Substation maintenance data management method and system based on data analysis

Abstract

The invention belongs to the technical field of data analysis. The method comprises the following steps of: converting discrete textualized technical entities and relationships in a maintenance technical document into a group of mathematical constraint sets with internal logic by constructing a semantic geometric conformation library, so that a computable geometric representation is constructed for each possible technical semantic state of equipment; the establishment of the evolution path set further explicitly models an empirical rule contained in historical maintenance data into evolution logic and probability between different technical semantic states, and reveals a typical mode and an internal mechanism of equipment state recession; by calculating the docking affinity of the real-time data probe and the semantic geometric conformation, the current complex state, formed by multi-dimensional sensor data, of the equipment can be accurately mapped to a specific semantic state; this state is then projected to a set of evolutionary paths. The maintenance instruction generated by the invention not only has timeliness, but also comprises traceability explanation of a fault evolution path.

Description

Substation Maintenance Data Management Method and System Based on Data Analysis Technical Field

[0001] This invention belongs to the field of data analysis technology, specifically relating to a substation maintenance data management method and system based on data analysis. Background Technology

[0002] In modern power grid systems, substations serve as critical hubs, and their safe and stable operation is of paramount importance. To ensure equipment reliability, maintenance departments accumulate massive amounts of multi-source heterogeneous data, mainly comprising two categories: one is structured time-series data collected in real time by sensors and monitoring systems, presented as precise numerical values, such as current, voltage, temperature, and vibration signals; the other is unstructured technical documents in natural language form, containing a wealth of engineering principles and expert experience, such as equipment design manuals, historical maintenance work orders, and standard operating procedures. Currently, data-driven intelligent maintenance technology aims to utilize this data to achieve early warning and accurate diagnosis of potential equipment faults. However, existing technical methods typically process these two types of data separately or only perform superficial correlations, for example, using machine learning models to analyze statistical patterns in structured data or using natural language processing technology to retrieve information from text. This approach ignores the complexity of equipment fault mechanisms, leading to a core technical bottleneck for existing diagnostic systems when facing practical engineering problems: how to resolve the deep-seated contradiction between the contextual dependence of fault characteristics in equipment operating data and the static nature of diagnostic models.

[0003] This contradiction stems from two interconnected levels. First, the expression of equipment fault characteristics is not static but highly context-dependent. An anomaly in the same physical quantity can indicate drastically different fault modes and risk levels under varying equipment loads, ambient temperatures and humidity, and years of operation. Existing static diagnostic models typically treat all operating conditions equally during training, failing to capture this dynamic dependency. This leads to a significant decrease in the model's accuracy under specific operating conditions, resulting in false positives or false negatives. Second, the inherent laws governing this contextual dependency are often deeply explained within unstructured technical documents. Technical manuals detail the design margins and performance boundaries of equipment under different operating conditions. However, this textual engineering knowledge cannot be directly absorbed and utilized by existing models to dynamically guide and constrain their analysis of structured data. Consequently, the model's diagnostic logic becomes disconnected from the physical reality of the equipment, generating unreasonable maintenance instructions. Summary of the Invention

[0004] The purpose of this invention is to provide a data analysis-based substation maintenance data management method and system. This method changes the substation maintenance decision-making from relying on threshold judgments based on isolated data points to a comprehensive intelligent diagnostic level based on understanding of technical mechanisms, historical experience deduction, and real-time status matching. The generated maintenance instructions are not only timely, but also include a traceable explanation of the fault evolution path.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: Firstly, a substation maintenance data management method based on data analysis, characterized by the following steps: S101, acquiring historical maintenance data of substation equipment in the target substation and maintenance technical documents associated with the substation equipment; S102, extracting technical entities and entity relationships from the maintenance technical documents, combining the technical entities and entity relationships to generate a set of mathematical constraints for representing preset technical semantics, and constructing a semantic geometric construct library; S103, combining the analysis of maintenance technical documents and historical maintenance data to construct a system containing the evolution of preset technical semantics. S104. Obtain the equipment time-series data of the substation equipment in real time and process the equipment time-series data into data probes that reflect the current equipment status of the substation equipment; S105. Obtain a set of docking targets by calculating the docking affinity scores between the data probes and each semantic geometric conformation in the semantic geometric conformation library; S106. Project the docking targets onto the evolution path set, identify the target evolution path that matches the docking target in the evolution path set, and evaluate the equipment risk of the substation equipment based on the target evolution path; S107. Combine the equipment risk and the target evolution path to generate maintenance instructions for the substation equipment.

[0006] Further, in step S102, extracting technical entities and entity relationships from the maintenance technical documents, generating a set of mathematical constraints to represent preset technical semantics, and constructing a semantic geometric construct library includes the following steps: processing the maintenance technical documents using a preset entity recognition model to extract four types of technical entities: equipment components, fault phenomena, physical parameters, and operating instructions; analyzing the co-occurrence relationships of technical entities in the context of the maintenance technical documents using a relation extraction model to generate semantic relationship triples between technical entities; loading all semantic relationship triples into the initial graph data structure, and constructing an initial equipment maintenance knowledge graph through knowledge fusion and disambiguation processing of the initial graph data structure; configuring one or more sets of corresponding mathematical constraint sets for the preset technical semantic nodes in the equipment maintenance knowledge graph; and structurally associating and storing the preset technical semantic nodes and mathematical constraint sets to construct a semantic geometric construct library.

[0007] Further, in step S105, a set of docking targets is obtained by calculating the docking affinity scores between the data probe and each semantic geometric construct in the semantic geometric construct library, including the following steps: mapping the data probe to a preset multidimensional feature space, generating a dynamic target hypersurface for each semantic geometric construct in the multidimensional feature space according to the mathematical constraint set corresponding to the semantic geometric construct; calculating the weighted geometric distance from the feature points of the data probe mapped in the multidimensional feature space to each target hypersurface; normalizing and inversely processing the weighted geometric distance to obtain the docking affinity score, and taking the data probe and semantic geometric constructs with docking affinity scores higher than a preset docking threshold as a set of docking targets.

[0008] Furthermore, the data probe is mapped to a preset multidimensional feature space. In the multidimensional feature space, a dynamic target hypersurface is generated for each semantic geometric configuration based on the set of mathematical constraints corresponding to the semantic geometric configuration. This includes the following steps: extracting context state variables representing the macroscopic operating background of the current substation equipment from the data probe; constructing a dynamic Riemann metric tensor based on the context state variables for the preset multidimensional feature space, wherein the Riemann metric tensor is used to define the local geometric curvature of the multidimensional feature space under the macroscopic operating background; mapping the set of mathematical constraints to the multidimensional feature space defined by the local geometric curvature; solving for the geodesic hypersurface corresponding to the set of mathematical constraints in the multidimensional feature space; and using the geodesic hypersurface as the dynamic target hypersurface.

[0009] Further, the weighted geometric distance from the feature points mapped by the data probe in the multidimensional feature space to each target hypersurface is calculated, including the following steps: decomposing the mathematical constraint set into a set of independent feature constraint rules, each feature constraint rule corresponding to a dimension in the multidimensional feature space; for each feature constraint rule, calculating the deviation of the feature value of the data probe in the corresponding dimension from the target interval defined by the feature constraint rule to obtain the unidimensional feature satisfaction; extracting the association constraint weights representing the correlation between different feature dimensions from the semantic geometric conformation library; using the association constraint weights to perform a weighted summation of the unidimensional feature satisfaction of all dimensions, and using the weighted summation result as the weighted geometric distance from the feature points mapped by the data probe in the multidimensional feature space to the target hypersurface.

[0010] Further, in step S106, the docking target is projected onto the evolution path set, and the target evolution path matching the docking target is identified in the evolution path set. The equipment risk of the substation equipment is assessed based on the target evolution path, including the following steps: using a set of docking targets as initial activation nodes, node activation diffusion is performed in the evolution path set; the cumulative activation intensity received by each node in the evolution path set is calculated based on a preset diffusion propagation model; based on the cumulative activation intensity and the preset transition probability of the path connection edges in the evolution path set, one or more target evolution paths are identified in the evolution path set using a maximum probability path search algorithm; for each target evolution path, the path length of the target evolution path, the average docking affinity score of the nodes contained in the target evolution path, and the preset hazard level of the terminal node of the target evolution path are extracted; combining the path length, the average docking affinity score, and the preset hazard level of the terminal node, a risk assessment function is constructed, and the comprehensive risk score of the docking target relative to the substation equipment is calculated.

[0011] Furthermore, the cumulative activation intensity is calculated until convergence using the following formula: ,in, For the nodes of the evolution path The cumulative activation intensity at the (k+1)th iteration; An evolutionary propagation factor characterizing the degree of dependence on propagation within an evolutionary path set; The node is determined by the docking target. The initial activation strength; For all pointing nodes in the fault signal path network The set of upstream nodes; upstream node The cumulative activation intensity at the k-th iteration; For upstream nodes Pointing to node The preset transition probability of the path connecting edges; upstream node The sum of the preset transition probabilities of all outgoing path connecting edges.

[0012] Furthermore, by analyzing maintenance technical documents and historical maintenance data, an evolutionary path set containing the evolutionary relationships between preset technical semantics is constructed, including the following steps: constructing a network node set based on all technical semantic nodes in the semantic geometric construct library; statistically analyzing the chronological order and time interval of the technical semantics corresponding to different technical semantic nodes in the historical maintenance data, and calculating the temporal transition probability between network nodes in the network node set; extracting the logical causal relationship between network nodes by retrieving and analyzing the text statements describing the fault causal chain in the maintenance technical documents; integrating the temporal transition probability and the logical causal relationship to generate a comprehensive connection weight for each pair of network nodes, and constructing a weighted directed graph as the evolutionary path set based on the comprehensive connection weight.

[0013] In a second aspect, the present invention also provides a substation maintenance data management system based on data analysis, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the substation maintenance data management method based on data analysis as described in any one of the first aspects.

[0014] Thirdly, the present invention also provides a computer-readable storage medium storing instructions that, when executed by a processor, cause the processor to perform the substation maintenance data management method based on data analysis according to any one of the first aspects.

[0015] The beneficial effects of this invention are as follows: By constructing a semantic geometric construct library, this invention transforms the discrete textual technical entities and relationships in maintenance technical documents into a set of mathematical constraints with inherent logic, thereby constructing a computable geometric representation for each possible technical semantic state of the equipment. The construction of the evolution path set further explicitly models the empirical patterns contained in historical maintenance data as the evolutionary logic and probability between different technical semantic states, revealing typical patterns and underlying mechanisms of equipment state degradation. By calculating the docking affinity between real-time data probes and semantic geometric constructs, the complex state of the equipment, currently composed of multi-dimensional sensor data, can be accurately mapped to a specific semantic state. This state is then projected onto the evolution path set, achieving high-precision prediction and risk quantification assessment of the future state evolution trend of the equipment. This method shifts substation maintenance decision-making from threshold judgments relying on isolated data points to a comprehensive intelligent diagnostic level based on technical mechanism understanding, historical experience deduction, and real-time state matching. The generated maintenance instructions not only have timeliness but also include a traceable explanation of the fault evolution path. Attached Figure Description

[0016] Figure 1 is a flowchart illustrating a substation maintenance data management method based on data analysis in one embodiment of the present invention.

[0017] Figure 2 is a schematic diagram of the process of constructing a semantic geometric construct library in one embodiment of the present invention.

[0018] Figure 3 is a schematic diagram of the process of obtaining the docking target in one embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention are within the scope of protection of the present invention.

[0020] The terms "first," "second," etc., used in the specification and claims of this invention are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and the number of objects is not limited; for example, a first object can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0021] Figure 1 is a flowchart illustrating a substation maintenance data management method based on data analysis in one embodiment. It should be understood that although the steps in the flowchart of Figure 1 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; they can be executed in other orders. Furthermore, at least some of the steps in Figure 1 may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0022] As shown in Figure 1, a substation maintenance data management method based on data analysis specifically includes the following steps: S101. Obtain historical maintenance data of substation equipment in the target substation and maintenance technical documents associated with the substation equipment.

[0023] The first step involves comprehensive data collection and standardization, connecting and accessing multiple heterogeneous data sources. These include a computerized maintenance management system recording historical equipment operations, faults, and repairs, as well as an unstructured document library storing equipment design specifications, operating procedures, and fault analysis reports. Customized data extraction, transformation, and loading processes are deployed to integrate data from different sources and formats into a central data warehouse. During this process, historical maintenance data undergoes rigorous cleaning, handling missing values, correcting abnormal records, and aligning timestamps to ensure temporal consistency and accuracy. For unstructured text such as maintenance technical documents, natural language processing preprocessing techniques are employed, including word segmentation, part-of-speech tagging, and stop word removal, converting them into a structured or semi-structured format suitable for machine analysis. After data collection and preprocessing, a high-quality, multimodal fusion dataset is created. This dataset not only contains event sequences of equipment operation but also includes expert knowledge and operating procedures.

[0024] S102. Extract the technical entities and entity relationships from the maintenance technical documents, combine the technical entities and entity relationships to generate a set of mathematical constraints to represent the preset technical semantics, and construct a semantic geometric construct library.

[0025] After the initial data preparation is complete, expert knowledge needs to be extracted from massive amounts of technical documents, and a knowledge model that can be understood and computed by computers—namely, a semantic geometric construct library—needs to be constructed. First, a deep learning-based entity recognition model, such as an architecture combining bidirectional long short-term memory networks and conditional random fields, is applied to scan the document text, automatically identifying and classifying four preset technical entities: specific equipment components, clearly defined fault phenomena, quantified physical parameters, and standardized operating instructions. Then, an attention-based relation extraction model is used to analyze the contextual dependencies of the identified entities in sentences and paragraphs, generating a large number of semantic relation triples in the form of (entity A, relation, entity B). These triples are injected into an initial graph database, and a knowledge fusion algorithm is used to resolve synonymous entity referencing and relation conflicts, constructing a preliminary equipment maintenance knowledge graph. Finally, a set of mathematical constraints representing the inherent physical logic of the technical semantic nodes representing key equipment states or fault modes in the graph is configured. For example, the node <abnormal transformer oil temperature> is associated with the constraint {oil temperature > 90℃ ∧ total hydrocarbon content in gas chromatography > 50 μL / L}. By structurally storing these technical semantic nodes and their mathematical constraint sets, a semantic geometric construct library is ultimately formed, where each entry represents a technical semantic with a clear mathematical boundary.

[0026] S103. Combine the analysis of maintenance technical documents and historical maintenance data to construct an evolution path set that includes the evolutionary relationships between preset technical semantics.

[0027] The construction of the evolution path set begins with using all technical semantic nodes in the semantic geometry construct library as the node set of the network graph. Next, deep analysis of the timestamp information in historical maintenance data records is conducted. Statistical analysis of the time intervals and frequencies of successive occurrences of one technical semantic term following another in a large number of time series is used to calculate the temporal transition probabilities between nodes. Simultaneously, natural language processing techniques are used to perform semantic mining on maintenance technical documents, particularly retrieving statements describing fault mechanisms and causal chains, such as "If humidity is too high, insulation performance will decrease, leading to breakdown," extracting the logical causal relationships between nodes. Finally, the data-based temporal transition probabilities and text-based logical causal relationships are weighted and fused to generate a comprehensive connection weight for each potentially related pair of nodes. Based on all nodes and the calculated comprehensive connection weights, a weighted directed graph is constructed; this graph is the evolution path set, which probabilistically represents the potential paths for equipment to evolve from one state to another.

[0028] S104. Acquire the equipment timing data of substation equipment in real time, and process the equipment timing data into data probes that reflect the current equipment status of substation equipment.

[0029] To link the real-time operation of equipment with the knowledge base, the continuous raw equipment monitoring data must be transformed into a standardized data probe. This transformation process begins with the real-time acquisition of time-series data streams from substation monitoring and data acquisition systems and other online monitoring devices. These data include high-frequency sampled signals such as voltage, current, temperature, and vibration frequency. To capture the dynamic characteristics of the equipment status, a sliding time window technique is used to slice the data stream. Within each time window, a series of feature engineering processes are performed on the raw data. This includes not only calculating basic statistics such as mean, variance, and peak value, but also more complex signal processing techniques, such as applying wavelet transform to analyze the energy distribution of transient signals or calculating sample entropy to assess signal complexity. These extracted multi-dimensional features together constitute a feature vector. This feature vector is normalized to eliminate the influence of different physical dimensions, ultimately forming a data probe that can accurately represent the current status of the equipment in a high-dimensional feature space. This data probe condenses the core information of the equipment's operating status at a specific moment, enabling precise mathematical comparison and matching with the abstract technical semantics defined in the semantic geometry model library.

[0030] S105. A set of docking targets is obtained by calculating the docking affinity scores between the data probe and each semantic geometric conformation in the semantic geometric conformation library.

[0031] The process involves mapping a data probe composed of multidimensional features to a feature point in a predefined high-dimensional feature space. For each semantic geometric conformation in the semantic geometric conformation library, its associated set of mathematical constraints defines a theoretically ideal region or boundary in this high-dimensional space. This boundary is geometrically represented as a dynamic target hypersurface. The shape of this hypersurface is not fixed but dynamically adjusted according to certain macroscopic environmental variables (such as ambient temperature and grid-side load). Next, the geometric distance from the feature point to each target hypersurface needs to be calculated. Considering the different importance of different feature dimensions to equipment status assessment, a weighted geometric distance calculation method is adopted. After calculating the weighted geometric distance, it is converted into a docking affinity score between 0 and 1 through normalization and inverse processing (e.g., using an exponential decay function). The higher the score, the more closely the current equipment status matches the technical semantics represented by the semantic geometric conformation. Finally, all semantic geometric conformations with docking affinity scores exceeding a predefined docking threshold are selected and, together with the data probe, form a set of docking targets.

[0032] S106. Project the docking target onto the evolution path set, identify the target evolution path that matches the docking target in the evolution path set, and assess the equipment risk of the substation equipment based on the target evolution path.

[0033] In this process, the node corresponding to the docking target in the evolution path set is considered the initial activated node, and a corresponding docking affinity score is assigned as the initial activation intensity. A node activation diffusion process is then initiated, with the activation intensity propagating along the directed edges in the graph. The cumulative activation intensity received by each node depends on the activation intensity of its upstream node and the weight of the connecting edge. After propagation, a maximum probability path search algorithm, such as a variant of the Viterbi algorithm, is used to find the most probable path to a predefined high-risk terminal node, starting from the initial activated node. This identified path or multiple paths constitute the target evolution path. For each target evolution path, key risk indicators are extracted, including the path length (representing the time required for evolution), the average docking affinity score of all nodes on the path (representing the degree of fit between the current state and the path), and the preset hazard level of the terminal node. Finally, a comprehensive risk assessment function is constructed to calculate a quantified comprehensive risk score, providing a basis for decision-making.

[0034] S107. Generate maintenance instructions for substation equipment by combining equipment risk and target evolution path backtracking.

[0035] The system employs a backtracking analysis using the target evolution path, starting from the highest-risk endpoint and tracing back along the path to the current target node. In this backtracking process, each directed edge connecting two nodes represents a key state transition or causal relationship. For each state transition step on the path, the system automatically queries the knowledge graph built during the construction of the semantic geometric construct library. In the knowledge graph, each technical semantic node and relationship is associated with corresponding technical document fragments and standard operating procedures, containing operational instructions for preventing, diagnosing, or repairing the state. For example, if the backtracking path shows that <excessive moisture in insulating oil> is a key precondition leading to <partial discharge>, the system will retrieve standard operations related to handling insufficient moisture in insulating oil, such as oil filtration and replacing the desiccant. The operational instructions corresponding to all key steps on the backtracking path are collected and prioritized based on each step's position in the path, the probability of state transition, and the risk contribution of the corresponding node. Finally, the prioritized instructions are integrated into a clear and specific maintenance work order, clearly indicating the components to be inspected, the operations to be performed, and the recommended time window.

[0036] In one embodiment, referring to Figure 2, extracting technical entities and entity relationships from the maintenance technical document, generating a set of mathematical constraints to represent preset technical semantics, and constructing a semantic geometric construct library includes the following steps: S201. Processing the maintenance technical document using a preset entity recognition model to extract four types of technical entities: equipment components, fault phenomena, physical parameters, and operation instructions; S202. Analyzing the co-occurrence relationships of technical entities in the context of the maintenance technical document using a relation extraction model to generate semantic relationship triples between technical entities; S203. Loading all semantic relationship triples into the initial graph data structure, and constructing an initial equipment maintenance knowledge graph through knowledge fusion and disambiguation processing of the initial graph data structure; S204. Configuring one or more sets of corresponding mathematical constraint sets for the preset technical semantic nodes in the equipment maintenance knowledge graph; S205. Storing the preset technical semantic nodes and mathematical constraint sets in a structured association to construct a semantic geometric construct library.

[0037] In this implementation, the initial stage of knowledge system construction involves deep analysis and information extraction of unstructured maintenance technical documents, aiming to make implicit knowledge in the text explicit. This process utilizes a pre-trained named entity recognition model based on a hybrid architecture of Bidirectional Long Short-Term Memory (BiLSTM) and Conditional Random Field (CRF). First, the text content of the maintenance technical document is input into the BiLSTM layer. The network captures the contextual semantic information of each word from both forward and backward dimensions, generating vector representations rich in contextual features. These feature vectors are then passed to the CRF layer. The CRF layer learns the dependencies between label sequences, enabling the model to find the optimal label sequence globally, effectively avoiding labeling errors caused by local optima. The model's output probability can be expressed as: ,in It is the input text sequence. It is the predicted label sequence, where i is the position sequence and k is the feature function index. It is the transition characteristic function. It is a characteristic function. These are the corresponding weights. It is a normalization factor. After model processing, all text fragments in the document related to equipment components, fault phenomena, physical parameters, and operating instructions can be identified and marked.

[0038] After extracting technical entities, it's necessary to further explore the potential semantic relationships between these entities to construct a knowledge network structure. This step employs a relation extraction model, such as an attention-based graph neural network, to analyze the contextual co-occurrence relationships of identified entity pairs within the same sentence or paragraph. The model first encodes the sentence containing the entity pairs, generating a vector representation for each word. Then, it dynamically calculates the importance weights of other words in the sentence for determining the relationship between the two entities using an attention mechanism. By aggregating this weighted word information, the model can generate a comprehensive relation representation vector for the entity pairs. (Relationship judgment score) It can be calculated using a bilinear transformation: ,in: It is the entity i component. is the entity j component, and T is the transpose operation; and These are the vector representations of the two entities. It is a learnable weight matrix for a specific relation type r. It is a bias term. It is an activation function. By comparing the scores of different relationship types, the model can determine the most likely relationship type between two entities.

[0039] The raw data contains significant redundancy and ambiguity, necessitating knowledge fusion and disambiguation to ensure the quality of the knowledge graph. The core task of knowledge fusion is entity alignment, which involves identifying and merging different entity nodes pointing to the same object in the real world. For example, merging two nodes, "circuit breaker" and "high-voltage switch," into a single standardized node can be achieved using knowledge graph embedding technology. This technology maps each entity and relation to a low-dimensional, continuous vector space. By calculating the distance or similarity between vectors, the equivalence between entities can be determined. For instance, the scoring function of the TransE model... Used to measure the reasonableness of a triple, where: h is the head entity and t is the tail entity; These are the vector embeddings of the head entity, tail entity, and relation, respectively. The smaller the distance, the more likely the triple is to be true, and the vectors of similar entities will become closer. Meanwhile, disambiguation focuses on distinguishing entities with the same name but different meanings by analyzing the local neighborhood structure and connectivity of nodes in the graph to determine their true meaning.

[0040] To integrate symbolic knowledge graphs with quantified equipment monitoring data, specific mathematical definitions must be configured for the nodes representing key technical semantics within the graph. This step aims to transform abstract equipment states or fault phenomena into a set of mathematical constraints that can be directly verified by a computer. This configuration process typically combines domain expert knowledge with rules automatically mined from technical standards and procedural documents. For each selected preset technical semantic node, a set of multiple mathematical inequalities or logical expressions is defined. These constraints are directly related to measurable physical parameters such as temperature, pressure, and voltage. A constraint set not only includes simple threshold judgments but may also involve complex logical relationships (AND, OR, NOT) between multiple parameters. To handle the fuzziness of boundary states, a membership function from fuzzy logic can be introduced to describe the degree to which a parameter value conforms to a certain constraint. The formula for calculating the membership degree is as follows: Where x is the measured physical parameter value, and c is the ideal center value of that parameter. This controls the width of the fuzzy boundary.

[0041] Finally, the technical semantic nodes in the knowledge graph are efficiently and structurally associated and stored with their corresponding sets of mathematical constraints, forming the final semantic geometric construct library. The design goal of this library is to achieve rapid knowledge retrieval and computation. Its storage architecture typically employs an enhanced graph database or a hybrid data model, where each technical semantic node is treated as an independent construct object. Each construct object... They are all formally defined as a tuple, the structure of which can be represented as: ,in It is the unique identifier of this semantic node. It is a collection of entity attributes directly associated with the node. It stores all the in-and-out relationships of this node in the knowledge graph. This is the set of mathematical constraints bound to that node. When it is necessary to evaluate a specific technical semantics, it can be done through... Quickly locate the corresponding construct object and directly extract its set of mathematical constraints for subsequent calculation and analysis.

[0042] In one embodiment, referring to FIG3, obtaining a set of docking targets by calculating the docking affinity score between the data probe and each semantic geometric construct in the semantic geometric construct library includes the following steps: S301. Mapping the data probe to a preset multidimensional feature space, and generating a dynamic target hypersurface for each semantic geometric construct according to the mathematical constraint set corresponding to the semantic geometric construct in the multidimensional feature space; S302. Calculating the weighted geometric distance from the feature points of the data probe mapped in the multidimensional feature space to each target hypersurface; S303. Normalizing and inversely processing the weighted geometric distance to obtain the docking affinity score, and taking the data probe and semantic geometric constructs with docking affinity scores higher than a preset docking threshold as a set of docking targets.

[0043] In this embodiment, to achieve precise matching between the current device state and the abstract technical semantics, the real-time generated data probes are first projected into a predefined multi-dimensional feature space, making them specific feature points within that space. In this space, each semantic geometric construct derived from the semantic geometric construct library generates a dynamic target hypersurface based on its inherent set of mathematical constraints. This hypersurface is not static; rather, its shape and position adaptively change with the contextual state variables (e.g., ambient temperature, power grid load) representing the macroscopic operating background of the device. This dynamism is achieved by constructing a Riemannian metric tensor that depends on the contextual state variables. This tensor defines the local geometric structure of the feature space under a specific macroscopic background. The set of mathematical constraints is then mapped into this dynamically curved space. By solving a variational problem, geodesics that satisfy the constraints are found; the hypersurface formed by these geodesic families is the dynamic target hypersurface. After defining the feature points representing the data probes and the target hypersurface representing the ideal state in the feature space, the next step is to quantitatively calculate the degree of deviation between them, i.e., the weighted geometric distance.

[0044] Directly calculating the Euclidean distance from a point to a hypersurface fails to capture the differentiated importance of different feature dimensions for equipment status assessment and ignores the inherent relationships between dimensions. Therefore, a weighted strategy is adopted. First, the complex set of mathematical constraints is decomposed into a set of independent feature constraint rules for each single dimension, with each rule defining the ideal range or target value of the feature value for that dimension. Next, for each feature constraint rule, the deviation of the data probe's feature value in the corresponding dimension from the target range defined by the rule is calculated, resulting in a dimensionless single-dimensional feature satisfaction. More importantly, predefined association constraint weights are extracted from a semantic geometric construct library. These weights are stored in matrix form, quantifying the mutual influence relationships between different feature dimensions. Finally, a weighted summation formula is used to calculate the total geometric distance. This weighted geometric distance comprehensively considers the deviations of all dimensions and their mutual coupling effects, thus more accurately reflecting the true gap between the current equipment status and the specific technical semantics.

[0045] The calculated weighted geometric distance is a raw measure of deviation; a larger value indicates a more severe deviation. To transform it into a standardized affinity score, a normalization inverse process is required to generate a docking affinity score. This score should be between 0 and 1, with a score of 1 (perfect match) when the distance is 0 and a score approaching 0 (complete mismatch) as the distance approaches infinity. Specifically, this can be mapped using a negative exponential function. The formula for mapping the weighted geometric distance to the docking affinity score is as follows: In this formula, It is a weighted geometric distance. It is an adjustable sensitivity parameter. Its main function is scaling; it controls the independent variable d in the exponential function. wg The weights of the data probes are used to control the rate at which the score decays with increasing distance. After the score is calculated, it is compared with a preset docking threshold. Only semantic geometric conformations with docking affinity scores higher than this threshold are considered significantly relevant to the current device state. Finally, the data probes are combined with all the high-affinity semantic geometric conformations that pass the screening to form a set of docking targets.

[0046] In one embodiment, mapping a data probe to a preset multidimensional feature space and generating a dynamic target hypersurface for each semantic geometric construct based on the set of mathematical constraints corresponding to the semantic geometric construct in the multidimensional feature space includes the following steps: extracting context state variables representing the macroscopic operating background of the current substation equipment from the data probe; constructing a dynamic Riemann metric tensor for the preset multidimensional feature space based on the context state variables, wherein the Riemann metric tensor is used to define the local geometric curvature of the multidimensional feature space under the macroscopic operating background; mapping the set of mathematical constraints to the multidimensional feature space defined by the local geometric curvature; solving for the geodesic hypersurface corresponding to the set of mathematical constraints in the multidimensional feature space; and using the geodesic hypersurface as the dynamic target hypersurface.

[0047] In this embodiment, the data probe is a hybrid feature vector containing both the device's micro-state (e.g., winding temperature, partial discharge) and macro-context (e.g., ambient humidity, grid-side load level, system operating mode). By applying a domain-knowledge-based feature selection algorithm, variables that change slowly and are primarily influenced by the external environment or system-level scheduling can be identified and extracted. For example, parameters such as ambient temperature, cooling medium flow rate, and grid power flow are predefined as context state variables. These selected variables constitute a context state vector that captures the background stage of device operation. After extraction, a low-dimensional vector specifically describing the current macro-condition is obtained. This vector serves as a key input for adjusting the feature space geometry in subsequent steps, ensuring that the evaluation of device status is not performed in a static vacuum, but rather within a dynamic framework that fully considers actual operating conditions.

[0048] After obtaining the context state vector, a dynamically changing Riemannian metric tensor needs to be constructed for the predefined multidimensional feature space based on the context state variables. Mathematically, this metric tensor defines the distance between any two points in the feature space and the degree of local curvature of the space, making it no longer a static Euclidean space. Metric Tensor Each component is designed as a context state vector. The function. The specific construction method uses conformal transformation, that is... In this expression, It represents the feature mapping function that maps input data from the original space to a high-dimensional feature space. It is a Minkowski metric for the identity matrix, representing the flat space under the baseline condition; It is a scalar conformal factor calculated based on the context state vector, and its value is learned by a pre-trained deep neural network from historical data. When the macroscopic operating context changes... The changes that followed led to and metric tensor This also changes, causing the geometric curvature of the entire feature space to adjust dynamically. The space constructed in this way can more realistically reflect the nonlinear coupling relationships between feature variables under different operating conditions.

[0049] After defining a curved feature space that dynamically changes with the macroscopic background, the final step is to transform the set of abstract mathematical constraints derived from the semantic geometric construct library into a specific, geometrically-based, dynamic target hypersurface within this space. In the curved space, the optimal boundary state satisfying the constraints is no longer a simple plane or sphere, but a hypersurface composed of geodesics, which are generalizations of the shortest path between two points in this space. Solving for the geodesic hypersurface is a variational problem, namely, finding the set of paths that maximize a certain functional. Defined as the arc length integral of the path: .in, It is a parametric curve describing a path on a hypersurface. These are the parameters of the curve ( When describing a curve in space, points on the curve can be represented by the parameter λ. It is a metric tensor, where i is the spacetime x-coordinate and j is the spacetime y-coordinate. It is the derivative of the parameter q, representing a tiny change in the parameter. It is the differential of the parameter ⋋, representing a tiny change in the parameter. By solving the Euler-Lagrange equations corresponding to this functional, all geodesics that satisfy the boundary conditions of the mathematical constraint set under the current spatial geometry can be calculated. The hypersurface woven together by these geodesics is the final dynamic target hypersurface, which can provide a dynamic and geometrically ideal state boundary for each technical semantic under the current specific working condition.

[0050] In one implementation, calculating the weighted geometric distance from the feature points mapped by the data probe in the multidimensional feature space to each target hypersurface includes the following steps: decomposing the mathematical constraint set into a set of independent feature constraint rules, each feature constraint rule corresponding to a dimension in the multidimensional feature space; for each feature constraint rule, calculating the deviation of the feature value of the data probe in the corresponding dimension from the target interval defined by the feature constraint rule to obtain the unidimensional feature satisfaction; extracting the association constraint weights representing the correlation between different feature dimensions from the semantic geometric conformation library; using the association constraint weights to perform a weighted summation of the unidimensional feature satisfaction of all dimensions, and using the result of the weighted summation as the weighted geometric distance from the feature points mapped by the data probe in the multidimensional feature space to the target hypersurface.

[0051] In this embodiment, before calculating the geometric distance between the data probe and the target hypersurface, it is necessary to perform dimensionality reduction on the set of complex coupled mathematical constraints that define the hypersurface, and decompose it into a set of independent feature constraint rules for a single dimension. The purpose of this decomposition process is to transform a multi-variable, non-linear decision problem into a series of single-variable, linear decision sub-problems. For each set of mathematical constraints in the semantic geometric conformation library, the system will analyze its logical structure. For example, a composite constraint of the form {(feature A > a) ∧ (feature B < b)} is disassembled into two independent atomic rules: the first rule is "the target interval of feature A is (a, +∞)", where a is the left endpoint position of the interval and ∞ is the right endpoint of the interval being infinity; the second rule is "the target interval of feature B is (-∞, b)", where b is the right endpoint position of the interval. Each such generated feature constraint rule exactly corresponds to an axis in the multi-dimensional feature space and defines an ideal numerical category for that axis. After completing the decomposition of the feature constraint rules, for each rule, the deviation degree of the data probe in the corresponding dimension can be calculated independently, and this quantified result is called the single-dimensional feature satisfaction degree.

[0052] For the k-th dimension in the multi-dimensional feature space, the specific feature value of the data probe in this dimension is , and the corresponding feature constraint rule defines an ideal target interval , is the left endpoint of the interval [ak, bk], is the right endpoint of the interval [ak, bk]. The logic for calculating the deviation degree is: if the feature value falls within the target interval, it is considered to be fully satisfied in this dimension and the deviation degree is zero; if the feature value falls outside the interval, its deviation degree is proportional to the distance from this value to the nearest endpoint of the interval. To eliminate the influence of different dimension measures, the deviation degree needs to be normalized. The calculation of the single-dimensional feature deviation degree can be expressed as: . In this formula, is a dimensionless value that precisely quantifies the relative severity of the deviation of the current device state from the ideal category in the k-th feature dimension. This process traverses all feature dimensions and finally generates a deviation degree vector, and each element in the vector represents the degree of misalignment of the device state in a specific aspect.

[0053] After obtaining the single-dimensional feature deviation vectors for each dimension, they are combined into a comprehensive distance metric that reflects the correlation between dimensions, namely, weighted geometric distance. Simply adding the deviations of each dimension ignores the coupling effect between features; for example, the risk of both rising oil temperature and increasing gas content occurring simultaneously is far greater than their independent occurrence. Therefore, it is necessary to extract predefined correlation constraint weights from a semantic geometric construct library. These weights are represented by a symmetric matrix. The elements in the matrix exist in the form of... The interaction strength between the j-th and k-th features is quantified. Using this weight matrix, the final weighted geometric distance can be calculated using a quadratic formula: .in, and The first value is the previously calculated one-dimensional feature deviation; j represents the range of possible horizontal components, and k represents the range of possible vertical components. The result of this formula is not merely a simple summation of the deviations of each dimension, but rather a comprehensive consideration of the synergistic effect of pairwise combinations of all deviations. The resulting weighted geometric distance, as a scalar value, comprehensively and profoundly reflects the true distance between the current device state point represented by the data probe and the ideal state manifold represented by the dynamic target hypersurface.

[0054] In one embodiment, the method further includes the following steps: when there is no single score above the docking affinity score but two or more scores above a preset competition threshold, a competitive docking analysis is triggered, where the competition threshold is lower than the docking threshold; in response to the competitive docking analysis, a hypothetical secondary interference conformation and a hypothetical primary target conformation are selected iteratively from target docking targets above the competition threshold; prototype data forms corresponding to the secondary interference conformation are extracted from a semantic geometric conformation library, the prototype data forms are weighted based on the docking affinity score of the secondary interference conformation, and the weighted prototype data forms are subtracted from the data probe to generate a residual data probe; the docking affinity score between the residual data probe and the primary target conformation is recalculated to obtain a corrected affinity score; if the corrected affinity score is higher than the original docking affinity score between the data probe and the primary target conformation, the primary target conformation and the residual data probe are considered as a set of docking products.

[0055] In this embodiment, under certain complex device states, directly calculated docking affinity scores may not clearly point to a single fault or state. This manifests as the absence of any single score exceeding a preset high-confidence docking threshold, while two or more scores exceed a lower competition threshold. This situation indicates that the device may be in a superposition or transitional phase of multiple potential states, with interference between signals making a definitive judgment impossible. To resolve this ambiguity, a special competitive docking analysis process is triggered. The initiation condition for this process is a strict logical judgment: the maximum value in the affinity score set is lower than the docking threshold, and at least two elements in the set are higher than a preset, numerically lower competition threshold. In response to the triggering of competitive docking analysis, the system enters an iterative hypothesis testing loop aimed at identifying the most likely primary state from multiple competitors. This process first identifies all semantic geometric conformations with docking affinity scores higher than the competition threshold as a candidate set. Then, through a systematic pairing strategy, a hypothetical secondary interference conformation and a hypothetical primary target conformation are iteratively selected from this candidate set.

[0056] A typical iterative strategy is as follows: First, the candidate set is sorted in descending order of its original affinity score. In the first iteration, the conformation with the highest score is selected as the primary target, and the next highest score is selected as the secondary interference. In subsequent iterations, the primary target conformation remains unchanged, while the secondary interference conformations are replaced by the remaining conformations in the candidate set. This constructs a series of "primary-secondary" hypothesis pairs. Each pair represents a theoretical possibility that the physical phenomenon represented by the secondary conformation interferes with the signal characteristics of the primary conformation. After determining a set of hypothetical "primary-secondary" conformation pairs, an attempt is made to remove the influence of the secondary interference conformations from the original data. To this end, the prototype data form associated with the hypothetical secondary interference conformation is first retrieved from a semantic geometric conformation library. This prototype data form is a standardized feature vector representing the data performance of the secondary conformation in its most typical case. Subsequently, the prototype data form is weighted based on the original docking affinity score of the secondary interference conformation. This weighting operation aims to simulate the actual intensity of the currently observed interference signal. Finally, by subtracting the weighted prototype data from the original data probe using vector subtraction, a residual data probe is generated. The mathematical expression for this process is: In this formula, It is the generated residual data probe. It is a raw data probe. It is the prototype data form of the secondary interference conformation. It is a score based on minor conformation affinity. The nonlinear weighting function.

[0057] After generating the residual data probe, it is necessary to re-evaluate the degree of matching between the new data and the hypothesized primary target conformation to verify whether the previous interference hypothesis holds true. This evaluation process is consistent with the initial calculation of the affinity score, but the input data probe is changed to the residual data probe. The residual data probe is mapped to a multidimensional feature space, and the weighted geometric distance from its feature points to the dynamic target hypersurface corresponding to the primary target conformation is calculated and converted into a corrected affinity score. The new score reflects the degree of agreement between the remaining signal features in the original data and the primary target conformation after excluding potential interference factors. In the final step of the iteration process, the corrected affinity score is compared directly with the original affinity score calculated from the original data probe and the primary target conformation. If the corrected affinity score is significantly higher than the original score, it indicates that the previous hypothesis holds true: that is, the secondary interference conformation does indeed mask or weaken the signal features of the primary target conformation, successfully highlighting the main contradiction through signal subtraction. This improvement in score provides strong validation of the hypothesis. In this scenario, the hypothesized primary target conformation is confirmed and output as a set of docking products along with residual data probes generated through signal subtraction. This docking product not only identifies the most likely primary state of the current device but also provides residual data, freed from the influence of that state. This residual data can be used for further analysis to determine if other, less significant problems exist. If the corrected score does not improve or even decreases, the current hypothesis is rejected, and the next iteration begins, testing other "primary-minor" conformation combinations until the solution that maximizes the affinity score is found.

[0058] In one implementation, projecting the docking target onto an evolution path set, identifying target evolution paths matching the docking target within the evolution path set, and assessing the equipment risk of substation equipment based on the target evolution paths include the following steps: using a set of docking targets as initial activated nodes, performing node activation diffusion in the evolution path set; calculating the cumulative activation intensity received by each node in the evolution path set based on a preset diffusion propagation model; identifying one or more target evolution paths in the evolution path set using a maximum probability path search algorithm based on the cumulative activation intensity and the preset transition probability of the path connection edges in the evolution path set; for each target evolution path, extracting the path length of the target evolution path, the average docking affinity score of the nodes contained in the target evolution path, and the preset hazard level of the terminal node of the target evolution path; and constructing a risk assessment function by combining the path length, average docking affinity score, and preset hazard level of the terminal node to calculate the comprehensive risk score of the substation equipment corresponding to the docking target.

[0059] In this implementation, to predict the future evolution trend of the current device state, the docking target is first used as the initial signal source, and activation diffusion is performed in a pre-constructed evolution path set consisting of technical semantic nodes and weighted directed edges. This process simulates the dynamics of faults or state anomalies propagating within the device's associated components and system. Specifically, the node corresponding to the docking target is assigned an initial activation strength, which is typically taken directly from its docking affinity score. Subsequently, the activation strength propagates along the directed edges in the graph from these initial nodes to their adjacent nodes. The weight of each edge represents the probability or intensity of the state transitioning from the source node to the target node; the activation strength decreases or increases during propagation based on the edge weight. This diffusion process is not completed in one go but through multiple iterations, allowing the activation signal to reach regions farther from the initial node. After the activation diffusion process begins, the cumulative activation strength finally received by each node in the evolution path set after multiple iterations needs to be calculated, thereby quantifying the degree to which each potential future state is affected by the current initial state.

[0060] Specifically, an iterative update-based algorithm can be used, and the cumulative activation strength is calculated using the following formula until convergence: ,in, For the nodes of the evolution path The cumulative activation intensity at the (k+1)th iteration; An evolutionary propagation factor characterizing the degree of dependence on propagation within an evolutionary path set; The node is determined by the docking target. The initial activation strength; For all pointing nodes in the fault signal path network The set of upstream nodes; upstream node The cumulative activation intensity at the k-th iteration; For upstream nodes Pointing to node The preset transition probability of the path connecting edges; upstream node The sum of the preset transition probabilities of all outgoing path connecting edges, p is the summation index (also known as a dummy index or virtual index) in the summation operation. After a sufficient number of iterations until the system converges, each node will obtain a stable cumulative activation strength value, which comprehensively reflects the path reachability and connection tightness between it and the initial docking target.

[0061] Based on the calculated cumulative activation strength of each node and the preset transition probabilities of the edges in the evolution path set, the maximum probability path search algorithm is then used to identify the most probable evolution path from the current state (initial activated node) to a high-risk terminal state (such as equipment downtime or major failure) in complex networks. The maximum probability path search algorithm starts from one or more initial activated nodes and explores all possible paths. During the exploration process, the comprehensive evaluation value of a path is not merely a simple product of the transition probabilities of all its connecting edges, but also incorporates the cumulative activation strength of each node on the path as a dynamic weight. Overall score It can be defined as: .in It is the edge The transition probability, It is the horizontal component of the vertices. It is the vertical component of the nodes (vertices). It usually represents a set, which can contain multiple elements; It is the cumulative activation strength of node k. It refers to the path length. The goal of the algorithm is to find one or more paths that maximize the overall score; these paths are called the objective evolution paths.

[0062] After identifying one or more target evolution paths, a quantitative risk assessment is required. First, the path length, i.e., the number of edges or nodes contained in the path, is extracted. Path length is usually negatively correlated with the timescale of risk occurrence; the shorter the path, the shorter the time required for state deterioration, and the more urgent the risk. Second, the average docking affinity score of the nodes contained in the target evolution path is calculated. This indicator reflects the overall degree of fit between the entire evolution path and the current actual state of the equipment; the higher the average score, the more likely the predicted path is the actual direction of the current state's development. Finally, the pre-defined hazard level of the terminal nodes of the target evolution path is extracted from the knowledge base. Through the extraction of information from these three dimensions, each target evolution path is transformed into a triple consisting of path length, average affinity score, and terminal hazard level, providing a structured input for the final comprehensive risk calculation. Finally, the risk information from different dimensions is integrated into a single, intuitive quantitative result. Risk Assessment Function The construction of the risk assessment function requires comprehensive consideration of the contributions of three indicators and assigning them appropriate weights. The risk assessment function can be designed as follows: In this function, L is the path length. It is the average docking affinity score. It is the preset hazard level of the end node. and These are weighting coefficients corresponding to the reciprocal of the pathway length and the average affinity score, respectively, used to balance the importance of both in the assessment. The pathway length is taken as the reciprocal because risk is inversely proportional to path length. This function sums the urgency and probability by weights and then multiplies by the severity of the consequences, which conforms to the basic logic of risk engineering: risk = probability × consequence.

[0063] In one implementation, constructing an evolutionary path set containing the evolutionary relationships between preset technical semantics by analyzing maintenance technical documents and historical maintenance data includes the following steps: constructing a network node set based on all technical semantic nodes in the semantic geometric construct library; statistically analyzing the chronological order and time interval of the technical semantics corresponding to different technical semantic nodes in the historical maintenance data, and calculating the temporal transition probability between network nodes in the network node set; extracting the logical causal relationship between network nodes by retrieving and analyzing the text statements describing the fault causal chain in the maintenance technical documents; fusing the temporal transition probability and the logical causal relationship to generate a comprehensive connection weight for each pair of network nodes, and constructing a weighted directed graph as the evolutionary path set based on the comprehensive connection weight.

[0064] In this embodiment, the first step in constructing a network model capable of depicting the evolution path of equipment states is to establish the basic building blocks of the network, namely, the set of nodes. This process directly utilizes an existing semantic geometric construct library as a blueprint. Each technical semantic entry in this library, whether representing a specific fault phenomenon or a key equipment state, is directly mapped to an independent node in the evolution path set. By traversing the entire semantic geometric construct library and extracting the unique identifier of each entry, a complete set of nodes with clear technical connotations is formed. Based on the establishment of network nodes, it is necessary to quantify the dynamic correlation strength between nodes based on historical experience, i.e., to calculate the temporal transition probability. This process deeply mines massive amounts of historical maintenance and operation data, which are stored in the form of event sequences with precise timestamps.

[0065] By performing a sliding window scan on time series data, the chronological order and time intervals of events corresponding to different technical semantics are systematically statistically analyzed. For any pair of nodes (node ​​a, node b), the frequency of the event corresponding to node b occurring within a preset time window after the event corresponding to node a occurs is counted. (Time series transition probability) The calculation considers not only the frequency of occurrence but also the time decay effect, meaning that the shorter the time interval, the higher the transition probability. The calculation can be performed using the following formula: .in, This represents the total number of times the event corresponding to node a occurred. The summation term iterates through all instances where event b occurred after event a. It is the actual time interval between the two occurrences at the i-th time. It is a time decay constant used to control the decay rate.

[0066] Relying solely on statistical data may miss rare but important causal relationships. Therefore, it is necessary to mine expert knowledge from maintenance technical documents to extract logical causal relationships between nodes. This process utilizes natural language processing techniques, particularly dependency parsing and semantic role labeling, to perform a deep scan of a large amount of stored technical manuals, fault analysis reports, and other texts. The system focuses on retrieving statements containing explicit causal indicators (such as "cause," "lead to," "is due to") or specific sentence structures. Once a text fragment describing a causal chain between two technical semantics (node ​​u, node v) is identified, a quantitative strength score is assigned to this logical relationship based on a series of preset rules. (Logical causal relationship strength) The calculation can be modeled as follows: In this formula, It is a complete statement containing u and v. It is a feature function used to evaluate the strength of the k-th linguistic feature in a sentence (such as the explicitness of keywords, the tightness of syntactic structure), These are the learnable weights corresponding to the feature.

[0067] Finally, the temporal transition probabilities from data statistics and the logical causal relationships from text mining are effectively fused to generate a comprehensive connection weight for each pair of network nodes, and the final weighted directed graph is constructed based on this weight. The purpose of the fusion is to combine the objectivity of data statistics with the depth of expert knowledge to obtain a more comprehensive and reliable relational measure. The fusion process employs a non-linear weighting strategy to balance the contributions of the two information sources. For a connection from node x to node y, its comprehensive connection weight is... It can be calculated using the following formula: .in, It is the time transition probability from x to y. It corresponds to the strength of the logical causal relationship. It is a fusion hyperparameter between 0 and 1, used to adjust the relative importance of the two. This geometric mean fusion method ensures that a higher weight is only generated when both pieces of evidence indicate a correlation. Based on all nodes and the calculated comprehensive connection weights between each pair of nodes, a complete weighted directed graph is constructed. This graph is the evolution path set, and its structure and weight distribution comprehensively reflect the multi-source information of the device state evolution.

[0068] The present invention also discloses a substation maintenance data management system based on data analysis, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the substation maintenance data management method based on data analysis described above.

[0069] The processor can be a central processing unit (CPU). Of course, depending on the actual use, other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. can also be used. The general-purpose processor can be a microprocessor or any conventional processor, etc., and this invention does not limit it.

[0070] The memory can be an internal storage unit of a computer device, such as a hard disk or RAM, or an external storage device, such as a plug-in hard disk, smart memory card (SMC), secure digital card (SD), or flash memory card (FC) provided on the computer device. Furthermore, the memory can be a combination of an internal storage unit and an external storage device. The memory is used to store computer programs and other programs and data required by the computer device. The memory can also be used to temporarily store data that has been output or will be output. This invention does not limit this.

[0071] The present invention also discloses a computer-readable storage medium storing instructions that, when executed by a processor, configure the processor to perform the data analysis-based substation maintenance data management method described in any of the above embodiments.

[0072] The computer program can be stored in a machine-readable medium. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or certain middleware. The machine-readable medium includes any entity or device capable of carrying computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the machine-readable medium includes, but is not limited to, the above-mentioned components.

[0073] The substation maintenance data management method based on data analysis described in the above embodiments is stored in the computer-readable storage medium and loaded and executed on the processor to facilitate the storage and application of the above method.

[0074] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of protection of the present invention is limited to these examples; within the framework of the present invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of one or more embodiments of the present invention as described above, which are not provided in detail for the sake of brevity.

[0075] One or more embodiments of this invention are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of this invention. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments of this invention should be included within the protection scope of this invention.

Claims

1. A substation maintenance data management method based on data analysis, characterized in that... The process includes the following steps: S101. Obtain historical maintenance data of substation equipment in the target substation and the associated maintenance technical documents; S102. Extract technical entities and entity relationships from the maintenance technical documents, combine technical entities and entity relationships to generate a set of mathematical constraints to characterize the preset technical semantics, and construct a semantic geometric construct library; S103. Combine the analysis of maintenance technical documents and historical maintenance data to construct an evolution path set containing the evolutionary relationships between preset technical semantics; S104. Obtain real-time equipment time-series data of substation equipment and process the equipment time-series data into data probes reflecting the current equipment status of substation equipment; S105. Obtain a set of docking targets by calculating the docking affinity scores between the data probes and each semantic geometric construct in the semantic geometric construct library; S106. Project the docking targets onto the evolution path set, identify the target evolution path matching the docking targets in the evolution path set, and assess the equipment risk of substation equipment based on the target evolution path; S107. Combine equipment risk and target evolution path backtracking to generate maintenance instructions for substation equipment.

2. The substation maintenance data management method based on data analysis according to claim 1, characterized in that... In step S102, extracting technical entities and entity relationships from the maintenance technical documents, generating a set of mathematical constraints to represent the preset technical semantics, and constructing a semantic geometric construct library includes the following steps: processing the maintenance technical documents using a preset entity recognition model to extract four types of technical entities: equipment components, fault phenomena, physical parameters, and operation instructions; analyzing the co-occurrence relationships of technical entities in the context of the maintenance technical documents using a relation extraction model to generate semantic relation triples between technical entities; loading all semantic relation triples into the initial graph data structure, and constructing an initial equipment maintenance knowledge graph through knowledge fusion and disambiguation processing of the initial graph data structure; configuring one or more sets of corresponding mathematical constraints for the preset technical semantic nodes in the equipment maintenance knowledge graph; The preset technical semantic nodes and mathematical constraint sets are structured and associated for storage, forming a semantic geometric construct library.

3. The substation maintenance data management method based on data analysis according to claim 1, characterized in that... In step S105, a set of docking targets are obtained by calculating the docking affinity scores between the data probe and each semantic geometric construct in the semantic geometric construct library. This includes the following steps: mapping the data probe to a preset multidimensional feature space, and generating a dynamic target hypersurface for each semantic geometric construct in the multidimensional feature space according to the set of mathematical constraints corresponding to the semantic geometric construct. The weighted geometric distance from the feature points mapped by the data probe in the multidimensional feature space to each target hypersurface is calculated. The weighted geometric distance is normalized and inversely processed to obtain the docking affinity score. The data probe and the semantic geometric conformations with docking affinity scores higher than the preset docking threshold are regarded as a group of docking targets.

4. The substation maintenance data management method based on data analysis according to claim 3, characterized in that, The process involves mapping a data probe to a predefined multidimensional feature space, generating a dynamic target hypersurface for each semantic geometric construct based on the set of mathematical constraints corresponding to the semantic geometric construct in the multidimensional feature space, including the following steps: extracting context state variables representing the macroscopic operating background of the current substation equipment from the data probe; constructing a dynamic Riemann metric tensor based on the context state variables for the predefined multidimensional feature space, wherein the Riemann metric tensor is used to define the local geometric curvature of the multidimensional feature space under the macroscopic operating background; mapping the set of mathematical constraints to the multidimensional feature space defined by the local geometric curvature; solving for the geodesic hypersurface corresponding to the set of mathematical constraints in the multidimensional feature space; and using the geodesic hypersurface as the dynamic target hypersurface.

5. The substation maintenance data management method based on data analysis according to claim 3, characterized in that, The weighted geometric distance from the feature points of the data probe mapped in the multidimensional feature space to each target hypersurface is calculated by the following steps: decomposing the set of mathematical constraints into a set of independent feature constraint rules, each feature constraint rule corresponding to a dimension in the multidimensional feature space; for each feature constraint rule, calculating the deviation of the feature value of the data probe in the corresponding dimension from the target interval defined by the feature constraint rule to obtain the single-dimensional feature satisfaction; and extracting the association constraint weights representing the relationship between different feature dimensions from the semantic geometric conformation library. The unidimensional feature satisfaction of all dimensions is weighted and summed using the association constraint weights, and the result of the weighted sum is used as the weighted geometric distance from the feature points mapped by the data probe in the multidimensional feature space to the target hypersurface.

6. The substation maintenance data management method based on data analysis according to claim 1, characterized in that... In step S106, the docking target is projected onto the evolution path set, the target evolution path matching the docking target is identified in the evolution path set, and the equipment risk of the substation equipment is evaluated based on the target evolution path, including the following steps: using a set of docking targets as initial activation nodes, node activation diffusion is carried out in the evolution path set; The cumulative activation intensity received by each node in the evolution path set is calculated based on a pre-defined diffusion propagation model. Based on the cumulative activation intensity and the preset transition probability of the path connection edges in the evolution path set, the maximum probability path search algorithm is used to identify one or more target evolution paths in the evolution path set; for each target evolution path, the path length of the target evolution path, the average docking affinity score of the nodes contained in the target evolution path, and the preset hazard level of the terminal node of the target evolution path are extracted. By combining the path length, average docking affinity score, and preset hazard level of the terminal node, a risk assessment function is constructed to calculate the comprehensive risk score of the docking target for the strain power station equipment.

7. The substation maintenance data management method based on data analysis according to claim 6, characterized in that, The cumulative activation intensity is calculated using the following formula until convergence: ,in, For the nodes of the evolution path The cumulative activation intensity at the (k+1)th iteration; An evolutionary propagation factor characterizing the degree of dependence on propagation within an evolutionary path set; The node is determined by the docking target. The initial activation strength; For all pointing nodes in the fault signal path network The set of upstream nodes; upstream node The cumulative activation intensity at the k-th iteration; For upstream nodes Pointing to node The preset transition probability of the path connecting edges; upstream node The sum of the preset transition probabilities of all outgoing path connecting edges.

8. The substation maintenance data management method based on data analysis according to claim 6, characterized in that, An evolutionary path set containing the evolutionary relationship between preset technical semantics is constructed by analyzing maintenance technical documents and historical maintenance data. The steps include: constructing a network node set based on all technical semantic nodes in the semantic geometric model library; statistically analyzing the chronological order and time interval of the technical semantics corresponding to different technical semantic nodes in the historical maintenance data, and calculating the temporal transition probability between network nodes in the network node set. By retrieving and analyzing textual statements describing the causal chain of faults in maintenance technical documents, logical causal relationships between network nodes are extracted; by fusing temporal transition probabilities and logical causal relationships, comprehensive connection weights are generated for each pair of network nodes, and a weighted directed graph is constructed based on the comprehensive connection weights as an evolution path set.

9. A data analysis-based substation maintenance data management system for implementing the method of claim 1, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the substation maintenance data management method based on data analysis as described in any one of claims 1 to 8.

10. A computer-readable storage medium storing instructions thereon, characterized in that, When executed by a processor, this instruction causes the processor to be configured to perform the substation maintenance data management method based on data analysis according to any one of claims 1 to 8.

Images