Methods, apparatus, computer equipment, readable storage media, and program products for generating stator partial discharge detection reports.

Abstract

This application relates to a method, apparatus, computer equipment, computer-readable storage medium, and computer program product for generating a stator partial discharge detection report, belonging to the field of power technology, and can improve the efficiency of generating stator partial discharge detection reports. The method includes: establishing a knowledge graph of stator partial discharge; acquiring discharge features corresponding to real-time partial discharge data; comparing the discharge features with entity features of multiple entities in the knowledge graph; determining entities matching the discharge features based on the comparison results; determining a target fault mode based on the entities matching the discharge features; determining a decision support path including a fault causal chain and a fault maintenance strategy based on the target fault mode and the relationships in the knowledge graph; and describing the decision support path in natural language using a large language model to obtain a stator partial discharge fault detection report.

Description

Technical Field

[0001] This application relates to the field of power technology, and in particular to a method, apparatus, computer equipment, computer-readable storage medium, and computer program product for generating a report on stator partial discharge detection. Background Technology

[0002] With the development of intelligent monitoring technology for power equipment status, online partial discharge monitoring technology combining multi-sensor data acquisition and machine learning algorithms has emerged. In the traditional process, stator discharge signals and generator operating parameters are collected using high-frequency current transformers and ultra-high-frequency sensors. Then, signal processing techniques such as wavelet transform are used to extract discharge features. Finally, machine learning models such as support vector machines are used to identify the discharge type and output a preliminary report containing statistical data and simple alarms. However, the generated report is simplistic and fails to efficiently provide a comprehensive fault analysis report that can directly assist in operation and maintenance decisions. Summary of the Invention

[0003] Therefore, it is necessary to provide a method, apparatus, computer equipment, computer-readable storage medium, and computer program product for generating a stator partial discharge detection report to address the aforementioned technical problems.

[0004] In a first aspect, this application provides a method for generating a stator partial discharge detection report, comprising:

[0005] A knowledge graph of stator partial discharge is established; the knowledge graph includes multiple entities related to stator partial discharge faults and the relationships between these entities; the multiple entities include fault mode entities, equipment parameter entities, maintenance instance entities, and industry standard entities;

[0006] Obtain the discharge features corresponding to real-time partial discharge data, compare the discharge features with the entity features of multiple entities in the knowledge graph, determine the entity that matches the discharge features based on the comparison results, and determine the target fault mode based on the entity that matches the discharge features.

[0007] Based on the target failure mode and the correlation in the knowledge graph, a decision support path including the failure causal chain and failure maintenance strategy is determined.

[0008] The decision support path is described in natural language using a large language model to obtain a fault detection report of stator partial discharge.

[0009] In one embodiment, the discharge feature is compared with the entity features of a plurality of entities in the knowledge graph, and the entity matching the discharge feature is determined based on the comparison result. The target fault mode is then determined based on the entity matching the discharge feature, including:

[0010] The discharge features and the entity features of each entity are transformed into a unified feature vector space, and the transformed discharge features are combined with the transformed entity features to obtain joint features;

[0011] In the feature vector space, the correlation score of each joint feature is calculated;

[0012] Based on the correlation scores of each joint feature, the degree of correlation between each entity and the discharge feature relative to other entities is determined, and entities whose correlation degree meets preset conditions are obtained.

[0013] If the entity is the device parameter entity, then based on the association relationships related to the device parameter entity, a target fault mode matching the discharge characteristics is determined; and,

[0014] If the entity is the fault mode entity, then the fault mode indicated by the fault mode entity is taken as the target fault mode.

[0015] In one embodiment, the entity features are obtained through the following steps:

[0016] Obtain the entity vector of each entity and the relationship vector of the association between each entity;

[0017] For each set of associations, the head entity and tail entity associated with the association, the sum of the association vector corresponding to the association and the head entity vector corresponding to the head entity is obtained. Based on the difference between the sum and the tail entity vector corresponding to the tail entity, the difference between the entity vector obtained by transforming the head entity vector through the association and the tail entity vector is determined.

[0018] A loss value is determined based on the multiple differences, and each entity vector and each relation vector is iteratively adjusted based on the loss value until a preset iteration termination condition is met. Each entity feature is obtained based on the current entity vectors.

[0019] In one embodiment, after acquiring real-time partial discharge data and extracting the discharge features of the real-time partial discharge data, the method further includes:

[0020] The discharge features are input into a pre-trained risk level determination model; the risk level determination model includes the probabilistic dependency between each risk level and each discharge feature; the probabilistic dependency indicates the conditional probability of the discharge feature under different risk levels, and the current occurrence probability of each risk level;

[0021] The risk level determination model outputs the current risk level of the stator partial discharge based on the probability dependency and the discharge characteristics.

[0022] In one embodiment, the determination of the risk quantification parameter corresponding to the current risk level includes:

[0023] Based on preset weights, the standardized discharge characteristics and the real-time partial discharge data corresponding to the standardized discharge characteristics are weighted to determine the risk score under the current risk level.

[0024] Determine the degree of deviation of the risk score from the preset benchmark risk score, and determine the risk quantification parameter corresponding to the current risk level based on the degree of deviation.

[0025] In one embodiment, before obtaining the stator partial discharge fault detection report by describing the decision support path in natural language using a large language model, the method further includes:

[0026] Based on the discharge characteristics and the decision support path, the keywords corresponding to the fields of the preset structured template are determined, and a structured fault detection report is generated based on the keywords.

[0027] Before obtaining the stator partial discharge fault detection report by describing the decision support path in natural language using a large language model, the method further includes:

[0028] Obtain the translated content of the fault detection report in the target language, and display the fault detection report based on the translated content.

[0029] Secondly, this application also provides a device for generating a stator partial discharge detection report, comprising:

[0030] A partial discharge knowledge graph building module is used to build a knowledge graph of stator partial discharge; the knowledge graph includes multiple entities related to stator partial discharge faults and the relationships between these entities; the multiple entities include fault mode entities, equipment parameter entities, maintenance instance entities, and industry standard entities.

[0031] The target fault mode determination module is used to acquire the discharge features corresponding to real-time partial discharge data, compare the discharge features with the entity features of multiple entities in the knowledge graph, determine the entity that matches the discharge features based on the comparison results, and determine the target fault mode based on the entity that matches the discharge features.

[0032] The decision support path determination module is used to determine a decision support path, including the fault causal chain and the fault repair strategy, based on the target fault mode and the correlation in the knowledge graph.

[0033] The fault detection report generation module is used to describe the decision support path in natural language using a large language model to obtain a fault detection report of stator partial discharge.

[0034] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0035] A knowledge graph of stator partial discharge is established; the knowledge graph includes multiple entities related to stator partial discharge faults and the relationships between these entities; the multiple entities include fault mode entities, equipment parameter entities, maintenance instance entities, and industry standard entities;

[0036] Obtain the discharge features corresponding to real-time partial discharge data, compare the discharge features with the entity features of multiple entities in the knowledge graph, determine the entity that matches the discharge features based on the comparison results, and determine the target fault mode based on the entity that matches the discharge features.

[0037] Based on the target failure mode and the correlation in the knowledge graph, a decision support path including the failure causal chain and failure maintenance strategy is determined.

[0038] The decision support path is described in natural language using a large language model to obtain a fault detection report of stator partial discharge.

[0039] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0040] A knowledge graph of stator partial discharge is established; the knowledge graph includes multiple entities related to stator partial discharge faults and the relationships between these entities; the multiple entities include fault mode entities, equipment parameter entities, maintenance instance entities, and industry standard entities;

[0041] Obtain the discharge features corresponding to real-time partial discharge data, compare the discharge features with the entity features of multiple entities in the knowledge graph, determine the entity that matches the discharge features based on the comparison results, and determine the target fault mode based on the entity that matches the discharge features.

[0042] Based on the target failure mode and the correlation in the knowledge graph, a decision support path including the failure causal chain and failure maintenance strategy is determined.

[0043] The decision support path is described in natural language using a large language model to obtain a fault detection report of stator partial discharge.

[0044] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:

[0045] A knowledge graph of stator partial discharge is established; the knowledge graph includes multiple entities related to stator partial discharge faults and the relationships between these entities; the multiple entities include fault mode entities, equipment parameter entities, maintenance instance entities, and industry standard entities;

[0046] Obtain the discharge features corresponding to real-time partial discharge data, compare the discharge features with the entity features of multiple entities in the knowledge graph, determine the entity that matches the discharge features based on the comparison results, and determine the target fault mode based on the entity that matches the discharge features.

[0047] Based on the target failure mode and the correlation in the knowledge graph, a decision support path including the failure causal chain and failure maintenance strategy is determined.

[0048] The decision support path is described in natural language using a large language model to obtain a fault detection report of stator partial discharge.

[0049] The aforementioned method, apparatus, computer equipment, computer-readable storage medium, and computer program product for generating stator partial discharge detection reports establish a knowledge graph of stator partial discharge. This knowledge graph includes multiple entities related to stator partial discharge faults and the relationships between these entities. These entities include fault mode entities, equipment parameter entities, maintenance instance entities, and industry standard entities. The method acquires discharge features corresponding to real-time partial discharge data, compares these features with the entity features of the multiple entities in the knowledge graph, determines the entities matching the discharge features based on the comparison results, and identifies the target fault mode based on the matching entities. Based on the target fault mode and the relationships in the knowledge graph, a decision support path, including fault causal chains and fault maintenance strategies, is determined. Finally, a natural language description of the decision support path is performed using a large language model to obtain the stator partial discharge fault detection report. In this application, a knowledge graph containing fault modes and their associations is constructed to lay a solid knowledge foundation for fault analysis. Subsequently, real-time extracted discharge features are intelligently matched with entities in this graph to accurately pinpoint the target fault mode consistent with the current operating condition. Based on this, the inherent logical relationships within the knowledge graph are further utilized to deduce decision support paths, revealing the fault evolution mechanism or recommended response strategies. Finally, a large language model automatically transforms the fault handling path into an easily understandable natural language detection report, completing a closed loop from raw data to intelligent detection and improving the quality and efficiency of stator partial discharge detection reports. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 This is a flowchart illustrating a method for generating a stator partial discharge detection report in one embodiment;

[0052] Figure 2 This is a flowchart illustrating a method for generating a stator partial discharge detection report in another embodiment;

[0053] Figure 3 This is a flowchart illustrating a method for generating a stator partial discharge detection report in yet another embodiment;

[0054] Figure 4 This is a structural block diagram of a stator partial discharge detection report generation device in one embodiment;

[0055] Figure 5This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0056] 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.

[0057] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0058] In one embodiment, such as Figure 1 As shown, a method for generating a stator partial discharge detection report 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 further to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0059] Step S102: Establish a knowledge graph of stator partial discharge.

[0060] The knowledge graph includes, but is not limited to, multiple entities related to stator partial discharge faults and the relationships between these entities. These entities include, but are not limited to, fault mode entities, equipment parameter entities, maintenance instance entities, and industry standard entities.

[0061] Knowledge graphs are semantic networks based on graph data structures. They store knowledge in the form of nodes and edges. Nodes represent entities in the real world—anything that can be distinguished from other things, such as a specific fault mode, a device parameter, or an industry standard. Edges represent objectively existing relationships between entities; for example, there can be a relationship between the fault mode entity "insulation aging" and the device parameter entity "overheating winding." The basic building block of a knowledge graph is typically a triple of head entity-relationship-tail entity.

[0062] For example, a knowledge graph can be constructed using pre-defined rules by experts. Specifically, for instance, firstly, one or more experts in generator fault diagnosis jointly define the schema layer of the knowledge graph, including determining the types of entities to be included in the graph (such as the aforementioned fault mode entities, equipment parameter entities, maintenance instance entities, and industry standard entities) and the types of relationships (such as fault results, associated environmental factors, repair solutions, and standard violations). Subsequently, based on their professional knowledge and existing documentation, the expert team or knowledge engineers use semi-automated tools to input specific knowledge instances (such as fault mode: stator slot discharge; equipment parameter: rated voltage 35kV) into the knowledge graph database in the form of triples, forming the initial knowledge graph.

[0063] Step S104: Obtain the discharge features corresponding to the real-time partial discharge data, compare the discharge features with the entity features of multiple entities in the knowledge graph, determine the entities that match the discharge features based on the comparison results, and determine the target fault mode based on the entities that match the discharge features.

[0064] The discharge characteristics can be one or more sets of quantitative indicators extracted from the raw, continuous, real-time partial discharge data waveforms acquired by the sensor to facilitate quantitative analysis and comparison. These characteristics can be obtained as numerical values, vectors, or pattern diagrams through signal processing algorithms (such as Fast Fourier Transform, Wavelet Transform, Hilbert Transform, etc.) and statistical analysis. For example, discharge characteristics include, but are not limited to, the amplitude of the discharge pulse, pulse count, signal rise time, energy integral value, statistical characteristics of the discharge phase distribution spectrum (such as skewness and kurtosis), and spectral energy distribution of the discharge signal.

[0065] Entity features can be a set of descriptive attributes or values ​​associated with each entity node. These features define the specific characteristics of the entity within the domain knowledge. In the embodiments of this application, entity features can be symbolic descriptions (such as text labels), numerical values, or more complex structured data (such as a typical feature template). For example, for a fault mode entity (such as stator slot discharge), its entity features can be a set of typical discharge characteristics associated with that fault mode, representing a baseline value or range; for a device parameter entity (such as generator rated voltage), its entity features are the specific numerical value or category of that parameter (e.g., 35kV).

[0066] For example, the terminal continuously receives raw electrical signal waveforms from a high-frequency current transformer (HFCT) or an ultra-high frequency (UHF) sensor coupled to the generator stator. The signal processing unit within the terminal processes these waveform data in real time. For instance, it extracts energy features of different frequency bands through wavelet packet decomposition, or constructs a discharge phase distribution map through phase analysis technology, and calculates statistical features such as shape factor and asymmetry from the map, thereby forming a real-time discharge feature vector containing multiple dimensions of features.

[0067] Optionally, the entity feature of each fault mode entity in the knowledge graph can be a standardized fault feature template. This template can be a multi-dimensional vector, with each dimension corresponding to a typical discharge feature value. When the terminal obtains the real-time discharge feature vector, it calculates the distance or similarity between this real-time vector and each fault feature template stored in the knowledge graph. This distance can be Euclidean distance, Mahalanobis distance, or the similarity can be cosine similarity. The terminal identifies the fault mode entity corresponding to the fault feature template with the smallest distance (or the highest similarity) as the matching entity and uses it as the target fault mode.

[0068] Optionally, the matching process is not limited to directly matching fault mode entities. If the terminal determines that a device parameter entity (e.g., entity characteristic: current winding temperature > 30°C) highly matches the real-time discharge characteristics, the terminal will further query the knowledge graph to find fault mode entities (e.g., insulation thermal aging) that have a causal or strong correlation with the device parameter entity, and use them as the inferred target fault mode.

[0069] Step S106: Based on the target failure mode and the relationships in the knowledge graph, a decision support path including the failure causal chain and failure maintenance strategy is determined.

[0070] The decision support path can be a subgraph extracted or inferred from the main knowledge graph or a structured data object (such as JSON or XML). This path uses a defined target fault mode entity as its core node and includes a series of upstream (fault causes) and downstream (fault repair strategies) entities and relationship chains connected to this core node through specific associations. This provides maintenance personnel with a panoramic view of the fault, greatly assisting them in fault confirmation, root cause analysis, and repair decisions. In one embodiment, for example, the decision support could be: increased discharge pulse density → [fault result] → insulation aging → [violation] → industry standard A → [repair plan] → replacement of insulation material.

[0071] For example, the terminal uses a pre-defined set of query templates in a structured graph query language (e.g., SPARQL or Cypher). Upon receiving a target failure mode, the terminal fills the identifier of that failure mode into the corresponding query template. For instance, a template might be designed to query all entity paths connected to the [target failure mode] through failure outcome relationship chains, and simultaneously query all entities connected to the [target failure mode] through repair scheme relationships. The terminal then executes this fully populated query statement into the knowledge graph database. The query results returned by the database are themselves a structured dataset containing information on all matching nodes and paths, which the terminal directly organizes into decision support paths.

[0072] In one specific embodiment, the terminal starts from the insulation aging entity node in the knowledge graph and performs reverse causal tracing. It can find a fault outcome relationship pointing to insulation aging and discover that its upstream associated entity is increased discharge pulse density. This chain from cause / phenomenon to outcome (increased discharge pulse density → [fault outcome] → insulation aging) constitutes the fault causal chain part in the decision support path. This reveals to the user the direct representation of the fault outcome, insulation aging.

[0073] In some embodiments, the terminal can also explore forward relationships starting from the insulation aging entity node. First, a violation relationship is discovered, pointing to an industry standard entity, namely industry standard A. This indicates that the current insulation aging condition may no longer meet the requirements of this industry standard. Next, the terminal can find a repair solution relationship starting from the insulation aging node or the industry standard A node. This relationship points to a repair instance entity or a specific repair action, which could be replacing the insulation material. This chain from fault to standard to solution (insulation aging → [violation] → industry standard A → [repair solution] → replacement of insulation material) constitutes the fault repair strategy part of the decision support path.

[0074] Step S108: The decision support path is described in natural language using a large language model to obtain a fault detection report of stator partial discharge.

[0075] The large language model can be a deep learning-based model, such as a neural network model with a Transformer architecture. Its characteristics include a massive number of model parameters (typically billions or more), and it is pre-trained on extremely large text corpora, thereby acquiring rich linguistic rules, factual knowledge, and a certain degree of reasoning ability. In this application, examples that can be used include, but are not limited to, models based on architectures such as GPT (Generative Pre-trained Transformer) or BART / T5, which have been fine-tuned or possess context learning capabilities. Through their Natural Language Generation (NLG) capabilities, these models understand the logical, causal, and strategic information contained in decision support paths input in specific formats (such as structured text, JSON, or vectorized representations), and generate grammatically correct and logically clear natural language text based on this information.

[0076] For example, the terminal first serializes the entire decision support path, for instance, by converting it into a detailed, descriptive text string or a JSON object. Then, using this serialized string as context, the terminal constructs a complete instruction prompt, such as: "You are a generator fault diagnosis expert. Please write a detailed fault detection report based on the following structured information: [Insert serialized decision support path here]". This prompt is directly sent to a general-purpose large language model. The large language model understands all the input information at once and directly generates a complete fault detection report text containing all parts in an end-to-end manner.

[0077] In this embodiment, a knowledge graph containing fault modes and their associations is constructed to lay a solid knowledge foundation for fault analysis. Subsequently, real-time extracted discharge features are intelligently matched with entities in this graph to accurately pinpoint the target fault mode matching the current operating condition. Based on this, the inherent relational logic of the knowledge graph is further utilized to deduce a decision support path, revealing the fault evolution mechanism or recommended response strategies. Finally, a large language model automatically transforms the fault handling path into an easily understandable natural language detection report, completing the closed loop from raw data to intelligent detection and improving the quality and efficiency of stator partial discharge detection reports.

[0078] In an exemplary embodiment, the discharge characteristics are compared with the entity characteristics of multiple entities in the knowledge graph. Based on the comparison results, entities matching the discharge characteristics are determined. Based on the entities matching the discharge characteristics, the target fault mode is determined, including:

[0079] The discharge features and entity features of each entity are transformed into a unified feature vector space, and the transformed discharge features are combined with the transformed entity features to obtain joint features. In the feature vector space, the correlation score of each joint feature is calculated. Based on the correlation score of each joint feature, the correlation degree of each entity with the discharge features relative to other entities is determined, and entities with correlation degrees that meet preset conditions are obtained. If the entity is a device parameter entity, the target fault mode matching the discharge feature is determined based on the correlation relationship of the device parameter entity. If the entity is a fault mode entity, the fault mode indicated by the fault mode entity is taken as the target fault mode.

[0080] The feature vector space can be a multi-dimensional mathematical space, where each dimension corresponds to a specific quantitative feature. This allows information with different physical meanings and data types (such as real-time discharge features from sensors and entity features from knowledge graphs) to be converted into a unified format (i.e., feature vectors), thereby supporting standardized quantitative comparisons and algebraic operations.

[0081] Joint features can be new feature vectors formed by combining two or more independent feature vectors. One possible combination method is vector concatenation, which connects two vectors end-to-end to form a new vector with a longer dimension that contains all the information of the two original vectors.

[0082] The correlation score can be an unnormalized scalar value, an intermediate result in the correlation calculation process, reflecting the degree of matching between a knowledge graph entity and the current real-time discharge feature. A higher value generally indicates a higher degree of matching. The correlation degree can be a normalized value, typically distributed between 0 and 1, and the sum of all correlation degrees for a comparison set (e.g., all relevant neighbor entities) is 1. It is used to distinguish the similarity of different entities relative to the discharge feature.

[0083] For example, the terminal performs a transformation of the discharge features and the entity features of each entity into a unified feature vector space, and concatenates the transformed discharge features with the transformed entity features to obtain joint features. A set of optimized parameterized mathematical operations is applied to each joint feature to calculate its correlation score. After obtaining the original correlation scores of all entities paired with the current discharge feature, the terminal performs a normalization process, for example, transforming all the original scores through a function that highlights relative size (such as the Softmax function) to generate a set of standardized correlation weights. The terminal then obtains one or more entities with the highest correlation or that meet a preset threshold as the matching result. In a specific embodiment, the correlation between the discharge feature vector and any entity feature vector can be realized using the attention mechanism formula (1). The calculation.

[0084]

[0085] in, It is a shared linear transformation matrix used to transform features to a unified eigenvector space. It is the discharge characteristic vector. It is the entity feature vector. It is a learnable attention parameter vector. It is the set of neighboring nodes. Other neighboring nodes, These are the entity feature vectors of other neighboring nodes. This is achieved through analysis of the joint features. The overall operation constitutes the computational processing of joint features, the result of which is an original association score. The original association scores calculated for all neighboring entities are summed, and the entire fractional structure is a Softmax function used for normalization, thereby calculating the final association degree of a node relative to all other neighboring nodes.

[0086] After obtaining the entity with the highest relevance or that meets the preset conditions through any of the above methods, the following logical judgment is executed:

[0087] If the entity that meets the conditions is a device parameter entity (for example, matching the winding temperature entity, and its state is too high), the terminal will further query the knowledge graph to find a fault mode entity that has a preset association with the device parameter entity (such as fault result or high correlation), and will return the insulation thermal aging entity. At this time, the terminal will take this inferred insulation thermal aging as the final target fault mode.

[0088] If the entity that meets the conditions is itself a fault mode entity (for example, directly matched to the end corona discharge entity), then the terminal directly uses the fault mode indicated by the entity (i.e., end corona discharge) as the target fault mode for this detection.

[0089] In this embodiment, by using a matching method based on feature vector space and combined with attention mechanism, it is possible to capture deeper and non-linear relationships between features, dynamically learn the importance of different features and neighboring entities, thereby significantly improving the accuracy and robustness of fault mode recognition.

[0090] In one exemplary embodiment, the features of each entity are obtained through the following steps:

[0091] Obtain the entity vectors of each entity and the relation vectors of the relationships between each entity; for each set of relationships, the head entity and the tail entity associated with the relationship, obtain the sum of the relationship vector corresponding to the relationship and the head entity vector corresponding to the head entity; based on the difference between the sum and the tail entity vector corresponding to the tail entity, determine the difference between the entity vector obtained by transforming the head entity vector through the relationship and the tail entity vector; determine the loss value based on multiple differences, and iteratively adjust each entity vector and each relation vector based on the loss value until the preset iteration termination condition is met; obtain the features of each entity based on the current entity vectors.

[0092] Specifically, the terminal will traverse all entity and relation types in the knowledge graph and initialize a vector of a specified dimension for each entity and relation. For any correct triple in the knowledge graph (e.g., head entity h, relation r, tail entity t), their corresponding vectors (h, r, t) should approximately satisfy a simple translation relation in the feature vector space, that is, the head entity vector plus the relation vector should be approximately equal to the tail entity vector, as shown in formula (2).

[0093]

[0094] in, Represents a knowledge graph.

[0095] For each correct triplet, the terminal calculates the sum of the head entity vector and the relation vector, and then calculates the difference or distance between this sum and the tail entity vector. This difference or distance represents the model's prediction error for this correct fact; ideally, this difference should be close to zero. Specifically, a total loss value is calculated using a preset loss function, as shown in formula (3). Then, the terminal uses an optimization algorithm (e.g., stochastic gradient descent) to make minor adjustments to each entity vector and relation vector involved in the calculation based on the magnitude and gradient direction of the loss value, so that the adjusted vectors can make the total loss value... The loss decreases until a preset iteration termination condition is met (e.g., the loss value no longer decreases significantly, or a preset number of iterations is reached). Once the iteration process meets the termination condition, all current entity vectors are considered stable and optimized, and they will be solidified as the final entity features for subsequent steps such as fault mode recognition.

[0096]

[0097] in, It is actually a preset interval parameter, which the terminal can minimize through an optimization algorithm. In this process, the entity vectors and relation vectors are iteratively adjusted.

[0098] In this embodiment, the symbolic, discrete knowledge graph is transformed into a continuous vector space model capable of mathematical operations. This not only enables efficient real-time reasoning of the knowledge graph, but more importantly, it endows the system with the ability to process unknown faults and predict missing relationships based on semantic similarity, thereby improving detection efficiency.

[0099] In an exemplary embodiment, after acquiring real-time partial discharge data and extracting the discharge characteristics of the real-time partial discharge data, the method further includes:

[0100] The discharge characteristics are input into a pre-trained risk level determination model. The risk level determination model includes the probabilistic dependencies between each risk level and each discharge characteristic. The probabilistic dependencies indicate the conditional probability of the discharge characteristics under different risk levels, as well as the current probability of occurrence of each risk level. Based on the probabilistic dependencies and discharge characteristics, the risk level determination model outputs the current risk level of stator partial discharge.

[0101] Among them, the pre-trained risk level determination model is a machine learning model based on probabilistic graphical theory. Specifically, it can be a Bayesian network that, when receiving real-time discharge characteristics, can perform probabilistic reasoning based on learned knowledge to infer the true risk level.

[0102] Probabilistic dependencies are the specific mathematical expression of the knowledge within the model. They formally define the mutual influence and correlation strength among various variables (risk level and discharge characteristics) in the system. This specifically includes two parts: first, conditional probability, which quantifies the probability of a specific discharge characteristic occurring at a given risk level, reflecting the influence of the risk level on the discharge characteristic; and second, the current probability of occurrence of each risk level (also known as prior probability in Bayesian theory), which represents the inherent probability of occurrence of each risk level based on historical statistics or expert experience before any real-time observation data is available.

[0103] In one specific embodiment, after extracting real-time discharge features, the terminal performs a risk level determination step. The pre-trained risk level determination model is specifically a Bayesian network. The network topology is designed such that nodes representing four predefined risk levels (low, medium, high, and emergency) serve as parent nodes, while nodes representing various discharge features extracted from real-time data (e.g., pulse density, skewness, etc. in feature vector F) serve as their child nodes. Directed edges from risk level nodes to discharge feature nodes represent the probabilistic dependencies between them.

[0104] During the pre-training phase, the terminal analyzes historical data to populate the conditional probability table for each discharge feature node. This table stores the conditional probability of the discharge feature taking different values ​​(such as high, medium, and low) at each specific risk level. At the same time, it also determines the prior probability for each risk level node.

[0105] During real-time detection, when the terminal acquires a new set of discharge features, it uses the Bayesian network to perform inference, calculates the posterior probability of each risk level, and outputs the risk level with the highest posterior probability as the current risk level. Specifically, as shown in formula (4), the posterior probability of a high risk level is calculated.

[0106]

[0107] Where F is the discharge characteristic vector, It is the conditional probability of a high or low risk level. It is the probability that the current risk level is high. This indicates the available risk levels, such as low, medium, high, and emergency. It is the conditional probability under risk level c. It is the probability that the current risk level is c.

[0108] In this embodiment, by using prior knowledge based on historical statistics or expert experience (i.e., prior probability of risk level) and real-time evidence from the current operating conditions (i.e. discharge characteristics), and employing a Bayesian network equal probability model, the discharge characteristics can be accurately mapped to a specific risk level, which can be used to instruct maintenance personnel to better handle the current risk fault.

[0109] In one exemplary embodiment, the determination of risk quantification parameters corresponding to the current risk level includes:

[0110] Based on preset weights, the standardized discharge characteristics and the corresponding real-time partial discharge data are weighted to determine the risk score under the current risk level; the degree of deviation of the risk score from the preset benchmark risk score is determined, and the risk quantification parameters corresponding to the current risk level are determined based on the degree of deviation.

[0111] The risk score can be a comprehensive scalar value, obtained by weighted summation of multiple standardized risk impact factors (including discharge characteristics and equipment status parameters). Its function is to compress and integrate multi-dimensional, fragmented risk information into a single, uniformly assessable value. The preset benchmark risk score can be a key threshold pre-defined through statistical analysis of historical data. Physically, it represents the critical point where risk shifts from one order of magnitude to another, serving as a reference benchmark for measuring the current risk score. The risk quantification parameter can be a continuous value, typically normalized to a specific range (e.g., 0 to 1). It is not a discrete level label but a quantitative description of the severity of the risk. In some embodiments, it represents the membership degree or confidence level of the current state belonging to a risk set.

[0112] Specifically, firstly, the terminal performs a risk score determination step. This step, based on preset weights, linearly weights and sums multiple standardized risk factors to determine the risk score. Risk factors exemplarily include standardized discharge characteristics (e.g., pulse density, skewness, etc.) and standardized equipment state parameters (e.g., winding temperature, load rate, etc.). Subsequently, the terminal performs a risk quantification parameter determination step. In an exemplary implementation, this step is achieved through a non-linear membership function. This function aims to determine the deviation of the risk score from a preset benchmark risk score, and based on the degree of this deviation and a sensitivity parameter controlling the degree of change, the final risk quantification parameter is determined. In one embodiment, the calculation process can be represented by the following formula (5):

[0113]

[0114] in, This is the risk score calculated in the previous step. To pre-set a baseline risk score, Both of these sensitivity parameters, which control the steepness of the curve, are predetermined by analyzing and calibrating historical data.

[0115] In this embodiment, the above steps enable the fuzzy risk level to be represented by quantitative risk parameters, allowing maintenance personnel to have a more objective understanding of the fault risk, enabling the entire system to have early warning and fault prediction capabilities, and improving the efficiency of fault handling.

[0116] In an exemplary embodiment, before obtaining the stator partial discharge fault detection report by describing the decision support path in natural language using a large language model, the method further includes:

[0117] Based on discharge characteristics and decision support paths, the keywords corresponding to the fields of the preset structured template are determined, and a structured fault detection report is generated based on the keywords.

[0118] The pre-defined structured template can be a predefined report document framework, technically a text file with placeholders, an Extensible Markup Language (XML) object, or a JSON object. It provides a unified and standardized format and structure for the final fault detection report, ensuring the standardization and consistency of the report content organization. Keywords can be core information fragments extracted from discharge characteristics and decision support paths that can be precisely filled into the placeholders in the structured template, such as a specific fault mode name, a quantified risk level value, or a brief description of maintenance measures.

[0119] In one specific embodiment, the terminal parses the determined decision support path, quickly extracts key information such as fault location (LOC) and risk level (RISK), and fills these keywords into a predefined report framework containing placeholders such as {LOC} and {RISK}, thereby rapidly generating the structured summary part of the report.

[0120] Optionally, for analytical content that needs to be elaborated in detail in the report, such as a detailed interpretation of the causal chain of the failure, the terminal will call a pre-trained text generation model based on the Transformer architecture, as shown in formula (6). This model receives the graph encoding vector of the knowledge graph path as its conditional input and generates text based on the autoregressive principle.

[0121]

[0122] Specifically, the process by which this model generates report paragraphs follows a probabilistic chain rule, that is, throughout the entire word group... Next, predict and generate the next word in the report text one by one. ), its generation probability It depends on two core pieces of information, the previously generated preceding content ( To ensure smooth writing and the initial input vector serving as the basis for the facts ( To ensure the accuracy of the content, the model continuously selects and concatenates the highest-probability words in each step, generating logically coherent and detailed natural language paragraphs. Finally, the terminal assembles the structured information filled in by the template with the narrative paragraphs generated by deep learning to form a complete and appropriately detailed final fault detection report.

[0123] In this embodiment, a deterministic keyword filling method is used for critical factual data such as fault names and risk levels that cannot be wrong. This can completely avoid the risk of information errors or omissions that may occur during the generation process of the large language model, and ensure the accuracy of the core conclusions of the report.

[0124] Accordingly, in an exemplary embodiment, after obtaining the stator partial discharge fault detection report by describing the decision support path in natural language using a large language model, the method further includes:

[0125] Obtain the translated content of the fault detection report in the target language, and display the fault detection report based on the translated content.

[0126] The target language is another human language that the user expects the original fault detection report to be translated into, such as English, German, or Japanese. Its selection typically depends on the language preferences of the report's end recipient or the language requirements of their work environment. The translated content can be all or part of the textual information in the original fault detection report, presented in the target language as an equivalent semantic expression after translation.

[0127] Specifically, the terminal needs to determine the target language, which can be selected by the user through the interface or automatically determined based on preset regional configuration. After determining the target language, the terminal takes the full text of the generated Chinese fault detection report or key paragraphs as input, along with explicit translation instructions, such as "Please translate the following content into [target language name, such as English]: [Chinese report content]". The large language model will then output the corresponding translated text according to the instructions.

[0128] Optionally, to ensure the accuracy and consistency of translations of technical terms in the power equipment field and to avoid ambiguities that may arise from generic translation models, the terminal can further incorporate a pre-defined domain terminology database when executing any of the above translation methods. This database contains professional terms in fields such as power engineering and fault diagnosis, along with their standard translations in different target languages. During or after translation, the terminal can use this database to proofread and correct the translation results. For example, a term that might be translated as "aging" in a generic translation can be precisely corrected using the database to the corresponding professional expressions for insulation degradation or insulation aging in the target language.

[0129] After obtaining the translated content in the target language, the terminal displays the fault detection report based on this translation. For example, it displays the translated report text on the user interface, or it generates a report file in the target language for the user to download or view.

[0130] In this embodiment, by combining a domain terminology database, mistranslation or inconsistent translation of professional terms by general translation tools can be effectively avoided, ensuring the professionalism and rigor of the technical report. This helps to avoid operational errors caused by misunderstandings.

[0131] In one exemplary embodiment, such as Figure 2 As shown, this application provides a method for generating a detection report of stator partial discharge. First, a knowledge graph is constructed based on multi-source heterogeneous data (online monitoring data, maintenance cases, and standard documents) to build a knowledge graph in the field of generator stator partial discharge, supporting incremental knowledge updates. After obtaining the knowledge graph, a graph neural network and a rule engine are combined to realize fault mode recognition and causal chain reasoning. Based on the realized fault modes and causal chains, the knowledge graph paths and template constraints are integrated to generate a fault report containing data evidence and reasoning logic.

[0132] Specifically, such as Figure 3 Shown, including:

[0133] Step S301: Knowledge Graph Construction and Entity Feature Learning. The terminal first constructs a knowledge graph specifically for generator stator partial discharge. This graph systematically defines various entity types, including fault modes, equipment parameters, maintenance examples, and industry standards, and clarifies the logical relationships between them. To ensure that this symbolic knowledge can be effectively utilized by subsequent models, the terminal further employs graph embedding technology to learn a low-dimensional, dense vector representation for each entity and relation in the graph. This learning process iteratively optimizes the algebraic operations on entity vectors and relation vectors in the vector space to accurately reflect their true logical connections within the graph. The resulting stable entity vectors serve as the entity features used in subsequent steps.

[0134] S302, Real-time Partial Discharge Data Acquisition and Corresponding Discharge Feature Extraction. After the system enters online monitoring mode, the terminal acquires raw partial discharge monitoring data streams in real time from sensors (such as high-frequency current transformers or ultra-high-frequency sensors) installed on the generator stator. Then, the terminal uses a series of signal processing algorithms, such as wavelet transform, Fourier analysis, or statistical moment calculation, to extract a set of discharge feature vectors from these raw and complex signal waveforms. These features can quantify the key characteristics of the current partial discharge phenomenon, including pulse amplitude, number, phase distribution characteristics, energy, and waveform parameters.

[0135] S303, Risk Level Assessment and Quantification. After obtaining real-time discharge characteristics, the terminal inputs them into a pre-trained risk level determination model (e.g., a Bayesian network). This model internally stores the conditional probabilities of various discharge characteristics occurring under different risk levels (e.g., low, medium, high, emergency) and the prior probability of each risk level. By integrating these probabilistic dependencies with the currently observed discharge characteristics, the model performs probabilistic inference and outputs a preliminary discrete current risk level. For further refined assessment, the terminal also calculates a continuous risk quantification parameter: First, the standardized discharge characteristics and related equipment status parameters are weighted to obtain a risk score; then, a nonlinear function based on historical data is used to evaluate the degree of deviation and trend of this risk score relative to a preset benchmark risk score, thereby obtaining a more refined and continuous risk quantification value.

[0136] S304, Fault Mode Recognition Based on Knowledge Graph. This is the core diagnostic stage, where the real-time discharge features extracted by the terminal are compared with the entity features of each entity within the knowledge graph. Specifically, the terminal first maps the real-time discharge feature vector and the feature vectors of each entity to a unified feature vector space, and then combines the real-time discharge feature vector with each entity feature vector to form joint features. Next, the terminal processes these joint features using a set of adjustable preset parameters optimized from historical data, calculating the original correlation score corresponding to each joint feature. These scores are then normalized to obtain the final correlation degree of each entity with the current discharge feature relative to other entities. The terminal selects the entity with the highest correlation degree or that meets specific conditions, and ultimately determines one or more target fault modes based on the type of the matched entity (e.g., directly matching a fault mode entity, or matching a device parameter entity and then inferring through graph relationships).

[0137] S305, Construction of Decision Support Path. After identifying the target failure mode, the terminal uses this as a starting point to conduct in-depth mining and path construction within the knowledge graph. It traces the causes and solutions along predefined relationships (such as failure outcome, cause, solution, and reference standards) in the knowledge graph, performing reverse causal tracing to construct the failure causal chain and forward solution exploration to construct a failure repair strategy. These two parts together constitute a structured decision support path that comprehensively explains the possible causes, development logic, and recommended countermeasures or reference standards of the target failure mode.

[0138] S306, Generation and Display of Fault Detection Report. This process first includes a template filling step: based on discharge characteristics and decision support paths, the terminal extracts key information as keywords and fills them into the corresponding fields of a preset structured template to quickly generate a structured summary of the report. Subsequently, the terminal calls a large language model, using the obtained complete decision support path as core input, and through autoregressive text generation, describes its content in detail and coherently in natural language. The generated natural language description is integrated with the aforementioned structured information to form the final fault detection report. If necessary, the terminal can also translate this report into a user-specified target language for display, thus realizing a complete and intelligent fault detection and reporting process.

[0139] Optionally, to verify the effectiveness of this application, an exemplary test was conducted. In this test, the method of this application was applied to a set of actual generator stator partial discharge cases and compared with traditional diagnostic methods to evaluate the improvement in diagnostic accuracy. The improvement in diagnostic accuracy is calculated as shown in formula (7). Specifically, the number of cases correctly diagnosed by the new method ( The number of correctly diagnosed cases compared to the old method () The difference between () and the total number of test cases () Multiply by 100%. Through actual testing, the method of this invention, compared to traditional methods, shows an improvement in diagnostic accuracy ( The figure reached over 25%.

[0140]

[0141] This improved accuracy means that fault modes can be identified more reliably, reducing false positives and false negatives, and providing a more solid technical guarantee for subsequent equipment maintenance and operation decisions.

[0142] This embodiment also demonstrates significant advantages in report generation efficiency. The total time for report generation ( As shown in formula (8), it mainly consists of two parts: one is the fault matching and decision path generation time ( ), and secondly, the natural language report generation time ( In the test, the time for fault matching and decision path generation was controlled within 1 second. <1s), while the generation time of natural language reports is controlled within 5 seconds. <5s). Therefore, the entire fault detection report can be generated within seconds, which is in stark contrast to the traditional method that usually requires hours or even longer for manual analysis and report writing. This greatly improves the response speed of fault diagnosis and reporting, and saves valuable time for rapid decision-making and timely handling of faults.

[0143]

[0144] In this embodiment, no manual intervention is required from data collection to report generation, and the response time is shortened from several hours to within minutes. This enables timely detection and handling of faults, effectively preventing equipment damage and downtime caused by faults, and improving the reliability and stability of the equipment. At the same time, by associating fault mechanisms and maintenance suggestions through a knowledge graph, the report logic is clear and conforms to the cognitive habits of maintenance personnel.

[0145] 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 in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0146] Based on the same inventive concept, this application also provides a stator partial discharge detection report generation apparatus for implementing the stator partial discharge detection report generation method described above. The solution provided by this apparatus is similar to the implementation described in the above method; therefore, the specific limitations in one or more stator partial discharge detection report generation apparatus embodiments provided below can be found in the limitations of the stator partial discharge detection report generation method described above, and will not be repeated here.

[0147] In one exemplary embodiment, such as Figure 4 As shown, a device for generating a stator partial discharge detection report is provided, comprising: a partial discharge knowledge graph establishment module 410, a target fault mode determination module 420, a decision support path determination module 430, and a fault detection report generation module 440, wherein:

[0148] The partial discharge knowledge graph establishment module 410 is used to establish a knowledge graph of stator partial discharge; the knowledge graph includes multiple entities related to stator partial discharge faults and the relationships between the entities; the multiple entities include fault mode entities, equipment parameter entities, maintenance instance entities and industry standard entities.

[0149] The target fault mode determination module 420 is used to acquire the discharge features corresponding to real-time partial discharge data, compare the discharge features with the entity features of multiple entities in the knowledge graph, determine the entity that matches the discharge features based on the comparison results, and determine the target fault mode based on the entity that matches the discharge features.

[0150] The decision support path determination module 430 is used to determine a decision support path including the fault causal chain and the fault repair strategy based on the target fault mode and the correlation in the knowledge graph.

[0151] The fault detection report generation module 440 is used to describe the decision support path in natural language using a large language model to obtain a fault detection report of stator partial discharge.

[0152] In one embodiment, the target fault mode determination module 420 is further configured to convert the discharge feature and the entity features of each entity into a unified feature vector space, and combine the converted discharge feature with the converted entity features to obtain joint features; calculate the correlation score of each joint feature in the feature vector space; determine the correlation degree of each entity with the discharge feature relative to other entities based on the correlation score of each joint feature, and obtain entities whose correlation degree meets preset conditions; if the entity is the device parameter entity, determine the target fault mode matching the discharge feature based on the correlation relationship related to the device parameter entity; and if the entity is the fault mode entity, take the fault mode indicated by the fault mode entity as the target fault mode.

[0153] In one embodiment, the partial discharge knowledge graph establishment module 410 is further configured to obtain entity vectors of each entity and relation vectors of the association relationships between each entity; for each set of association relationships, the head entity and the tail entity associated with the association relationship, obtain the sum of the association relationship vector corresponding to the association relationship and the head entity vector corresponding to the head entity; determine the difference between the entity vector obtained by transforming the head entity vector through the association relationship and the tail entity vector based on the difference between the sum of the sums and the tail entity vector corresponding to the tail entity; determine a loss value based on multiple differences; iteratively adjust each entity vector and each relation vector based on the loss value until a preset iteration termination condition is met; and obtain each entity feature based on the current entity vectors.

[0154] In one embodiment, the target fault mode determination module 420 is further configured to input the discharge features into a pre-trained risk level determination model; the risk level determination model includes a probabilistic dependency relationship between each risk level and each discharge feature; the probabilistic dependency relationship indicates the conditional probability of the discharge feature under different risk levels, and the current occurrence probability of each risk level; the risk level determination model outputs the current risk level of the stator partial discharge based on the probabilistic dependency relationship and the discharge features.

[0155] In one embodiment, the target fault mode determination module 420 is further configured to, based on preset weights, weight the standardized discharge characteristics and the real-time partial discharge data corresponding to the standardized discharge characteristics to determine the risk score under the current risk level; determine the degree of deviation of the risk score from the preset benchmark risk score, and determine the risk quantification parameter corresponding to the current risk level based on the degree of deviation.

[0156] In one embodiment, the fault detection report generation module 440 is further configured to determine the keywords corresponding to the fields of the preset structured template based on the discharge characteristics and the decision support path, and generate a structured fault detection report based on the keywords; obtain the translated content of the fault detection report in the target language, and display the fault detection report based on the translated content.

[0157] Each module in the aforementioned stator partial discharge detection report generation device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0158] In one exemplary embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 5As shown, the computer device includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, Near Field Communication (NFC), or other technologies. When executed by the processor, the computer program implements a method for generating a stator partial discharge detection report.

[0159] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0160] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0161] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0162] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0163] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0164] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0165] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0166] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present application. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, and these modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present application shall be determined by the appended claims.

Claims

1. A method for generating a detection report of stator partial discharge, characterized in that, The method includes: A knowledge graph of stator partial discharge is established; the knowledge graph includes multiple entities related to stator partial discharge faults and the relationships between these entities; the multiple entities include fault mode entities, equipment parameter entities, maintenance instance entities, and industry standard entities; The process involves: acquiring discharge features corresponding to real-time partial discharge data; transforming the discharge features and entity features of each entity into a unified feature vector space; and combining the transformed discharge features with the transformed entity features to obtain joint features. The entity feature acquisition steps include: acquiring entity vectors of each entity and relation vectors of the relationships between entities; for each set of relationships, the head entity associated with the relationship, and the tail entity, acquiring the sum of the relationship vector corresponding to the relationship and the head entity vector corresponding to the head entity; determining the difference between the entity vector obtained by transforming the head entity vector through the relationship and the tail entity vector based on the difference between the sum and the tail entity vector; determining a loss value based on multiple differences; iteratively adjusting each entity vector and each relation vector based on the loss value until a preset iteration termination condition is met; and obtaining each entity feature based on the current entity vectors. In the feature vector space, the correlation score of each joint feature is calculated; Based on the correlation scores of each joint feature, the degree of correlation between each entity and the discharge feature relative to other entities is determined, and entities whose correlation degree meets preset conditions are obtained. If the entity is the device parameter entity, then based on the association relationships related to the device parameter entity, a target fault mode matching the discharge characteristics is determined; and, If the entity is the fault mode entity, then the fault mode indicated by the fault mode entity is taken as the target fault mode. Based on the target failure mode and the correlation in the knowledge graph, a decision support path including the failure causal chain and failure maintenance strategy is determined. The decision support path is described in natural language using a large language model to obtain a fault detection report of stator partial discharge.

2. The method according to claim 1, characterized in that, After acquiring real-time partial discharge data and extracting the discharge features of the real-time partial discharge data, the method further includes: The discharge features are input into a pre-trained risk level determination model; the risk level determination model includes the probabilistic dependency between each risk level and each discharge feature; the probabilistic dependency indicates the conditional probability of the discharge feature under different risk levels, and the current occurrence probability of each risk level; The risk level determination model outputs the current risk level of the stator partial discharge based on the probability dependency and the discharge characteristics.

3. The method according to claim 2, characterized in that, The current risk level corresponds to the determination of risk quantification parameters, including: Based on preset weights, the standardized discharge characteristics and the real-time partial discharge data corresponding to the standardized discharge characteristics are weighted to determine the risk score under the current risk level. Determine the degree of deviation of the risk score from the preset benchmark risk score, and determine the risk quantification parameter corresponding to the current risk level based on the degree of deviation.

4. The method according to any one of claims 1 to 3, characterized in that, Before obtaining the stator partial discharge fault detection report by describing the decision support path in natural language using a large language model, the method further includes: Based on the discharge characteristics and the decision support path, the keywords corresponding to the fields of the preset structured template are determined, and a structured fault detection report is generated based on the keywords. After obtaining the stator partial discharge fault detection report by describing the decision support path in natural language using a large language model, the method further includes: Obtain the translated content of the fault detection report in the target language, and display the fault detection report based on the translated content.

5. A device for generating a detection report of stator partial discharge, characterized in that, The device includes: A partial discharge knowledge graph building module is used to build a knowledge graph of stator partial discharge; the knowledge graph includes multiple entities related to stator partial discharge faults and the relationships between these entities; the multiple entities include fault mode entities, equipment parameter entities, maintenance instance entities, and industry standard entities. The target fault mode determination module is used to acquire discharge features corresponding to real-time partial discharge data, transform the discharge features and entity features of each entity into a unified feature vector space, and combine the transformed discharge features with the transformed entity features to obtain joint features. The entity feature acquisition steps include: acquiring the entity vectors of each entity and the relationship vectors of the association relationships between the entities; for each group of association relationships, the head entity and tail entity associated with the association relationship, acquiring the sum of the association relationship vector corresponding to the association relationship and the head entity vector corresponding to the head entity; and based on the difference between the sum and the tail entity vector corresponding to the tail entity, determining the entity vector obtained by transforming the head entity vector through the association relationship and the tail entity vector. The process involves identifying differences; determining loss values ​​based on these differences; iteratively adjusting entity vectors and relation vectors based on these loss values ​​until a preset iteration termination condition is met; obtaining entity features based on the current entity vectors; calculating the correlation score of each joint feature in the feature vector space; determining the correlation degree between each entity and the discharge feature relative to other entities based on the correlation score of each joint feature; and acquiring entities whose correlation degree meets a preset condition. If the entity is a device parameter entity, a target fault mode matching the discharge feature is determined based on the correlation relationship related to the device parameter entity. If the entity is a fault mode entity, the fault mode indicated by the fault mode entity is taken as the target fault mode. The decision support path determination module is used to determine a decision support path, including the fault causal chain and the fault repair strategy, based on the target fault mode and the correlation in the knowledge graph. The fault detection report generation module is used to describe the decision support path in natural language using a large language model to obtain a fault detection report of stator partial discharge.

6. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 3.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 3.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 3.

Images