Transformer substation main equipment fault diagnosis method and system, terminal and medium

Abstract

A transformer substation main equipment fault diagnosis method is characterized in that the method comprises the following steps: collecting transformer substation main equipment multi-modal total data, and obtaining a transformer substation main equipment knowledge graph through word segmentation, entity extraction, relation extraction and knowledge representation processes; extracting a multi-source semaphore related to the current transformer substation main equipment from the multi-modal full-quantity data of the transformer substation main equipment, obtaining a frequent item set in the multi-source semaphore by adopting an association rule mining algorithm, and constructing sensing data information according to the frequent item set; and a graph attention network is utilized to encode the triple in the transformer substation main equipment knowledge graph, a multi-hop reasoning mode is adopted to identify perception data information, and a transformer substation main equipment fault diagnosis result is obtained according to an output result.

Description

Technical Field

[0001] The present invention relates to the field of power systems, and more specifically, to a method, system, terminal and medium for diagnosing faults of main equipment in a transformer substation. Background Art

[0002] With the digitalization and intelligent development of power grids, the amount of operating data generated by substations continues to increase. These data come from a wide range of sources and in various formats. Traditional manual inspection and fault diagnosis modes have problems such as high cost and uncentralized data management. The digital twin substations currently under construction can provide auxiliary support to operation and maintenance managers in terms of operating status monitoring. However, there are certain limitations in the use and integration of data. Substation data management through information technology has become an important means to optimize digital twin substations.

[0003] In recent years, artificial intelligence and big data technologies have been rapidly iterating, and technologies such as natural language processing, knowledge graph, machine learning and data mining have become increasingly mature. Natural language processing technology can automatically parse texts such as substation fault records and maintenance logs, extract key information, and provide a data basis for the construction of an intelligent diagnosis expert database; in substation fault diagnosis, the intelligent diagnosis expert database can integrate multi-source heterogeneous data, build associations between entities such as equipment, faults, symptoms, and maintenance measures, and provide comprehensive knowledge support for fault diagnosis; through the analysis and mining of large amounts of data, machine learning algorithms can discover potential laws and patterns in the data, and provide prediction and decision support for fault diagnosis; data mining technology can extract useful information from massive data, and provide data support for the update and improvement of knowledge graphs.

[0004] The background technology of constructing knowledge graphs in substation fault diagnosis mainly stems from the power system's demand for efficient and accurate fault identification and processing, as well as the rapid development of technologies such as artificial intelligence and big data. Therefore, knowledge graph technology, with its unique advantages, is integrated with digital twin substations to achieve effective use and centralized management of data, which will play an increasingly important role in improving diagnostic accuracy, improving diagnostic efficiency, and supporting decision optimization.

[0005] Therefore, there is an urgent need for a method, system, terminal and medium for diagnosing faults of main equipment in a substation. Summary of the invention

[0006] In order to solve the shortcomings existing in the prior art, the present invention provides a substation main equipment fault diagnosis method, system, terminal and medium. The method adopts natural language processing, knowledge graph and deep learning technologies to mine the fault data characteristics of the substation main equipment, construct a knowledge graph and integrate it with the equipment twin model.

[0007] The present invention adopts the following technical solution.

[0008] In the first aspect of the present invention, there is provided a method for fault diagnosis of main substation equipment. The method includes the following steps: collecting multi-modal full-scale data of the main substation equipment, and obtaining a knowledge graph of the main substation equipment through processes such as word segmentation, entity extraction, relationship extraction, and knowledge representation; extracting multi-source signal quantities related to the current main substation equipment from the multi-modal full-scale data of the main substation equipment, using an association rule mining algorithm to obtain frequent item sets in the multi-source signal quantities, and constructing perception data information based on the frequent item sets; using a graph attention network to encode triples in the knowledge graph of the main substation equipment, identifying the perception data information by means of multi-hop reasoning, and obtaining a fault diagnosis result of the main substation equipment according to the output result.

[0009] Preferably, collecting multi-modal full-scale data of the main substation equipment and obtaining a knowledge graph of the main substation equipment through processes such as word segmentation, entity extraction, relationship extraction, and knowledge representation includes: defining key entity types and relationship types according to the requirements of substation fault diagnosis; the key entity types include equipment, fault types, symptoms, and maintenance measures; the relationship types include equipment - fault type, fault type - symptom, and symptom - maintenance measure.

[0010] Preferably, extracting multi-source signal quantities related to the current main substation equipment from the multi-modal full-scale data of the main substation equipment includes: collecting local operation information of the main substation equipment, and collecting perception layer operation information of the main substation equipment through perception layer devices; the local operation information and perception layer operation information include production data, status data, audio, video, images, fault records, maintenance logs, operation manuals, and expert experience of the main substation equipment; removing noise, irrelevant information, and duplicates from the data.

[0011] Preferably, using an association rule mining algorithm to obtain frequent item sets in the multi-source signal quantities and constructing perception data information based on the frequent item sets includes: using the FP-Growth algorithm as the association rule mining algorithm; constructing an FP-Tree of the multi-modal full-scale data of the main substation equipment through the association rule mining algorithm to obtain all frequent item sets of the main substation equipment.

[0012] Preferably, using a graph attention network to encode triples in the knowledge graph of the main substation equipment and identifying the perception data information by means of multi-hop reasoning includes: collecting the current operation state of the main substation equipment, and extracting the state of the corresponding entity in the knowledge graph of the main substation equipment from the current operation state; combining the corresponding actual states and querying the encoding result of the graph attention network.

[0013] Preferably, a multi-hop reasoning method is adopted to identify the sensed data information, and the substation main equipment fault diagnosis result is obtained according to the output result, including: in the multi-hop reasoning method, entities in the substation main equipment knowledge graph are classified by a trainable neural network; with each type as a node, new type nodes are added to the substation main equipment knowledge graph.

[0014] Preferably, a multi-hop reasoning method is adopted to identify the sensed data information, and the substation main equipment fault diagnosis result is obtained according to the output result, including: the multi-hop path is 2 hops or 3 hops.

[0015] In a second aspect of the present invention, there is provided a substation main equipment fault diagnosis system, which implements the substation main equipment fault diagnosis method in the first aspect of the present invention; wherein, the system includes a knowledge graph module, an association module, and a diagnosis module; wherein, the knowledge graph module is configured to collect multi-modal full-scale data of the substation main equipment, and obtain the substation main equipment knowledge graph through word segmentation, entity extraction, relationship extraction, and knowledge representation processes; the association module is configured to extract multi-source signal quantities related to the current substation main equipment from the multi-modal full-scale data of the substation main equipment, obtain frequent item sets in the multi-source signal quantities by using an association rule mining algorithm, and construct sensed data information based on the frequent item sets; the diagnosis module is configured to encode triples in the substation main equipment knowledge graph by using a graph attention network, identify the sensed data information by using a multi-hop reasoning method, and obtain the substation main equipment fault diagnosis result according to the output result.

[0016] In a third aspect of the present invention, there is provided a terminal, including a processor and a storage medium; the storage medium is used to store instructions; the processor is configured to operate according to the instructions to execute the steps of the method described in the first aspect of the present invention.

[0017] In a fourth aspect of the present invention, there is provided a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the steps of the method described in the first aspect of the present invention are implemented.

[0018] In a third aspect of the present invention, there is provided a terminal, including a processor and a storage medium; the storage medium is used to store instructions; the processor is configured to operate according to the instructions to execute the steps of the method described in the first aspect of the present invention.

[0019] In a fourth aspect of the present invention, there is provided a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the steps of the method described in the first aspect of the present invention are implemented.

[0020] The beneficial effects of the present invention are as follows. Compared with the prior art, in a substation main equipment fault diagnosis method, system, terminal and medium of the present invention, the method uses technologies such as natural language processing, knowledge graph and deep learning to mine the fault data characteristics of substation main equipment, construct a knowledge graph and fuse it with the device twin model. The present invention provides a knowledge graph construction and application system for fault diagnosis of substation main equipment oriented to digital twin substations. The system consists of modules such as data acquisition, data processing, graph construction, model mapping, early warning diagnosis, etc., and can realize the improvement of the fault diagnosis ability of digital twin substations.

[0021] In summary, the present invention provides a solution for establishing an expert library for fault diagnosis of main equipment in digital twin substations. Through advanced technical means, the operation and maintenance efficiency of substations is significantly improved, which has important strategic significance for the efficient operation of the power industry.

[0022] The beneficial effects of the present invention also include:

[0023] 1. The implementation of the present invention in the field of substation operation and maintenance significantly enhances the fault diagnosis ability of the main equipment of digital twin substations through knowledge graph and deep learning technologies. The present invention provides a data centralized management and equipment fault diagnosis system that integrates multi-source heterogeneous data such as equipment data, manual materials, and operation and maintenance logs. Its design uses deep learning technology to mine fault data, improve the reliability of diagnosis, significantly improve the digitalization level of substations, and provide an innovative data centralized management and equipment fault diagnosis solution for modern power systems.

[0024] 2. The knowledge graph technology can store and represent various knowledge of substation main equipment in a structured form, forming a rich knowledge network. This enables the system to quickly match and infer possible fault causes during the fault diagnosis process, improving the certainty of diagnosis. Thus, the present invention improves the accuracy and efficiency of fault diagnosis. The fault diagnosis system based on the knowledge graph and deep learning can combine historical data and real-time data to continuously evaluate the operation status of the equipment, provide a scientific basis for preventive maintenance, reduce unplanned downtime, and improve the reliability and availability of the equipment. Operation and maintenance personnel can reasonably arrange maintenance plans and resource allocations according to the suggestions provided by the system, improve the pertinence and effectiveness of operation and maintenance work, and reduce operation and maintenance costs.

[0025] 3. Based on the fault diagnosis requirements of substation main equipment, this invention establishes a Recurrent Convolutional Neural Network (RCNN), designs the structure and parameters of the deep learning model, trains the deep learning model using the preprocessed data, and adjusts the model parameters through the backpropagation algorithm so that the model can learn the characteristic patterns under normal operation and fault states of the equipment. The implementation of this invention is an important step in the digital transformation of the substation operation and maintenance field. By introducing advanced technologies such as knowledge graphs and deep learning, it promotes the development of substation operation and maintenance towards intelligence and automation. This not only improves the efficiency and reliability of substation operation and maintenance, but also provides strong support for the digital transformation of the power industry and promotes the sustainable development of the entire industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a schematic diagram of a substation main equipment fault diagnosis system of this invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0027] To make the objectives, technical solutions and advantages of this invention clearer and more accurate, the technical solutions of this invention will be described in detail through multiple specific embodiments below. The embodiments adopted by this invention are only used to explain this invention and do not limit the content of this invention.

[0028] Figure 1 It is a schematic diagram of a substation main equipment fault diagnosis system of this invention. As Figure 1 , in the first aspect of this invention, it relates to a substation main equipment fault diagnosis method, and the method includes the following steps: collecting multi-modal full-scale data of substation main equipment, and obtaining a knowledge graph of substation main equipment through processes such as word segmentation, entity extraction, relationship extraction, and knowledge representation; extracting multi-source signal quantities related to the current substation main equipment from the multi-modal full-scale data of substation main equipment, obtaining frequent item sets in the multi-source signal quantities using the association rule mining algorithm, and constructing perception data information based on the frequent item sets; encoding the triples in the knowledge graph of substation main equipment using a graph attention network, identifying the perception data information using a multi-hop reasoning method, and obtaining the substation main equipment fault diagnosis result according to the output result.

[0029] The fault diagnosis of substation main equipment requires collecting data from multiple sources. Through data cleaning, transformation, and word segmentation processing, the data is structured and integrated into a unified data platform for centralized data management.

[0030] The multimodal full-scale data of substation main equipment mainly includes structured data and unstructured data. Among them, structured data mainly refers to production data, status data, audio, video, images, etc. collected through substation main equipment information and sensing layer equipment. They have a fixed structure and are very convenient to call when analyzing and using. Unstructured data, on the other hand, has a very wide range of sources and no fixed structure, mainly including fault records, maintenance logs, operation manuals, expert experience, etc.

[0031] In the data preprocessing process, it is necessary to select the information with higher confidence in the above data and remove the noise, irrelevant information and duplicates in the data. Since there is no obvious separator between Chinese words, for text data, the present invention first performs word segmentation on it. Regarding the word segmentation problem of substation main equipment information description as a sequence labeling task, that is, labeling each character to indicate its position in the word, usually including the beginning of the word B, the middle of the word M, the end of the word E, and the word alone S, etc. Based on the processed corpus, the model parameters λ=(A, B, π) are obtained.

[0032] Among them, the matrix A represents the possible degree of conversion of main equipment information between different word position states, written as

[0033]

[0034] B, M, E, S are the input word positions, and P is the possible degree of conversion between each state.

[0035] The matrix B represents the possible degree of obtaining different observation values according to the current state, written as

[0036]

[0037] P(o 1 / B) represents the probability that the observation value o1 is the beginning of the state word B, and on is the nth observation value.

[0038] π is the initial state vector, representing the possible degree of the initial state word position information of the main equipment information.

[0039] Then, solve the text to be segmented to realize the labeling of word positions for word formation. The initialization method is as follows:

[0040] S 1 (i)=π i b i (o 1 )

[0041]

[0042] Among them, S 1 (i) is the maximum possible degree of transition in the state i at the 1st moment of equipment information, The maximum single path in state i at time 1, π i For each initial state, b i Is the degree of possibility with a value of o1.

[0043] The state transition method is as follows:

[0044]

[0045] Among them, a ij Represents the degree of possibility that the state at the next moment is j when the state at time t is i. St(i) is the maximum degree of possibility in state i of the device at time t, and bij is the degree of possibility that the observed value is oi when the state at time t is i and the state at the next moment is j. Is the single path with the highest probability in state i at time t.

[0046] Output the state sequence with the highest possibility of device information:

[0047]

[0048] Among them, Is the shortest path, I * Is the final path.

[0049] Store the text data after word segmentation in JSON format.

[0050] Based on the multi-modal full-scale data obtained through the above process, the entity extraction and relationship extraction of the main substation equipment are completed. According to the requirements of substation fault diagnosis, key entity types are defined, such as "equipment" (transformer, circuit breaker), "fault type" (overheating, short circuit), "symptom" (abnormal sound, oil leakage), "maintenance measures", etc. Entities are identified from the text using rules such as regular expressions and keyword matching. Use the Named Entity Recognition (NER) model to train the text to identify entities. Standardize the identified entities, such as unified naming and ambiguity elimination.

[0051] According to the logical connection between entities, relationship types are defined, such as "equipment - fault type" (faults that the equipment may have), "fault type - symptom" (specific manifestations of faults), "symptom - maintenance measures" (maintenance methods to be taken for symptoms), etc. Based on the method of supervised learning of RCNN, use the relationship extraction model to learn the relationship pattern from the labeled data.

[0052] RNN can handle long-term dependence relationships in sequence data. CNN extracts local features through convolution operations. In the RCNN structure adopted in the present invention, the output of the previous layer of RNN is fed into CNN to improve the performance of understanding context semantics.

[0053] The left and right context representations of each word can be obtained using forward and backward RNNs:

[0054]

[0055] where c l (w i ) represents the left context word vector of the i-th word, and c r (w i ) represents the right context word vector of the i-th word. W (l) is the hidden layer transformation matrix, and f is a non-linear activation function.

[0056] The loop structure can obtain all c l during the forward scan of the text and all c r during the backward scan of the text. Therefore, the output vector x i of the i-th word can be written as:

[0057] x i = [c l (w i ) ; e(w i ) ; c r (w i )]

[0058] The convolutional layer sets convolutional kernels with the same dimension as xi, a stride of 1, and uses tanh as the activation function to obtain a non-linear output result.

[0059] The pooling layer adopts the max pooling method to reduce the input dimension while retaining the most important semantic information in the text. The fully connected layer is fully connected to the pooling layer, and the softmax function is used to classify the output vector to discriminate the defective text category with the highest probability. The extracted relationships are verified to ensure the accuracy and rationality of the relationships.

[0060] Knowledge representation and fusion: Data integration is implemented based on each entity, and a database index is established to store the data. The Neo4j graph database is used to represent entities and relationships. The entity-attribute-relationship triples of the main substation equipment are imported, and the corpus is loaded to create entity relationship and attribute matching, forming an expert database for fault diagnosis of main substation equipment. The constructed knowledge graph data of the main substation equipment is imported into the relational database storage system to form an expert database for main equipment fault diagnosis. To improve the efficiency and performance of data query, indexes need to be established for the key fields in the database. An index is a special data structure in the database that can quickly locate specific records in the database table. By establishing an index, the scanning range during database query can be reduced, improving the query speed.

[0061] The present invention uses the main device name as digital "identity information" to support the associated mapping between the digital twin and the knowledge graph model. Through key technologies such as event extraction, knowledge fusion, and knowledge update, a knowledge graph of the main device status evaluation and fault diagnosis events with events as the core is constructed, storing the associated relationships between events and the argument information of the events themselves, breaking the isolation between events and entities formally, forming a global network that can roam and reason, and having numerous dimensions, realizing the knowledge mining and precipitation of the main device, and providing knowledge support for downstream tasks such as device operation status evaluation and fault diagnosis.

[0062] The application effect of the knowledge graph model is evaluated through a test set, an expert database for the main device fault diagnosis is established, and the expert database is iteratively optimized according to the evaluation results. The accuracy rate, recall rate and other indicators of the knowledge graph established above are evaluated through the test set to evaluate its effect in actual applications. Then, according to the evaluation results, the algorithms and parameters of entity extraction, relationship extraction and other links are adjusted to optimize the construction process of the knowledge graph and the knowledge access process of the main device fault diagnosis expert database.

[0063] Data mining, as an important field of computer science, aims to discover hidden patterns, associations and valuable information from large-scale datasets. As an excellent association rule mining algorithm, the FP-Growth algorithm can efficiently mine frequent item sets on large datasets by constructing a compact data structure and an efficient processing method. This article will deeply analyze the principle and advantages of the FP-Growth algorithm and introduce its cases in practical applications.

[0064] The health status of the main substation equipment is usually affected by signal quantities from multiple sources. Therefore, it is necessary to fully explore the associated relationships of multi-source data to accurately evaluate and diagnose the status of the equipment. Based on the above analysis, the present invention conducts research on multi-source heterogeneous association rule mining and analysis based on the FP-Growth algorithm.

[0065] The frequent item sets in multi-source signal quantities are obtained by using an association rule mining algorithm, and the perception data information is constructed based on the frequent item sets, including: using the FP-Growth algorithm as the association rule mining algorithm; constructing an FP-Tree of the multi-modal full-scale data of the main substation equipment through the association rule mining algorithm to obtain all the frequent item sets of the main substation equipment.

[0066] The FP-Growth algorithm uses a tree-like data structure, the frequent pattern tree, to store transaction records, thus achieving compressed storage of data. During the construction process, it is necessary to generate an item header table from the original data and then construct an FP-Tree through a node linked list. This algorithm does not need to eliminate infrequent item sets by traversing the transaction database, nor does it need to generate candidate item sets. Based on this algorithm, the association analysis of multi-source heterogeneous data is realized, and key data is extracted.

[0067] Classic association rule mining algorithms, but they also have some disadvantages: in large-scale data sets, the generation of candidate item sets and the calculation overhead of support degrees are relatively large, resulting in low algorithm efficiency. It is necessary to scan the data set multiple times, which has a large IO overhead, especially when the memory is limited. The generated candidate item sets may be very large and occupy a large amount of storage space. The FP-Growth (Frequent Pattern Growth) algorithm is a frequent item set mining algorithm based on the FP-Tree structure. The FP-Growth algorithm efficiently mines frequent item sets by constructing an FP-Tree, thus avoiding the process of generating candidate item sets.

[0068] In association rule mining, a frequent item set refers to an item set that appears in the data set more frequently than a pre-set threshold (support degree threshold). Frequent item sets are the basis of association rule mining and can be used to generate interesting association rules. The FP-Tree is the core data structure of the FP-Growth algorithm and is used to store frequent item sets and support degree counts. The FP-Tree consists of a root node, internal nodes, and leaf nodes. The root node does not store any information and is only used to connect different transaction paths. The internal nodes store element items and their corresponding support degree counts. The same element items in multiple transactions share a node, and their occurrence frequencies are counted through the count. The leaf nodes store element items.

[0069] The process of constructing an FP-Tree is as follows: Create a root node. For each transaction, insert element items into the FP-Tree in descending order of support degree. In the FP-Growth algorithm, the conditional pattern base refers to the prefix path ending with a certain element item. The conditional pattern base is used to construct a new conditional FP-Tree, thus realizing recursive mining of frequent item sets. The construction of the FP-Tree is the first stage of the FP-Growth algorithm. It mainly involves two scans of the data set: the first is used to count the support degree counts of each element item and sort them in descending order of support degree; the second is used to reconstruct the FP-Tree. During the construction of the FP-Tree, since the element items are already arranged in descending order of support degree, the same element items will appear adjacent, which makes the construction process of the FP-Tree very efficient. The finally constructed FP-Tree will be used in the second stage, that is, mining frequent item sets.

[0070] The root node of the FP-Tree does not store any information and is only used to connect different transaction paths. For each transaction, the items are inserted into the FP-Tree in descending order of support: for each transaction, according to the list of items in descending order of support, the items are inserted into the FP-Tree starting from the root node. If an item already exists in the FP-Tree, the support count of the corresponding node of this item is incremented. If the item does not exist in the FP-Tree, a new node is added to the tree to represent this item, and the support count is initialized to 1. The same items in multiple transactions share the same node in the FP-Tree. In this way, the FP-Tree realizes the compressed storage of frequent item sets. During the construction of the FP-Tree, since the items are already arranged in descending order of support, the same items will appear adjacent to each other, which makes the construction process of the FP-Tree very efficient. Thus, the present invention obtains the mining results of frequent item sets.

[0071] The FP-Growth algorithm compresses and stores the frequent item sets through the FP-Tree structure and does not need to generate candidate item sets, thus saving a large amount of storage space. In traditional association rule mining algorithms, such as the Apriori algorithm, in order to find frequent item sets, it is necessary to generate all possible candidate item sets and then perform support counting on the candidate item sets. Since the candidate item sets may be very large, this will occupy a large amount of storage space and computing resources. The FP-Growth algorithm replaces the process of generating candidate item sets by constructing an FP-Tree. The FP-Tree stores the frequent item sets in the form of a tree, thus avoiding the overhead of generating a large number of candidate item sets. Since the FP-Tree compresses and stores the same items, such a structure can represent large-scale frequent item sets in a smaller storage space, thus saving storage resources.

[0072] Preferably, a graph attention network is used to encode the triples in the knowledge graph of the main substation equipment, and a multi-hop reasoning method is adopted to identify the perception data information, including: collecting the current operating state of the main substation equipment, and extracting the state of the corresponding entity in the knowledge graph of the main substation equipment from the current operating state; combining the corresponding actual states and querying the encoding result of the graph attention network.

[0073] Graph Attention Networks (GAT) is a type of graph neural network that combines the attention mechanism. By learning the attention weights between nodes, it achieves weighted aggregation of neighbor node features, thereby improving the model's robustness to noisy neighbors and enhancing the model's interpretability. The core idea of GAT is to assign different weights to each node, and these weights are learned through the attention mechanism. In GAT, each node is regarded as the central node, and its feature vector is combined with the feature vectors of neighbor nodes, and an attention coefficient is calculated through a shared attention mechanism.

[0074] The advantage of GAT lies in its ability to adaptively assign weights to different neighbors, which is different from traditional graph convolutional networks (such as GCN) that usually use fixed weights. This adaptive ability makes GAT more flexible and powerful when dealing with graph data. In addition, GAT also introduces the multi-head attention mechanism, that is, multiple attention mechanisms are applied in parallel in the same layer, and then the results are combined. This method can capture richer features and improve the model's expressive ability.

[0075] In recent years, multi-hop reasoning has been widely studied to obtain more interpretable link prediction. However, we found in experiments that many of the paths given by these models are actually unreasonable, and there is little work on the evaluation of their interpretability. In this paper, we propose a unified framework to quantitatively evaluate the interpretability of multi-hop reasoning models to promote their development. Specifically, we define three metrics, including path recall, local interpretability, and global interpretability for evaluation, and design an approximation strategy to calculate these metrics using the interpretable scores of rules. In addition, we manually annotate all possible rules and establish a benchmark to detect the interpretability of multi-hop reasoning (BIMR). In the experiment, we verified the effectiveness of our benchmark test. In addition, we ran 9 representative baselines on the benchmark, and the experimental results show that the interpretability of current multi-hop reasoning models is not very satisfactory, which is 51.7% lower than the upper limit given by our benchmark. In addition, rule-based models are superior to multi-hop reasoning models in terms of performance and interpretability, which points out the direction for future research, that is, how to better incorporate rule information into multi-hop reasoning models.

[0076] Multi-hop reasoning in knowledge graphs (KGs) has been widely studied. It can not only infer new knowledge, but also provide reasoning paths that can explain the prediction results and make the model credible.

[0077] In the present invention, the knowledge graph can be represented by multiple triples. The head entity and the tail entity of a triple can be connected by a relationship. During the process of querying the knowledge graph through the sensed data information, the query is implemented starting from the sensed data information. Therefore, the sensed data information is an unchanged anchor point.

[0078] In one embodiment, the sensed data information is described as: Per = [Ver 1 , Ver 2 , …, Ver m , where Ver m is a kind of sensed data including the change amount. For example, by collecting the multi-modal full amount data of the main equipment of the substation at the current moment, the fault equipment address, fault record time, time of maintenance log, etc. in the fault record can be obtained.

[0079] Since the current multi-modal full amount data does not contain all types of data information. Or only some data contents in the current multi-modal full amount information change significantly. Therefore, this significant difference is extracted, and one or more most similar frequent item sets are extracted from all frequent item sets. During the process of constructing the frequent item sets, the equipment fault alarm information corresponding to a certain equipment and a certain fault can be mined. Therefore, potential classification of faults can be implemented by extracting the frequent item sets.

[0080] It is easy to think that since the knowledge graph of the main equipment of the substation includes all the information in the multi-modal full amount data. Therefore, if this graph is directly used for querying, it may lead to many redundancies in the final analysis result, or too many analysis paths. On this basis, the present invention provides importance weights for the relationships between the entities of the knowledge graph according to the data contents included in all frequent item sets, or deletes the invalid entities and invalid tuples that do not exist in the frequent items;

[0081] On this basis, the application effect of the knowledge graph model is evaluated through a test set, a main equipment fault diagnosis expert library is established, and iterative optimization of the expert library is realized according to the evaluation result. The accuracy rate, recall rate and other indexes of the knowledge graph established above are evaluated through the test set to evaluate its effect in actual application. Then, according to the evaluation result, the algorithms and parameters of the entity extraction, relationship extraction and other links are adjusted to optimize the construction process of the knowledge graph and the knowledge retrieval process of the main equipment fault diagnosis expert library.

[0082] For the state assessment of the main equipment in a substation, a multi-hop knowledge reasoning model based on a graph attention network is constructed according to the graph attention mechanism in the theoretical basis. Combining the perceptual data information extracted by association mining, multi-hop knowledge reasoning is performed on the state diagnosis knowledge graph. The specific implementation process of the graph attention network integrating the multi-hop path method is as follows: First, the entities and relationships on the multi-hop path are encoded by the graph attention network, and then the distributed representations of the encoded entities and relationships are input into the multi-hop reasoning model MultiHop for decoding and testing. The model framework is an encoder-decoder structure, where the graph attention model is the encoder and the multi-hop reasoning model is the decoder.

[0083] Before integrating the multi-hop path, it is necessary to use a graph traversal algorithm to obtain the multi-hop path corresponding to each triple, usually 2-hop or 3-hop. Based on the traversed multi-hop path, verify whether the triple satisfies the definition of the above valid tuple, that is, whether there is a multi-hop path that can reach the tail entity of the triple, and retain the valid tuple. Then, save each tuple and the corresponding multi-hop valid path, and at the same time remove the invalid tuples to construct a new data set. In addition, in order to improve the model's ability to distinguish valid paths and invalid paths, it is necessary to further construct the multi-hop invalid path corresponding to each tuple. Next, it is necessary to encode the multi-hop valid inference path and multi-hop invalid path of each tuple to obtain the encoded representations of the valid path and invalid path.

[0084] The multi-hop reasoning method is used to identify the perceptual data information, and the fault diagnosis result of the main equipment in the substation is obtained according to the output result, including: in the multi-hop reasoning method, the entities in the knowledge graph of the main equipment in the substation are classified by a trainable neural network; taking each type as a node, new type nodes are added to the knowledge graph of the main equipment in the substation.

[0085] The present invention uses a neural network to obtain the type of each piece of information in the perceptual data information. The neural network may include a convolutional layer, an activation layer, a hidden layer, and an output layer connected in sequence.

[0086] Among them, the convolutional layer performs a linear operation, and the hidden layer uses the result output by the convolutional layer as a weight, and adds the result convolution method to the output of the previous iteration of the hidden layer to obtain the current output of the hidden layer. Then, a convolutional operation is performed through an output layer, and finally the type of the entity is obtained.

[0087] Taking the type of each entity as a virtual node, it is added to the frequent item set after PF-growth condensation. Then a new knowledge graph is generated;

[0088] By training the encoder, the word vectors of entities and relationships in the knowledge graph triples can integrate the semantic and structural information in the multi-hop paths and learn to distinguish the paths that can reach the target entity from those that cannot. The distributed representations of the encoded entities and relationships use the improved multi-hop inference model MultiHop as the decoder for decoding and testing. The goal is to verify the generalization performance of the word embedding representations of the entities and relationships output by the encoder on the multi-hop inference model.

[0089] By combining the encoder and the decoder, a multi-hop knowledge inference model based on the graph attention network is constructed to realize the state evaluation of the main equipment in the substation and determine the health status of the equipment.

[0090] Based on the constructed expert library for the fault diagnosis of the main equipment, applications such as fault diagnosis and operation and maintenance assistance are developed. The expert library for the fault diagnosis of the main equipment is updated regularly, including adding new entities, relationships, correcting error information, etc., to ensure the timeliness and accuracy of the expert library for the fault diagnosis of the main equipment. Thereby, the fault warning and disposal auxiliary decision-making of the main equipment are output, and the prediction of the measurement data of the main equipment and the abnormal detection results are realized.

[0091] Through the above process, an expert library applicable to the fault diagnosis of the main equipment in the substation can be systematically constructed and integrated with the digital twin model, providing strong support for the operation and maintenance management of the substation. Based on the digital twin three-dimensional modeling technology, the fault warning results are presented in the equipment twin model, and the simulation deduction of abnormal disposal is carried out.

[0092] In the second aspect of the present invention, it relates to a fault diagnosis system for the main equipment in a substation. The system is implemented by using a fault diagnosis method for the main equipment in the substation in the first aspect of the present invention; wherein, the system includes a graph module, an association module, and a diagnosis module; wherein, the graph module is used to collect the multi-modal full-scale data of the main equipment in the substation and obtain the knowledge graph of the main equipment in the substation through the processes of word segmentation, entity extraction, relationship extraction, and knowledge representation; the association module is used to extract multi-source signal quantities related to the current main equipment in the substation from the multi-modal full-scale data of the main equipment in the substation, obtain the frequent item sets in the multi-source signal quantities by using the association rule mining algorithm, and construct the perception data information based on the frequent item sets; the diagnosis module is used to encode the triples in the knowledge graph of the main equipment in the substation by using the graph attention network, identify the perception data information in a multi-hop reasoning manner, and obtain the fault diagnosis result of the main equipment in the substation according to the output result.

[0093] In the third aspect of the present invention, it relates to a terminal, including a processor and a storage medium; the storage medium is used to store instructions; the processor is used to operate according to the instructions to execute the steps of the method described in the first aspect of the present invention.

[0094] In the fourth aspect of the present invention, it relates to a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the steps of the method described in the first aspect of the present invention are implemented.

[0095] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that there are still contents in the technical solutions of the present invention that can modify the specific implementation manners of the present invention or make equivalent replacements. Any modification or equivalent replacement without departing from the spirit and scope of the present invention shall be covered by the protection scope of the claims of the present invention.

Claims

1. A method for diagnosing faults of main equipment in a substation, characterized in that: The method comprises the following steps: Collect the full amount of multi-modal data of the main equipment of the substation, and obtain the knowledge graph of the main equipment of the substation through word segmentation, entity extraction, relationship extraction, and knowledge representation process; Extracting multi-source signals related to the current substation main equipment from the multi-modal full data of the substation main equipment, using an association rule mining algorithm to obtain frequent item sets in the multi-source signals, and constructing perception data information based on the frequent item sets; The graph attention network is used to encode the triplets in the knowledge graph of the substation main equipment, a multi-hop reasoning method is adopted to identify the perception data information, and the fault diagnosis result of the substation main equipment is obtained according to the output result.

2. A method for diagnosing faults of main equipment in a substation according to claim 1, characterized in that: The method collects the multi-modal full data of the main equipment of the substation, and obtains the knowledge graph of the main equipment of the substation through word segmentation, entity extraction, relationship extraction, and knowledge representation, including: According to the requirements of substation fault diagnosis, key entity types and relationship types are defined; The key entity types include equipment, fault type, symptom and repair measures; The relationship types include equipment-fault type, fault type-symptom, and symptom-maintenance measure.

3. A method for diagnosing faults of main equipment in a substation according to claim 2, characterized in that: The step of extracting multi-source signal quantities related to the current substation main equipment from the multi-modal full amount data of the substation main equipment includes: Collect local operation information of main equipment in substations, and collect perception layer operation information of main equipment in substations through perception layer equipment; The local operation information and the perception layer operation information include production data, status data, audio, video, image, fault record, maintenance log, operation manual, and expert experience of the main equipment of the substation; Remove noise, irrelevant information, and duplicates from your data.

4. A method for diagnosing faults of main equipment in a substation according to claim 3, characterized in that: The method of using an association rule mining algorithm to obtain frequent item sets in multi-source signals and constructing perception data information based on the frequent item sets includes: Use FP-Growth algorithm as association rule mining algorithm; Through the association rule mining algorithm, the FP-Tree of the multimodal full data of the substation main equipment is constructed to obtain all frequent item sets of the substation main equipment.

5. A method for diagnosing faults of main equipment in a substation according to claim 4, characterized in that: The method of using a graph attention network to encode the triples in the knowledge graph of the substation main equipment and using a multi-hop reasoning method to identify the perception data information includes: Collect the current operating status of the main equipment of the substation, and extract the status of the corresponding entity in the knowledge graph of the main equipment of the substation from the current operating status; The corresponding actual states are combined and a query is performed on the encoding result of the graph attention network.

6. A method for diagnosing faults of main equipment in a substation according to claim 5, characterized in that: The method of using a multi-hop reasoning method to identify the sensing data information and obtaining the fault diagnosis result of the main equipment of the substation according to the output result includes: In the multi-hop reasoning method, entities in the substation main equipment knowledge graph are classified by a trainable neural network; Taking each type as a node, a new type node is added to the substation main equipment knowledge graph.

7. A method for diagnosing faults of main equipment in a substation according to claim 6, characterized in that: The method of using a multi-hop reasoning method to identify the sensing data information and obtaining the fault diagnosis result of the main equipment of the substation according to the output result includes: The multi-hop path is 2 hops or 3 hops.

8. A fault diagnosis system for main equipment of a substation, characterized by: The system is implemented by using a substation main equipment fault diagnosis method according to any one of claims 1 to 7; wherein, The system includes a map module, a correlation module and a diagnosis module; wherein, The graph module is used to collect the multi-modal full data of the main equipment of the substation, and obtain the knowledge graph of the main equipment of the substation through word segmentation, entity extraction, relationship extraction, and knowledge representation process; The association module is used to extract multi-source signal quantities related to the current substation main equipment from the multi-modal full data of the substation main equipment, obtain frequent item sets in the multi-source signal quantities by using an association rule mining algorithm, and construct perception data information based on the frequent item sets; The diagnosis module is used to encode the triples in the knowledge graph of the substation main equipment using a graph attention network, identify the perception data information using a multi-hop reasoning method, and obtain the substation main equipment fault diagnosis result based on the output result.

9. A terminal comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the steps of the method according to any one of claims 1-7.

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

Images