Autonomous learning method, system and device based on power system agent cooperation and medium

Abstract

The application discloses an autonomous learning method and system based on power system agent cooperation, equipment and medium, relates to the technical field of artificial intelligence and knowledge management, and the method comprises the following steps: collecting user operation behavior data and interactive text data, calculating semantic coverage and behavior deviation to identify implicit knowledge candidate segments, generating guiding inquiries to obtain knowledge elements, aligning and generating fusion vectors in a joint semantic space through comparative learning of the knowledge elements and the operation behavior data to generate structured knowledge entries, establishing weighted association edges in an organization memory bank according to vector similarity and co-occurrence frequency to construct a graph structure, calculating a matching degree after receiving a current business scenario and propagating an activation value through the weighted association edges, and pushing a structured knowledge entry with the highest activation value. The application realizes autonomous learning and accurate pushing of implicit knowledge, and improves the intelligent level of power system knowledge management.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and knowledge management technology, and in particular to autonomous learning methods, systems, devices and media based on agent collaboration in power systems. Background Technology

[0002] In the operation and maintenance of power systems, a large amount of knowledge relies on human experience. This knowledge often exists implicitly in the practice of operators and is difficult to make explicit and pass on through traditional methods. As the scale and complexity of power systems continue to expand, the problem of knowledge loss caused by the retirement or turnover of experienced operators is becoming increasingly serious. The training cycle for new employees is long and it is difficult for them to quickly master key experience, which seriously affects the safe and stable operation of power systems.

[0003] Existing knowledge management systems primarily rely on manual organization and documentation for knowledge accumulation, but this approach has significant limitations. Operators often struggle to fully articulate their decision-making rationale and judgment logic when performing business operations, resulting in a large amount of tacit knowledge remaining uncaptured. Traditional knowledge bases employ static storage, lacking dynamic modeling of the relationships between knowledge items and the ability to flexibly push relevant knowledge based on business scenarios, leading to low knowledge utilization efficiency. Furthermore, existing methods lack proactive guidance mechanisms during knowledge acquisition, failing to specifically mine key knowledge elements or verify the effectiveness and applicability of knowledge. Summary of the Invention

[0004] The purpose of this invention is to provide a method, system, device, and medium for autonomous learning based on power system agent collaboration, aiming to solve at least one of the technical problems existing in the prior art.

[0005] The technical solution of this invention is: an autonomous learning method based on power system agent collaboration, comprising:

[0006] Collect user action data and interaction text data during business operations;

[0007] Calculate the semantic coverage of interactive text data and the behavioral deviation of operation behavior data from the preset knowledge base. When the semantic coverage is lower than the preset coverage threshold or the behavioral deviation is higher than the preset deviation threshold, extract contextual information as implicit knowledge candidate fragments.

[0008] Generate guiding questions for implicit knowledge candidate fragments and extract knowledge elements based on user responses;

[0009] Knowledge elements and operational behavior data are aligned in the joint semantic space through comparative learning to generate fusion vectors as structured knowledge entries;

[0010] Establish an organizational memory and calculate the similarity between the fusion vector of the structured knowledge item and the vector of the existing knowledge item in the organizational memory. Based on the vector similarity and co-occurrence frequency, establish weighted association edges in the graph structure and store the structured knowledge item and the weighted association edges in the organizational memory.

[0011] Receive the current business scenario and calculate the matching degree between the current business scenario and the fusion vector of the structured knowledge entries in the organization memory. Based on the matching degree and the weighted association edge propagation activation value, push the structured knowledge entry with the highest activation value.

[0012] Calculate the semantic coverage of interactive text data and the behavioral deviation of operational behavior data from a preset knowledge base. When the semantic coverage is lower than a preset coverage threshold or the behavioral deviation is higher than a preset deviation threshold, extract contextual information as implicit knowledge candidate fragments, including:

[0013] After mapping interactive text data to semantic space, the coverage is obtained by retrieving it from the preset knowledge base. When the operation object corresponding to the knowledge item within the coverage is different from the current operation object, the corresponding interactive text data is determined to be a valid sample. The proportion of valid samples is calculated to obtain the semantic coverage.

[0014] Collect operation behavior data of different users under the same business scenario to construct an operation behavior set, count the execution frequency of operation steps in the operation behavior set to obtain the frequency distribution, extract operation steps whose execution frequency exceeds the preset frequency threshold as the benchmark operation steps, and calculate the proportion of the benchmark operation steps in the current operation behavior data to obtain the behavior deviation degree.

[0015] When the semantic coverage is lower than the preset coverage threshold, the interactive text data involved in different operation objects are extracted as context information. When the behavior deviation is higher than the preset deviation threshold, the interactive text data in the time period corresponding to the baseline operation steps that did not appear are extracted as context information.

[0016] The contextual information is associated with the current operation behavior data and the baseline operation steps that have not appeared to form implicit knowledge candidate fragments.

[0017] The process of generating guided questions based on tacit knowledge candidate fragments and extracting knowledge elements based on user responses includes:

[0018] Extract the operation object and execution steps from the implicit knowledge candidate fragments, and combine the attribute features of the operation object and the temporal relationship of the execution steps to form a scene description vector;

[0019] Retrieve existing knowledge entries from the preset knowledge base that are less than the preset distance threshold from the scene description vector but are different from the category of the operation object, and statistically analyze the distribution of knowledge elements in the existing knowledge entries to obtain a set of knowledge element types.

[0020] The difference between the knowledge element types already contained in the implicit knowledge candidate fragment and the set of knowledge element types is calculated to obtain the type to be completed. For the type to be completed, the corresponding query expression is extracted from the existing knowledge entries and the operation object that replaces the query expression is generated to produce a guiding query.

[0021] The system sends guiding questions to users and receives their responses. It then parses the noun phrases, conditional clauses, and judgment statements in the user responses and associates them with the operation objects, execution steps, and execution results in the implicit knowledge candidate fragments to obtain knowledge elements. These knowledge elements include object knowledge elements, conditional knowledge elements, and verification knowledge elements.

[0022] The knowledge elements and operational behavior data are aligned in the joint semantic space through contrastive learning to generate fusion vectors as structured knowledge items, including:

[0023] After encoding knowledge elements and operational behavior data, a joint semantic space is constructed through comparative learning.

[0024] Extract the constraints corresponding to the conditional knowledge elements, locate the execution steps that satisfy the constraints in the operational behavior data as alignment anchors, and align the knowledge elements and operational behavior data in the joint semantic space based on the alignment anchors.

[0025] The comparison difference between the verification knowledge element in the aligned knowledge element and the execution result in the aligned operation behavior data is calculated. When the comparison difference exceeds the preset difference threshold, the attributes that are inconsistent with the verification knowledge element are extracted from the execution result as implicit constraints to supplement the conditional knowledge element and re-encode it to obtain the supplemented knowledge element.

[0026] Based on the temporal position of the alignment anchor points, the supplemented knowledge elements and the aligned operation behavior data are divided into temporal segments and the bidirectional coverage is calculated. The temporal segments with bidirectional coverage exceeding the preset coverage threshold are selected to generate a fusion vector as a structured knowledge item.

[0027] The process involves establishing an organizational memory and calculating the similarity between the fusion vector of structured knowledge entries and the vectors of existing knowledge entries in the organizational memory. Based on vector similarity and co-occurrence frequency, weighted association edges are established in the graph structure. The structured knowledge entries and these weighted association edges are then stored in the organizational memory.

[0028] An organizational memory bank is built using a graph structure, where nodes are structured knowledge entries and edges are weighted association edges;

[0029] Calculate the vector distance between the fusion vector of structured knowledge entries and the vectors of existing knowledge entries in the organizational memory;

[0030] The overlap between the constraints corresponding to the conditional knowledge elements of structured knowledge entries and existing knowledge entries is extracted as constraint consistency. Vector similarity is obtained by correcting the vector distance based on constraint consistency.

[0031] The co-occurrence frequency is defined as the number of times that structured knowledge items and existing knowledge items are found to be consistent with each other in historical operational behavior data.

[0032] Existing knowledge items with vector similarity greater than a preset similarity threshold are selected as candidate knowledge items. The vector similarity is adjusted based on the differences between the candidate knowledge items and the verification knowledge elements of the structured knowledge items. The association strength is obtained by weighting the adjusted vector similarity with the co-occurrence frequency.

[0033] Based on the correlation strength, weighted association edges are established between structured knowledge entries and candidate knowledge entries in the graph structure, and the structured knowledge entries and weighted association edges are stored in the organizational memory.

[0034] The system receives the current business scenario and calculates the matching degree between the current business scenario and the fusion vector of structured knowledge entries in the organizational memory. Based on the matching degree and the activation value propagated by the weighted association edge, it pushes the structured knowledge entries with the highest activation values, including:

[0035] Receive the current business scenario and extract the constraints corresponding to the conditional knowledge elements;

[0036] Iterate through the constraints of structured knowledge entries in the organization memory, and filter out the structured knowledge entries as to be matched when all the constraints of the structured knowledge entries exist in the constraints of the current business scenario.

[0037] Calculate the matching degree between the fusion vector of the current business scenario and the knowledge item to be matched;

[0038] The historical confidence score is calculated as the proportion of knowledge elements verified in historical operational behavior data that match the knowledge elements verified in the current business scenario. The matching score is then adjusted based on the historical confidence score.

[0039] The corrected matching degree is used as the initial activation value and assigned to the knowledge entry to be matched.

[0040] Extract adjacent structured knowledge entries connected by weighted association edges, and compare the verification knowledge elements of adjacent structured knowledge entries with the verification knowledge elements of the current business scenario;

[0041] Whether to propagate is determined based on the differences in the verified knowledge elements. When propagating, the activation value is propagated to the adjacent structured knowledge items according to the association strength of the weighted association edges and the activation value is accumulated.

[0042] Push the structured knowledge item with the highest activation value.

[0043] This invention provides an autonomous learning system based on power system agent collaboration, comprising:

[0044] The data collection module is used to collect user action data and interactive text data during business operations.

[0045] The recognition module is used to calculate the semantic coverage of interactive text data and the behavioral deviation of operation behavior data from the preset knowledge base. When the semantic coverage is lower than the preset coverage threshold or the behavioral deviation is higher than the preset deviation threshold, the context information is extracted as a candidate fragment of implicit knowledge.

[0046] The query module is used to generate guiding queries for implicit knowledge candidate fragments and obtain knowledge elements based on user responses;

[0047] The alignment module is used to align knowledge elements and operational behavior data in the joint semantic space through contrastive learning to generate fusion vectors as structured knowledge entries.

[0048] The storage module is used to establish an organizational memory and calculate the similarity between the fusion vector of the structured knowledge entry and the vector of the existing knowledge entry in the organizational memory. Based on the vector similarity and co-occurrence frequency, weighted association edges are established in the graph structure, and the structured knowledge entry and the weighted association edges are stored in the organizational memory.

[0049] The push module is used to receive the current business scenario and calculate the matching degree between the current business scenario and the fusion vector of the structured knowledge entries in the organization memory. Based on the matching degree and the activation value of the weighted association edge propagation, the module pushes the structured knowledge entry with the highest activation value.

[0050] The present invention also provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of any of the preceding methods.

[0051] The present invention also provides a computer-readable storage medium storing computer program instructions, which, when executed by a processor, implement the steps of any of the preceding methods.

[0052] This invention accurately identifies knowledge base gaps through dual indicators of semantic coverage and behavioral deviation, proactively discovering tacit knowledge that traditional methods struggle to capture. A guided inquiry mechanism specifically supplements knowledge elements, avoiding omissions in manually compiled information and improving the completeness and accuracy of knowledge acquisition. Contrastive learning is employed to align knowledge elements with operational behavior data in a joint semantic space, generating fusion vectors that achieve deep integration of explicit knowledge and tacit experience, enhancing the richness of knowledge representation. Weighted association edges are established using vector similarity and co-occurrence frequency to construct a graph structure for organizing the knowledge base, dynamically modeling the relationships between knowledge items and supporting networked knowledge management. An activation value propagation mechanism based on matching degree and weighted association edges intelligently pushes the most relevant knowledge items according to the current business scenario, significantly improving the accuracy and timeliness of knowledge push, reducing user retrieval costs, and providing timely and effective decision support for power system operation and maintenance personnel. Attached Figure Description

[0053] Figure 1 A flowchart illustrating the autonomous learning method based on power system agent collaboration provided in an embodiment of the present invention;

[0054] Figure 2 This is a schematic diagram of the structure of the autonomous learning system based on power system agent collaboration according to an embodiment of the present invention. Detailed Implementation

[0055] like Figure 1 As shown, the autonomous learning method based on power system agent collaboration includes the following steps:

[0056] Step 101: Collect user action data and interaction text data for performing business operations.

[0057] Step 102: Calculate the semantic coverage of the interactive text data and the behavioral deviation of the operation behavior data from the preset knowledge base. When the semantic coverage is lower than the preset coverage threshold or the behavioral deviation is higher than the preset deviation threshold, extract the context information as a candidate fragment of implicit knowledge.

[0058] In some embodiments of the present invention, step 102 may specifically include the following sub-steps:

[0059] Sub-step 1021: After mapping the interactive text data to the semantic space, the coverage is retrieved in the preset knowledge base. When the operation object corresponding to the knowledge item within the coverage is different from the current operation object, the corresponding interactive text data is determined to be a valid sample. The proportion of valid samples is calculated to obtain the semantic coverage.

[0060] Sub-step 1022: Collect operation behavior data of different users under the same business scenario to construct an operation behavior set, count the execution frequency of operation steps in the operation behavior set to obtain the frequency distribution, extract operation steps whose execution frequency exceeds the preset frequency threshold as the benchmark operation steps, and calculate the proportion of the benchmark operation steps in the current operation behavior data to obtain the behavior deviation.

[0061] Sub-step 1023: When the semantic coverage is lower than the preset coverage threshold, extract the interactive text data involved in different operation objects as context information; when the behavior deviation is higher than the preset deviation threshold, extract the interactive text data in the time period corresponding to the baseline operation steps that did not appear as context information.

[0062] Sub-step 1024: Associate the context information with the current operation behavior data and the baseline operation steps that have not appeared to form implicit knowledge candidate fragments.

[0063] First, operational behavior data and interactive text data of power system operators are collected during their daily work. Operational behavior data includes specific operational steps, objects being operated on, and operation times in scenarios such as power dispatching, equipment maintenance, and fault handling. Interactive text data includes communication records and operation logs between operators and other personnel during operations. This data is automatically recorded and stored by data collection agents deployed within the power system.

[0064] Based on the collected data, the semantic coverage of the interactive text data and the pre-set knowledge base is calculated. First, the interactive text data is mapped to a semantic space using a semantic mapping model. This model employs a semantic encoder pre-trained in the power industry domain, representing the text as a high-dimensional vector. Knowledge entries with semantically similar meanings to the interactive text data are retrieved from the pre-set knowledge base to determine the coverage area. When the operation object corresponding to a knowledge entry within the coverage area differs from the current operation object, the corresponding interactive text data is determined as a valid sample. For example, if an operator uses a method applicable to another transformer model when handling a fault in a certain type of transformer, the interactive text in this case is marked as a valid sample. The semantic coverage is calculated by dividing the number of valid samples by the total number of samples. The number of valid samples is the number of samples within the coverage area whose operation object differs from the current operation object, and the total number of samples is the total number of samples in all interactive text data.

[0065] An operational behavior set is constructed by collecting operational behavior data from different operators within the same power business scenario. The execution frequency of each operational step in the set is statistically analyzed to obtain a frequency distribution. The frequency distribution records the number of times each operational step appears in the entire operational behavior set. Operational steps with an execution frequency exceeding a preset frequency threshold are extracted as baseline operational steps. These baseline operational steps represent standard operating procedures that are routine and widely adopted in the power system. The behavior deviation is calculated as 1 minus the proportion of baseline operational steps appearing in the current operational behavior data. The number of baseline operational steps is the total number of baseline operational steps included in the current operational behavior data, and the total number of baseline operational steps is the total number of all baseline operational steps in this business scenario.

[0066] When the calculated semantic coverage is lower than a preset coverage threshold or the behavioral deviation is higher than a preset deviation threshold, it indicates that the current operation or interaction text contains potential implicit knowledge that needs to be extracted. For cases where the semantic coverage is lower than the preset coverage threshold, interactive text data related to different operation objects is extracted as contextual information. This interactive text data typically contains specific processing methods used by operators for particular devices or scenarios, methods that are not fully covered in the preset knowledge base. Specifically, this can be achieved by comparing the operation object identifier in the extracted knowledge entry with the current operation object identifier. If they do not match but the semantic similarity is higher than a set threshold, the interactive text is marked and extracted.

[0067] For behaviors deviating above a preset deviation threshold, the interactive text data within the corresponding time period of the missing baseline operation step is extracted as contextual information. This interactive text data may contain the reasons why the operator skipped or substituted the standard operation step and the alternative solutions. This is achieved by aligning the timeline to locate the time period during which the missing baseline operation step should have been executed, and then extracting the interactive text within a certain time window before and after that period. The size of the time window can be flexibly set according to different business scenarios, but it can generally be set to twice the standard execution time of the baseline operation step.

[0068] The extracted context information is associated with the current operational behavior data and the missing baseline operational steps to form implicit knowledge candidate fragments. The association process is achieved through timestamp matching and semantic relevance analysis. Timestamp matching ensures the temporal consistency between context information and operational behavior, while semantic relevance analysis ensures the content-related relevance between the context information and operational behavior. Semantic relevance analysis employs a similarity calculation method optimized for the power industry, considering the weights of specific factors such as power industry terminology, equipment names, and operational procedures. The resulting implicit knowledge candidate fragments are stored in a structured manner, containing fields such as operational scenario description, triggering conditions, operational objects, operational step sequences, and context information, constituting a complete implicit knowledge representation.

[0069] In practical applications, the generated tacit knowledge candidate fragments also need to be verified. Verification methods include historical case retrospective verification and expert review verification. Historical case retrospective verification applies the tacit knowledge candidate fragments to historical problem cases to test their effectiveness in solving problems; expert review verification involves domain experts reviewing the tacit knowledge candidate fragments to evaluate their accuracy and applicability. Verified tacit knowledge candidate fragments will be added to the power system knowledge base, forming a continuously updated knowledge system.

[0070] To ensure the reliability and timeliness of tacit knowledge, a knowledge update mechanism is also in place. When new operational behavior data and interactive text data are collected, their similarity to stored tacit knowledge is automatically calculated. For highly similar cases, the weight of that knowledge is increased; for cases with conflicts, they are marked as items to be verified, awaiting further data or expert confirmation. This dynamic update mechanism enables the knowledge base to adapt to changes and developments in the power system.

[0071] This invention effectively identifies and extracts tacit knowledge by automating the analysis of the behavioral patterns and interactive texts of power system operators, solving the problem of capturing and transmitting tacit knowledge in traditional power system knowledge management. This method reduces reliance on expert experience, improves the completeness and accessibility of power system knowledge, promotes the accumulation and inheritance of power system operation and maintenance experience, enhances the safety and stability of power system operation, and simultaneously improves the efficiency of new employee training and the power system's emergency response capabilities.

[0072] Step 103: Generate guiding questions for implicit knowledge candidate fragments and obtain knowledge elements based on user responses.

[0073] In some embodiments of the present invention, step 103 may specifically include the following sub-steps:

[0074] Sub-step 1031: Extract the operation object and execution steps from the implicit knowledge candidate fragments, and combine the attribute features of the operation object and the temporal relationship of the execution steps to form a scene description vector;

[0075] Sub-step 1032: Retrieve existing knowledge entries in the preset knowledge base that are less than the preset distance threshold but have different categories of operation objects from the scene description vector, and statistically analyze the distribution of knowledge elements in the existing knowledge entries to obtain a set of knowledge element types.

[0076] Sub-step 1033: Calculate the difference between the knowledge element types already contained in the implicit knowledge candidate fragments and the set of knowledge element types to obtain the type to be completed. For the type to be completed, extract the corresponding query expression from the existing knowledge entries and replace the operation object of the query expression to generate a guiding query.

[0077] Sub-step 1034: Send a guiding question to the user and receive the user's response. Parse the noun phrases, conditional clauses and judgment statements in the user's response, and associate them with the operation object, execution steps and execution results in the implicit knowledge candidate fragments to obtain knowledge elements. The knowledge elements include object knowledge elements, conditional knowledge elements and verification knowledge elements.

[0078] This method extracts operation objects and execution steps from implicit knowledge candidate fragments. Operation objects are equipment, components, or systems within a power system, such as transformers, circuit breakers, and relay protection devices. Execution steps are the specific sequence of operational behaviors performed by operators when handling related tasks. The extraction method employs rule-based entity recognition technology, matching and extracting data through a predefined dictionary of power system equipment and operation verbs. For operation objects, attributes such as model, parameters, and operating status are extracted; for execution steps, information such as operational actions, target states, and temporal relationships is extracted. The attribute features of the operation objects and the temporal relationships of the execution steps are combined to form a scene description vector. The scene description vector is a multi-dimensional feature representation, containing equipment feature dimensions and operation feature dimensions. The equipment feature dimension includes equipment type, physical parameters, and operating parameters; the operation feature dimension includes operation type, operation sequence, and operation conditions. These features are combined with weights to form a unified vector representation.

[0079] The system retrieves existing knowledge entries from a pre-defined knowledge base that are less than a pre-defined distance threshold from the scene description vector but differ in the category of the operation object. The retrieval employs a similarity calculation method, calculating the cosine distance between the scene description vector and the vectors of each knowledge entry in the knowledge base. When the distance between two vectors is less than the pre-defined distance threshold, it indicates that the two scenarios are similar in operation procedures and characteristics, despite the different operation objects. The distribution of various knowledge elements among the existing knowledge entries that meet the criteria is statistically analyzed to obtain a set of knowledge element types. Knowledge element types may include equipment parameter elements, operating condition elements, operating step elements, fault characteristic elements, safety precaution elements, etc. The statistical analysis process calculates the frequency of occurrence of each type of knowledge element in the retrieved knowledge entries to determine its importance weight.

[0080] The difference between the knowledge element types already contained in the implicit knowledge candidate fragment and the set of knowledge element types is calculated to obtain the type to be completed. The difference calculation is achieved through set operations, removing existing knowledge element types from the implicit knowledge candidate fragment from the set of knowledge element types; the remaining part is the type to be completed. For the type to be completed, the corresponding query expression is extracted from existing knowledge entries, and the operation object in the query expression is replaced to generate a guiding query. The query expression is a standardized query template for a specific knowledge element type. The replacement process identifies object placeholders in the query template and replaces them with the actual operation object in the current implicit knowledge candidate fragment. For example, if the type to be completed is an operation condition element, the query expression extracted from existing knowledge entries might be "Under what conditions is it necessary to perform operation Y on X?". Replacing X with the current operation object, such as "high-voltage circuit breaker," and Y with the current execution step, such as "emergency tripping," the generated guiding query would be "Under what conditions is it necessary to perform an emergency tripping operation on the high-voltage circuit breaker?"

[0081] The generated guided queries are sent to users through a user interface, and the system receives user responses. Users are typically experienced power system operators. After a user responds, the system parses the noun phrases, conditional clauses, and judgment statements in the response. The parsing process uses natural language processing technology to identify key semantic units in the user response. Noun phrases mainly contain information such as equipment names, parameter values, and status descriptions; conditional clauses mainly contain information such as operation triggering conditions, environmental constraints, and preconditions; judgment statements mainly contain information such as operation result verification, safety checks, and quality standards.

[0082] The parsed semantic units are associated with the operation objects, execution steps, and execution results in the implicit knowledge candidate fragments, resulting in complete knowledge elements. Noun phrases are associated with operation objects to form object knowledge elements, describing the detailed characteristics and parameters of the operation objects; conditional clauses are associated with execution steps to form conditional knowledge elements, describing the preconditions and constraints of the execution steps; and judgment statements are associated with execution results to form verification knowledge elements, describing how to determine whether the operation was successful and whether the result meets expectations. The association process is achieved through semantic relevance calculation, matching the parsed semantic units with the corresponding parts in the implicit knowledge candidate fragments.

[0083] For object-related knowledge elements, the focus is on information such as equipment model, specifications, and special operating conditions mentioned in the user's response; for conditional knowledge elements, the focus is on information such as the timing of operation, environmental requirements, and safety prerequisites described in the user's response; and for verification knowledge elements, the focus is on information such as the testing methods, normal indicators, and abnormal phenomena mentioned in the user's response. These extraction rules are built based on knowledge in the power system domain to ensure that the extracted knowledge elements conform to power system operating specifications.

[0084] The extracted knowledge elements are integrated with candidate tacit knowledge fragments to form complete tacit knowledge entries. The integration process follows a structured knowledge representation model, organizing knowledge elements into an organic whole according to their type and relationships. After verification, the integrated tacit knowledge entries can be added to the power system knowledge base for subsequent use.

[0085] This invention, through guided inquiry to acquire knowledge elements, effectively solves the problem of incompleteness in tacit knowledge extraction, improving the accuracy and applicability of knowledge. This method makes the tacit knowledge of experienced operators explicit, enabling effective knowledge transfer and reuse. The power system knowledge base built based on this method is more comprehensive and practical, providing strong support for the safe operation and fault handling of power systems.

[0086] Step 104: Align knowledge elements and operational behavior data in the joint semantic space through comparative learning to generate fusion vectors as structured knowledge entries.

[0087] In some embodiments of the present invention, step 104 may specifically include the following sub-steps:

[0088] Sub-step 1041 involves encoding knowledge elements and operational behavior data and then constructing a joint semantic space through comparative learning.

[0089] Sub-step 1042: Extract the constraints corresponding to the conditional knowledge elements, locate the execution steps that satisfy the constraints in the operation behavior data as alignment anchors, and align the knowledge elements and operation behavior data in the joint semantic space based on the alignment anchors.

[0090] Sub-step 1043: Calculate the comparison difference between the verification knowledge element in the aligned knowledge element and the execution result in the aligned operation behavior data. When the comparison difference exceeds the preset difference threshold, extract the attributes that are inconsistent with the verification knowledge element from the execution result as implicit constraints to supplement the conditional knowledge element and re-encode it to obtain the supplemented knowledge element.

[0091] Sub-step 1044: Divide the supplemented knowledge elements and aligned operation behavior data into time segments according to the temporal position of the alignment anchor points and calculate the bidirectional coverage. Select time segments with bidirectional coverage exceeding the preset coverage threshold and fuse them to generate a fusion vector as a structured knowledge item.

[0092] Knowledge elements include object knowledge elements, conditional knowledge elements, and verification knowledge elements, all existing in text form. Operational behavior data includes information such as the operation object, the sequence of operation steps, and execution time, exhibiting structured and temporal characteristics. For knowledge elements, a semantic encoder is used to map them into vector representations. The semantic encoder is built based on a pre-trained language model, optimized for the power industry, and can accurately capture the semantic information of power industry terminology and expressions. For operational behavior data, a behavior encoder is used to map it into vector representations. The behavior encoder considers the temporal and structural nature of operational behaviors, and can express the logical relationships and dependencies between operation steps. The encoded knowledge element vectors and operational behavior vectors reside in different vector spaces.

[0093] The goal of contrastive learning is to bring semantically similar knowledge element vectors and operational behavior vectors closer together in the joint semantic space, while widening the distance between semantically dissimilar vectors. This is achieved by constructing a dual-tower model: one tower handles knowledge elements, and the other tower handles operational behavior data. The output vectors of both towers are mapped to the same-dimensional joint semantic space through a projection layer. During training, a contrastive loss function is used to optimize the model parameters. The contrastive loss function calculates the difference between the similarity scores of positive and negative sample pairs. Positive sample pairs represent knowledge elements and operational behavior data from the same business scenario, while negative sample pairs represent knowledge elements and operational behavior data from different business scenarios. By minimizing the contrastive loss, the model learns the semantic correspondence between knowledge elements and operational behavior data.

[0094] After the joint semantic space is constructed, it is necessary to align knowledge elements and operational behavior data within this space. The core of the alignment process is determining alignment anchor points. Constraints are extracted from the conditional knowledge elements; these constraints describe the preconditions required to perform a specific operation. The temporal positions of execution steps that satisfy the constraints in the operational behavior data are located as alignment anchor points. This location process is achieved by calculating the similarity between the constraint vector and the execution step vectors at each temporal position in the operational behavior data. The similarity calculation uses the cosine similarity method, calculating the dot product of two vectors divided by the product of their magnitudes. A higher similarity value indicates that the execution step better conforms to the description of the constraints. The position of the execution step with the highest similarity is selected as the alignment anchor point. If multiple alignment anchor points exist, they are arranged in temporal order to form an anchor point sequence. Based on the alignment anchor points, knowledge elements and operational behavior data are aligned in the joint semantic space, ensuring a one-to-one correspondence between them in the temporal dimension.

[0095] After alignment, the difference between the verification knowledge elements in the aligned knowledge elements and the execution results in the aligned operational behavior data is calculated. Verification knowledge elements describe the expected results or verification criteria after operation execution, while execution results are the actual results recorded in the operational behavior data. The difference is quantified by calculating the distance between the verification knowledge element vector and the execution result vector. The difference degree can be calculated by subtracting the cosine similarity of the two vectors; the larger the difference degree value, the greater the difference between the verification knowledge element and the execution result. When the difference degree exceeds a preset difference threshold, attributes inconsistent with the verification knowledge elements are extracted from the execution results. The extraction process identifies inconsistent parts by comparing the semantic representations of the verification knowledge elements and the execution results. These inconsistent attributes may be due to special conditions or constraints in the operation process and are an important part of tacit knowledge. The extracted inconsistent attributes are used as implicit constraints, supplemented to the conditional knowledge elements, and re-encoded to obtain the supplemented knowledge elements.

[0096] The supplemented knowledge elements and aligned operational behavior data are divided into time segments based on the temporal position of the alignment anchor points. Each time segment is a fragment of knowledge elements and operational behavior data defined by the alignment anchor points, and each time segment contains one or more knowledge elements and corresponding operational behavior data. For each time segment, bidirectional coverage is calculated, reflecting the degree of semantic mutual coverage between knowledge elements and operational behavior data. Bidirectional coverage is calculated as a weighted average of the degree to which knowledge elements cover operational behavior data and the degree to which operational behavior data covers knowledge elements. Coverage is determined by calculating the similarity matrix between vectors and taking the maximum value. Time segments with bidirectional coverage exceeding a preset coverage threshold are selected, indicating that the knowledge elements and operational behavior data in these time segments have a high degree of semantic consistency.

[0097] Knowledge elements and operational behavior data from selected time series segments are fused to generate a fusion vector. The fusion process employs a weighted summation method, assigning weights based on the importance of the knowledge elements and operational behavior data. The importance of knowledge elements is determined by their centrality in the joint semantic space, while the importance of operational behavior data is determined by their criticality within the operational process. The fusion vector preserves the semantic information of the knowledge elements and the structural characteristics of the operational behavior data, forming a more comprehensive and accurate knowledge representation. The fusion vector is then stored as structured knowledge entries in the power system knowledge base.

[0098] By aligning knowledge elements and operational behavior data in a joint semantic space through contrastive learning and generating fusion vectors, a structured representation and storage of tacit knowledge is achieved. This method captures the tacit knowledge of power system operators and transforms it into shareable and reusable explicit knowledge, effectively solving the problem of expressing and transmitting tacit knowledge in traditional knowledge management. The knowledge base built based on this method is highly professional and practical, providing strong support for the safe operation and fault handling of power systems, and improving the overall operation and maintenance level and emergency response capabilities of power systems.

[0099] Step 105: Establish an organizational memory and calculate the similarity between the fusion vector of the structured knowledge item and the vector of the existing knowledge item in the organizational memory. Based on the vector similarity and co-occurrence frequency, establish weighted association edges in the graph structure and store the structured knowledge item and the weighted association edges in the organizational memory.

[0100] In some embodiments of the present invention, step 105 may specifically include the following sub-steps:

[0101] Sub-step 1051: Establish an organizational memory bank with a graph structure, where nodes are structured knowledge entries and edges are weighted association edges;

[0102] Sub-step 1052: Calculate the vector distance between the fusion vector of the structured knowledge entry and the vector of the existing knowledge entry in the organization memory.

[0103] Sub-step 1053: Extract the overlap between the constraints corresponding to the conditional knowledge elements of the structured knowledge entries and the existing knowledge entries as constraint consistency, and correct the vector distance based on constraint consistency to obtain vector similarity.

[0104] Sub-step 1054: The number of times structured knowledge items and existing knowledge items co-occur in historical operational behavior data to verify the consistency of knowledge elements is used as the co-occurrence frequency;

[0105] Sub-step 1055: Select existing knowledge items with vector similarity greater than a preset similarity threshold as candidate knowledge items. Adjust the vector similarity according to the difference between the candidate knowledge items and the verification knowledge elements of the structured knowledge items. Calculate the association strength by weighting the adjusted vector similarity with the co-occurrence frequency.

[0106] Sub-step 1056: Based on the association strength, establish weighted association edges between structured knowledge entries and candidate knowledge entries in the graph structure, and store the structured knowledge entries and weighted association edges in the organization memory.

[0107] The basic structure of the organizational memory is a graph, where nodes represent structured knowledge entries and edges represent weighted association edges. Each structured knowledge entry contains object knowledge elements, conditional knowledge elements, verification knowledge elements, and a fusion vector. Weighted association edges represent the strength of the association between two knowledge entries, determined by vector similarity and co-occurrence frequency. The graph structure uses an adjacency list for storage, facilitating quick retrieval of related knowledge entries. Initially, the organizational memory may contain only a small number of knowledge entries; as the learning process continues, the number of knowledge entries gradually increases, and the graph structure becomes richer and more complete.

[0108] Before incorporating newly generated structured knowledge entries into the organizational memory, it is necessary to calculate their vector distance with existing knowledge entries in the memory. Vector distance is an indicator that measures the semantic similarity between two knowledge entries. The calculation uses Euclidean distance to determine the distance between the fused vector of the newly generated structured knowledge entry and the vector of an existing knowledge entry. Euclidean distance is a commonly used distance metric in vector spaces; for two vectors, the Euclidean distance is the square root of the sum of the squares of the differences between their corresponding elements. The smaller the calculated vector distance, the more semantically similar the two knowledge entries are. Considering the specialized and specific nature of power system knowledge, relying solely on vector distance may not fully reflect the actual degree of correlation between knowledge entries; therefore, it is necessary to incorporate domain-specific information for correction.

[0109] The constraints corresponding to the conditional knowledge elements of structured knowledge entries and existing knowledge entries are extracted, and the overlap between these constraints is calculated as constraint consistency. The constraints in the conditional knowledge elements describe the preconditions or environmental requirements for performing operations and are an important basis for judging whether two knowledge entries are applicable to similar scenarios. The constraint consistency calculation method decomposes the constraints of two knowledge entries into basic constraint terms, calculates the number of common constraint terms divided by the total number of constraint terms. The higher the constraint consistency, the more similar the applicable scenarios of the two knowledge entries. Vector similarity is obtained by correcting the vector distance based on constraint consistency. The correction process considers the weight of constraint consistency; when constraint consistency is high, the vector similarity is appropriately increased; when constraint consistency is low, the vector similarity is appropriately decreased. Through this correction, the vector similarity can more accurately reflect the actual similarity between knowledge entries.

[0110] The co-occurrence frequency (COF) is the number of times structured knowledge items and existing knowledge items co-occur in historical operational behavior data when their verification knowledge elements are consistent. Verification knowledge elements describe the expected results or verification criteria after an operation is performed; consistency in verification knowledge elements implies that the operational goals are the same. COF reflects how often two knowledge items appear simultaneously in practical applications and is an important indicator of the degree of association between knowledge items. A higher COF indicates a closer association between the two knowledge items in practice. Calculating COF requires analyzing a large amount of historical operational behavior data, extracting operational records with the same operational goals, and counting the number of times structured knowledge items and existing knowledge items co-occur.

[0111] Existing knowledge items with vector similarity greater than a preset similarity threshold are selected as candidate knowledge items. The preset similarity threshold is a parameter pre-set by the system to filter existing knowledge items with high relevance to the new knowledge item. Knowledge items with vector similarity exceeding the threshold are considered to be closely related to the new knowledge item and require further processing. The vector similarity is adjusted based on the differences in verification knowledge elements between the candidate knowledge items and the structured knowledge items. These differences reflect the degree of difference in the operational goals of the two knowledge items; the greater the difference, the lower the similarity should be; the smaller the difference, the higher the similarity should be. The adjusted vector similarity more accurately reflects the similarity between knowledge items at the operational goal level. The association strength is calculated by weighting the adjusted vector similarity with co-occurrence frequency. The formula for calculating association strength is:

[0112] Association strength = α × adjusted vector similarity + (1-α) × normalized co-occurrence frequency, where α is a weighting coefficient ranging from 0 to 1, used to balance the importance of vector similarity and co-occurrence frequency in the association strength calculation; normalized co-occurrence frequency is a value that maps the co-occurrence frequency to the interval between 0 and 1.

[0113] Based on the calculated association strength, weighted association edges are established between structured knowledge entries and candidate knowledge entries in the organizational memory's graph structure. The weight of each weighted association edge represents the association strength, reflecting the closeness of the relationship between the two knowledge entries. A stronger association strength indicates a closer relationship between the two knowledge entries, resulting in higher relevance during knowledge retrieval and recommendation. Structured knowledge entries are treated as new nodes, and weighted association edges are treated as new edges, stored in the organizational memory. The storage process includes updating the node set and edge set of the graph structure and updating relevant index information to ensure rapid retrieval and access to newly added knowledge entries and their associations.

[0114] By establishing an organizational memory and storing structured knowledge entries along with weighted edges, a systematic organization and management of power system knowledge is achieved. The graph-based organizational memory effectively captures the complex relationships between knowledge entries, supports multi-dimensional knowledge retrieval and reasoning, and provides more comprehensive and accurate knowledge services. As knowledge entries accumulate, the organizational memory continuously enriches and improves, gradually forming a knowledge network covering various professional fields within the power system, providing strong knowledge support for the safe and efficient operation of the power system.

[0115] This invention utilizes a graph-structured organized memory to achieve structured storage and association mining of power system knowledge. By calculating vector similarity and co-occurrence frequency, weighted association edges are established between knowledge items, forming a knowledge network rich in semantic information. This method not only makes tacit knowledge explicit and structured but also captures the inherent relationships between knowledge items, supporting accurate knowledge retrieval, associative reasoning, and knowledge discovery. This improves the efficiency of power system operation and maintenance, enhances fault handling capabilities, and strengthens the knowledge assurance level for safe power grid operation.

[0116] Step 106: Receive the current business scenario and calculate the matching degree between the current business scenario and the fusion vector of the structured knowledge entries in the organization memory. Based on the matching degree and the activation value of the weighted association edge propagation, push the structured knowledge entry with the highest activation value.

[0117] In some embodiments of the present invention, step 106 may specifically include the following sub-steps:

[0118] Sub-step 1061: Receive the current business scenario and extract the constraints corresponding to the conditional knowledge elements;

[0119] Sub-step 1062: Traverse the constraints of structured knowledge entries in the organization memory. When all the constraints of a structured knowledge entry exist in the constraints of the current business scenario, it is selected as a knowledge entry to be matched.

[0120] Sub-step 1063: Calculate the matching degree between the fusion vector of the current business scenario and the knowledge item to be matched;

[0121] Sub-step 1064: Calculate the proportion of the knowledge elements verified in the historical operation behavior data that are consistent with the knowledge elements verified in the current business scenario as the historical confidence score, and adjust the matching score based on the historical confidence score.

[0122] Sub-step 1065: Assign the corrected matching degree as the initial activation value to the knowledge entry to be matched.

[0123] Sub-step 1066: Extract adjacent structured knowledge entries connected by weighted association edges, and compare the verification knowledge elements of adjacent structured knowledge entries with the verification knowledge elements of the current business scenario.

[0124] Sub-step 1067: Determine whether to propagate based on the differences in the verified knowledge elements. When propagating, propagate the activation value to the adjacent structured knowledge entries based on the association strength of the weighted association edges and accumulate them to obtain the activation value.

[0125] Sub-step 1068: Push the structured knowledge entry with the highest activation value.

[0126] Business scenarios typically include information such as environmental constraints and operational objectives. It is necessary to extract the constraints corresponding to the conditional knowledge elements from these scenarios. Constraints are key elements describing the characteristics of the business scenario, including equipment status, power grid operation mode, load level, and weather conditions. The process of extracting constraints employs semantic analysis methods, breaking down the business scenario description into basic semantic units and identifying constraints related to power system operation. The extracted constraints form a constraint set, which serves as the basis for subsequent matching processes.

[0127] The process iterates through all structured knowledge entries in the organizational memory, checking whether the constraints of each knowledge entry are included in the constraints of the current business scenario. The determination of constraint inclusion needs to consider semantic equivalence, not just exact string matching. Two constraints may differ in expression but may be semantically equivalent; for example, "voltage is within the normal range" and "voltage value is within the allowable deviation range" are semantically equivalent. During the determination, the constraints of each knowledge entry are compared one by one with the constraints of the current business scenario, calculating semantic similarity. When the similarity exceeds a preset threshold, the constraint is considered to exist in the current business scenario. When all constraints of a knowledge entry exist in the constraints of the current business scenario, the knowledge entry is selected as a candidate for matching. Candidate knowledge entries are those that match the constraints of the current business scenario and have high applicability from a constraint perspective.

[0128] The matching degree between the current business scenario and the fused vector of the knowledge item to be matched is calculated. The matching degree reflects the similarity between the current business scenario and the knowledge item in the semantic space and is an important indicator for judging the applicability of the knowledge item. The matching degree is calculated using the cosine similarity method, which is to calculate the dot product of the two vectors and divide it by the product of their magnitudes. For the current business scenario vector A and the fused vector B of the knowledge item to be matched, the matching degree calculation formula is:

[0129] ,

[0130] Where n is the vector dimension, A i and Bi Let A and B be the values ​​of vectors A and B in the i-th dimension, respectively. The matching degree ranges from -1 to 1. The closer the value is to 1, the more similar the two vectors are, meaning the higher the fit between the current business scenario and the knowledge item.

[0131] The historical confidence score is calculated as the proportion of knowledge elements in historical operational behavior data that match the verified knowledge elements in the current business scenario. Verified knowledge elements describe the expected results or verification standards after an operation is executed, and historical confidence score reflects the historical applicability of the knowledge item in similar business scenarios. During the statistical process, historical scenarios similar to the current business scenario are retrieved from the historical operational behavior database, corresponding operation records are extracted, and the application and results of the knowledge item to be matched in these scenarios are analyzed. The historical confidence score is obtained by dividing the number of operations with consistent verified knowledge elements by the total number of operations. The matching score is then adjusted based on the historical confidence score. The adjustment method is to take a weighted average of the original matching score and the historical confidence score; the higher the historical confidence score, the greater the improvement in the original matching score; conversely, the lower the historical confidence score, the greater the reduction in the original matching score. The adjusted matching score better reflects the reliability of the knowledge item in practical applications.

[0132] The revised matching degree is assigned as the initial activation value to the knowledge item to be matched. The initial activation value is the initial state of the knowledge item during the activation and propagation process, reflecting the direct relevance of the knowledge item to the current business scenario. The allocation of the initial activation value is based on the revised matching degree; the higher the matching degree, the larger the initial activation value. This allocation method ensures that knowledge items directly related to the current business scenario receive a high initial activation value, thus having a greater influence in the subsequent propagation process.

[0133] Adjacent structured knowledge entries connected by weighted associative edges are extracted, and their verification knowledge elements are compared with those of the current business scenario. Adjacent structured knowledge entries are those directly connected to the knowledge entry to be matched in the graph structure of the organization memory. The comparison of verification knowledge elements employs a semantic similarity calculation method, calculating the semantic similarity between the verification knowledge elements of adjacent knowledge entries and those of the current business scenario. Higher semantic similarity indicates that the operational goals of adjacent knowledge entries are closer to the goals of the current business scenario.

[0134] Verifying differences in knowledge elements reflects the degree of difference between the operational goals of knowledge items and the goals of the current business scenario. When the semantic similarity of verified knowledge elements exceeds a preset propagation threshold, the operational goals of adjacent knowledge items are considered to be consistent with the goals of the current business scenario, and activation value propagation is allowed. When the semantic similarity is lower than the preset propagation threshold, the operational goals are considered to be significantly different, and propagation is not allowed. When propagation is deemed possible, the activation value is propagated to adjacent structured knowledge items and accumulated based on the association strength of weighted association edges. The activation value calculation method is to multiply the activation value of the source knowledge item by the association strength to obtain the activation value propagated to the target knowledge item. The final activation value of the target knowledge item is the sum of its initial activation value and all incoming activation values. The propagation process can be repeated in multiple rounds, with each round covering one layer of adjacent nodes in the graph structure. After multiple rounds of propagation, an activation value distribution network is formed, reflecting the comprehensive relevance of knowledge items to the current business scenario.

[0135] The structured knowledge item with the highest activation value is pushed to operators or the decision-making system. This highest-activated knowledge item is the most suitable for the current business scenario, taking into account direct matching, historical confidence, and the influence of related knowledge. The push process involves integrating the detailed content, applicable conditions, operational steps, and expected results of the knowledge item into an easily understandable and applicable format, and delivering it to the recipient through a user interface or API. The pushed knowledge items can directly guide operators' decisions and actions, improving operational accuracy and efficiency. Simultaneously, multiple knowledge items with high activation values ​​can be pushed as alternatives for operators or the decision-making system to consider and choose from.

[0136] This invention achieves intelligent recommendation based on power system knowledge by calculating the matching degree between the current business scenario and knowledge entries in the organizational memory and using the activation value propagation mechanism. This method considers constraint matching, vector similarity, historical confidence, and the correlation between knowledge entries, comprehensively evaluating the applicability of knowledge entries to the current business scenario. It can provide timely, accurate, and effective knowledge support for power system operation and maintenance, improve the safe operation level and emergency response capabilities of the power system, reduce human error, enhance the safety and stability of the power grid, and promote the effective application and inheritance of power system knowledge.

[0137] like Figure 2 As shown, the autonomous learning system based on power system agent collaboration includes:

[0138] The data collection module 201 is used to collect user action data and interactive text data during business operations.

[0139] The recognition module 202 is used to calculate the semantic coverage of interactive text data and the behavioral deviation of operation behavior data from the preset knowledge base. When the semantic coverage is lower than the preset coverage threshold or the behavioral deviation is higher than the preset deviation threshold, the context information is extracted as a candidate fragment of implicit knowledge.

[0140] The query module 203 is used to generate guiding queries for implicit knowledge candidate fragments and obtain knowledge elements based on user responses;

[0141] Alignment module 204 is used to align knowledge elements and operational behavior data in the joint semantic space through contrastive learning to generate fusion vectors as structured knowledge entries;

[0142] Storage module 205 is used to establish an organizational memory and calculate the similarity between the fusion vector of the structured knowledge entry and the vector of the existing knowledge entry in the organizational memory. Based on the vector similarity and co-occurrence frequency, weighted association edges are established in the graph structure, and the structured knowledge entry and the weighted association edges are stored in the organizational memory.

[0143] The push module 206 is used to receive the current business scenario and calculate the matching degree between the current business scenario and the fusion vector of the structured knowledge entries in the organization memory. Based on the matching degree and the activation value of the weighted association edge propagation, the push module 206 pushes the structured knowledge entry with the highest activation value.

[0144] One technical solution provided in this embodiment of the invention is an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps in any of the aforementioned methods.

[0145] One technical solution provided in this embodiment of the invention is a computer-readable storage medium storing a computer program, wherein the processor executes the computer program to implement the steps in any of the aforementioned methods.

[0146] The specific embodiments described above are preferred embodiments of the present invention and are not intended to limit the specific scope of the present invention. The scope of the present invention includes, but is not limited to, these specific embodiments. All equivalent changes made in accordance with the shape and structure of the present invention are within the protection scope of the present invention.

Claims

1. A self-learning method based on agent-based collaboration in power systems, characterized in that, include: Collect user action data and interaction text data during business operations; Calculate the semantic coverage of interactive text data and the behavioral deviation of operation behavior data from the preset knowledge base. When the semantic coverage is lower than the preset coverage threshold or the behavioral deviation is higher than the preset deviation threshold, extract contextual information as implicit knowledge candidate fragments. Generate guiding questions for implicit knowledge candidate fragments and extract knowledge elements based on user responses; Knowledge elements and operational behavior data are aligned in the joint semantic space through comparative learning to generate fusion vectors as structured knowledge entries; Establish an organizational memory and calculate the similarity between the fusion vector of the structured knowledge item and the vector of the existing knowledge item in the organizational memory. Based on the vector similarity and co-occurrence frequency, establish weighted association edges in the graph structure and store the structured knowledge item and the weighted association edges in the organizational memory. Receive the current business scenario and calculate the matching degree between the current business scenario and the fusion vector of the structured knowledge entries in the organization memory. Based on the matching degree and the weighted association edge propagation activation value, push the structured knowledge entry with the highest activation value.

2. The method according to claim 1, characterized in that, Calculate the semantic coverage of interactive text data and the behavioral deviation of operational behavior data from a preset knowledge base. When the semantic coverage is lower than a preset coverage threshold or the behavioral deviation is higher than a preset deviation threshold, extract contextual information as implicit knowledge candidate fragments, including: After mapping interactive text data to semantic space, the coverage is obtained by retrieving it from the preset knowledge base. When the operation object corresponding to the knowledge item within the coverage is different from the current operation object, the corresponding interactive text data is determined to be a valid sample. The proportion of valid samples is calculated to obtain the semantic coverage. Collect operation behavior data of different users under the same business scenario to construct an operation behavior set, count the execution frequency of operation steps in the operation behavior set to obtain the frequency distribution, extract operation steps whose execution frequency exceeds the preset frequency threshold as the benchmark operation steps, and calculate the proportion of the benchmark operation steps in the current operation behavior data to obtain the behavior deviation degree. When the semantic coverage is lower than the preset coverage threshold, the interactive text data involved in different operation objects are extracted as context information. When the behavior deviation is higher than the preset deviation threshold, the interactive text data in the time period corresponding to the baseline operation steps that did not appear are extracted as context information. The contextual information is associated with the current operation behavior data and the baseline operation steps that have not appeared to form implicit knowledge candidate fragments.

3. The method according to claim 1, characterized in that, The process of generating guided questions based on tacit knowledge candidate fragments and extracting knowledge elements based on user responses includes: Extract the operation object and execution steps from the implicit knowledge candidate fragments, and combine the attribute features of the operation object and the temporal relationship of the execution steps to form a scene description vector; Retrieve existing knowledge entries from the preset knowledge base that are less than the preset distance threshold from the scene description vector but are different from the category of the operation object, and statistically analyze the distribution of knowledge elements in the existing knowledge entries to obtain a set of knowledge element types. The difference between the knowledge element types already contained in the implicit knowledge candidate fragment and the set of knowledge element types is calculated to obtain the type to be completed. For the type to be completed, the corresponding query expression is extracted from the existing knowledge entries and the operation object that replaces the query expression is generated to produce a guiding query. The system sends guiding questions to users and receives their responses. It then parses the noun phrases, conditional clauses, and judgment statements in the user responses and associates them with the operation objects, execution steps, and execution results in the implicit knowledge candidate fragments to obtain knowledge elements. These knowledge elements include object knowledge elements, conditional knowledge elements, and verification knowledge elements.

4. The method according to claim 1, characterized in that, The knowledge elements and operational behavior data are aligned in the joint semantic space through contrastive learning to generate fusion vectors as structured knowledge entries, including: After encoding knowledge elements and operational behavior data, a joint semantic space is constructed through comparative learning. Extract the constraints corresponding to the conditional knowledge elements, locate the execution steps that satisfy the constraints in the operational behavior data as alignment anchors, and align the knowledge elements and operational behavior data in the joint semantic space based on the alignment anchors. The comparison difference between the verification knowledge element in the aligned knowledge element and the execution result in the aligned operation behavior data is calculated. When the comparison difference exceeds the preset difference threshold, the attributes that are inconsistent with the verification knowledge element are extracted from the execution result as implicit constraints to supplement the conditional knowledge element and re-encode it to obtain the supplemented knowledge element. Based on the temporal position of the alignment anchor points, the supplemented knowledge elements and the aligned operation behavior data are divided into temporal segments and the bidirectional coverage is calculated. The temporal segments with bidirectional coverage exceeding the preset coverage threshold are selected to generate a fusion vector as a structured knowledge item.

5. The method according to claim 1, characterized in that, The process involves establishing an organizational memory and calculating the similarity between the fusion vector of structured knowledge entries and the vectors of existing knowledge entries in the organizational memory. Based on vector similarity and co-occurrence frequency, weighted association edges are established in the graph structure. The structured knowledge entries and these weighted association edges are then stored in the organizational memory. An organizational memory bank is built using a graph structure, where nodes are structured knowledge entries and edges are weighted association edges; Calculate the vector distance between the fusion vector of structured knowledge entries and the vectors of existing knowledge entries in the organizational memory; The overlap between the constraints corresponding to the conditional knowledge elements of structured knowledge entries and existing knowledge entries is extracted as constraint consistency. Vector similarity is obtained by correcting the vector distance based on constraint consistency. The co-occurrence frequency is defined as the number of times that structured knowledge items and existing knowledge items are found to be consistent with each other in historical operational behavior data. Existing knowledge items with vector similarity greater than a preset similarity threshold are selected as candidate knowledge items. The vector similarity is adjusted based on the differences between the candidate knowledge items and the verification knowledge elements of the structured knowledge items. The association strength is obtained by weighting the adjusted vector similarity with the co-occurrence frequency. Based on the correlation strength, weighted association edges are established between structured knowledge entries and candidate knowledge entries in the graph structure, and the structured knowledge entries and weighted association edges are stored in the organizational memory.

6. The method according to claim 1, characterized in that, The system receives the current business scenario and calculates the matching degree between the current business scenario and the fusion vector of structured knowledge entries in the organizational memory. Based on the matching degree and the activation value propagated by the weighted association edge, it pushes the structured knowledge entries with the highest activation values, including: Receive the current business scenario and extract the constraints corresponding to the conditional knowledge elements; Iterate through the constraints of structured knowledge entries in the organization memory, and filter out the structured knowledge entries as to be matched when all the constraints of the structured knowledge entries exist in the constraints of the current business scenario. Calculate the matching degree between the fusion vector of the current business scenario and the knowledge item to be matched; The historical confidence score is calculated as the proportion of knowledge elements verified in historical operational behavior data that match the knowledge elements verified in the current business scenario. The matching score is then adjusted based on the historical confidence score. The corrected matching degree is used as the initial activation value and assigned to the knowledge entry to be matched. Extract adjacent structured knowledge entries connected by weighted association edges, and compare the verification knowledge elements of adjacent structured knowledge entries with the verification knowledge elements of the current business scenario; Whether to propagate is determined based on the differences in the verified knowledge elements. When propagating, the activation value is propagated to the adjacent structured knowledge items according to the association strength of the weighted association edges and the activation value is accumulated. Push the structured knowledge item with the highest activation value.

7. A self-learning system based on agent-based collaboration in a power system, used to implement the method described in any one of claims 1-6, characterized in that, include: The data collection module is used to collect user action data and interactive text data during business operations. The recognition module is used to calculate the semantic coverage of interactive text data and the behavioral deviation of operation behavior data from the preset knowledge base. When the semantic coverage is lower than the preset coverage threshold or the behavioral deviation is higher than the preset deviation threshold, the context information is extracted as a candidate fragment of implicit knowledge. The query module is used to generate guiding queries for implicit knowledge candidate fragments and obtain knowledge elements based on user responses; The alignment module is used to align knowledge elements and operational behavior data in the joint semantic space through contrastive learning to generate fusion vectors as structured knowledge entries. The storage module is used to establish an organizational memory and calculate the similarity between the fusion vector of the structured knowledge entry and the vector of the existing knowledge entry in the organizational memory. Based on the vector similarity and co-occurrence frequency, weighted association edges are established in the graph structure, and the structured knowledge entry and the weighted association edges are stored in the organizational memory. The push module is used to receive the current business scenario and calculate the matching degree between the current business scenario and the fusion vector of the structured knowledge entries in the organization memory. Based on the matching degree and the activation value of the weighted association edge propagation, the module pushes the structured knowledge entry with the highest activation value.

8. An electronic device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions that, when executed by a processor, implement the steps of the method as described in any one of claims 1 to 6.

Images