, drug information, complications, etc. Create new nodes to represent this information and establish connections with the original nodes. After calculation and screening, finally add 100 new nodes and 300 new edges to the knowledge graph to form the expanded knowledge graph.

[0053] Based on the expansion weight, the current scene features are extracted from the expanded knowledge graph, and the center node related to the current scene is determined, such as the current patient node; the graph attention network is applied, and the neighbor node information within a certain hop is aggregated from the center node; in the aggregation process, the weight of the neighbor node is determined by the expansion weight of the edge; the weighted average of each node representation obtained by aggregation is calculated, and the weight is the expansion weight of the node; and the weighted average result is taken as the current scene feature.

[0054] In practical application, taking a patient node as the center, the neighbor node information within 2 hops is aggregated, including symptom, examination result, disease, drug, etc. Assuming that there are 30 neighbor nodes, each node is represented by a 150-dimensional vector, and the context representation of 30 nodes is obtained after aggregation by the graph attention network. The 150-dimensional current scene feature vector is obtained by weighted average according to the node expansion weight.

[0055] The current scene feature, the historical feature and the relationship feature in the expanded knowledge graph are combined to construct a scene feature vector, and the historical scene related to the current scene is retrieved from the knowledge graph; the feature representation of the historical scene is extracted; the relationship feature between the current scene and the historical scene is extracted, such as time difference, environment difference, etc.; and the current scene feature, the historical feature and the relationship feature are spliced to form a complete scene feature vector.

[0056] In the intelligent medical scenario, the current patient scene feature is a 150-dimensional vector, the related historical patient scene feature is also a 150-dimensional vector, and the relationship feature is a 50-dimensional vector. After splicing, a 350-dimensional complete scene feature vector is obtained, which comprehensively describes the current patient scene and its association with the historical scene.

[0057] The similarity between the scene feature vector and the historical scene feature vector is calculated using the extended weight, and a complete scene feature vector is obtained; 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, and the weight is the extended weight of the feature dimension; the similarity is sorted to determine the most similar historical scene set.

[0058] In practical applications, assume that the knowledge graph stores 5000 historical medical scenes, each represented by a 350-dimensional feature vector. The similarity between the current scene and the 5000 historical scenes is calculated, resulting in a range of 0 to 1. After sorting, the top 10 historical scenes have similarities of [0.95, 0.92, 0.88, 0.86, 0.83, 0.81, 0.79, 0.77, 0.75, 0.72], indicating that these historical scenes are highly similar to the current scene.

[0059] Similar historical scenes are identified from the extended knowledge graph based on similarity, a similarity threshold is set, and historical scenes with a similarity higher than the threshold are filtered; the complete information of the filtered historical scenes is extracted from the knowledge graph, including node features, relationship features, and decision information; the similar historical scenes are sorted by similarity to form a similar scene set.

[0060] In the intelligent medical scene, set the similarity threshold to 0.8, and filter out 5 historical scenes with a similarity greater than 0.8 from the 5000 historical scenes. Extract the complete information of these 5 scenes, including patient features, disease features, treatment plans, etc. Sort by similarity to form a similar medical scene set to provide a reference for subsequent decision-making.

[0061] From the similar historical scenes, the corresponding decision features are extracted based on the matching coefficient. For each similar historical scene, extract its decision features, such as treatment plans, drug selection, follow-up plans, etc.; combine the similarity and matching coefficient of the similar historical scenes to calculate the weight of the decision features; and weight the decision features of each historical scene based on the weight to obtain the comprehensive decision features.

[0062] In practical applications, the decision features of the 5 similar historical scenes are each 100-dimensional vectors, the similarities are [0.95, 0.92, 0.88, 0.86, 0.83], and the matching coefficients are [0.85, 0.82, 0.79, 0.76, 0.72]. The decision weight of each scene is calculated as the product of the similarity and the matching coefficient normalized to [0.25, 0.23, 0.21, 0.20, 0.11]. According to the decision weight, the 5 decision features are weighted and averaged to obtain a 100-dimensional comprehensive decision feature.

[0063] The decision features are fused with the scene feature vector to obtain a decision weight, the scene feature vector is input into an attention network to generate an attention weight, each dimension of the decision features is weighted according to the attention weight to highlight important decision factors in the current scene, and the weighted decision features are normalized to obtain a decision weight vector representing the priority of each decision option.

[0064] In the intelligent medical scene, a 350-dimensional scene feature vector is input into an attention network to generate a 100-dimensional attention weight vector representing the attention degree of each dimension of the decision features. The attention weight is multiplied by the 100-dimensional decision features to obtain weighted decision features. After normalization, a decision weight vector is obtained, indicating the priority of each treatment plan.

[0065] The decision weight and the weighted function value of the matching coefficient are 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 by linear combination, and the weight is set according to the specific application. The weighted function value is used as a query vector to calculate the similarity with each row of the knowledge antibody feature matrix. The most matching feature mode is identified according to the similarity. The corresponding historical decision information is extracted from the feature mode.

[0066] In actual application, the decision weight accounts for 70%, and the matching coefficient accounts for 30%. The weighted function value is calculated. The value is used to calculate the similarity with the 20 feature modes of the knowledge antibody feature matrix to obtain a similarity array. Assuming that the highest similarity is 0.91, corresponding to the 3rd feature mode, indicating that the historical decision information under this mode is most suitable for the current scene.

[0067] Based on the feature matching result, the historical decision information related to the current scene is obtained. From the most matching feature mode, the corresponding historical decision record is retrieved. The decision content, decision reason, execution method and effect evaluation information are extracted. The historical decision information is adaptively adjusted according to the current scene characteristics. A decision suggestion report is generated, including recommended decision, reason explanation and expected effect.

[0068] In the intelligent medical scene, 10 historical decision records are retrieved from the most matching feature mode, each record containing diagnosis results, treatment plans, drug recommendations and prognosis. The core content of these records is extracted, such as drug A dose of 10 mg, three times a day; treatment lasts for 7 days; combined with physiotherapy twice a week, etc. The drug dose and treatment frequency are adjusted according to the special circumstances of the current patient (such as age, weight, liver and kidney function, etc.). Finally, a decision suggestion report is generated, recommending the most suitable personalized treatment plan for the current patient.

[0069] 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 (devices, components, fault types, etc.) and 800 edges (composition relationships, fault associations, etc.). When an abnormal vibration of a key device is detected, the system executes the above process: introduce the current scene features (vibration frequency, device load, environment temperature, etc.) to calculate the connection weights between nodes; build a knowledge antibody feature matrix containing 30 fault modes; calculate the matching coefficients of the device nodes and the feature modes, and find that the matching degree with the bearing wear mode is as high as 0.87; based on the extended weights, the knowledge graph is expanded to 500 nodes and 1200 edges, and historical maintenance records, spare parts information, etc. are added; extract the current scene features and compare them with the historical scenes to identify 5 historical cases with a similarity of more than 0.85; extract the decision features from these cases, and get the decision weight after fusion; finally, match to the best maintenance scheme: replace the specific model bearing and check the lubrication system at the same time, with a predicted maintenance time of 4 hours, which can avoid a major failure after 72 hours. The scheme is highly consistent with the historical successful cases, and the expected effect is reliable.

[0070] In an optional implementation, the scene features and the historical decision information are jointly input into a reasoning module of the knowledge graph to generate a decision scheme set, and an execution effect score of the decision scheme set is calculated, including: The scene features and the historical decision are input into a reasoning module of the knowledge graph, a nonlinear state equation set is constructed in the reasoning module of the knowledge graph, a system evolution trajectory is obtained by solving the nonlinear state equation set, and dynamic evolution features of the knowledge graph are extracted based on a state sequence in the system evolution trajectory; The reasoning module of the knowledge graph reasons the scene features and the historical decision information based on a time sequence variation law of the dynamic evolution features to generate a plurality of candidate decision schemes and form a decision scheme set; A gradient vector of the system evolution trajectory is calculated, a plurality of control points are identified from the gradient vector, the control points are taken as reference points for score calculation, and an evaluation value corresponding to each candidate decision scheme in the decision scheme set is calculated based on state information of the control points; A neuron activation intensity corresponding to the evaluation value is calculated, a plasticity connection matrix representing dynamic association 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, and an execution effect score of the candidate decision scheme is calculated based on the network resonance state.

[0071] As shown in Figure 2 The method includes: The scene features and historical decisions are taken as inputs of the reasoning module of the knowledge graph, a nonlinear state equation set is constructed in the reasoning module of the knowledge graph, the scene features are extracted as an n-dimensional vector F, each component representing a specific scene feature value; the historical decision information is encoded as an m-dimensional vector D, each component representing a specific decision parameter value; the two vectors are spliced to form a complete input vector I; in the knowledge graph reasoning module, a state variable vector S is defined to represent the state value of the key node; a nonlinear state equation set is constructed to describe the relationship between the rate of change of the state variable with time, the current state, the input vector and time.

[0072] In the intelligent transportation scenario, the scene feature vector contains 20 components, such as the current road traffic volume (250 vehicles / hour), the average vehicle speed (45 km / h), the weather condition code (sunny day = 1, current value = 1), the time period code (late peak = 3, current value = 3), etc.; the historical decision vector contains 15 components, such as the signal timing scheme parameters (red light duration = 60 seconds, green light duration = 40 seconds), the 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 state of each key node in the road network, such as the intersection queue length (35 meters), the traffic efficiency (0.75), the safety risk index (0.15), etc.

[0073] The system evolution trajectory is obtained by solving the nonlinear state equation set, the dynamic evolution features of the knowledge graph are extracted based on the state sequence in the system evolution trajectory, the initial state is set to correspond to the current actual observed system state; a suitable numerical integration method is selected, such as the fourth-order Runge-Kutta method, to solve the nonlinear state equation set; the integration step size and the total simulation time are set, the state values at each time are calculated to obtain the state sequence; the state sequence is organized in matrix form; the first-order difference and the second-order difference of the state variable are calculated; the statistical features and the time domain features of the state sequence are extracted; and the above features are combined to form a dynamic evolution feature vector.

[0074] In the intelligent transportation scenario, the initial state is set to the current measured traffic state. The step size h = 10 seconds and the total simulation time T = 3600 seconds (1 hour) are selected, the fourth-order Runge-Kutta method is used to solve the state equation set, and a state sequence matrix containing 360 rows (time points) and 25 columns (state variables) is obtained. The mean value of the queue length is 28 meters, the standard deviation is 12 meters, the maximum value is 52 meters, and the minimum value is 10 meters; the mean value of the traffic efficiency is 0.82, the standard deviation is 0.08, the maximum value is 0.95, and the minimum value is 0.65. These features together form a dynamic evolution feature vector with a dimension of 250.

[0075] The reasoning module of the knowledge graph infers the scene features and historical decision information based on the time sequence variation law of the dynamic evolution characteristics, generates multiple candidate decision schemes and forms a decision scheme set, defines a target state to represent an ideal state expected to be achieved, constructs a state prediction model based on the dynamic evolution characteristic vector, defines a decision parameter space containing all possible decision parameter combinations, generates a series of candidate decision parameter combinations in the decision parameter space, predicts the terminal state of the system under each candidate decision parameter combination, calculates the deviation between the terminal state and the target state, and selects the top N candidate decision parameter combinations with smaller deviations to form the decision scheme set.

[0076] In the intelligent traffic scenario, the target state is set to a queue length of no more than 20 meters, a traffic efficiency of no less than 0.9, and a safety risk index of no more than 0.1. Based on the extracted dynamic evolution characteristics, a prediction model is constructed. The decision parameter space includes all combinations of signal timing schemes and lane allocation strategies. Through searching the parameter space, 200 candidate decision parameter combinations are generated, and the top 8 decision schemes with the smallest deviations are selected to form the decision scheme set. For example, scheme 1 is a red light of 65 seconds, a green light of 45 seconds, 2 straight lanes, and 2 left turn lanes; scheme 2 is a red light of 55 seconds, a green light of 55 seconds, 3 straight lanes, and 1 left turn lane.

[0077] A gradient vector of the system evolution trajectory is calculated, a plurality of control points are identified from the gradient vector, the control points are used as reference points for score calculation, the derivative (gradient) of each state variable in the system evolution trajectory with respect to time is calculated to form a gradient vector, the norm of the gradient vector is calculated, local maximum and minimum points of the norm of the gradient vector are found, and time points of sign changes of components of the gradient vector are found. Based on the above feature points, M key time points are selected as control points.

[0078] In the intelligent traffic scenario, the gradient vector of the 1-hour evolution trajectory is calculated, 5 local maximum points of the norm are identified, located at the 28th minute, the 35th minute, the 42nd minute, the 49th minute, and the 56th minute, respectively, 3 local minimum points and time points of sign changes of gradients of key state variables are identified. Considering these feature points comprehensively, 6 control points are selected, located at the 15th minute, the 28th minute, the 35th minute, the 42nd minute, the 49th minute, and the 56th minute, respectively.

[0079] Based on the state information of each candidate decision scheme in the decision scheme set at the control points, the corresponding evaluation values are calculated, the state of the system at the control points is calculated for each decision scheme in the decision scheme set, a plurality of evaluation indexes such as state deviation, resource consumption, and operation complexity are defined, weights are assigned to each evaluation index, and the comprehensive evaluation value at each control point is calculated.

[0080] In the intelligent transportation scenario, the evaluation indicators include the deviation of the queue length from the target value (20 meters) (weight 0.3), the deviation of the traffic efficiency from the target value (0.9) (weight 0.3), the deviation of the safety risk index from the target value (0.1) (weight 0.2), and the signal light adjustment frequency (weight 0.1) and the lane change frequency (weight 0.1). Taking scheme 1 as an example, at the 15th minute control point, the comprehensive evaluation value is calculated as 0.563. Similarly, the evaluation values of scheme 1 at other control points are calculated to obtain the evaluation value vector [0.563, 0.482, 0.415, 0.387, 0.402, 0.425].

[0081] Based on the corresponding evaluation values, the corresponding neuron activation intensity is calculated, a nonlinear activation function such as the Sigmoid function is defined, the evaluation values are normalized, the corresponding neuron activation intensity is calculated, and the neuron activation intensity vector of the decision scheme at all control points is obtained.

[0082] In the intelligent transportation scenario, the Sigmoid function is selected as the activation function, and the parameters a = 5 and β = 0.5. For the evaluation value 0.563 of scheme 1 at the 15th minute control point, the neuron activation intensity is calculated as 0.73. Similarly, the neuron activation intensities of scheme 1 at other control points are calculated to obtain the activation intensity vector [0.73, 0.79, 0.83, 0.85, 0.84, 0.82]. The activation intensity vector of scheme 3 is [0.82, 0.85, 0.88, 0.90, 0.89, 0.87].

[0083] A plastic connection matrix representing the dynamic association strength between control points is constructed, an M × M matrix is created to represent the connection strength between M control points, the time interval between any two control points is calculated, the correlation coefficient of the state vector at the control point is calculated, the connection strength calculation formula is defined considering the decay effect of the time interval, and the matrix is normalized.

[0084] In the intelligent transportation scenario, a 6 × 6 plastic connection matrix is formed between the 6 control points, the Pearson correlation coefficient of the state vectors at two control points is calculated, such as the correlation coefficient between the first and second control points is 0.85. Assuming that the decay coefficient γ = 0.02, the connection strength W12 = 0.85 × exp(-0.02 × 13) = 0.65. The connection strength of all control point pairs is calculated to obtain the complete plastic connection matrix, and the first row is [0, 0.65, 0.48, 0.32, 0.25, 0.18]. The matrix is normalized to ensure that the sum of each row is 1.

[0085] The neuron activation intensity is iteratively calculated with the plasticity connection matrix to obtain a stable network resonance state, the initial state is set as the neuron activation intensity vector of the decision scheme; the iteration update rule is defined; the iteration termination condition is set; the iteration is calculated until the termination condition is reached, and the final activation intensity vector is obtained.

[0086] 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]. Set 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].

[0087] Based on the network resonance state, the execution effect score of the candidate decision scheme is calculated, and the importance weight is 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; and the scores of all candidate decision schemes are sorted.

[0088] In the intelligent transportation scenario, the importance weights of the 6 control points are set to [0.1, 0.15, 0.2, 0.25, 0.2, 0.1]. For scheme 1, the network resonance state is [0.95, 0.96, 0.96, 0.95, 0.94, 0.93], and the execution effect 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. The execution effect of scheme 3 is the best, with a score of 0.978.

[0089] In the intelligent manufacturing scenario, the method is applied to optimize the production line scheduling, and 10 schemes are selected from 200 candidate decision schemes for final evaluation. By calculating the neuron activation intensity, constructing the plasticity connection matrix and iterating to the network resonance state, the execution effect scores of each scheme are obtained. The highest score scheme is 0.985, and its core strategy is: first small batch trial production for 2 hours, adjust parameters according to quality feedback; use maximum capacity for 4 hours in the middle stage; and gradually reduce the production capacity while strengthening quality inspection in the last 2 hours. This scheme is expected to improve production efficiency by 15%, reduce defective rate by 8%, and reduce comprehensive cost by 12%.

[0090] In the intelligent logistics scenario, the method is applied to optimize the distribution route, and through network resonance calculation, a scheme with a highest score of 0.967 is obtained, and the core strategy is: adopting a regional block distribution strategy, using dynamic path planning in each block region; avoiding congested road sections during peak hours; and adopting a time window clustering mode for the distribution points. The scheme is expected to reduce the distribution time by 18%, reduce the fuel consumption by 15%, and improve the customer satisfaction by 12%.

[0091] By the above method, the scene features and historical decision information are input into the knowledge graph reasoning module to generate a complete technical solution of a decision scheme set and calculating the execution effect score. The method combines the nonlinear system theory and neural network dynamics organically to establish an end-to-end solution from scene features to decision score, and provides a scientific basis and technical support for intelligent decision-making.

[0092] In an optional implementation, the knowledge graph reasoning module is used to infer the scene features and the historical decision information based on the time sequence variation law of the dynamic evolution features, to generate a plurality of candidate decision schemes and form a decision scheme set, which includes: The scene features and the historical decision information are constructed into a feature vector representing time sequence characteristics, time sequence evolution analysis is performed on the feature vector to obtain an evolution law of the feature vector, and dynamic variation features are extracted from the evolution law; The feature vector is input into the knowledge graph reasoning module for feature space conversion to obtain a converted feature expression, a stable region of the converted feature expression is determined according to the evolution law of the feature vector, a plurality of feature parameters representing system stability are obtained by constraining the converted feature expression using the stable region; The knowledge graph reasoning module is used to construct a feature weight distribution based on the feature parameters, and a variation trend of the converted feature expression is calculated according to the feature weight distribution; The dynamic variation features are input into the knowledge graph reasoning module, the converted feature expression is inferred based on the variation trend to obtain an inference result, and a plurality of candidate decision schemes are generated according to the inference result under the constraint of the stable region; Evaluation indexes including scheme reliability, scheme diversity and information gain are calculated for the candidate decision schemes, a dynamically changing screening threshold is determined based on the evolution law of the feature vector, the evaluation indexes are compared with the screening threshold, and the candidate decision schemes satisfying the screening threshold are selected to form a decision scheme set.

[0093] As shown in Figure 3 The method includes: The scene features and historical decision information are constructed as feature vectors representing timing characteristics, static attributes and dynamic attributes in the scene features are extracted, the static attributes include environmental parameters, resource states, constraint conditions, etc., the dynamic attributes include time series data, state change rate, evolution trend, etc.; decision parameters, execution timing, effect evaluation and other key data in the historical decision information are extracted; the static attributes are normalized to ensure comparability of features with different dimensions; the dynamic attributes are time-series sampled to retain data at key time points; the processed static attributes and dynamic attributes are organized into feature vectors according to a predefined structure.

[0094] In the smart grid load scheduling scenario, the scene features include static attributes such as grid topology, substation capacity, line load, and dynamic attributes such as load curve, voltage fluctuation, power factor; the historical decision information includes power generation scheduling strategy, load transfer operation and emergency response measures in the past 24 hours. Static attributes such as substation capacity (original value 500MVA) are normalized to 0.75, and line load rate (original value 65%) is normalized to 0.65; dynamic attributes such as load curve are sampled at 30-minute intervals in the past 24 hours to obtain 48 time point data; these processed features are organized into a 256-dimensional feature vector, of which the first 64 dimensions represent static attributes and the last 192 dimensions represent dynamic attributes.

[0095] The evolution law of the feature vector is obtained by time series evolution analysis of the feature vector, the dynamic change feature is extracted from the evolution law, the feature vectors at multiple time points are collected to form a time series data set; the difference between adjacent time point feature vectors is calculated to obtain the change rate; the change rate is averaged by sliding window to reduce noise influence; the mutation points of the change rate are identified, which usually correspond to the transition of system state; the autocorrelation coefficient of each dimension of the feature vector is calculated to identify the periodic change mode; the evolution law model is constructed according to the change rate, mutation point, periodicity and frequency characteristics; the key features representing the dynamic behavior of the system are extracted from the evolution law model to form the dynamic change feature.

[0096] In the smart grid load scheduling scenario, the feature vectors of the past 7 days are collected, with 48 time points per day, totaling 336 samples. The difference between adjacent time points is calculated, and a 5-length sliding window is applied to average the change rate. It is identified that the power grid load has obvious mutation points near 07:30, 12:00 and 18:30 every day, corresponding to the morning peak, midday peak and evening peak. The autocorrelation coefficients of each dimension of the feature vector are calculated, and it is found that the load curve has a clear 24-hour periodicity, with a correlation coefficient of 0.92 at a 24-hour delay. According to these analyses, a load evolution model is constructed, and 8 key dynamic change features are extracted, including the intra-day load growth rate (maximum 0.08 / hour), peak duration (average 3.5 hours), peak-to-valley ratio (average 1.8), etc.

[0097] The feature vector is input into the reasoning module of the knowledge graph for feature space conversion to obtain the converted feature expression. The feature mapping layer of the knowledge graph is constructed, which includes multiple nonlinear transformation units. The feature vector is input into the feature mapping layer to obtain high-dimensional feature expression through nonlinear transformation. Dimensionality reduction techniques are applied to project the high-dimensional feature expression into a low-dimensional semantic space. In the semantic space, similar feature vectors are mapped to similar positions, forming a clustering structure. The topological relationship and distance information in the semantic space are extracted to construct the converted feature expression.

[0098] In the smart grid load scheduling scenario, the feature mapping layer of the knowledge graph includes three layers of nonlinear transformation units with node numbers of 512, 256 and 128. The 256-dimensional feature vector is input into the feature mapping layer to obtain a 128-dimensional high-dimensional feature expression through nonlinear transformation. The t-SNE dimensionality reduction technique is applied to project the 128-dimensional feature expression into a 3-dimensional semantic space. In the semantic space, the feature vectors of weekdays form a cluster with a coordinate center of approximately (0.4, 0.6, 0.5); the feature vectors of weekends form another cluster with a coordinate center of approximately (-0.3, -0.4, 0.2); and the feature vectors under special weather conditions form a third cluster with a coordinate center of approximately (0.1, -0.2, -0.6). The topological relationship and internal structure of these clusters are extracted to construct a 64-dimensional converted feature expression.

[0099] The stable region of the converted feature expression is determined according to the evolution rule of the feature vector, the converted feature expression is constrained by using the stable region, a plurality of feature parameters representing system stability are obtained, and the change range and change mode of the feature vector in the time sequence evolution process are analyzed; in the converted feature space, the distribution region of the historical data is marked; the region of the system stable operation in the historical data is identified, and is defined as a stable region; the geometric characteristics of the stable region, such as the center point, the boundary and the shape parameter, are calculated; the constraint condition is constructed according to the geometric characteristics of the stable region, and the reasonable range of the feature expression is limited; the converted feature expression is projected into the constraint condition, and the feature expression meeting the stability requirement is obtained; and the key parameters are extracted from the constrained feature expression as the feature parameters representing system stability.

[0100] In the intelligent power grid load scheduling scene, the change range of the feature vector in the past 7 days is analyzed, and the load fluctuation is between 45% and 75% of the total capacity. In the converted 3-dimensional semantic space, the distribution of 336 historical data points is marked. The region of the system stable operation is identified, which is approximately an ellipsoid with the center point coordinates (0.2, 0.1, 0.0) and the semi-axis length (0.5, 0.4, 0.3). The constraint condition is constructed based on the ellipsoid, and it is required that the new feature expression must fall within the ellipsoid or be no more than 0.1 units away from the surface. The current feature expression (0.3, 0.2, -0.1) is projected into the constraint condition, and the feature expression (0.3, 0.2, -0.1) meeting the stability requirement is obtained, which is already in the stable region. Five key parameters are extracted from the constrained feature expression: the distance to the center point (0.14), the coordinates projected onto the principal axis (0.25, 0.15, -0.10) and the stability margin (0.36), which represent the stability state of the system.

[0101] The reasoning module of the knowledge graph constructs a feature weight distribution based on the feature parameters, calculates the change trend of the converted feature expression according to the feature weight distribution, inputs the feature parameters into the weight generation network of the knowledge graph; in the weight generation network, the importance score of each dimension is calculated according to the value of the feature parameter; the importance score is normalized to obtain the feature weight distribution; the converted feature expression is weighted according to the feature weight distribution; the gradient direction of the weighted feature expression is calculated, which represents the most likely change direction of the system; in combination with the observed change mode in the historical data, the position of the feature expression at the future time point is predicted; the change trend, including the change direction, the rate and the acceleration, is extracted from the predicted position sequence.

[0102] In the smart grid load scheduling scenario, five feature parameters are input into the weight generation network, which contains two layers with node numbers of 32 and 64 respectively. According to the values of the feature parameters, the importance scores of each dimension of the 64-dimensional feature representation are calculated. The scores are normalized by softmax to obtain the feature weight distribution, and the weight values range from 0.005 to 0.042. According to the weight distribution, the feature representation is weighted to highlight the influence of important dimensions. The gradient direction of the weighted feature representation in the 3-dimensional semantic space is calculated to obtain the vector (0.05, 0.03, -0.02), which represents the most likely change direction of the system. Combined with the change pattern in the historical data, the position sequence of the feature representation in the next 6 hours is predicted. From these positions, the change trend is extracted: the direction remains relatively stable, and the rate gradually increases from the initial 0.017 / hour to the final 0.057 / hour, indicating that the system load is growing at an accelerating rate.

[0103] The dynamic change feature is input into the reasoning module of the knowledge graph, and the reasoning result is obtained by reasoning the converted feature representation based on the change trend. Under the constraint of the stable region, multiple candidate decision schemes are generated based on the reasoning result. The extracted dynamic change feature and the converted feature representation are combined to form complete input information. In the reasoning module of the knowledge graph, the knowledge nodes related to the current scene are activated. Based on the activated knowledge nodes and the input information, the reasoning rules are applied for forward deduction. Considering the change trend of the system, the state changes of the system that may be caused by different decisions are predicted. The predicted state changes are evaluated, and the state transition paths that meet the target are selected. Based on each valid state transition path, the corresponding decision scheme is generated. It is checked whether the decision scheme will cause the system to leave the stable region, and the schemes that do not meet the stability constraint are eliminated. The decision schemes that meet all the constraint conditions are retained to form the candidate decision scheme set.

[0104] In the smart grid load scheduling scenario, 8 dynamic features are combined with 64-dimensional feature expression to input the inference module of the knowledge graph. In the knowledge graph, 150 knowledge nodes related to power load scheduling, energy distribution and stability control are activated. Based on these knowledge nodes and input information, expert rules of power system scheduling are applied for reasoning. Considering the accelerating trend of load growth, the state changes of the system in the next 3 hours under different scheduling strategies are predicted. 15 possible scheduling strategies are generated, such as increasing the output of conventional generator units, starting standby power sources, reducing non-critical loads, etc. The effectiveness of each strategy is evaluated, and 10 effective state transition paths are selected, corresponding to 10 candidate decision schemes. It is checked whether these schemes will cause the system to leave the stable region, and it is found that 2 schemes may cause line overload under high load conditions, pushing the system out of the stable region, so these 2 schemes are excluded. Finally, 8 candidate decision schemes that meet all constraint conditions are retained, such as scheme 1 "increase the output of No. 3 generator unit by 30 MW, and reduce the non-critical load in B area by 15 MW", scheme 2 "start standby gas turbine to provide 40 MW emergency capacity", etc.

[0105] For the candidate decision schemes, evaluation indexes including scheme reliability, scheme diversity and information gain are calculated, the screening threshold of dynamic change is determined based on the evolution law of feature vector, the evaluation indexes are compared with the screening threshold, the candidate decision schemes that meet the screening threshold are selected to constitute the decision scheme set, for each candidate decision scheme, its reliability is evaluated based on historical data and expert knowledge, and a reliability score is calculated; the similarity between decision schemes is analyzed, the diversity index is calculated to ensure that the scheme set has sufficient coverage; the information gain of each decision scheme is calculated to evaluate its contribution to the current decision problem; the comprehensive evaluation index is calculated by combining reliability, diversity and information gain; according to the characteristics of state change, a dynamically adjusted screening threshold is set, so that the threshold changes adaptively with the system state; the evaluation indexes of each decision scheme are compared with the current screening threshold; the decision schemes whose evaluation indexes exceed the screening threshold are selected to constitute the final decision scheme set.

[0106] In the smart grid load scheduling scenario, eight candidate decision schemes are evaluated. Based on historical data and power system expert knowledge, the reliability score of each scheme is evaluated, such as scheme 1 score 0.85, scheme 2 score 0.78. The cosine similarity matrix between schemes is calculated, the similarity between scheme 1 and scheme 3 is 0.72, indicating that the two schemes are similar; the similarity between scheme 2 and scheme 4 is 0.35, indicating that the two schemes are quite different. The information gain of each scheme is calculated, scheme 1 is 0.42, scheme 2 is 0.56. Comprehensive three indicators, the comprehensive evaluation index of scheme 1 is 0.67, and the comprehensive evaluation index of scheme 2 is 0.73. According to the system state, set the reliability threshold to 0.75, the diversity threshold to 0.4, the information gain threshold to 0.45, and the comprehensive evaluation threshold to 0.65. Compare the evaluation index of each scheme with the threshold, finally select scheme 1, scheme 2 and scheme 7 to form the decision scheme set, which are "increase the output of No. 3 generator set by 30MW, and reduce the non-critical load in B area by 15MW", "start the standby gas turbine to provide 40MW emergency capacity" and "adjust the power grid topology and transfer the load from the high load area to the low load area".

[0107] Through the above method, the technical goal of using the reasoning module of the knowledge graph to reason the scene characteristics and historical decision information based on the time sequence change rule of the dynamic evolution characteristics, generating multiple candidate decision schemes and forming a decision scheme set is achieved. This method has wide application prospects in complex system decision-making scenarios such as smart grid load scheduling, and can effectively improve the scientificity and adaptability of decision-making.

[0108] In an optional implementation, an optimal decision scheme is selected based on the execution effect score, and an execution result of the optimal decision scheme is used as a reward signal of a reinforcement learning algorithm, and the reasoning module comprises: A decision ranking value is constructed based on the result of the execution effect score of the decision scheme, a phase difference of multiple neurons is calculated based on the decision ranking value to construct a phase synchronization matrix, different oscillation mode characteristics are cooperatively mapped using the phase synchronization matrix to obtain a neural oscillation mode prediction value corresponding to the decision scheme, and the decision ranking value is cooperatively mapped with the neural oscillation mode prediction value through the phase synchronization matrix to obtain a decision comprehensive evaluation value, and an optimal decision scheme is selected based on the decision comprehensive evaluation value; An immediate reward value is constructed by comparing the actual execution result of the optimal decision scheme with the decision comprehensive evaluation value, a neural oscillation synchronization change value is calculated based on the phase synchronization matrix, the immediate reward value is cooperatively mapped with the neural oscillation synchronization change value through the phase synchronization matrix to obtain a compensation reward value, and the compensation reward value is time-series accumulated to obtain a cumulative reward value, and the cumulative reward value is input into a reinforcement learning algorithm as a reward signal; Based on the reinforcement learning algorithm, a gradient value of an inference module parameter is calculated, a constraint difference value between the inference module parameter and a target neural oscillation mode is calculated according to the phase synchronization matrix, and the inference module is dynamically adjusted by collaborative mapping of the gradient value and a historical update value of the inference module parameter.

[0109] A decision ranking value is constructed based on the result of the execution effect score of the decision scheme, the execution effect scores of all candidate decision schemes are collected, and the scores are ranked from high to low; the ranked scores are normalized so that the highest score corresponds to a normalized value of 1 and the lowest score corresponds to a normalized value of 0; each decision scheme is assigned a ranking weight, and the decision schemes with higher ranking obtain higher weights; the normalized value of the score is multiplied by the ranking weight to obtain a weighted score; the weighted score is converted into a decision ranking value to ensure that the decision ranking value reflects the relative advantages and disadvantages of the decision scheme.

[0110] In the intelligent medical diagnosis scenario, assuming that there are 5 candidate decision schemes with execution effect scores of 0.92, 0.87, 0.83, 0.76 and 0.65. The scores are normalized to obtain normalized values of 1.00, 0.81, 0.67, 0.41 and 0.00 respectively. The ranking weights are assigned as [0.40, 0.30, 0.15, 0.10, 0.05], and the weighted scores are calculated as 0.40, 0.24, 0.10, 0.04 and 0.00 respectively. The decision ranking values are converted to [1.00, 0.60, 0.25, 0.10, 0.00], which clearly reflect the relative order of the 5 decision schemes.

[0111] A phase difference matrix is constructed according to the phase difference of a plurality of neurons calculated from the decision ranking value, a group of neural oscillators is defined, each oscillator corresponds to a key variable in the decision process; the phase value of each oscillator is initialized, 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, the key variable corresponding to the decision scheme with high ranking value obtains 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 a phase synchronization strength; the phase synchronization strength is organized in matrix form to form a phase synchronization matrix.

[0112] In the intelligent medical diagnosis scenario, 8 neural oscillators are defined, corresponding to 8 key variables such as patient symptoms, vital signs, test results, image features, disease probability, treatment plan, drug selection, and prognosis prediction. The phase values of the oscillators are initialized as [0.2, 0.5, 0.3, 0.6, 0.4, 0.7, 0.1, 0.8]. According to the ranking values [1.00, 0.60, 0.25, 0.10, 0.00], adjust the phase change rate, and calculate the new phase values of the 8 oscillators as [0.8, 0.7, 0.9, 0.6, 0.5, 0.3, 0.4, 0.2]. Calculate the phase difference between any two oscillators, such as the phase difference between the first and second oscillators is 0.1. Apply the phase synchronization function to calculate the synchronization strength, such as the synchronization strength between the first and second oscillators is 0.9. Construct an 8x8 phase synchronization matrix, the first row is [1.0, 0.9, 0.7, 0.6, 0.5, 0.3, 0.4, 0.2].

[0113] Use the phase synchronization matrix to cooperatively map different oscillation mode features to obtain the neural oscillation mode prediction value corresponding to the decision scheme. Collect the observed neural oscillation mode features in the historical decision-making process; input the features of the current decision-making problem into the pattern recognition network to predict the possible oscillation mode; calculate the similarity between the predicted mode and the historical mode to identify the most matched historical mode; extract the feature parameters of the most matched historical mode; combine the historical mode features with the current phase synchronization matrix and perform feature mapping through matrix multiplication or convolution operation; extract the key features from the mapping result to form the neural oscillation mode prediction value.

[0114] In the intelligent medical diagnosis scenario, 100 typical neural oscillation modes are collected from the historical database, and each mode is described by 20 feature parameters. Input the features of the current diagnosis problem into the pattern recognition network to predict 3 possible oscillation modes, with similarity to the historical mode of 0.88, 0.72 and 0.65 respectively. Extract the feature parameters of the historical mode with the highest similarity, including alpha wave frequency (10 Hz), beta wave frequency (20 Hz), alpha wave amplitude (15 μV), beta wave amplitude (8 μV), and alpha-beta phase coupling strength (0.6). Combine these feature parameters with the current 8x8 phase synchronization matrix and perform mapping through matrix multiplication to obtain a 64-dimensional feature vector. Extract the key components from the vector to form a 12-dimensional neural oscillation mode 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].

[0115] The decision ranking value is cooperatively mapped with the neural oscillation pattern prediction value through the phase synchronization matrix to obtain a decision comprehensive evaluation value, and the decision ranking value is expressed in vector form; the decision ranking value is transformed through the phase synchronization matrix to obtain a ranking mapping value; the dot product of the ranking mapping value and the neural oscillation pattern prediction value is calculated to obtain a 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 field, a weight coefficient is set to balance the influence of the ranking factor and the oscillation mode factor; 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.

[0116] In the intelligent medical diagnosis scene, the decision ranking value [1.00, 0.60, 0.25, 0.10, 0.00] is expressed as a 5-dimensional vector. The ranking value is transformed through an 8x8 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 the mapping value and the 12-dimensional neural oscillation pattern prediction value is calculated to obtain a preliminary evaluation score of 3.25. The score is adjusted by applying a Sigmoid activation function to obtain 0.96. The ranking factor weight is set to 0.6 and the oscillation mode factor weight is set to 0.4, and the weighted evaluation score is calculated as 0.6x1.0 + 0.4x0.96 = 0.984. Repeat the above calculation for all candidate schemes to obtain the decision comprehensive evaluation values of the 5 schemes as [0.984, 0.731, 0.412, 0.286, 0.153]. The evaluation value of scheme 1 is the highest, and it is selected as the optimal decision scheme.

[0117] An immediate reward value is constructed by comparing the actual execution result of the optimal decision scheme with the decision comprehensive evaluation value, the optimal decision scheme is executed, and the actual execution result is obtained; a set of evaluation indicators are defined to measure the quality of the execution result; the actual score of the execution result is calculated according to the evaluation indicators; the actual score is compared with the expected decision comprehensive evaluation value to calculate the difference between the two; a positive or negative immediate reward is given according to the difference; the absolute size of the difference is considered, the larger the difference, the greater the amplitude of the reward or punishment; the calculated reward value is limited within a predetermined range to prevent the influence of extreme values.

[0118] In the intelligent medical diagnosis scenario, the optimal decision scheme (Scheme 1) is executed, the actual execution result is obtained, and the evaluation indexes include diagnosis accuracy (0.95), treatment effectiveness (0.88), side effect degree (0.15), medical resource consumption (0.25), and patient satisfaction (0.90). According to these indexes, the actual score is calculated as 0.92. Comparing the actual score (0.92) with the expected decision comprehensive evaluation value (0.984), the difference is calculated as -0.064. Since the actual score is slightly lower than the expected value, a small negative immediate reward is given, and the immediate reward value is set as -0.05.

[0119] The specific implementation of calculating the neural oscillation synchronization change value based on the phase synchronization matrix, cooperatively mapping the immediate reward value through the phase synchronization matrix and the neural oscillation synchronization change value to obtain the compensation reward value, time-series accumulating the compensation reward value to obtain the cumulative reward value, inputting the cumulative reward value into the reinforcement learning algorithm as a reward signal, and calculating the gradient value of the inference module parameter based on the reinforcement learning algorithm is as follows.

[0120] The dynamic adjustment of the inference module is demonstrated through an intelligent logistics distribution scenario. There are 6 candidate distribution schemes in the logistics distribution decision system, and the execution effect scores are [0.88, 0.85, 0.79, 0.76, 0.72, 0.68] respectively. The decision ranking value is constructed as [1.00, 0.85, 0.55, 0.40, 0.20, 0.00]. Ten neural oscillators are defined, corresponding to distribution time, route length, vehicle quantity, loading rate, fuel consumption, traffic condition, weather influence, customer time window, distribution priority, and warehouse inventory, etc. key variables. A 10x10 phase synchronization matrix is constructed to extract the neural oscillation mode prediction value. The decision ranking value is mapped through the phase synchronization matrix, and the decision comprehensive evaluation values of each scheme are calculated as [0.925, 0.842, 0.631, 0.483, 0.296, 0.142], and Scheme 1 is selected as the optimal scheme.

[0121] After the optimal solution is executed, the actual score is 0.90, which is -0.025 different from the expected 0.925, and the immediate reward value is -0.02. The neural oscillation synchronization change value is calculated as 0.65, indicating that the delivery network coordination has improved. The compensation reward value -0.008 is obtained by synergistically mapping the immediate reward value -0.02 and the synchronization change value 0.65. The compensation reward value is time-cumulative, combined with the historical reward [0.015, 0.020, -0.005, 0.010, 0.012], and the cumulative reward value 0.009 is calculated. The cumulative reward value is input into the policy gradient algorithm, and the state space includes delivery area characteristics, vehicle state and order information, and the action space includes route planning strategy, vehicle allocation strategy and time window adjustment strategy. The gradient value of the reasoning module parameter is calculated, and the delivery decision model is optimized. After 10 rounds of iterative optimization, the system's small order merging strategy weight during peak hours is improved from 0.35 to 0.48, and the delivery efficiency is improved by 12%.

[0122] In an alternative embodiment, the optimal decision scheme is optimized using the reinforcement learning algorithm based on the adjusted reasoning module to generate a new decision scheme, which includes: A chaotic feature matrix is constructed based on the output of the adjusted reasoning module, a reward value is generated according to the chaotic feature matrix, a state distribution of the decision scheme is determined based on the reward value, and a decision optimization sequence is generated based on the state distribution; The decision optimization sequence is input into the reasoning module, and the parameters of the reasoning module are iteratively updated according to the reward value to obtain updated reasoning module parameters; The decision scheme is expanded and optimized using the updated reasoning module parameters to obtain a multi-dimensional decision set, and the optimization index of each decision scheme in the multi-dimensional decision set is calculated based on the chaotic feature matrix, and a new decision scheme is selected from the multi-dimensional decision set according to the optimization index.

[0123] The optimal decision scheme is optimized using the reinforcement learning algorithm based on the adjusted reasoning module to generate a new decision scheme. In this process, a chaotic feature matrix is constructed based on the output of the adjusted reasoning module. Specifically, the vector data output by the reasoning module is feature extracted to obtain key dimension information such as decision information, environmental state, expected result, etc. Through a nonlinear mapping function, these multi-dimensional features are converted into state points in a chaotic feature space to form a feature matrix with chaotic dynamic characteristics. For example, for a decision problem of an intelligent transportation system, the traffic flow, vehicle density, weather conditions, etc. data output by the reasoning module can be converted into an 80x80 chaotic feature matrix, and each element value in the matrix is in the interval [-1, 1], representing the mapping value of the feature in the chaotic space.

[0124] When generating the reward value according to the chaotic feature matrix, a feature sensitivity evaluation method is adopted, and for each element in the chaotic feature matrix, the influence degree of the element on the decision result is calculated, and a feature with a high influence degree is given a higher weight. Through weighted summation, a comprehensive reward value is calculated. In specific implementation, a reward function can be set, which receives the chaotic feature matrix as input and outputs a scalar value as the reward of the current decision. For example, in the resource scheduling scenario, if the decision scheme can reduce resource consumption while meeting the task completion time, it will obtain a higher reward value, if the resource utilization rate is increased by 5%, the reward value is increased by 0.2; if the task completion time is shortened by 10%, the reward value is increased by 0.3.

[0125] The state distribution of the decision scheme based on the reward value is realized by probability mapping, and the reward value is converted into a probability distribution using a softmax function, so that states with high reward values have higher selection probabilities. In actual operation, the probability value of each state is calculated to form a state probability vector. For example, for five candidate states, the probability distribution after normalization is [0.15, 0.25, 0.35, 0.10, 0.15], indicating that the third state has the highest selection probability.

[0126] When generating the decision optimization sequence based on the state distribution, the Monte Carlo tree search method is adopted, starting from the initial state, and according to the state distribution, multiple samplings are performed to explore the decision path. Each sampling generates a decision sequence, and the cumulative reward of the sequence is recorded. After multiple rounds of sampling, the decision sequence with the highest cumulative reward is selected as the optimization result. In specific implementation, the sampling number can be set to 1000, and a decision sequence with a length of 20 is generated each time, from which the top 5 sequences with the highest cumulative reward values are selected to form the final decision optimization sequence.

[0127] After inputting the decision optimization sequence into the reasoning module, the parameters of the reasoning module are iteratively updated according to the reward value, the policy gradient method is adopted to calculate the gradient of the reward value with respect to the model parameters, and the gradient ascent method is used to update the parameters, so that the model parameters are adjusted in the direction that can obtain higher rewards. In the actual updating process, a parameter update rule with a learning rate of 0.01 is used, and learning rate decay is performed every 100 iterations with a decay coefficient of 0.95. In this way, the reasoning module can gradually adapt to the decision environment and generate better decision schemes.

[0128] When expanding and optimizing the decision scheme using the updated reasoning module parameters to obtain a multi-dimensional decision set, a parameter perturbation technique is used to add random perturbations of different degrees to the updated reasoning module parameters to generate multiple variant models. Each variant model generates different decision schemes for the same decision problem, thereby forming a multi-dimensional decision set. In specific implementation, 30 variant models can be generated, and the parameter perturbation range of each model is controlled within ±5% of the original parameters, ensuring that the generated decision schemes have a certain diversity while maintaining basic effectiveness.

[0129] When calculating the optimization index of each decision scheme in the multi-dimensional decision set based on the chaos characteristic matrix, the performance of the decision is evaluated in combination with multiple dimensions. The evaluation dimensions include resource consumption, time efficiency, stability, adaptability, etc. For each dimension, a corresponding evaluation function is designed to map the decision scheme to the chaos characteristic matrix and calculate its score in each dimension. By weighted combination of the dimension scores, a comprehensive optimization index is formed. For example, in the intelligent power grid load balancing decision, the resource utilization rate weight can be set to 0.4, the system stability weight to 0.3, and the load balancing degree weight to 0.3, and the comprehensive score of each decision scheme is calculated.

[0130] When selecting a new decision scheme from the multi-dimensional decision set according to the optimization index, not only the single optimal solution is considered, but also the diversity of the decision scheme. The Pareto optimal principle is used to select a group of non-dominated solutions from the multi-dimensional decision set to form a Pareto front. From the Pareto front, the most suitable scheme is selected as the new decision scheme according to the current system state and decision requirements. In specific implementation, the top 3 decision schemes with the highest comprehensive scores can be selected to form a candidate set, and then the final decision is made in combination with the specific requirements of the current system. For example, under different load conditions, different decision schemes with different emphases are selected: for light load, a scheme with resource utilization rate priority is selected; for high load, a scheme with system stability priority is selected. This dynamic selection mechanism can make the decision system better adapt to complex and changing environments.

[0131] The method further includes: The knowledge acquisition module realizes multi-source data acquisition through an interface visual configuration method. The operation and maintenance personnel can set API interface parameters on the configuration interface, including interface address, authentication information, request method, request header, and request body, etc. The uploaded documents support PDF, Word, Excel, Markdown, etc. The video materials support mainstream formats such as MP4 and AVI. When setting a timed collection task, the collection frequency can be specified, such as every hour, every day, every week, or a custom time interval. For example, a data center operation and maintenance team configures the Git repository API of the data center source code, sets to automatically pull the latest code at 2 a.m. every day, and uploads 500 historical fault solution documents and 50 fault handling video tutorials, establishing a comprehensive knowledge acquisition channel.

[0132] The knowledge content extraction module uses corresponding large model processing techniques for different types of materials. For text materials, a natural language processing large model is used to extract key information and divide long texts into knowledge segments containing complete semantics, each segment containing problem description and solution. For picture materials, an image recognition large model is used to extract text information, chart data, and visual elements in the picture and convert them into structured descriptions. For voice materials, a speech recognition large model is used to convert audio into text and label the speaker, key content, and timestamp. For video materials, key frames and explanation content are extracted by combining video frame analysis and speech recognition, and a text summary with time index is generated. In actual application, for a fault handling video containing data in the ETL task failure, the system extracts the operation steps, error log screenshot content, and engineer's oral explanation in the video to form a complete knowledge entry.

[0133] The knowledge construction and labeling module performs structured processing on the extracted knowledge. Based on a semantic similarity algorithm, the system clusters knowledge entries with a similarity greater than 0.85 to form knowledge clusters. For each knowledge entry, 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 relevance between knowledge entries and identify causal relationships, inclusion relationships, and similarity relationships. Key data is labeled, including problem type, severity, involved components, solution difficulty, and applicable scenarios. For example, for the knowledge content of a data access component, the system labels attributes such as "data source type = relational database", "error type = connection timeout", "severity = high", and "solution method = check network connection and database parameter configuration", facilitating subsequent accurate retrieval.

[0134] The knowledge system storage module uses a dual-track storage strategy. The unstructured knowledge system storage part uses a distributed document database to save source code, technical documents, fault cases, and Internet solutions in their original format, and creates full-text and vector indexes for each document. The structured knowledge system storage part uses a graph database to construct a table relationship graph and a problem solution semantic graph. The table relationship graph records the association between data tables in the data center, including field mapping, foreign key constraints, and business dependencies. The semantic graph connects problem descriptions, fault causes, solution steps, and verification methods to form a complete problem-solving knowledge network. In a certain data center environment, the system stores 5000 source code files, 2000 table structure definitions, and 1500 historical problem and solution records. Through the graph database, it establishes more than 10000 entity nodes and 30000 relationship edges, achieving efficient storage and fast retrieval of knowledge.

[0135] The problem description module provides a visual interface, supports multiple input methods, and operation and maintenance personnel can describe the problem details through a text input box; supports uploading error screenshots, log files and other auxiliary information; provides voice input function, automatically converted to text; the interface integrates form guidance, helping users fill in structured information such as problem type, occurrence time, and impact range; the system records the user's historical queries, facilitating problem tracking and correlation analysis. For example, an operation and maintenance personnel inputs "Data integration task A1023 fails continuously, error log shows HDFS connection timeout" on the interface, and uploads an error screenshot. The system integrates these information to form a complete problem description.

[0136] The knowledge retrieval module retrieves relevant knowledge based on the problem description, converts the problem description into a vector representation, and calculates the similarity to each entry in the knowledge base; adopts a hybrid retrieval strategy, combining keyword matching and semantic understanding to improve retrieval accuracy; for problems containing pictures, extract the picture content for multi-modal retrieval; combine user historical interaction records and problem solving feedback to optimize retrieval ranking; query related entities and relationships from the graph database to supplement the retrieval results. For the above HDFS connection timeout problem, the system retrieves 10 relevant knowledge entries, including 3 similar cases, 2 HDFS configuration documents and 5 related troubleshooting methods, with similarity of 0.92, 0.87, 0.85, 0.79, 0.78, 0.75, 0.72, 0.70, 0.68, 0.65.

[0137] The prompt word packaging module converts the retrieval results into prompt words that can be efficiently processed by the large model. The module analyzes the retrieved knowledge content and extracts key information points; organizes the information into structured prompts according to the preset prompt word templates; adds context information such as system environment and related component configurations; adds task instructions to clearly require the large model to provide specific solution steps; sets the output format requirements to standardize the presentation form of the solution. For example, the system packages the retrieved HDFS connection timeout related knowledge into the prompt word: "Background: Data integration task A1023 of data center reports HDFS connection timeout error. Related knowledge: 1. Common causes of HDFS connection timeout include network problems, configuration errors, and permission issues; 2. Integration task log shows that connection attempt failed after 3 times; 3. In similar cases, 80% are caused by firewall rules. Task: Based on the above information, analyze the possible causes and provide detailed troubleshooting steps, format as 'cause analysis' and'solution steps'." The large model management module realizes efficient management and application of model resources. The module supports new model configuration, including selection of a basic model type, setting of parameters, and deployment of an environment; provides model fine-tuning functions to adapt to specific domain knowledge through a small amount of data samples; supports model training to improve the professional ability of the model in the operation and maintenance field by using a large amount of labeled data; realizes unified API management of the model to provide standardized interfaces for models with different capabilities; monitors model performance indicators, including response time, accuracy, and resource occupation, to realize dynamic scheduling of the model. In actual deployment, the system configures a general large language model for text understanding and generation, a special image recognition model for screenshot analysis, and a domain expert model for solving complex technical problems, and coordinates the calling of these model resources through a unified API interface.

[0138] The multiple reasoning module performs deep analysis based on the large model and generates a solution. A preliminary solution is first generated based on the retrieved knowledge and the problem description; the completeness and applicability of the preliminary solution are evaluated to identify information that needs to be supplemented; through multiple rounds of interaction, the knowledge base is requested to supplement information, and the solution is optimized; a final solution with detailed steps and clear operations is generated; the reasoning process and evidence chain are recorded to support the explainability of the solution. For the HDFS connection timeout problem, the system first generates a preliminary solution containing network checks and configuration verification, evaluates that the firewall configuration information of a specific environment is missing, supplements the relevant knowledge through secondary retrieval, and finally gives a complete solution of "checking whether the firewall rules prevent 8020 port communication" as the primary troubleshooting step, which accurately hits the root cause of the problem.

[0139] A data engineer finds that the daily incremental synchronization task of the data warehouse has failed continuously, and the system log shows "data source connection exception". The engineer inputs the problem description and uploads the error log screenshot through the intelligent operation and maintenance assistant interface. The system vectorizes the problem and retrieves 7 similar cases and 3 relevant documents from the knowledge base, with similarity from 0.95 to 0.75. The knowledge retrieval result shows that 80% of similar problems are related to data source network connection, 15% are related to account permissions, and 5% are related to source system changes. The system encapsulates prompt words and calls the large model for multiple rounds of reasoning. After generating a preliminary solution, it is found that the configuration information of a specific data source is missing, and the relevant content is automatically supplemented from the knowledge base. The engineer executes the suggestion and finds that the problem is caused by the IP address change of the source system, which is successfully solved by updating the configuration. The entire process from problem submission to solution takes only 15 minutes, which is 80% shorter than the traditional manual troubleshooting time.

[0140] In a second aspect, the embodiment of the present application provides a data center intelligent operation and maintenance system based on a multi-modal large model, comprising: The first unit is configured to acquire multi-modal data in a data center operation and maintenance scenario, construct node features of a knowledge graph based on the multi-modal data, and extract scene features of the multi-modal data; construct node relationships of the knowledge graph based on the scene features, analyze the correlation strength between the nodes, and construct connection weights between the nodes; The second unit is configured to expand the knowledge graph by using the node features and the connection weights, match similar scenes in the knowledge graph based on the scene features, and identify historical decision information related to the current scene; input the scene features and the historical decision information into an inference module of the knowledge graph, generate a decision scheme set, and calculate execution effect scores of the decision scheme set; The third unit is configured to select an optimal decision scheme based on the execution effect scores, use an execution result of the optimal decision scheme as a reward signal of a reinforcement learning algorithm, and dynamically adjust the inference module; optimize the optimal decision scheme by using the reinforcement learning algorithm based on the adjusted inference module, and generate a new decision scheme. The fourth unit is configured to generate an operation and maintenance operation instruction by using the new decision scheme, execute the operation and maintenance operation instruction according to a preset execution process, and record an execution process and an execution result of the decision scheme.

[0141] In a third aspect, an electronic device is provided, including: a processor; a memory for storing processor-executable instructions; The processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0142] In a fourth aspect, a computer-readable storage medium is provided, which stores computer program instructions. The computer program instructions are executed by a processor to implement the method described above.

[0143] The present application can be a method, device, system and / or computer program product. The computer program product can include a computer readable storage medium having computer readable program instructions loaded thereon for implementing various aspects of the present application.

[0144] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for intelligent operation and maintenance of a data platform based on a multi-modal large model, characterized in that, The method comprises the following steps: acquiring multi-modal data in a data center operation scene, constructing node features of a knowledge graph based on the multi-modal data, and extracting scene features of the multi-modal data; constructing node relationships of the knowledge graph based on the scene features, analyzing the correlation strength between the nodes, and constructing connection weights between the nodes; extending the knowledge graph using the node features and the connection weights, matching similar scenes in the knowledge graph based on the scene features, identifying historical decision information related to the current scene, inputting the scene features and the historical decision information into an inference module of the knowledge graph, generating a decision scheme set, and calculating the execution effect score of the decision scheme set; selecting an optimal decision scheme based on the execution effect score, taking the execution result of the optimal decision scheme as a reward signal of a reinforcement learning algorithm, dynamically adjusting the inference module, optimizing the optimal decision scheme using the reinforcement learning algorithm based on the adjusted inference module, and generating a new decision scheme; generating an operation instruction based on the new decision scheme, executing the operation instruction according to a preset execution process, and recording the execution process and the execution result of the decision scheme.

2. The method of claim 1, wherein, constructing node features of a knowledge graph based on the multi-modal data, and extracting scene features of the multi-modal data; constructing node relationships of the knowledge graph based on the scene features, analyzing the correlation strength between the nodes, and constructing connection weights between the nodes, which comprises: extracting scene features of the multi-modal data, mapping the scene features to node features of a knowledge graph, and constructing node features of the knowledge graph; performing semantic correlation analysis on the node features of the knowledge graph based on the scene features, constructing connection relationships between the nodes, calculating the correlation strength between the nodes according to the connection relationships, and taking the correlation strength as the connection weights between the nodes.

3. The method of claim 1, wherein, extending the knowledge graph using the node features and the connection weights, matching similar scenes in the knowledge graph based on the scene features, and identifying historical decision information related to the current scene, which comprises: introducing a scene feature vector into a node feature vector to calculate the connection weights between the nodes, establishing a knowledge antibody feature matrix containing multiple feature modes using the node feature vector and the connection weights between the nodes; calculating a matching coefficient between the node feature vector and the feature modes in the knowledge antibody feature matrix based on the knowledge antibody feature matrix; calculating an expansion weight based on the connection weights between the nodes and the matching coefficient, dynamically expanding the initial knowledge graph based on the expansion weight to obtain an expanded knowledge graph; extracting a current scene feature from the expanded knowledge graph based on the expansion weight, and combining the current scene feature with historical features and relationship features in the expanded knowledge graph to construct a scene feature vector; The extended weight is used to calculate the similarity between the scene feature vector and the historical scene feature vector, to identify similar historical scenes from the extended knowledge graph according to the similarity, and to extract corresponding decision features from the similar historical scenes based on the matching coefficient; The decision features and the scene feature vector are fused to obtain a decision weight, and the decision weight and the weighted function value of the matching coefficient are input into the knowledge antibody feature matrix for feature matching, and the historical decision information related to the current scene is obtained based on the feature matching result.

4. The method of claim 1, wherein, The scene feature and the historical decision information are input into the reasoning module of the knowledge graph to generate a decision scheme set, and the execution effect score of the decision scheme set is calculated, including: The scene feature and the historical decision are input into the reasoning module of the knowledge graph, and a nonlinear state equation set is constructed in the reasoning module of the knowledge graph; the system evolution trajectory is obtained by solving the nonlinear state equation set, and the dynamic evolution feature of the knowledge graph is extracted based on the state sequence in the system evolution trajectory; The reasoning module of the knowledge graph is used to reason the scene feature and the historical decision information based on the time sequence variation law of the dynamic evolution feature, generate a plurality of candidate decision schemes and form a decision scheme set; The gradient vector of the system evolution trajectory is calculated, a plurality of control points are identified from the gradient vector, the control points are used as reference points for score calculation, and the corresponding evaluation value is calculated based on the state information of each candidate decision scheme in the decision scheme set at the control point; The corresponding neuron activation intensity is calculated based on the corresponding evaluation value, a plasticity connection matrix representing the dynamic correlation strength between the control points is constructed, and the neuron activation intensity and the plasticity connection matrix are iteratively calculated to obtain a stable network resonance state, and the execution effect score of the candidate decision scheme is calculated based on the network resonance state.

5. The method of claim 4, wherein, The reasoning module of the knowledge graph is used to reason the scene feature and the historical decision information based on the time sequence variation law of the dynamic evolution feature, generate a plurality of candidate decision schemes and form a decision scheme set, including: The scene feature and the historical decision information are constructed into a feature vector representing the time sequence characteristics, the evolution law of the feature vector is obtained by performing time sequence evolution analysis on the feature vector, and the dynamic change feature is extracted from the evolution law; The feature vector is input into the reasoning module of the knowledge graph to perform feature space conversion to obtain a converted feature expression, the stable region of the converted feature expression is determined according to the evolution law of the feature vector, the converted feature expression is constrained by using the stable region, and a plurality of feature parameters representing system stability are obtained; The reasoning module of the knowledge graph is used to construct a feature weight distribution based on the feature parameters, and the change trend of the converted feature expression is calculated according to the feature weight distribution; input the dynamic change feature into an inference module of the knowledge graph, perform reasoning on the converted feature expression based on the change trend to obtain a reasoning result, and generate a plurality of candidate decision schemes according to the reasoning result under the constraint of the stable region; evaluate the candidate decision schemes to obtain evaluation indexes including scheme reliability, scheme diversity and information gain, determine a dynamically changing screening threshold based on an evolution rule of the feature vector, compare the evaluation indexes with the screening threshold, and select candidate decision schemes satisfying the screening threshold to form a decision scheme set.

6. The method of claim 1, wherein, select an optimal decision scheme based on the execution effect score, use an execution result of the optimal decision scheme as a reward signal of a reinforcement learning algorithm, and dynamically adjust the inference module to include: construct a decision ranking value based on a result of the execution effect score of the decision scheme, calculate a phase difference of a plurality of neurons based on the decision ranking value to construct a phase synchronization matrix, cooperatively map different oscillation mode features by using the phase synchronization matrix to obtain a neural oscillation mode prediction value corresponding to the decision scheme, cooperatively map the decision ranking value and the neural oscillation mode prediction value by using the phase synchronization matrix to obtain a decision comprehensive evaluation value, and select an optimal decision scheme based on the decision comprehensive evaluation value; compare an actual execution result of the optimal decision scheme with the decision comprehensive evaluation value to construct an instant reward value, calculate a neural oscillation synchronization change value based on the phase synchronization matrix, cooperatively map the instant reward value and the neural oscillation synchronization change value by using the phase synchronization matrix to obtain a compensation reward value, time-series accumulate the compensation reward value to obtain a cumulative reward value, and input the cumulative reward value into the reinforcement learning algorithm as a reward signal; calculate a gradient value of an inference module parameter based on the reinforcement learning algorithm, calculate a constraint difference value between the inference module parameter and a target neural oscillation mode based on the phase synchronization matrix, and cooperatively map the gradient value and a historical update value of the inference module parameter to dynamically adjust the inference module.

7. The method of claim 1, wherein, based on the adjusted inference module, optimize the optimal decision scheme by using the reinforcement learning algorithm, and generate a new decision scheme to include: construct a chaotic feature matrix based on an output of the adjusted inference module, generate a reward value based on the chaotic feature matrix, determine a state distribution of a decision scheme based on the reward value, and generate a decision optimization sequence based on the state distribution; input the decision optimization sequence into the inference module, iteratively update parameters of the inference module based on the reward value, and obtain updated inference module parameters; expand and optimize the decision scheme by using the updated inference module parameters to obtain a multi-dimensional decision set, calculate an optimization index of each decision scheme in the multi-dimensional decision set based on the chaotic feature matrix, and select a new decision scheme from the multi-dimensional decision set based on the optimization index.

8. The data center intelligent operation and maintenance system based on a multi-modal large model, for implementing the method of any one of the preceding claims 1-7, characterized in that, the first unit is configured to acquire multi-modal data in a data center operation and maintenance scenario, construct node features of a knowledge graph based on the multi-modal data, and extract scene features of the multi-modal data; ​ construct a node relationship of a knowledge graph based on the scenario feature, analyze a correlation strength between the nodes, and construct a connection weight between the nodes; a second unit configured to expand the knowledge graph by using the node feature and the connection weight, match similar scenarios in the knowledge graph based on the scenario feature, identify historical decision information related to the current scenario, input the scenario feature and the historical decision information into an inference module of the knowledge graph together, generate a decision scheme set, and calculate an execution effect score of the decision scheme set; a third unit configured to select an optimal decision scheme based on the execution effect score, take an execution result of the optimal decision scheme as a reward signal of a reinforcement learning algorithm, dynamically adjust the inference module, optimize the optimal decision scheme by using the reinforcement learning algorithm based on the adjusted inference module, and generate a new decision scheme; a fourth unit configured to generate an operation and maintenance operation instruction by using the new decision scheme, execute the operation and maintenance operation instruction according to a preset execution process, and record an execution process and an execution result of the decision scheme.

9. An electronic device, comprising: comprise: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to invoke the instructions stored in the memory to execute the method in any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon computer program instructions, wherein, The computer program instructions, when executed by the processor, implement the method in any one of claims 1 to 7.

Images