Superconducting power knowledge system construction method and system based on lightweight large model

Abstract

The invention relates to a superconducting power knowledge system construction method and system based on a lightweight large model. According to the method, multi-source data is utilized for unified representation, deep semantic information is obtained through a lightweight large model, knowledge extraction is achieved, and a basic knowledge graph is constructed; further forming a knowledge graph which can be dynamically updated through knowledge fusion; and in combination with real-time data, adaptive multi-hop reasoning is executed, and a diagnosis and decision result with a causal chain is generated. According to the method, the semantic understanding and reasoning ability of a lightweight large model is taken as a core, the structural expression advantage of a knowledge graph is combined, the limitation of updating lagging and fixed rule reasoning rigidness of a traditional static knowledge base is broken through, and precise cognition and causal chain inference of a complex system state are achieved. The method has universality and mobility, and is suitable for knowledge modeling and intelligent diagnosis of various electric power scenes and other complex industrial systems. The method is verified by taking a superconducting power system as an example, and the feasibility and effectiveness of the construction path are proved.

Description

Technical Field

[0001] This invention belongs to the field of superconducting power technology, and in particular relates to a method for constructing a superconducting power knowledge system based on a lightweight large model. Background Technology

[0002] With the deepening application of artificial intelligence technology in the field of power system fault diagnosis, fault analysis using a combination of knowledge graphs and large models has become an important research direction. Especially in advanced power networks with high complexity and strong coupling, such as superconducting power systems, how to construct an effective knowledge system from massive monitoring data and textual knowledge, and achieve accurate fault reasoning, is the key to ensuring their safe and stable operation.

[0003] Traditional technologies typically combine a pre-built static knowledge base with a rule engine. Specifically, a structured knowledge base is first manually constructed based on domain expert experience and historical data, defining entities such as devices, faults, and symptoms, along with their fixed relationships. When a system fault occurs, the diagnostic program, based on real-time input data, matches and infers within the static knowledge base according to predefined logical rules or finite graph traversal paths, ultimately outputting a diagnostic conclusion. This approach relies on the completeness and accuracy of prior knowledge.

[0004] However, the core knowledge base of the aforementioned existing technologies is static and difficult to update and expand automatically and in a timely manner with changes in system topology, equipment iteration, or the emergence of new failure modes, resulting in a lagging knowledge system. Secondly, the reasoning path based on fixed rules lacks adaptability and cannot dynamically adjust the depth and breadth of reasoning according to the complexity and correlation of real-time failures. When facing the "cascading failures" with spatiotemporal propagation characteristics unique to superconducting power systems, it is very easy to make misjudgments or miss key failure chains due to the rigidity of the reasoning path. Summary of the Invention

[0005] Therefore, it is necessary to provide a method and system for constructing a superconducting power knowledge system based on a lightweight large model, which can automatically construct and dynamically update knowledge and has intelligent and adaptive reasoning capabilities, in order to address the above-mentioned technical problems.

[0006] Firstly, this application provides a method for constructing a superconducting mechanical knowledge system based on a lightweight large model, including:

[0007] Feature extraction is performed based on monitoring data from superconducting power systems to generate standardized feature sets.

[0008] Model training is performed based on textual data and standardized feature sets in the power sector to generate a lightweight large model with multi-hop reasoning capabilities.

[0009] Knowledge extraction is performed based on a lightweight large model to generate an initial knowledge graph;

[0010] Knowledge fusion processing is performed based on the initial knowledge graph to generate a dynamic knowledge graph of superconducting forces.

[0011] Multi-hop reasoning is performed based on the superconducting dynamic knowledge graph and real-time fault data to generate fault diagnosis results.

[0012] In one embodiment, the model is trained based on textual data from the power sector and a standardized feature set to generate a lightweight large model with multi-hop inference capabilities, including:

[0013] A basic language model is generated by pre-training text data from the power industry.

[0014] Based on standardized feature sets, the parameters of the basic language model are optimized to generate a domain-adaptive model;

[0015] Model compression is performed based on pre-trained large models and domain adaptation models to generate lightweight large models.

[0016] In one embodiment, model compression is performed based on a pre-trained large model and a domain adaptation model to generate a lightweight large model, including:

[0017] The output probability distribution is aligned based on the output probability distribution of the pre-trained large model to obtain the optimized output layer parameters.

[0018] Feature representation alignment is performed based on the intermediate layer feature representation of the pre-trained large model to obtain the optimized hidden layer parameters;

[0019] Attention pattern alignment is performed based on the attention distribution matrix of the pre-trained large model to obtain optimized attention parameters;

[0020] The model parameters are updated based on the optimized output layer parameters, optimized hidden layer parameters, and optimized attention parameters to generate a lightweight large model; the loss function used for attention pattern alignment is:

[0021]

[0022] in, This is the loss value. To focus on the number of heads, and Let be the query matrix and key matrix of the i-th attention head in the lightweight large model, respectively. and Let be the query matrix and key matrix of the i-th attention head of the pre-trained large model, respectively. The dimension of the key vector. It is the Frobenius norm.

[0023] In one embodiment, knowledge extraction is performed based on a lightweight large model to generate an initial knowledge graph, including:

[0024] Pattern design is performed based on a domain knowledge base to generate a knowledge graph pattern layer;

[0025] Entity recognition is performed based on a lightweight large model to generate an entity set;

[0026] Relation extraction is performed based on a lightweight large model to generate a set of relation triples;

[0027] The initial knowledge graph is generated by constructing a graph based on the knowledge graph schema layer, entity set, and relation triple set.

[0028] In one embodiment, relation extraction is performed based on a lightweight large model to generate a set of relation triples, including:

[0029] Sequence labeling is performed based on text encoding representation to generate entity location information;

[0030] Relationship classification is performed based on graph neural networks to generate semantic relationships between entities;

[0031] Confidence assessment is performed based on a path ranking algorithm to generate relationship confidence scores;

[0032] The process of generating relation triples is performed based on entity location information, semantic relationships between entities, relation confidence scores, and temporal constraints, resulting in a set of relation triples.

[0033] In one embodiment, multi-hop inference processing is performed based on a superconducting dynamic knowledge graph and real-time fault data to generate fault diagnosis results, including:

[0034] Graph matching is performed based on the superconducting dynamic knowledge graph and real-time fault data to generate initial inference nodes;

[0035] Path selection is performed based on the initial inference nodes and path evaluation function to generate the selected inference path;

[0036] Anomaly detection is performed based on the selected inference path and quality monitoring indicators, and anomaly detection results are generated.

[0037] Based on the selected inference path and the anomaly detection results, the path is corrected to generate a corrected inference path.

[0038] Fault diagnosis is performed based on the corrected inference path, and fault diagnosis results are generated.

[0039] In one embodiment, path selection is performed based on the initial inference node and the path evaluation function to generate a selected inference path, including:

[0040] Based on the current inference state and the initial inference node, feature extraction is performed to generate a path feature vector;

[0041] Value is calculated based on path feature vectors and a value assessment function to generate the expected value of the path; the value assessment function is:

[0042]

[0043] in, For expected value, For query With entity semantic similarity, For path confidence level For entities The contribution of knowledge increment, For path length, , , , These are the weighting coefficients;

[0044] Path selection is performed based on the expected value of the path to generate the selected inference path.

[0045] In one embodiment, anomaly detection is performed based on path feature vectors and quality monitoring indicators to generate anomaly detection results, including:

[0046] Trend analysis is performed on the confidence sequence of the selected inference path to generate the confidence change trend;

[0047] Consistency testing is performed based on multi-source evidence data to generate an evidence consistency score;

[0048] Stagnation detection is performed based on inference progress data to generate inference progress indicators;

[0049] Anomaly detection results are generated by comprehensively evaluating the confidence trend, evidence consistency score, and reasoning progress index.

[0050] Secondly, this application also provides a superconducting power knowledge system construction system based on a lightweight large model, including:

[0051] The feature extraction module is used to extract features from monitoring data of the superconducting power system and generate a standardized feature set.

[0052] The model training module is used to train models based on text data and standardized feature sets in the power industry, generating lightweight large models with multi-hop inference capabilities.

[0053] The knowledge extraction module is used to extract knowledge based on a lightweight large model and generate an initial knowledge graph.

[0054] The knowledge fusion module is used to perform knowledge fusion processing based on the initial knowledge graph to generate a dynamic knowledge graph of superconducting power.

[0055] The fault diagnosis module is used to perform multi-hop reasoning based on the superconducting dynamic knowledge graph and real-time fault data to generate fault diagnosis results.

[0056] The aforementioned method and system for constructing a knowledge system for superconducting power based on a lightweight large model constructs a standardized feature set by processing monitoring data of the superconducting power system. This standardized feature set, along with text data from the power sector, is then used to train a lightweight large model with multi-hop reasoning capabilities. Furthermore, based on this lightweight large model, knowledge is automatically extracted from multi-source information to construct an initial knowledge graph. This initial knowledge graph is then continuously evolved into a dynamic knowledge graph through knowledge fusion processing. Finally, combined with real-time fault data, this dynamic knowledge graph supports a multi-hop reasoning process that can adaptively adjust the reasoning path. This achieves the technical effect of accurate and adaptive diagnosis of complex faults, especially cascading faults with spatiotemporal propagation characteristics. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0058] Figure 1 A flowchart illustrating a method for constructing a superconducting power knowledge system based on a lightweight large model, provided by this invention.

[0059] Figure 2 This is a flowchart illustrating a method for generating fault diagnosis results in an optional embodiment of the present invention.

[0060] Figure 3 This is a schematic diagram of the structure of a superconducting mechanical knowledge system construction system based on a lightweight large model, provided by the present invention. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0062] In one embodiment, such as Figure 1 As shown, a method for constructing a superconducting power knowledge system based on a lightweight large model is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, and to a system including both a terminal and a server, and implemented through the interaction between the terminal and the server. In this embodiment, the method includes the following steps S101 to S105:

[0063] S101. Based on the monitoring data of the superconducting power system, feature extraction is performed to generate a standardized feature set.

[0064] Optionally, the monitoring data of the superconducting power system comes from sensors deployed at key nodes of the system, covering electrical parameters, thermal parameters, and mechanical state parameters of the superconducting equipment. Electrical parameters include voltage and current; thermal parameters include equipment body temperature and cooling system temperature; and mechanical state parameters include vibration signals and displacement data. Feature extraction requires appropriate methods for different types of data: for time-series monitoring data, key features reflecting the equipment's operating status are extracted through time-domain analysis or frequency-domain analysis. Time-domain analysis includes calculating the signal's mean, variance, and peak value; frequency-domain analysis includes Fourier transform and wavelet transform. For non-time-series data, feature selection algorithms, such as analysis of variance and mutual information methods, are used to screen features strongly correlated with the equipment's state.

[0065] After feature extraction, standardization is required. The z-score standardization method is used, and the formula is: ,in, These are the standardized eigenvalues. These are the original eigenvalues. This is the mean of the feature across all monitored samples. This is the standard deviation of the feature across all monitored samples, thereby eliminating the influence of dimensional differences among different features.

[0066] S102. Based on text data and standardized feature sets in the power industry, train the model to generate a lightweight large model with multi-hop reasoning capabilities.

[0067] Optionally, the text data in the power sector includes academic literature related to superconducting power, equipment operation and maintenance records, fault handling reports, industry standards and specifications, etc., which need to be preprocessed before use: a specialized word segmentation tool for the power sector is used to identify professional terms such as "superconducting coil" and "liquid nitrogen cooling system," and stop words without actual semantic meaning are removed. Then, the words are converted into vectors using Word2Vec or BERT encoding. Model training employs transfer learning combined with distillation techniques: a general-purpose large model is pre-trained using power sector text data to master domain semantics; a standardized feature set is introduced, and an attention mechanism is used to establish the association between text vectors and numerical features; knowledge distillation is used to transfer complex model knowledge to a simplified Transformer architecture, reducing the number of encoder-decoder layers. Simultaneously, multi-hop inference tasks are introduced to train the model's inference capabilities, such as constructing samples based on a causal chain of "fault phenomenon - intermediate cause - root cause."

[0068] S103. Extract knowledge based on a lightweight large model to generate an initial knowledge graph.

[0069] Optionally, knowledge extraction revolves around three elements: entities, relationships, and attributes. Entity extraction utilizes the model's named entity recognition capabilities to identify entities such as "superconducting transformer" and "inter-turn short-circuit fault" from preprocessed text and standardized feature sets, classifying them into types such as equipment, fault, and parameter. Relationship extraction learns semantic associations between entities through the model and uses a relationship classification method to match entity pairs with relationships such as "fault association" and "parameter influence." Attribute extraction extracts key attributes of entities, such as the rated capacity and cooling method of "superconducting transformer," and associates attribute values. The extraction results are organized in RDF format into triples of "entity-relationship-entity" or "entity-attribute-attribute value" to form an initial knowledge graph.

[0070] S104. Based on the initial knowledge graph, perform knowledge fusion processing to generate a superconducting dynamic knowledge graph.

[0071] Optionally, knowledge fusion includes four stages: entity alignment, attribute fusion, conflict resolution, and incremental update. Entity alignment uses semantic similarity calculation, where the model calculates the cosine similarity of entity name descriptions and, combined with industry rules, merges synonymous entities such as "superconducting cable" and "superconducting power cable." Attribute fusion integrates the average value of multiple attribute values ​​of the same entity, such as the rated current of equipment in different records, by weighting the average value according to the accuracy of the data acquisition equipment. Conflict resolution uses the model's semantic understanding and domain rules to filter out correct information from contradictory knowledge (such as the opposite cause of the same fault). Incremental update establishes a real-time data access mechanism, where new monitoring or text data triggers the extraction and fusion process to achieve dynamic knowledge expansion.

[0072] S105. Perform multi-hop reasoning processing based on the superconducting dynamic knowledge graph and real-time fault data to generate fault diagnosis results.

[0073] Optionally, real-time fault data is preprocessed first, using the aforementioned feature extraction and standardization methods to convert data such as voltage fluctuations and temperature anomalies during a fault into standard feature vectors. Multi-hop inference employs a Graph Neural Network (GNN) algorithm, converting the dynamic knowledge graph into a graph structure, where entities are nodes, relationships are edges, and attributes are node features. Real-time feature vectors are associated with fault entity nodes in the graph, and information propagates along the relationship edges in multiple steps through the GNN message passing mechanism, such as "voltage fluctuation → associated fault phenomenon → inter-turn short circuit → fault cause → insulation aging." The matching degree between potential fault nodes and real-time features is calculated, and an attention mechanism is introduced to weight different inference paths. Finally, the fault type, root cause, involved equipment, and handling measures are selected based on the matching degree to form a diagnostic result.

[0074] The aforementioned method for constructing a superconducting power knowledge system based on a lightweight, large-scale model begins with feature extraction from superconducting power system monitoring data. It then integrates textual data from the power sector to train a lightweight multi-hop inference model. Based on the model, an initial knowledge graph is generated, which is then processed to construct a dynamic knowledge graph. Finally, real-time fault data is combined to generate diagnostic results. This method effectively addresses the problems of lagging traditional static knowledge bases and rigid, fixed-rule inference paths, significantly improving the timeliness and completeness of the superconducting power knowledge system. Simultaneously, it enhances the accuracy and timeliness of fault diagnosis, providing reliable technical support for the safe and stable operation of superconducting power systems.

[0075] In one embodiment, the model is trained based on textual data from the power sector and a standardized feature set to generate a lightweight large model with multi-hop inference capabilities, including:

[0076] S201. Based on text data in the power industry, perform pre-training processing to generate a basic language model.

[0077] Optionally, systematic preprocessing of text data in the power field needs to be carried out first: use a special word segmentation tool in the power field, such as a power term word segmentation model based on BERT, to identify professional terms such as "superconducting magnet, cryogenic refrigeration system", and剔除 common stop words without actual semantics such as "of, in" and invalid text fragments such as "under commissioning, to be detected"; then use the WordPiece word segmentation algorithm to segment the processed text into sub-word units, and use the Embedding layer to map the sub-words into vector representations of a fixed dimension. In the pre-training stage, the Transformer architecture is used as the basic model framework, and masked language model (MLM) and sentence relationship prediction (NSP) tasks are designed: the MLM task randomly masks some sub-words in the text and allows the model to predict the masked words according to the context; the NSP task judges whether two text fragments are consecutive sentences. During the training process, the AdamW optimizer is used to minimize the prediction loss through gradient descent, so that the model gradually learns the semantic logic and term associations in the power field, and finally generates a basic language model with the ability to understand the semantics of the basic field.

[0078] S202. Optimize the parameters of the basic language model based on the standardized feature set to generate a domain adaptation model.

[0079] Optionally, in the parameter optimization stage, first construct a feature-text association mapping module: associate each feature vector in the standardized feature set with the corresponding feature description text in the power field text data, such as the description text corresponding to the "superconducting coil temperature feature", and calculate the similarity weight between the feature vector and the text vector through the attention mechanism to establish the mapping relationship between numerical features and semantic information. Subsequently, embed this mapping module into the encoder layer of the basic language model, with the feature prediction task as the optimization goal: input some standardized feature vectors, let the model predict the missing feature vectors according to the learned semantic associations, and use the mean square error loss function: , where is the mean square error loss value, is the number of feature dimensions, is the true standardized feature value, is the model-predicted feature value, measuring the prediction deviation. During the optimization process, fix the underlying parameters of the basic language model and only update the parameters of the upper encoder and mapping module. Adjust the parameter weights through multiple rounds of iteration to enable the model to accurately combine the numerical information and text semantics of the standardized features, and finally generate a domain adaptation model.

[0080] S203. Perform model compression processing based on the pre-trained large model and the domain adaptation model to generate a lightweight large model.

[0081] Optionally, the pre-trained large model is a general-domain large model with the same architecture as the basic language model, i.e., containing a complete encoder and decoder layer. The domain adaptation model is an optimized model with domain adaptation capabilities. Model compression adopts a technical approach combining knowledge distillation and structured pruning: First, a "teacher-student model" training framework is constructed, using the fusion output of the pre-trained large model and the domain adaptation model as the "teacher model" output. Through weighted summation, i.e., the weights are set according to the model's accuracy in the domain task, the prediction results of the two models for the same input are integrated to form a comprehensive knowledge representation containing general knowledge and domain knowledge. The domain adaptation model is used as the "student model," and the KL divergence loss between the student model output and the teacher model output is minimized. ,in, This represents the KL divergence loss value. The probability distribution output by the teacher model. The probability distribution output by the student model enables the transfer of knowledge from the teacher model to the student model. Subsequently, structured pruning is performed: the contribution of attention heads and neurons in each layer of the student model is analyzed, and attention heads and redundant neurons with contributions below a threshold are removed, while the core feature module—text association—is retained. After compression, model accuracy is restored through fine-tuning training, ultimately generating a lightweight large model with reduced parameters and improved inference speed.

[0082] In the above embodiments, pre-training of text data in the power sector is performed first. The parameters of the basic language model are optimized through standardized feature sets. Then, compression processing is performed by combining the pre-trained large model with a domain-adaptive model to gradually build a lightweight large model. This embodiment first enables the model to master the basic semantics of the domain, then improves its adaptability to monitoring data features, and finally balances performance and resource consumption through compression. This effectively solves the problems of poor domain adaptability of general-purpose large models and high computational cost of dedicated models. The resulting lightweight large model can accurately understand superconducting power sector knowledge, integrate monitoring data features, and quickly respond to real-time task requirements.

[0083] In one embodiment, model compression is performed based on a pre-trained large model and a domain adaptation model to generate a lightweight large model, including:

[0084] S301. Align the probability distribution based on the output probability distribution of the pre-trained large model to obtain the optimized output layer parameters.

[0085] Optionally, the pre-trained large model needs to infer within a subset of subdivided scenarios in the superconducting power domain text dataset. The selected subset should align with the specific task requirements of subsequent knowledge extraction and fault diagnosis. During initial model inference, labeled samples from the corresponding subdivided scenarios should be input synchronously to ensure that the probability distribution is consistent with the benchmark. During probability distribution alignment, a gradient pruning operation is added on top of parameter adjustment via backpropagation. The pruning threshold is set based on the variance of the domain data features to avoid gradient explosion due to differences in the distribution of subdivided scenario data. Loss convergence judgment should be combined with the subdivided task accuracy on the validation set. Parameter updates should only be stopped when the difference between the losses of two rounds is less than the threshold and the accuracy fluctuation is within a very small range.

[0086] S302. Align the feature representations based on the intermediate layer feature representations of the pre-trained large model to obtain the optimized hidden layer parameters.

[0087] Optionally, the selection of intermediate layers in the pre-trained large model needs to be determined through domain task validation. By comparing the performance metrics of different layers of the Transformer architecture in the superconducting fault association identification task, a small number of layers with the best performance and stability are selected as the core alignment layers. When updating the hidden layer parameters of the initial model, a local learning rate strategy is adopted. The learning rate is adjusted for sub-layers with high correlation to fault feature extraction (determined through feature importance evaluation) to accelerate the adaptation speed of key features. After feature representation alignment, the cosine similarity between the initial model and the core alignment layer features of the large model needs to be calculated to ensure a high level of similarity and avoid local feature alignment deviations.

[0088] S303. Based on the attention distribution matrix of the pre-trained large model, perform attention pattern alignment to obtain optimized attention parameters.

[0089] Optionally, before calculating the attention distribution matrix, the attention heads of the pre-trained large model and the initial model need to be functionally adapted. Attention heads are divided into different functional types according to task requirements, and pattern matching of heads with the same function is prioritized during alignment. When calculating the loss function, a weighting factor is added to assign higher weights to the loss values ​​of key functional heads, enhancing the alignment accuracy of core attention patterns. The SGD (Stochastic Gradient Descent) optimizer needs to introduce a momentum parameter to accelerate convergence, taking into account the sparsity of domain data. After loss convergence, the matrix difference rate after alignment of each attention head is calculated to ensure that the difference rate is at a low level.

[0090] S304. Based on the optimized output layer parameters, optimized hidden layer parameters, and optimized attention parameters, the model parameters are updated to generate a lightweight large model; the loss function used for attention pattern alignment is:

[0091]

[0092] in, This is the loss value. To focus on the number of heads, and Let be the query matrix and key matrix of the i-th attention head in the lightweight large model, respectively. and Let be the query matrix and key matrix of the i-th attention head of the pre-trained large model, respectively. The dimension of the key vector. It is the Frobenius norm.

[0093] Optionally, parameter validity verification needs to be supplemented with quantitative standards, setting evaluation dimensions for the model performance after replacing output layer, hidden layer, and attention parameters respectively, to ensure that the performance of the replaced model meets the requirements of the domain task. A parameter compatibility check is added during parameter updates, calculating the similarity of the numerical distribution between the replaced parameters and the remaining parameters of the initial model to avoid adaptation conflicts between parameter modules. Temporary model fine-tuning adopts a layered fine-tuning strategy, setting different learning rates for the output layer, attention layer, and hidden layer to reduce core parameter oscillations; after fine-tuning, performance acceptance metrics are added, requiring lightweight model performance in terms of both response time and accuracy to ensure adaptability to real-time diagnostic scenarios.

[0094] In the above embodiments, based on the original three-dimensional alignment, optimizations such as subdivided scenario adaptation, core layer selection, attention head functional division, and quantitative verification standards further enhance the domain specificity and performance stability of parameter alignment. This embodiment solves potential problems such as poor data generalization, insufficient weight of key modules, and adaptation conflicts in the original alignment process. It enables the lightweight large model to retain the domain capabilities of the pre-trained large model while more accurately matching the subdivided needs of real-time knowledge extraction and fault diagnosis in superconducting power systems, further enhancing the model's deployment value in real-world scenarios.

[0095] In one embodiment, knowledge extraction is performed based on a lightweight large model to generate an initial knowledge graph, including:

[0096] S401. Design patterns based on domain knowledge base to generate knowledge graph pattern layers.

[0097] Optionally, the domain knowledge base needs to select authoritative resources in the field of superconducting power, covering equipment technical manuals, fault diagnosis standards, industry specifications, etc., and the pattern design needs to focus on the core components and business logic of the superconducting power system. During the design process, first, a core class hierarchy is defined: with "superconducting power system" as the top-level class, subclasses such as "superconducting equipment," "fault type," and "monitoring parameters" are added. Subclasses are further subdivided according to the actual system architecture, such as "superconducting equipment" being subdivided into "superconducting transformer" and "superconducting cable." Then, the attributes and relationships of the classes are defined: for the "superconducting equipment" class, attributes such as "rated power" and "cooling method" are set; for the "fault type" and the "superconducting equipment" class, domain-specific relationships such as "occurred at" and "caused by" are defined. Simultaneously, the pattern layer needs to reserve extension interfaces for attributes and relationships to adapt to any newly added equipment or fault types.

[0098] S402. Entity recognition is performed based on a lightweight large model to generate an entity set.

[0099] Optionally, during entity recognition, the lightweight large model needs to input multi-source text from the superconducting power field, such as operation and maintenance logs, fault reports, and technical documents. It should focus on handling domain-specific compound and polysemous terms: for compound terms like "superconducting magnet coil," the model needs to identify its complete entity boundary through semantic context to avoid splitting and recognizing it; for polysemous terms like "parameter," it needs to determine its specific meaning and classify it based on the text context, such as "monitoring parameters" and "equipment parameters." After recognition, entities are classified and mapped, matching each entity to the corresponding class in the knowledge graph pattern layer. For example, "liquid nitrogen cooling system" is matched to the subclass "superconducting equipment - cooling device." Duplicate entities are also removed, such as merging "superconducting power cable" and "superconducting cable" into the same entity.

[0100] S403. Extract relations based on a lightweight large model and generate a set of relation triples.

[0101] Optionally, relation extraction needs to focus on semantic associations specific to the superconducting power domain. Lightweight large models need to capture two types of core relations from the text: one is static relations between entities (such as the "attribution" relationship between "superconducting transformer" and "substation"), and the other is dynamic relations (such as the "occurrence" relationship between "inter-turn short-circuit fault" and "superconducting coil," and the "symptom prediction" relationship between "temperature anomaly" and "inter-turn short-circuit fault"). During extraction, the model needs to filter invalid relations using domain semantic logic. For example, the comparison between "superconducting equipment" and "ordinary transformer" mentioned incidentally in the text is removed because it is not a core domain relation, and the extracted relations are organized into triples in the format of "entity 1-relationship-entity 2." Simultaneously, semantic validation is performed on the extracted triples to ensure that the relations conform to the relation types defined in the knowledge graph schema layer. For example, "symptom prediction" needs to match the preset relationship between the schema layer's "fault symptoms" and "fault types."

[0102] S404. Construct the knowledge graph based on the knowledge graph schema layer, entity set, and relation triple set to generate the initial knowledge graph.

[0103] Optionally, the construction process first confirms the entity-schema layer mapping: It verifies the correctness of the match between each entity in the entity set and the schema layer class. If an entity exists, such as the newly added "superconducting current limiting device" which has no corresponding class, the schema layer subclass needs to be temporarily expanded and recorded for later optimization. Next, it aligns the triplet-schema layer relationships: ensuring that the relationship of each triplet is consistent with the preset relationship in the schema layer. Differences in relationship descriptions, such as "cause" versus "leads to" in the schema layer, are uniformly mapped. Then, a graph database storage structure is used to build the knowledge graph: entities are used as nodes, and node attributes are associated with class attributes defined in the knowledge graph schema layer, such as the "superconducting transformer" node being associated with the "rated power" attribute value. Triplet relationships are used as edges to connect the corresponding entity nodes. After construction, a structure verification is performed to check for isolated nodes (entities without any relational edges) or invalid edges (triples with mismatched relationships), and preliminary cleanup is carried out.

[0104] In the above embodiments, through a domain-adaptive pattern layer design, lightweight large models are used to accurately extract domain entities and core relationships. Then, an initial knowledge graph is generated through structured construction and multi-dimensional verification. This embodiment solves the problems of pattern layer generalization, inaccurate entity recognition, and irrelevance of relationship extraction in traditional knowledge graph construction, making the initial knowledge graph conform to the business logic and semantic characteristics of the superconducting power domain, and also possessing a standardized structure and consistent hierarchical relationships.

[0105] In one embodiment, relation extraction is performed based on a lightweight large model to generate a set of relation triples, including:

[0106] S501. Perform sequence labeling based on text encoding representation to generate entity location information.

[0107] Optionally, the text encoding should employ an encoding method adapted to the terminology structure of the superconducting power field. For multi-character professional terms within the field, such as "superconducting magnet cooling system," the encoding process should preserve the correlation characteristics of characters within the term to avoid term segmentation deviations due to overly coarse encoding granularity. Sequence labeling adopts a labeling system tailored to the entity types within the field. In addition to conventional "equipment, fault" tags, new sub-tags such as "superconducting-specific components, cryogenic system parameters" are added to ensure that the labeling results can accurately distinguish field-specific entities. During the labeling process, the labeling results are validated in real-time using a domain terminology dictionary. When discrepancies exist between the labeled entity boundaries and the term boundaries in the dictionary, a backtracking correction of the encoding vector is triggered to ensure that the generated entity location information (including the entity's start and end indices in the text) perfectly matches the actual expression of the domain terminology.

[0108] S502. Relationship classification is performed based on graph neural networks to generate semantic relationships between entities.

[0109] Optionally, when constructing the input graph of the graph neural network, entities in the superconducting power domain need to be grouped according to their functional relationships. For example, "superconducting cable, cable insulation layer, insulation aging fault" should be grouped into the same functional group, so that the initial association of nodes in the graph conforms to the actual system structure of the domain. The message passing mechanism of the graph neural network needs to incorporate domain semantic rules. For example, when calculating the message weight between nodes, priority should be given to strengthening the weight of core domain associations such as "equipment-its specific parameters" and "fault-corresponding symptoms," while weakening the message passing strength between unrelated entities. The relation classification output needs to cover the semantic relation types specific to the superconducting power domain, such as "abnormalities in the cryogenic system lead to equipment failure" and "superconducting coil parameters affect the system's power transmission efficiency," rather than simply using broad relation definitions such as "association" and "belonging to."

[0110] S503. Confidence assessment is performed based on the path sorting algorithm to generate relationship confidence scores.

[0111] Optionally, the path selection in the path ranking algorithm should be based on the design of potential related paths in the knowledge graph of the superconducting power field. Paths with validated validity within the domain should be prioritized as the ranking benchmark, such as "abnormal equipment operating parameters → component performance degradation → equipment failure." Redundant paths without practical domain significance, such as "equipment model → document number → failure type," should be excluded. When calculating path weights, domain expert experience should be introduced to set path importance coefficients, assigning higher weights to core related paths to ensure that the path contribution aligns with the actual domain correlation strength. During the confidence assessment process, domain-specific confidence threshold judgment logic should be set. When the path support of a certain type of relationship, such as "failure of a superconducting-specific component," reaches a certain condition, the threshold can be appropriately adjusted to retain more high-value, low-frequency relationships.

[0112] S504. Based on entity location information, semantic relationships between entities, relationship confidence scores, and temporal constraints, perform relation triple generation processing to generate a set of relation triples.

[0113] Optionally, timing constraints need to be set in conjunction with the time characteristics of the superconducting power system's operation. For example, regarding "fault development timing," the "fault precursor" entity relationship must be generated earlier than the "fault occurrence" entity relationship to avoid triples with contradictory timing logic, such as "equipment failure occurs → cryogenic system parameter anomaly." A dual "confidence-timing" screening mechanism is used when generating triples. Relationships with confidence levels below the domain's set threshold are first filtered out, and then remaining relationships that do not conform to timing constraints are eliminated. This ensures that the final generated triples possess both high reliability and conform to the time patterns of the superconducting system's operation, such as the format "entity 1 - semantic relationship - entity 2." Simultaneously, the generated triples are organized according to domain functional modules, such as "superconducting equipment module, cryogenic control system module," so that the structure of the triple set matches the actual architecture of the superconducting power system.

[0114] In the above embodiment, the location of domain entities is first accurately located through sequence labeling, then the specific semantic relationships between entities are classified using graph neural networks, followed by the evaluation of relationship confidence through path sorting, and finally, structured triples are generated by combining temporal constraints. This embodiment effectively solves the shortcomings of general triple generation methods in terms of domain adaptability and logical consistency, and the final generated set of relational triples can accurately and completely carry the core knowledge of the superconducting field.

[0115] In an optional embodiment, such as Figure 2 As shown, multi-hop inference processing is performed based on the superconducting dynamic knowledge graph and real-time fault data to generate fault diagnosis results, including:

[0116] S601. Graph matching is performed based on the superconducting dynamic knowledge graph and real-time fault data to generate initial inference nodes.

[0117] Optionally, the superconducting power dynamic knowledge graph includes various types of entities and their relationships, such as equipment, faults, and operating parameters. Before matching, feature mapping needs to be performed on the real-time fault data to establish a correspondence between the electrical and thermal features in the fault data and the attribute fields of the entities in the graph. For example, the fault data "abnormal coil temperature" corresponds to the "operating temperature" attribute of the "superconducting coil" entity in the graph. During the graph matching process, an attribute weight hierarchical strategy is adopted, prioritizing the matching of core attributes directly related to the fault, such as fault type and equipment identifier, and then matching auxiliary attributes, such as runtime and maintenance records. At the same time, the dynamic update time of the entities in the graph is taken into account, and entities with the latest attributes matching the fault data are selected as candidate nodes. Through this hierarchical matching strategy, the entity most closely related to the real-time fault can be accurately located, avoiding the initial node from deviating from the actual fault scenario due to disordered attribute matching.

[0118] S602. Based on the initial inference node and the path evaluation function, perform path selection and generate the selected inference path.

[0119] Optionally, the design of the path evaluation function needs to incorporate the domain logic of fault propagation in superconducting power systems. Core evaluation dimensions include the relevance of entity relationships within the path (e.g., direct causal relationships like "fault trigger - parameter anomaly" have higher weight than indirect relationships), the reasonableness of the path length (avoiding excessively long paths that lead to inference redundancy, while excluding fault propagation links that might be missed by excessively short paths), and the dynamic activity of entities within the path (prioritizing paths composed of entities with recent attribute updates in the dynamic graph). Path selection starts from the initial inference node, traversing its directly associated entities to form a candidate path set. The evaluation function calculates the comprehensive score of each candidate path, and paths with scores higher than a threshold are selected as the chosen inference paths.

[0120] S603. Based on the selected inference path and quality monitoring indicators, perform anomaly detection and generate anomaly detection results.

[0121] Optionally, quality monitoring indicators are constructed around the logical consistency and attribute timeliness of the path: Logical consistency indicators are used to detect contradictions in entity relationships within the path, such as conflicting relationships like "device A is fault-free" and "device A triggered a fault alarm," which are verified by comparing the relationship's rationality with the domain rule base in the dynamic graph; attribute timeliness indicators are used to check whether entity attributes in the path match the time dimension of real-time fault data, such as if the "latest status" attribute of an entity in the path is updated earlier than the fault occurrence time, indicating a timeliness anomaly. During detection, indicators are verified node-by-node and relationship-by-relationship for each selected path, marking path segments with logical contradictions or timeliness anomalies, forming detection results that include the anomaly type and location.

[0122] S604. Based on the selected inference path and the anomaly detection results, perform path correction to generate a corrected inference path.

[0123] Optionally, path correction should adopt differentiated strategies based on the anomaly type: For logically contradictory anomalies, refer to historical valid fault paths and domain expert rules in the dynamic graph, delete contradictory relationship nodes and add logically consistent intermediate association nodes. For example, if the path has a "fault-no cause" contradiction, add reasonable cause nodes such as "insulation aging". For attribute timeliness anomalies, retrieve the latest attribute values ​​of the corresponding entities in the dynamic graph to update the path nodes, and re-verify the logical coherence of the updated path. During the correction process, the adjusted path needs to be verified again by quality monitoring indicators to ensure that the corrected path has no new anomalies and conforms to the fault propagation logic. Through this targeted correction operation, the defects of the selected path can be repaired, the deficiencies of the initial path in terms of logical completeness and information timeliness can be made up for, and a reasoning path that is more in line with the actual fault scenario can be formed.

[0124] S605. Perform fault diagnosis based on the corrected inference path and generate fault diagnosis results.

[0125] Optionally, the diagnostic process requires extracting core diagnostic information from the revised reasoning path: identifying the fundamental fault entity at the path's starting point (e.g., "inter-turn short circuit in a superconducting transformer"), tracing the fault propagation chain within the path (e.g., "inter-turn short circuit → localized overheating → overload of the cooling system"), and determining the scope of related equipment affected by the fault. Simultaneously, by combining the intensity of abnormal features in real-time fault data, the severity of the fault is graded, such as classifying fault levels based on the degree to which parameters deviate from normal ranges. This is then correlated with historical handling suggestions for similar faults in the dynamic graph, forming a diagnostic result that includes the fault type, root cause, scope of impact, severity level, and initial response direction.

[0126] In the above embodiments, fault diagnosis is finally completed through graph matching of superconducting power dynamic knowledge graph and real-time fault data, followed by path selection, anomaly detection, and path correction. This embodiment addresses potential issues such as node deviations, invalid paths, and logical contradictions during reasoning by implementing targeted operations at each step, such as hierarchical matching to ensure accurate starting points, domain assessment to screen effective paths, indicator verification to identify defects, difference correction to improve paths, and information extraction to generate diagnostics. Ultimately, this achieves accurate and comprehensive fault diagnosis of superconducting power systems, adapting to their highly complex and dynamic operational characteristics.

[0127] In one embodiment, path selection is performed based on the initial inference node and the path evaluation function to generate a selected inference path, including:

[0128] S701. Based on the current inference state and the initial inference node, feature extraction is performed to generate a path feature vector.

[0129] Optionally, the current inference state includes partially constructed inference path fragments, temporary associations between nodes, and incompletely matched entity attribute states. This state needs to be structured first to clarify the associated entity pairs, the types of relationships to be supplemented, and the attributes of the initial inference node that were not matched. During feature extraction, two core pieces of information need to be integrated: first, the key attribute features of the initial inference node, such as device type and historical fault record attributes, which are converted into fixed-dimensional attribute feature sub-vectors through vector encoding; second, the dynamic features of the current inference state, such as the distribution of relationship types in the formed path fragments and the association strength of the attributes to be matched, which are converted into state feature sub-vectors using statistical encoding. Subsequently, the attribute feature sub-vectors and state feature sub-vectors are fused according to domain weight ratios to form a path feature vector containing the essential attributes of the initial node and the dynamic associations of the current inference.

[0130] S702. Calculate the expected value of the path based on the path feature vector and the value assessment function; the value assessment function is:

[0131]

[0132] in, For expected value, For query With entity semantic similarity, For path confidence level For entities The contribution of knowledge increment, For path length, , , , These are the weighting coefficients.

[0133] Optionally, For query With entity Semantic similarity is used to quantify the degree of semantic matching between two things. Typically, this is for real-time fault queries, such as "superconducting coil temperature anomaly associated fault," physical... This refers to superconducting forces entities in the reasoning path, such as "inter-turn short circuits in superconducting coils." Its calculation formula uses cosine similarity, i.e. ,in, The query text was generated by encoding a pre-trained BERT model in the power industry. The semantic vector is 3D, and the encoding process preserves the semantic associations of technical terms such as "superconductivity" and "temperature anomaly". For entities The name and attribute description text, such as "inter-turn short circuit in superconducting coil: accompanied by local overheating and insulation layer damage", are generated by encoding the same model. 3D semantic vector; The dot product of two vectors indicates that the vectors are in the same direction. , These are the L2 norms of the two vectors, reflecting the vector magnitudes. The result of this formula is in the range of [0,1]. The closer the value is to 1, the closer the semantic match between the query and the entity. For example, the semantic similarity between the query "abnormal temperature" and the entity "inter-turn short circuit" will be significantly higher than that between the query and the entity "cooling system operating normally", because the former's attribute description has a direct semantic relationship with the query.

[0134] For path The confidence level is used to quantify the reliability of the relationships between each link in the reasoning path. Composed of an "entity-relationship-entity" chain structure, such as superconducting coil → inter-turn short circuit → local overheating → cooling system overload, each relationship corresponds to a confidence score generated during the knowledge extraction stage, originating from the lightweight large model's judgment of the relationship's authenticity. Its calculation formula is the arithmetic mean of the confidence scores of all relationships within the path, i.e. ,in, For path The total number of relations contained, such as the example path above containing 3 relations, i.e. , For the first The confidence level of each relationship, ranging from [0,1], is determined by the probability value output when the lightweight large model extracts the relationship. The higher the probability, the more reliable the relationship. This formula avoids the excessive influence of a single low-confidence relationship on the overall path by averaging the reliability of each link. The result ranges from [0,1], and the higher the value, the more reliable the relationship chain of the path. For example, it includes "superconducting coil → inter-turn short circuit" ( "Inter-turn short circuit → local overheating" "Local overheating → Cooling system overload" The confidence level of a path involving three relationships is the arithmetic mean of the three relationships, which comprehensively reflects the overall reliability level of the path.

[0135] For entities The knowledge increment contribution is used to quantify the entity. The extent to which new domain knowledge is added to current reasoning, including newly added fault features of entities, previously uncollected relationships, and updated operational parameter constraints, all derived from real-time updates of the superconducting dynamic knowledge graph. Its calculation formula is a weighted sum of timeliness weight and knowledge increment weight, i.e. ,in, This is a timeliness weight, with a value range of [0,1]. For entities The timeliness quantification value is calculated using the following formula: , For the current reasoning time, For entities In dynamic knowledge graphs, the most recent update time uses a consistent time unit to ensure consistent computational dimensions. The result range is [0,1]. The closer the entity update time is to the current inference time, the better. The closer the value is to 1; For entities The quantitative value of knowledge increment is calculated using the following formula: , For entities Compared to the last update, this update includes more new attributes or relationships, such as the addition of the "fault trigger threshold in low-temperature environments" attribute. Add 1, For entities The current total number of attributes and relationships, The result range is [0,1]. The higher the percentage of newly added knowledge, the better. The larger the value; The final result ranges from [0,1]. Higher values ​​indicate greater supplementary knowledge value the entity provides for the current reasoning. For example, the "superconducting transformer" entity, which was recently updated and added with fault-related attributes, has a higher value. The value will be significantly higher than that of entities that have not been updated for a long time and have no new knowledge, providing richer decision-making basis for reasoning.

[0136] For path The length of the path is used to quantify the redundancy of the path. An excessively long path can easily introduce irrelevant entities, such as "cooling system overload → computer room temperature and humidity change → air conditioning operation status," causing inference to deviate from the core fault. An excessively short path may miss critical propagation links, such as only including "superconducting coil → cooling system fault" without including the intermediate link of "inter-turn short circuit." Its calculation formula is the difference between the number of entities in the path and 1, i.e. ,in, For path The total number of entities included, such as the path "superconducting coil → inter-turn short circuit → local overheating" containing 3 entities. ,but This formula directly reflects the number of relations in the path, that is, the path length = the number of relations, and the result is a non-negative integer, which can objectively quantify the complexity of the path.

[0137] In the actual value calculation process, it is necessary to first extract the original data of the above parameters through the path feature vector, such as parsing the query text and entity text, the relationship confidence score contained in the path, entity update records, and the number of entities in the path from the feature vector, and then substituting them into the formula of each parameter to calculate the value. , , , The specific numerical values ​​are then weighted and summed according to the value assessment function to obtain the expected path value. The overall approach quantifies key parameters to avoid the influence of subjective judgments on path value assessment, ensuring the generated results are accurate and accurate. It can objectively reflect the adaptability of the path to superconducting power fault reasoning.

[0138] S703. Based on the expected value of the path, perform path selection processing to generate the selected inference path.

[0139] Optionally, path selection requires first sorting all candidate paths by their expected value in descending order, and then selecting paths with values ​​higher than a set baseline as a preliminary candidate set. The baseline is determined based on the value distribution of historically valid inference paths in the superconducting power field. Within the preliminary candidate set, path diversity needs further consideration: when multiple paths have similar values, paths covering different fault propagation directions should be prioritized, such as internal equipment fault paths and system-level cascading fault paths, to avoid missing fault-related information due to path homogenization. Simultaneously, the path with the highest value in the preliminary candidate set undergoes logical verification, referencing the domain rule base of the superconducting power dynamic knowledge graph, to ensure that the path has no obvious logical contradictions before being included in the selected inference path; if slight logical deviations exist, fine-tuning based on value is required, such as reducing the confidence weight of contradictory segments before determining whether to include them.

[0140] In the above embodiments, starting with feature fusion extraction from the current inference state and the initial node, the comprehensive value of the path is quantified through a multi-dimensional value evaluation function, and then the final inference path is determined through a multi-layer selection strategy. This embodiment solves the problems of single feature dimension, one-sided value evaluation, and insufficient path coverage in traditional path selection. By combining the real-time inference state and domain priority settings, the selected inference path can not only closely meet the target requirements of superconducting power fault diagnosis, but also have reliable and comprehensive attributes, further adapting to the highly complex fault inference scenarios of superconducting power systems.

[0141] In one embodiment, anomaly detection is performed based on path feature vectors and quality monitoring indicators to generate anomaly detection results, including:

[0142] S801. Perform trend analysis based on the confidence sequence of the selected inference path to generate the confidence change trend.

[0143] Optionally, the confidence sequence of the selected inference path is a set of confidence scores ordered according to the fault propagation logic. , Indicates the first The confidence scores of each path node are calculated using a sliding window method combined with quantitative indicators for trend analysis. First, the sliding window size is defined, adaptively determined based on the total number of path nodes. Then, the slope of the confidence score trend within each window is calculated by sliding the window, using the following formula: ,in The average index of nodes within the window. The mean of the built-in confidence level for the window. This reflects the rate of change of the window's built-in confidence level, with positive indicating an increase, negative indicating a decrease, and near-zero indicating stability. It also calculates the amplitude of fluctuation in the window's built-in confidence level. The stability of the quantified confidence level. This is achieved through comparison. Compared with preset trend threshold , With fluctuation threshold ,like (Critical node window) is then marked as "Reliability Decline Trend". This is then marked as "abnormal fluctuation trend," and finally, all window analysis results are integrated to generate a confidence level change trend.

[0144] S802. Perform consistency checks based on multi-source evidence data and generate evidence consistency scores.

[0145] Optionally, multi-source evidence data includes real-time monitoring feature vector X and historical case feature vector sets. The domain rule constraint set R. Consistency testing consists of three steps: The first step calculates the similarity between real-time data and historical cases using the cosine similarity formula: ( Take the maximum value. The second step is to verify the fit between real-time data and domain rules, and define the rule satisfaction level. The relevant rules are matched and filtered from R according to fault type; the third step is to generate a consistency score by weighted summation, with the formula as follows: ,in, (Real-time and historical matching weights) and (Rule fit weight) is a non-negative coefficient set based on domain experience, and . The value range is [0,1], and the higher the score, the stronger the consistency of the multi-source evidence.

[0146] S803. Based on the inference progress data, perform stall detection and generate inference progress indicators.

[0147] Optionally, the reasoning progress data includes the preset total number of reasoning stages. Number of completed steps The current stage has a preset standard duration. Actual time spent in the current stage Stagnation detection is achieved through two core quantitative indicators: one is the progress completion rate. The first is the overall reasoning progress; the second is the deviation rate of time consumption in each step. Quantify the efficiency deviation of the current link, that is This indicates that the time taken exceeds the standard; the larger the deviation, the more serious the deviation. The inference progress indicator uses a comprehensive quantitative value: ,in, Assigned progress weights. Preset progress threshold. ,like and If so, it is determined to be a reasoning standstill.

[0148] S804. Based on the confidence trend, evidence consistency score and reasoning progress index, a comprehensive evaluation is conducted to generate anomaly detection results.

[0149] Optionally, the comprehensive evaluation uses a weighted summation formula to integrate the results of the first three quantitative steps, defining a comprehensive anomaly score: ,in: The value is a quantitative measure of trend anomalies, derived from... and Transformation hour, =1, otherwise =0; 1−C is the quantitative value of evidence inconsistency, where C is the consistency score. The larger the value, the more inconsistent the evidence. This is a quantitative value for schedule anomalies. and hour, =1, otherwise =0; Weights are assigned to different anomaly types, based on the impact of domain-specific failures, such as weights for reliability anomalies. The highest, and Preset abnormal threshold (High priority) and (Low priority), if If it is judged as a "high-priority anomaly," such as a sudden drop in confidence and contradictory evidence; if If it is judged as "medium priority anomaly", such as only minor inconsistencies in evidence; if The result is determined to be "no anomaly". The final result generates a comprehensive anomaly score, including anomaly level (high-priority anomaly, medium-priority anomaly, no anomaly), anomaly type (trend anomaly, evidence anomaly, progress anomaly). This comprehensive quantitative assessment integrates multi-dimensional anomaly indicators to achieve accurate classification and location of anomalies in the reasoning process, avoiding the oversights of single-dimensional assessments and ensuring the objectivity and comprehensiveness of anomaly detection.

[0150] In the above embodiments, analysis is conducted from three core dimensions: the confidence trend of the reasoning path, the consistency of multi-source evidence, and the progress of the reasoning process. Then, a comprehensive evaluation is performed by associating the reasoning value of each dimension, forming an anomaly detection chain of single-dimensional analysis followed by multi-dimensional correlation evaluation. This embodiment, through domain-adaptive design at each step, captures anomaly clues at the reliability, evidence, and process levels during the reasoning process. Finally, a comprehensive evaluation integrates these clues to form a complete judgment. This effectively solves the problem of traditional anomaly detection focusing only on a single dimension and being disconnected from the reasoning logic, ensuring accurate identification of various anomalies in the superconducting power reasoning process and guaranteeing the reliability of the reasoning results and the efficiency of the reasoning process.

[0151] The aforementioned method and system for constructing a knowledge system for superconducting power based on a lightweight large model first extracts standardized features from monitoring data of the superconducting power system. This is then combined with text data from the power sector for training, and optimized through parameter alignment and model compression to create a lightweight large model with multi-hop reasoning capabilities. Next, knowledge is extracted from this model as its core, forming a dynamic knowledge graph through pattern design, entity recognition, relation extraction, and graph construction. Finally, based on the dynamic graph and real-time fault data, graph matching is used to determine the starting point of reasoning, path selection, and optimization of the reasoning process. Furthermore, confidence trend analysis, multi-source evidence consistency testing, and reasoning stagnation detection are introduced to monitor reasoning anomalies. This technical solution addresses the lag problem of traditional static knowledge bases by leveraging the incremental update characteristics of the dynamic knowledge graph. It avoids the rigidity of fixed-rule reasoning paths by utilizing the multi-hop reasoning and dynamic path optimization capabilities of the lightweight large model. Combined with a reasoning anomaly detection mechanism, it reduces misjudgments of cascading faults in the superconducting power system and the omission of critical fault chains, effectively ensuring the safe and stable operation of the superconducting power system. In addition, this technical solution has strong universality and transferability, and can provide a general technical paradigm for the construction of knowledge graphs in various power sub-fields and complex industrial systems. It is not limited to the accurate diagnosis of cascade faults in superconducting power systems, but can also support the knowledge-based modeling and intelligent decision-making of complex problems in multiple fields, providing reusable core technical support for the dynamic and intelligent construction of domain knowledge graphs.

[0152] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0153] Based on the same inventive concept, this application also provides a system for implementing the above-mentioned method for constructing a superconducting power knowledge system based on a lightweight large model. The solution provided by this system is similar to the implementation scheme described in the above method. Therefore, the specific limitations of one or more embodiments of the system for constructing a superconducting power knowledge system based on a lightweight large model provided below can be found in the limitations of the method for constructing a superconducting power knowledge system based on a lightweight large model described above, and will not be repeated here.

[0154] In one exemplary embodiment, such as Figure 3As shown, a superconducting power knowledge system construction system 10 based on a lightweight large model is provided, including:

[0155] Feature extraction module 11 is used to extract features based on monitoring data of the superconducting power system and generate a standardized feature set;

[0156] Model training module 12 is used to train models based on text data and standardized feature sets in the power field, generating lightweight large models with multi-hop inference capabilities.

[0157] The knowledge extraction module 13 is used to extract knowledge based on a lightweight large model and generate an initial knowledge graph.

[0158] Knowledge fusion module 14 is used to perform knowledge fusion processing based on the initial knowledge graph to generate a superconducting dynamic knowledge graph.

[0159] The fault diagnosis module 15 is used to perform multi-hop reasoning based on the superconducting dynamic knowledge graph and real-time fault data to generate fault diagnosis results.

[0160] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the method for constructing a superconducting power knowledge system based on a lightweight large model as described above.

[0161] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the method for constructing a superconducting power knowledge system based on a lightweight large model as described above.

[0162] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0163] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for constructing a superconducting power knowledge system based on a lightweight large model, characterized in that, The method includes: Feature extraction is performed based on monitoring data from superconducting power systems to generate standardized feature sets. Model training is performed based on text data from the power sector and the standardized feature set to generate a lightweight large model with multi-hop reasoning capabilities. Knowledge extraction is performed based on the aforementioned lightweight large model to generate an initial knowledge graph; Based on the initial knowledge graph, knowledge fusion processing is performed to generate a superconducting dynamic knowledge graph. Based on the superconducting dynamic knowledge graph and real-time fault data, multi-hop reasoning is performed to generate fault diagnosis results.

2. The method according to claim 1, characterized in that, The process of training a model based on textual data from the power sector and the standardized feature set to generate a lightweight, large model with multi-hop inference capabilities includes: A basic language model is generated by pre-training the text data in the power sector. Based on the standardized feature set, the parameters of the basic language model are optimized to generate a domain-adaptive model; The lightweight large model is generated by performing model compression processing based on the pre-trained large model and the domain adaptation model.

3. The method according to claim 2, characterized in that, The process of compressing the pre-trained large model and the domain adaptation model to generate the lightweight large model includes: Based on the output probability distribution of the pre-trained large model, probability distribution alignment is performed to obtain the optimized output layer parameters; Based on the intermediate layer feature representations of the pre-trained large model, feature representation alignment is performed to obtain optimized hidden layer parameters; Attention pattern alignment is performed based on the attention distribution matrix of the pre-trained large model to obtain optimized attention parameters; The model parameters are updated based on the optimized output layer parameters, the optimized hidden layer parameters, and the optimized attention parameters to generate the lightweight large model; the loss function used for the attention pattern alignment processing is: in, The loss value. To focus on the number of heads, and Let be the query matrix and key matrix of the i-th attention head in the lightweight large model, respectively. and Let be the query matrix and key matrix of the i-th attention head of the pre-trained large model, respectively. The dimension of the key vector. It is the Frobenius norm.

4. The method according to claim 1, characterized in that, The step of extracting knowledge based on the lightweight large model to generate an initial knowledge graph includes: Pattern design is performed based on a domain knowledge base to generate a knowledge graph pattern layer; Entity recognition is performed based on the aforementioned lightweight large model to generate an entity set; Based on the aforementioned lightweight large model, relation extraction is performed to generate a set of relation triples; The initial knowledge graph is generated by constructing a graph based on the knowledge graph pattern layer, the entity set, and the relation triple set.

5. The method according to claim 4, characterized in that, The step of extracting relations based on the lightweight large model and generating a set of relation triples includes: Sequence labeling is performed based on text encoding representation to generate entity location information; Relationship classification is performed based on graph neural networks to generate semantic relationships between entities; Confidence assessment is performed based on a path ranking algorithm to generate relationship confidence scores; Based on the entity location information, the semantic relationship between the entities, the relationship confidence score, and the temporal constraints, a relationship triplet generation process is performed to generate the relationship triplet set.

6. The method according to claim 1, characterized in that, The multi-hop reasoning process based on the superconducting dynamic knowledge graph and real-time fault data to generate fault diagnosis results includes: Graph matching is performed based on the superconducting dynamic knowledge graph and the real-time fault data to generate initial inference nodes; Based on the initial inference node and the path evaluation function, a path selection is performed to generate the selected inference path; Anomaly detection is performed based on the selected inference path and quality monitoring indicators to generate anomaly detection results. Based on the selected inference path and the anomaly detection result, the path is corrected to generate a corrected inference path; Fault diagnosis is performed based on the revised inference path, and the fault diagnosis result is generated.

7. The method according to claim 6, characterized in that, The step of selecting a path based on the initial inference node and the path evaluation function to generate a selected inference path includes: Based on the current inference state and the initial inference node, feature extraction is performed to generate a path feature vector; Based on the path feature vector and the value assessment function, a value calculation is performed to generate the expected value of the path; the value assessment function is: in, For expected value, For query With entity semantic similarity, For path Confidence level, For entities The contribution of knowledge increment, For path length, , , , These are the weighting coefficients; Based on the expected value of the path, a path selection process is performed to generate the selected inference path.

8. The method according to claim 6, characterized in that, The anomaly detection based on the path feature vector and quality monitoring indicators, and the generation of anomaly detection results, include: Based on the confidence sequence of the selected inference path, a trend analysis is performed to generate a confidence change trend; Consistency testing is performed based on multi-source evidence data to generate an evidence consistency score; Stagnation detection is performed based on inference progress data to generate inference progress indicators; The anomaly detection result is generated by comprehensively evaluating the confidence change trend, the evidence consistency score, and the reasoning progress index.

9. A superconducting power knowledge system construction system based on a lightweight large model, characterized in that, The system includes: The feature extraction module is used to extract features from monitoring data of the superconducting power system and generate a standardized feature set. The model training module is used to train the model based on text data in the power field and the standardized feature set to generate a lightweight large model with multi-hop inference capabilities. The knowledge extraction module is used to extract knowledge based on the lightweight large model and generate an initial knowledge graph. The knowledge fusion module is used to perform knowledge fusion processing based on the initial knowledge graph to generate a superconducting dynamic knowledge graph. The fault diagnosis module is used to perform multi-hop reasoning processing based on the superconducting dynamic knowledge graph and real-time fault data to generate fault diagnosis results.

Images