A Fault Diagnosis Method and System for High-Voltage DC Transmission System Based on SER Event Sequence Link Prediction

Abstract

The present invention discloses a fault diagnosis method and system for a high-voltage direct current (HVDC) transmission system based on SER event sequence link prediction, which relates to the technical field of transmission safety of power systems. By introducing an event knowledge graph and integrating historical cases of multiple DC projects, an abnormal event knowledge graph of the HVDC transmission system with SER abnormal events as the core is constructed, breaking through the information islands between entities such as projects and faults and events, forming a global semantic network that can be traversed and reasoned, and realizing the deep coupling of business and knowledge. According to the experience of operation and maintenance experts, a sequence link prediction task is defined, a fault diagnosis data set for the HVDC transmission system is constructed, a pre-trained language model is introduced, and a fault diagnosis model and system for the HVDC transmission system based on SER event sequence link prediction are established. Hidden high-dimensional correlation features between projects, abnormal events and system faults are extracted to autonomously complete fault diagnosis, solving the problems of weak adaptability and low automation degree of existing methods.

Description

Technical Field

[0001] The present invention relates to the technical field of power system transmission safety, and more specifically, to a fault diagnosis method and system for a high-voltage direct current (HVDC) transmission system based on SER event sequence link prediction. Background Technique

[0002] Currently, with the rapid development and high-density access of renewable energy, new power systems are facing challenges such as uneven energy distribution and difficult fault analysis. As a core link of new power systems, high-voltage direct current (HVDC) transmission has advantages such as long transmission distance, large transmission capacity, and good economy.

[0003] The structure of HVDC transmission equipment is very complex and highly automated. When a fault occurs or there is an abnormal state, the system will generate a large amount of sequence of events recording (SER) data. SER data contains various status descriptions such as alarm groups, events, and levels, and is one of the important data for monitoring and evaluating the operating status of power systems, with characteristics such as high real-time, large data volume, and strong serialization. The converter station generates approximately 100,000 SER data per day, with potential risks such as missing important signals, misjudging the system status, and untimely maintenance operations. In 2022, when the maintenance personnel of a domestic 500 kV converter station viewed SER data as long as 19 pages, they failed to detect the event of the main transformer medium-voltage side switch tripping secretly in time, resulting in the continuous expansion of the accident scope. Therefore, by mining the fault characteristics hidden in a large amount of SER data, intelligent diagnosis of HVDC transmission system faults can be realized, ensuring the safe and reliable operation of the system.

[0004] Currently, the association rule mining method is mainly used to analyze SER data for semi-automated shallow feature extraction, mining strong association rules between SER abnormal events and system faults, and assisting maintenance personnel to manually review and judge the fault type according to established rules, such as the diagnosis of typical maintenance events in converter stations based on association rules, extraction of similar fault characteristics of converter station SER data, and SER data mining based on the FP-Growth algorithm. The above research aims to mine the association relationship between SER abnormal events and system faults, thereby assisting in evaluating whether the system is abnormal, but there are disadvantages such as severe dependence on manual work, poor real-time performance, and extremely low efficiency. The reasons are mainly reflected in the following aspects:

[0005] 1) Unable to extract hidden high-dimensional fault characteristics: Existing research only performs feature extraction and association rule analysis for abnormal signal types, unable to capture the rich semantic information and serialized structure characteristics of SER abnormal events, and the robustness and self-adaptability of the algorithm are weak;

[0006] 2) Lack of an extensible SER data analysis model: Due to the differences in line structures and signal naming methods among different DC projects, it is impossible to perform feature fusion on multiple projects and jointly model unified SER association rules.

[0007] 3) Unable to autonomously give diagnostic results: Existing research can only mine strong association rules between abnormal events and system failures, unable to achieve intelligent diagnosis of system failures. Operation and maintenance personnel still need to view SER data according to established rules, resulting in poor real-time performance and low efficiency.

[0008] Therefore, how to solve the above drawbacks is an urgent problem for those skilled in the art. Summary of the Invention

[0009] In view of this, the present invention provides a fault diagnosis method and system for a high-voltage DC transmission system based on SER event sequence link prediction. By introducing event knowledge graph (EKG) technology, fusing historical cases of multiple DC projects, constructing an abnormal event knowledge graph of the high-voltage DC transmission system with SER abnormal events as the core, breaking through the information islands between entities such as projects and faults and events, forming a globally semantic network that can roam and reason, and realizing the deep coupling of business and knowledge; according to SER data characteristics and operation and maintenance expert experience, defining a sequence link prediction (SLP) task, constructing a fault diagnosis data set for the high-voltage DC transmission system, introducing a pre-trained language model, establishing SER event sequence link prediction models SLP-PLM(Mul) and SLP-PLM(Bin) for fault diagnosis of the high-voltage DC transmission system, building a fault diagnosis system for the high-voltage DC transmission system based on SER event sequence link prediction, extracting hidden high-dimensional association features between projects, abnormal events and system failures, and autonomously completing fault diagnosis, solving problems such as weak adaptability and low automation of existing methods.

[0010] To achieve the above object, the present invention adopts the following technical solutions:

[0011] A fault diagnosis method for a high-voltage DC transmission system based on SER event sequence link prediction, comprising:

[0012] Introduce event knowledge graph technology into SER data analysis to construct an abnormal event knowledge graph of the high-voltage DC transmission system;

[0013] Based on SER data characteristics and fault diagnosis business logic, define a sequence link prediction task applicable to fault diagnosis of the high-voltage DC transmission system;

[0014] Based on the knowledge graph of abnormal events in the HVDC transmission system and its corresponding historical fault cases, according to the form of the sequence link prediction task, generate the evolution link sequence of DC projects and SER abnormal events under system faults. Taking fault cases as units, sample various types of fault data to obtain the fault diagnosis data set of the HVDC transmission system;

[0015] Establish an SER event sequence link prediction model for fault diagnosis of the HVDC transmission system;

[0016] Use the fault diagnosis data set of the HVDC transmission system to train the SER event sequence link prediction model to obtain a trained SER event sequence link prediction model;

[0017] Input the fault data of the HVDC transmission system to be tested into the SER event sequence link prediction model to obtain the fault diagnosis result of the HVDC transmission system.

[0018] Optionally, the process of constructing the knowledge graph of abnormal events in the HVDC transmission system is as follows:

[0019] Collect historical cases from multiple DC projects covering UHV, EHV, flexible and conventional HVDC, including SER files of each station and their corresponding fault types;

[0020] According to the SER data characteristics and the fault diagnosis business logic, design an ontology architecture of abnormal events knowledge graph for fault diagnosis of the HVDC transmission system. The ontology triple includes <project name, SOE alarm, abnormal event>, <abnormal event, followed by, abnormal event>, <abnormal event, diagnosis result, fault type>;

[0021] Use the TextRank algorithm to calculate the text similarity between abnormal events of abnormal levels in SER data and standard protection signals;

[0022] Check the abnormal event - protection signal text pairs with a similarity greater than 0.5 to obtain the abnormal event sequences of each case and their corresponding protection signal types;

[0023] According to the designed ontology architecture of the abnormal events knowledge graph, link and align the project, abnormal event sequence and fault type to obtain the knowledge graph of abnormal events in the HVDC transmission system.

[0024] Optionally, the calculation formula of the text similarity is:

[0025]

[0026] where t AE and t PS are the abnormal event text and the protection signal text respectively, and w is the word that appears in both text pairs.

[0027] Optionally, the application of the above definition to the sequence link prediction task for HVDC transmission system fault diagnosis is specifically as follows:

[0028] Define the knowledge graph as where \(E\) is the set of entities, is the set of relationships, is the set of triples, is the set of entity text descriptions;

[0029] The sequence link prediction task is defined as: Given an anchor entity sequence and a relationship predict the candidate entity where, is the set of anchor entity sequences, is the set of candidate entities; The sequence link prediction task is simply formalized as \((a_1,a_2,\ldots,a_{ k ,r,?)\), and is specifically divided into the sequence scoring function modeling \(\psi:\ and the candidate entity probability distribution modeling \(f:\ where,? represents the task objective of predicting the candidate entity \(c\), is the set of real numbers.

[0030] Optionally, the process of establishing the SER event sequence link prediction model for HVDC transmission system fault diagnosis specifically includes:

[0031] Introduce a pre-trained language model with language understanding ability, extract the semantic features of the SER event sequence text, and construct an SER event sequence link prediction model for HVDC transmission system fault diagnosis according to the SER event sequence link prediction modeling paradigm to realize the prediction of candidate entities for fault types.

[0032] Optionally, the SER event sequence link prediction modeling paradigm is divided into sequence scoring function modeling and candidate entity probability distribution modeling;

[0033] For the input anchor entity sequence and relationship The sequence scoring function modeling calculates the SER event sequence score \(s = \psi(a_1,a_2,\ldots,a_{ ,r,c k ,r,c * ) through the pre-trained language model for any candidate entity

[0034]

[0035] The sequence scoring function modeling paradigm transforms sequence link prediction into a binary classification problem, enabling the model to fully model the high-dimensional semantic relationships between entities and relationships and identify valid or invalid link sequences.

[0036] Modeling the probability distribution of candidate entities calculates the probability distribution of candidate entities under given conditions through a pre-trained language model:

[0037] s′ = P(c * a1,a2,…,a k ,r) = f(a1,a2,…,a k ,r);

[0038] Taking the candidate entity with the highest probability as the prediction result:

[0039]

[0040] The candidate entity probability distribution modeling paradigm views sequence link prediction as a multi-classification or sequence generation problem, and the model can directly capture the hidden correlation features among the anchor entity sequence, relationship, and candidate entity.

[0041] Optionally, under the sequence scoring function modeling paradigm, name the model SLP-PLM(Bin). For the link sequence (a1,a2,…,a k ,r,c), starting with the special token [CLS] and ending with the special token [SEP], tokenize all the node and relationship description texts in the sequence, and connect the upper and lower texts with the special token [SEP] to obtain the token sequence for fine-tuning the pre-trained language model; in the input sequence, each token is represented by the sum of the corresponding token embedding segment embedding and position embedding , that is, for token i, its feature representation is:

[0042]

[0043] Among them, for different node or relationship descriptions separated by the special token [SEP], odd elements share the same segment embedding e odd , while even elements share the same segment embedding e even ;

[0044] The pre-trained language model extracts the link sequence structure features and semantic features from the input sequence, and uses the hidden vector corresponding to the special token [CLS] in the last layer of the encoder or decoder as the aggregated feature for calculating the sequence score, where H is the hidden layer state dimension; the sequence scoring function is:

[0045] s = sigmoid(CW T );

[0046] Among them, is the binary classification weight matrix; during the fine-tuning process, the gradient descent method is used to optimize the pre-trained parameter weights and the weight matrix W, and the cross-entropy loss function is:

[0047]

[0048] Among them, is the set of positive examples of the link sequence, is the set of negative examples of the link sequence, y q ∈{0,1} is the negative / positive example label of the link sequence, s q1 , s q2 ∈[0,1] is the calculated score for negative and positive examples; since transforming the anchor entity a i in the positive example sequence will lead to the uncertainty of the candidate entity, a negative example sequence is obtained by randomly replacing the positive example candidate entity c:

[0049]

[0050] Optionally, in the candidate entity probability distribution modeling paradigm, the model is named SLP-PLM(Mul). Compared with the SLP-PLM(Bin) model, only the candidate entity c in the link sequence is deleted, that is, (a1, a2, …, a k , r) is used as the input sequence for fine-tuning the pre-trained language model, and the hidden vector C corresponding to the special token [CLS] of the last layer encoder or decoder is still used as the aggregated feature representation of the input sequence to predict the candidate entity under the current event sequence; the candidate entity probability distribution is:

[0051] s′ = softmax(CW′ T );

[0052] Among them, is the multi-classification weight matrix;

[0053] During the fine-tuning process, the cross-entropy loss function is:

[0054]

[0055] Among them, s′ qi ∈[0,1], and y′ qi is the indicator function of the candidate entity of the link sequence. When the predicted candidate entity is the true value, y′ qi = 1, otherwise, y′ qi = 0.

[0056] Optionally, inputting the fault data of the high-voltage DC transmission system to be measured into the SER event sequence link prediction model to obtain the fault diagnosis result of the high-voltage DC transmission system specifically includes:

[0057] Before the fault diagnosis of the high-voltage DC transmission system, first confirm the name of the high-voltage DC project for which the fault diagnosis is to be performed;

[0058] After confirming the high-voltage DC project, extract the abnormal events with non-normal levels and containing standard protection signals in the SER data;

[0059] After confirming the high-voltage DC project and extracting the SER abnormal events, construct a fault diagnosis link sequence and perform text tokenization on the link sequence;

[0060] After performing text tokenization on the link sequence, send the token sequence into the SER event sequence link prediction model to obtain the fault diagnosis result of the high-voltage DC transmission system.

[0061] A fault diagnosis system for a high-voltage DC transmission system based on SER event sequence link prediction includes:

[0062] A knowledge graph construction module, which introduces the event knowledge graph technology into the SER data analysis to construct an abnormal event knowledge graph of the high-voltage DC transmission system;

[0063] A task definition module, which defines a sequence link prediction task applicable to the fault diagnosis of the high-voltage DC transmission system based on the SER data characteristics and the fault diagnosis business logic;

[0064] A dataset construction module, which generates an evolution link sequence of the DC project and the SER abnormal events under system faults based on the abnormal event knowledge graph of the high-voltage DC transmission system and its corresponding historical fault instances, and samples various types of fault data in units of fault cases to obtain a fault diagnosis dataset for the high-voltage DC transmission system;

[0065] A model establishment module, which establishes an SER event sequence link prediction model for the fault diagnosis of the high-voltage DC transmission system;

[0066] A model training module, which uses the fault diagnosis dataset of the high-voltage DC transmission system to train the SER event sequence link prediction model to obtain a trained SER event sequence link prediction model;

[0067] A fault diagnosis module, which inputs the fault data of the high-voltage DC transmission system to be measured into the SER event sequence link prediction model to obtain the fault diagnosis result of the high-voltage DC transmission system.

[0068] As can be seen from the above technical solutions, compared with the prior art, the present invention discloses a fault diagnosis method and system for a high-voltage direct current (HVDC) transmission system based on SER event sequence link prediction, which has the following beneficial effects:

[0069] 1) Aiming at problems such as poor scalability of existing SER data analysis methods, the event knowledge graph technology is introduced, historical cases of multiple DC projects are integrated, and an abnormal event knowledge graph of the HVDC transmission system with SER abnormal events as the core is constructed, breaking through the information islands between entities such as projects and faults and events, forming a global semantic network that can be traversed and reasoned, with rich semantic information and structural features, significantly improving the representation ability and scalability of SER data, and realizing the deep coupling of business and knowledge.

[0070] 2) Aiming at problems such as weak adaptability and low automation degree of existing SER data analysis methods, according to the SER data characteristics and fault diagnosis business logic, the sequence link prediction (SLP) task is defined, a fault diagnosis data set for the HVDC transmission system is constructed, a pre-trained language model is introduced, and an SER event sequence link prediction model for fault diagnosis of the HVDC transmission system is established. By jointly modeling the high-dimensional fault features of multiple DC projects, the model can adaptively capture the fine-grained semantic information of projects, events and faults in the fault diagnosis data set and the deep correlation features hidden between them, and autonomously and accurately realize the fault diagnosis of the HVDC transmission system.

[0071] 3) A fault diagnosis system for the HVDC transmission system based on SER event sequence link prediction is built, improving the efficiency, accuracy of SER abnormal event analysis and the intelligent level of system operation and maintenance services, reducing the review burden of operation and maintenance personnel, and reducing the potential risks threatening the safe and stable operation of the power grid. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained according to the provided drawings without creative efforts.

[0073] Figure 1 It is a schematic diagram of the method flow provided by the present invention;

[0074] Figure 2 It is an example diagram of the abnormal event knowledge graph of the HVDC transmission system provided by the present invention;

[0075] Figure 3 It is an example diagram of the sequence link prediction task provided by the present invention;

[0076] Figure 4 Schematic diagram of the SER event sequence link prediction modeling method provided by the present invention;

[0077] Figure 5 Schematic diagram of the SLP-PLM(Bin) framework structure under the sequence scoring function modeling provided by the present invention;

[0078] Figure 6 Schematic diagram of the SLP-PLM(Mul) framework structure under the candidate entity probability distribution modeling provided by the present invention. Detailed implementation manners

[0079] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0080] An embodiment of the present invention discloses a fault diagnosis method for a high-voltage direct current transmission system based on SER event sequence link prediction, as Figure 1 shown, including:

[0081] Introduce the event knowledge graph technology into the SER data analysis to construct an abnormal event knowledge graph of the high-voltage direct current transmission system;

[0082] Define a sequence link prediction task applicable to the fault diagnosis of the high-voltage direct current transmission system based on the SER data characteristics and the fault diagnosis business logic;

[0083] Based on the abnormal event knowledge graph of the high-voltage direct current transmission system and its corresponding historical fault instances, generate an evolution link sequence of the DC project and the SER abnormal event under system faults according to the form of the sequence link prediction task. Sampling various fault data in units of fault cases to obtain a fault diagnosis data set for the high-voltage direct current transmission system;

[0084] Establish an SER event sequence link prediction model for the fault diagnosis of the high-voltage direct current transmission system;

[0085] Use the fault diagnosis data set of the high-voltage direct current transmission system to train the SER event sequence link prediction model to obtain a trained SER event sequence link prediction model;

[0086] Input the fault data of the high-voltage direct current transmission system to be measured into the SER event sequence link prediction model to obtain the fault diagnosis result of the high-voltage direct current transmission system.

[0087] In a specific embodiment, the process of constructing the abnormal event knowledge graph of the high-voltage direct current transmission system is as follows:

[0088] In a high-voltage direct current (HVDC) transmission system, SER records in real time system operation status information such as parameter violations, alarm events, and protection actions, and serially outputs six types of data: time, host, system, level, alarm group, and event. Among them, the "host" is the device server that records the event, the "system" represents the redundant system used, the "level" represents the operation status level of the current event, which is generally divided into "normal", "minor", "alarm", and "emergency", the "alarm group" is the device group or software that generates the event, and the "event" is the description of the current device operation status.

[0089] SER data has natural real-time and serialization characteristics and is often used to mine the evolution process of abnormal events under different DC projects and their mapping relationships with system failures. Taking a 500 kV converter station in China as an example, tens of thousands of SER data are generated daily on average, and the maximum can reach 500,000. Operation and maintenance personnel need to continuously monitor SER real-time information, judge the current system operation status, and timely discover the current fault type according to the order of protection action abnormal events generated by the alarm group, and take emergency disposal measures to avoid the continuous expansion of the accident scope.

[0090] Based on the SER data characteristics and the fault diagnosis business logic, the event knowledge graph technology is introduced. Taking "project name", "abnormal event", and "fault type" as node types, and "SOE alarm", "followed by", and "diagnosis result" as relationships, connect "project name" and "abnormal event", "abnormal event" and "abnormal event", "abnormal event" and "fault type" respectively to construct an abnormal event knowledge graph for the HVDC transmission system, as Figure 2 (Case) shown, where type I nodes represent the entity type of "project name", type II nodes represent the event type of "abnormal event", type III nodes represent the entity type of "fault type", and the "abnormal event" node is represented by the concatenation of "alarm group" and "event" in SER data. Compared with the traditional association rule representation method, the abnormal event knowledge graph breaks through the information islands among projects, events, and faults, forms a globally traversable and inferable network, has rich semantic information and structural characteristics, and significantly improves the representation ability and scalability of SER data. The detailed process is as follows:

[0091] Collect a large number of historical cases from multiple DC projects covering UHV, EHV, flexible, and conventional DC, including SER files of each station and their corresponding fault types;

[0092] According to the SER data characteristics and the fault diagnosis business logic, design an ontology architecture of an abnormal event knowledge graph for HVDC transmission system fault diagnosis. The ontology triple includes <project name, SOE alarm, abnormal event>, <abnormal event, followed by, abnormal event>, <abnormal event, diagnosis result, fault type>;

[0093] Using the TextRank algorithm, calculate the text similarity between abnormal events of non-normal levels in SER data and standard protection signals:

[0094]

[0095] Among them, t AE and t PS are the abnormal event text and the protection signal text respectively, and w is the word that appears in both text pairs.

[0096] Check the abnormal event - protection signal text pairs with a similarity greater than 0.5 to obtain the abnormal event sequences of each case and their corresponding protection signal types;

[0097] According to the designed ontology architecture of the abnormal event knowledge graph, link and align the project, abnormal event sequence, and fault type to obtain the abnormal event knowledge graph of the high - voltage direct - current transmission system.

[0098] In a specific embodiment, the sequence link prediction task applicable to the fault diagnosis of the high - voltage direct - current transmission system is specifically defined as:

[0099] The fault diagnosis of the high - voltage direct - current transmission system is to predict the possible system faults under the condition of a given DC project and the evolution process of SER abnormal events, that is, given the node link of "project name" and "abnormal event", predict the "fault type" node that the "diagnosis result" may point to, which is different from both the traditional link prediction (LP) task and the path query answering (PQA) task. Here, as Figure 3 shown, re - define the sequence link prediction (SLP) task applicable to the fault diagnosis of the high - voltage direct - current transmission system.

[0100] Define the knowledge graph as Among them, E is the entity set, is the relationship set, is the triple set, is the entity text description set;

[0101] The sequence link prediction (SLP) task is defined as: Given an anchor entity sequence and a relationship condition, predict the candidate entity Among them, is the anchor entity sequence set, is the candidate entity set; The sequence link prediction task is simply formalized as (a1, a2,..., a k, r,?), which is specifically divided into sequence scoring function modeling ψ: and candidate entity probability distribution modeling f: where? represents the task objective of predicting candidate entity c, which is a set of real numbers.

[0102] Compared with the traditional link prediction (LP) task, the input of the sequence link prediction task has more entities and more detailed pre-information; compared with the path query answering (PQA) task, the sequence link prediction task uses a more accurate entity sequence as the query path. Therefore, the sequence link prediction task integrates the task characteristics of link prediction and path query answering, and is a brand-new knowledge reasoning task.

[0103] In a specific embodiment, based on the knowledge graph of abnormal events in the HVDC transmission system and its corresponding historical fault instances, according to the form of the sequence link prediction task, an evolution link sequence of the DC project and SER abnormal events under system faults is generated to obtain a fault case link sequence dataset; taking cases as units, various types of fault data are sampled according to the ratio of the training set, validation set, and test set of 3:1:1 to obtain an HVDC transmission system fault diagnosis dataset.

[0104] In a specific embodiment, in order to make full use of the semantic features of the SER event sequence text, a pre-trained language model (PLM) is introduced, and candidate entity prediction is realized by fine-tuning the pre-trained model. Here, the pre-trained language model includes but is not limited to encoder-only models (such as BERT, RoBERTa, etc.), autoregressive (decoder-only) models (such as Qwen, Llama, etc.), and encoder-decoder models (such as GLM, BART, etc.). Based on the pre-trained language model with language understanding ability, an SER event sequence link prediction model SLP-PLM for HVDC transmission system fault diagnosis is established. The SER event sequence link prediction modeling paradigm is divided into sequence scoring function modeling and candidate entity probability distribution modeling, as Figure 4 shown.

[0105] For the input anchor entity sequence and relationship The sequence scoring function modeling calculates the SER event sequence score s = ψ(a1, a2,..., a , r, c k ) through the pre-trained language model for any candidate entity * , and selects the candidate entity with the highest score as the prediction result:

[0106]

[0107] The sequence scoring function modeling paradigm transforms sequence link prediction into a binary classification problem, enabling the model to fully model the high-dimensional semantic relationship between entities and relationships and identify valid or invalid link sequences.

[0108] Modeling the candidate entity probability distribution calculates the candidate entity probability distribution under given conditions through a pre-trained language model:

[0109] s′ = P(c * |a1,a2,…,a k ,r) = f(a1,a2,…,a k ,r);

[0110] The candidate entity with the highest probability is taken as the prediction result:

[0111]

[0112] The candidate entity probability distribution modeling paradigm regards sequence link prediction as a multi-classification or sequence generation problem. The model can directly capture the hidden correlation features among the anchor entity sequence, relationship, and candidate entities, avoiding traversing and scoring all candidate entities.

[0113] Under the sequence scoring function modeling paradigm, the model is named SLP-PLM(Bin), and the model framework is as Figure 5 shown. For the link sequence (a1,a2,…,a k ,r,c), starting with the special token [CLS] and ending with the special token [SEP], all nodes and relationship description texts in the sequence are tokenized, and the upper and lower texts are connected with the special token [SEP] to obtain the token sequence for fine-tuning the pre-trained language model; in the input sequence, each token is jointly represented by the sum of the corresponding token embedding segmentation embedding and position embedding , that is, for token i, its feature representation is:

[0114]

[0115] Among them, for different node or relationship descriptions separated by the special token [SEP], odd elements share the same segmentation embedding e odd , while even elements share the same segmentation embedding e even ;

[0116] The pre-trained language model extracts the link sequence structure features and semantic features from the input sequence, and uses the hidden vector corresponding to the special token [CLS] in the last layer of the encoder or decoder as the aggregation feature for calculating the sequence score, where H is the hidden state dimension;

[0117] The sequence scoring function is:

[0118] s = sigmoid(CW T );

[0119] Where, is the binary classification weight matrix; during the fine-tuning process, the pre-trained parameter weights and the weight matrix W are optimized using the gradient descent method;

[0120] The cross-entropy loss function is:

[0121]

[0122] Where, is the set of positive examples of the link sequence, is the set of negative examples of the link sequence, y q ∈ {0, 1} is the negative example / positive example label of the link sequence, s q1 , s q2 ∈ [0, 1] are the calculated scores for negative and positive examples;

[0123] Since transforming the anchor entity a i in the positive example sequence will lead to the uncertainty of the candidate entity, a negative example sequence is obtained by randomly replacing the positive example candidate entity c:

[0124]

[0125] Under the candidate entity probability distribution modeling paradigm, the model is named SLP-PLM(Mul), and the model framework is as Figure 6 shown. Compared with the SLP-PLM(Bin) model, only the candidate entity c in the link sequence is deleted, that is, (a1, a2,..., a k , r) is used as the input sequence for fine-tuning the pre-trained language model, and the hidden vector C corresponding to the special token [CLS] in the last layer of the encoder or decoder is still used as the aggregated feature representation of the input sequence to predict the candidate entity under the current event sequence;

[0126] The candidate entity probability distribution is:

[0127] s' = softmax(CW' T );

[0128] Where, is the multi-classification weight matrix;

[0129] During the fine-tuning process, the cross-entropy loss function is:

[0130]

[0131] Where, s'qi ∈[0,1], and y′ qi is an indicator function for candidate entities in the link sequence. When the predicted candidate entity is the true value, y′ qi = 1; otherwise, y′ qi = 0.

[0132] In a specific embodiment, inputting the fault data of the high-voltage DC power transmission system to be measured into the SER event sequence link prediction model to obtain the fault diagnosis result of the high-voltage DC power transmission system specifically includes:

[0133] High-voltage DC project confirmation

[0134] Before the fault diagnosis of the high-voltage DC power transmission system, first confirm the name of the high-voltage DC project for which the fault diagnosis is to be performed. The system provides an entry for high-voltage DC project confirmation for the user to select or enter the name of the high-voltage DC project. Here, the high-voltage DC project specified by the user should belong to the scope of high-voltage DC projects covered by the SER event sequence link prediction model. If not, the user needs to be prompted.

[0135] SER abnormal event extraction

[0136] After confirming the high-voltage DC project, extract the abnormal events with non-normal levels and containing standard protection signals from the SER data. Here, the standard protection signal should belong to the specified high-voltage DC project, and the abnormal event is represented by the concatenation of the "alarm group" and "event" in the SER data. The system provides an entry for uploading SER data or a network interface for the user to select offline or online data. The methods for determining whether an abnormal event contains a standard protection signal include, but are not limited to, word segmentation matching, fuzzy matching, semantic matching and other discrimination methods.

[0137] Link sequence text word segmentation

[0138] After confirming the high-voltage DC project and extracting the SER abnormal events, construct a fault diagnosis link sequence and perform text word segmentation on the link sequence. If the constructed SER event sequence link prediction model belongs to the sequence scoring function modeling paradigm, the link sequence should include the project name, abnormal event and fault category, as Figure 5 shown; if the constructed SER event sequence link prediction model belongs to the candidate entity probability distribution modeling paradigm, the link sequence should be composed of the project name and abnormal event links, as Figure 6 shown. The text word segmentation method should be applicable to the SER event sequence link prediction model, including but not limited to sub-word level word segmentation, character level word segmentation, hybrid word segmentation and other methods.

[0139] Fault diagnosis of high-voltage DC power transmission system based on SER event sequence link prediction

[0140] After performing word segmentation on the linked sequence text, the token sequence is fed into the SER event sequence link prediction model to obtain the fault diagnosis result of the HVDC transmission system. To facilitate users' analysis and judgment, the system should display the N types of fault types with the highest confidence levels in the diagnosis result and their confidence levels, and provide an entry for selecting the number of displayed fault types.

[0141] A fault diagnosis system for HVDC transmission systems based on SER event sequence link prediction, including:

[0142] A knowledge graph construction module that introduces event knowledge graph technology into SER data analysis to construct an abnormal event knowledge graph for the HVDC transmission system;

[0143] A task definition module that defines a sequence link prediction task suitable for fault diagnosis of HVDC transmission systems based on SER data characteristics and fault diagnosis business logic;

[0144] A dataset construction module that generates an evolution link sequence of the DC project and SER abnormal events under system faults based on the abnormal event knowledge graph of the HVDC transmission system and its corresponding historical fault instances, samples various fault data in units of fault cases, and obtains a fault diagnosis dataset for the HVDC transmission system;

[0145] A model establishment module that establishes an SER event sequence link prediction model for fault diagnosis of HVDC transmission systems;

[0146] A model training module that trains the SER event sequence link prediction model using the fault diagnosis dataset of the HVDC transmission system to obtain a trained SER event sequence link prediction model;

[0147] A fault diagnosis module that inputs the fault data of the HVDC transmission system to be measured into the SER event sequence link prediction model to obtain the fault diagnosis result of the HVDC transmission system.

[0148] In this specification, each embodiment is described in a progressive manner. The key point of each embodiment is to illustrate the differences from other embodiments. The same or similar parts among the embodiments can be referred to each other. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method part.

[0149] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the broadest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for fault diagnosis of a high voltage direct current transmission system based on SER event sequence link prediction, characterized in that: include: Introduce event knowledge graph technology into SER data analysis to build a knowledge graph of abnormal events in the HVDC transmission system; Based on SER data features and fault diagnosis business logic, a sequence link prediction task suitable for HVDC system fault diagnosis is defined; Based on the knowledge graph of abnormal events in the HVDC transmission system and its corresponding historical fault instances, the evolution link sequence of DC engineering and SER abnormal events under system faults is generated according to the form of sequence link prediction task. Various types of fault data are sampled based on fault cases to obtain the fault diagnosis data set of the HVDC transmission system. Establish a SER event sequence link prediction model for HVDC system fault diagnosis; The SER event sequence link prediction model is trained using the HVDC system fault diagnosis dataset to obtain a trained SER event sequence link prediction model. Input the fault data of the HVDC transmission system to be tested into the SER event sequence link prediction model to obtain the fault diagnosis result of the HVDC transmission system; The process of establishing the SER event sequence link prediction model for high voltage direct current transmission system fault diagnosis specifically includes: A pre-trained language model with language understanding ability is introduced to extract the semantic features of SER event sequence text. Based on the SER event sequence link prediction modeling paradigm, a SER event sequence link prediction model for high-voltage direct current transmission system fault diagnosis is constructed to achieve fault type candidate entity prediction. The SER event sequence link prediction modeling paradigm is divided into sequence scoring function modeling and candidate entity probability distribution modeling.

2. A method for fault diagnosis of a high voltage direct current transmission system based on SER event sequence link prediction according to claim 1, characterized in that: The process of constructing the knowledge graph of abnormal events in the HVDC transmission system is as follows: Collect historical cases from multiple DC projects covering UHV, EHV, flexible and conventional DC, including SER files of each station and their corresponding fault types; According to the SER data characteristics and fault diagnosis business logic, the abnormal event knowledge graph ontology architecture for HVDC transmission system fault diagnosis is designed. The ontology triples include <project name, SOE alarm, abnormal event>, <abnormal event, immediately after, abnormal event>, <abnormal event, diagnosis result, fault type>; Using the TextRank algorithm, the text similarity between abnormal events of non-normal levels and standard protection signals in SER data is calculated; Check the abnormal event-protection signal text pairs with similarity greater than 0.5 to obtain the abnormal event sequence of each case and its corresponding protection signal type; According to the designed abnormal event knowledge graph ontology architecture, the engineering, abnormal event sequences and fault types are linked and aligned to obtain the abnormal event knowledge graph of the high-voltage direct current transmission system.

3. A method for fault diagnosis of a high voltage direct current transmission system based on SER event sequence link prediction according to claim 2, characterized in that: The calculation formula of the text similarity is: Among them, t AE ,t PS are the abnormal event text and the protection signal text respectively, and w is the word that appears in the text pair at the same time.

4. The method for fault diagnosis of a high voltage direct current transmission system based on SER event sequence link prediction according to claim 1, characterized in that: The definition is applicable to the sequential link prediction task of HVDC system fault diagnosis as follows: Define the knowledge graph as Among them, ε is the entity set, is a set of relations, is a set of triples, A collection of entity text descriptions; The sequence link prediction task is defined as: given an anchor entity sequence and relationship Under the condition, predict the candidate entity in, is a set of anchor entity sequences, is a set of candidate entities; the sequence link prediction task is simply formalized as (a1, a2, …, a k ,r,? ), specifically divided into sequence scoring function modeling and candidate entity probability distribution modeling Among them, ? represents the task goal of predicting candidate entity c, is a set of real numbers.

5. The method for fault diagnosis of a high voltage direct current transmission system based on SER event sequence link prediction according to claim 1, characterized in that: For the input anchor entity sequence and relationship Sequence scoring function modeling calculates any candidate entity through a pre-trained language model The SER event sequence score s = ψ(a1, a2, ..., a k ,r,c * ), select the candidate entity with the highest score as the prediction result: The sequence scoring function modeling paradigm transforms sequence link prediction into a binary classification problem, enabling the model to fully model the high-dimensional semantic relationship between entities and relations and identify valid or invalid link sequences; Candidate entity probability distribution modeling calculates the probability distribution of candidate entities under given conditions through the pre-trained language model: s′=P(c * |a1,a2,…,a k ,r)=f(a1,a2,…,a k ,r); Take the candidate entity with the highest probability as the prediction result: The candidate entity probability distribution modeling paradigm regards sequence link prediction as a multi-classification or sequence generation problem. The model can directly capture the hidden association features between anchor entity sequences, relations and candidate entities.

6. A method for fault diagnosis of a high voltage direct current transmission system based on SER event sequence link prediction according to claim 5, characterized in that: Under the sequential scoring function modeling paradigm, the model is named SLP-PLM(Bin). For the link sequence (a1, a2, …, a k ,r,c), with the special token [CLS] as the head and the special token [SEP] as the tail, all the node and relationship description texts in the sequence are segmented, and the context text is connected with the special token [SEP] to obtain the token sequence for fine-tuning the pre-trained language model; In the input sequence, each token is embedded by the corresponding token Segmentation Embedding and position embedding The sum of is used to represent, that is, for token i, Its characteristics are expressed as: Among them, for different node or relationship descriptions segmented by special token [SEP], odd elements share the same segmentation embedding e odd , while even elements share the same segmentation embedding e even ; The pre-trained language model extracts the link sequence structure features and semantic features from the input sequence, and converts the hidden vector of the last layer encoder or decoder corresponding to the special token [CLS] As the aggregate feature used to calculate the sequence score, H is the hidden state dimension; the sequence scoring function is: s=sigmoid(CW T ); in, is the binary classification weight matrix; in the fine-tuning process, the gradient descent method is used to optimize the pre-training parameter weights and weight matrix W, and the cross entropy loss function is: in, is the set of positive examples of linked sequences, is the set of negative examples of the linked sequence, y q ∈{0,1} is the negative / positive label of the link sequence, s q1 ,s q2 ∈[0,1] is the score for negative and positive examples; due to the transformation of anchor entity a in the positive example sequence i This will lead to uncertainty in the candidate entities, and the negative example sequence is obtained by randomly replacing the positive candidate entity c:

7. A method for fault diagnosis of a high voltage direct current transmission system based on SER event sequence link prediction according to claim 6, characterized in that: Under the candidate entity probability distribution modeling paradigm, the model is named SLP-PLM(Mul). Compared with the SLP-PLM(Bin) model, only the candidate entity c in the link sequence is deleted, that is, (a1, a2, …, a k ,r) is used as the input sequence for fine-tuning the pre-trained language model, and the hidden vector C corresponding to the special token [CLS] of the last layer encoder or decoder is still used as the aggregated feature representation of the input sequence to predict the candidate entities under the current event sequence; the probability distribution of the candidate entities is: s′=softmax(CW′ T ); in, is the multi-classification weight matrix; During fine-tuning, the cross entropy loss function is: Among them, s′ qi ∈[0,1], and y′ qi is the indicator function of the link sequence candidate entity, when the predicted candidate entity is true value, y′ qi =1, otherwise, y′ qi =0.

8. The method for fault diagnosis of a high voltage direct current transmission system based on SER event sequence link prediction according to claim 1, characterized in that: The step of inputting the fault data of the high-voltage direct current transmission system to be tested into the SER event sequence link prediction model to obtain the fault diagnosis result of the high-voltage direct current transmission system specifically includes: Before diagnosing a high-voltage DC transmission system fault, first confirm the name of the high-voltage DC project to be diagnosed; After confirming the HVDC project, extract abnormal events of non-normal level and containing standard protection signals from SER data; After confirming the HVDC project and extracting SER abnormal events, a fault diagnosis link sequence is constructed and text segmentation is performed on the link sequence; After performing link sequence text segmentation, the token sequence is fed into the SER event sequence link prediction model to obtain the fault diagnosis results of the HVDC transmission system.

9. A high voltage direct current transmission system fault diagnosis system based on SER event sequence link prediction, characterized in that: A method for fault diagnosis of a high voltage direct current transmission system based on SER event sequence link prediction as claimed in any one of claims 1 to 8, comprising: The knowledge graph construction module introduces event knowledge graph technology into SER data analysis to build a knowledge graph of abnormal events in the HVDC transmission system; The task definition module defines the sequence link prediction task applicable to the fault diagnosis of the HVDC transmission system based on the SER data characteristics and fault diagnosis business logic; The dataset construction module generates a link sequence of the evolution of DC engineering and SER abnormal events under system faults based on the knowledge graph of abnormal events in the HVDC transmission system and its corresponding historical fault instances according to the form of sequence link prediction tasks. It samples various types of fault data based on fault cases to obtain a fault diagnosis dataset for the HVDC transmission system. Model building module, which builds the SER event sequence link prediction model for HVDC system fault diagnosis; The model training module uses the high-voltage direct current transmission system fault diagnosis data set to train the SER event sequence link prediction model to obtain a trained SER event sequence link prediction model; The fault diagnosis module inputs the fault data of the HVDC transmission system to be tested into the SER event sequence link prediction model to obtain the fault diagnosis result of the HVDC transmission system.

Images