Power grid main equipment operation and maintenance optimization method based on multi-modal data and knowledge graph

Abstract

The invention relates to a power grid main equipment operation and maintenance optimization method based on multi-modal data and a knowledge graph, and the method comprises the following steps: S1, obtaining multi-modal data related to the operation and maintenance of power grid main equipment, and carrying out the preprocessing of the multi-modal data; s2, constructing a knowledge graph by using a graph construction method; s3, expanding fault mode representation and probability distribution semantic enhancement characteristics based on the constructed knowledge graph, and describing fault causes and influence chains thereof in combination with causal inference to obtain a final knowledge graph; s4, combining the final knowledge graph and the fault prediction model to analyze the equipment state in real time, and predicting potential faults through historical faults and real-time operation and maintenance data; and S5, optimizing a maintenance plan and reducing downtime according to the equipment fault probability and the residual life predicted by the machine learning model and the potential fault chain deduced by the knowledge graph. The reliability and the operation and maintenance management level of the main equipment of the power grid can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of power grid operation and maintenance, and in particular to a method for optimizing power grid main equipment operation and maintenance based on multimodal data and knowledge graphs. Background Art

[0002] As the scale of the power system continues to expand and the complexity of equipment increases, the operational safety and stability of the main power grid equipment is particularly important for power supply security. However, the traditional power grid equipment operation and maintenance mainly relies on regular maintenance or manual processing after failure. This method has problems such as high maintenance costs, failure to detect potential failures in time, and difficulty in optimizing downtime. It is difficult to meet the high reliability and intelligence requirements of modern power grids.

[0003] In recent years, with the rapid development of big data, artificial intelligence and Internet of Things technologies, the sampling accuracy and monitoring granularity of power grid equipment operation have been significantly improved. At the same time, increasingly complex multi-source heterogeneous data (such as sensor data, fault records, maintenance logs, infrared images, etc.) provide rich data support for equipment status perception and fault diagnosis. However, the heterogeneity, dispersion and semantic inconsistency of multimodal data increase the difficulty of data processing and analysis. In addition, the failure modes are complex and the correlation between devices is strong. It is difficult to predict potential failures in a timely and accurate manner by simply relying on a single data source or a single model. Summary of the invention

[0004] In order to solve the above problems, the purpose of the present invention is to provide a power grid main equipment operation and maintenance optimization method based on multimodal data and knowledge graph, which can effectively improve the reliability and operation and maintenance management level of the power grid main equipment.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] A method for optimizing operation and maintenance of main equipment of a power grid based on multimodal data and knowledge graph comprises the following steps:

[0007] Step S1: Acquire multimodal data related to the operation and maintenance of the main equipment of the power grid, including structured data, unstructured data, sensor data and image data, and pre-process them;

[0008] Step S2: Use the graph construction method to build a knowledge graph, integrate data in different forms into the knowledge graph model, and map data in different forms to the same semantic space, using graph embedding technology to ensure consistency in form;

[0009] Step S3: Based on the constructed knowledge graph, the fault mode representation and probability distribution semantic enhancement features are expanded, combined with causal inference, to characterize the fault cause and its impact chain, and obtain the final knowledge graph;

[0010] Step S4: Analyze the equipment status in real time by combining the final knowledge graph and the fault prediction model, and predict potential faults through historical faults and real-time operation and maintenance data;

[0011] Step S5: Based on the equipment failure probability and remaining life predicted by the machine learning model and the potential failure chain inferred by the knowledge graph, the maintenance plan is optimized to reduce downtime.

[0012] Further, preprocessing is as follows:

[0013] Extract equipment status, fault location and related information from unstructured data and convert them into structured representation;

[0014] For structured data, remove noise (outliers, erroneous records), fill missing values, and complete feature normalization;

[0015] For sensor data, extract key time series features and reduce data redundancy through sampling, noise filtering, and feature extraction methods;

[0016] For image data, bilateral filtering is used to remove noise, the fault area in the image is located based on the target detection model, and the ResNet deep neural network is used for feature extraction to convert image features into feature vectors.

[0017] Furthermore, step S2 is specifically as follows:

[0018] Step S21: Using a graph construction method, data in different forms are integrated into the knowledge graph model to construct a knowledge graph G;

[0019] Step S22: Map the constructed knowledge graph G to a shared low-dimensional semantic vector space through graph embedding technology;

[0020] Step S23: By learning the alignment function, the representations of different modalities are embedded into a shared semantic space, maximizing the embedding consistency of the same device instance in different modalities and minimizing the cross-modal embedding difference.

[0021] In this embodiment, step S21 is specifically as follows:

[0022] Define a graph G = (V, E) consisting of a set of nodes V = {v1, v2, ..., v i …,v n} and edge set E = {(v i ,v j ,r)}

[0023] Where: v i ,v j∈V is a node, representing the core entity in multimodal data (such as equipment, fault, status, etc.); r∈R is an edge, representing the relationship between entities; E is an edge;

[0024] Convert attribute fields in structured data into node attributes; convert time series anomalies after sensor data processing into event nodes and connect them to corresponding device nodes; convert defect areas extracted after image recognition into fault feature nodes;

[0025] Finally, a multimodal fusion knowledge graph is constructed, represented as Gmulti = (V, E, R), which contains entities and relations extracted from multimodal data.

[0026] In this embodiment, step S22 is specifically as follows:

[0027] The goal of graph embedding is to learn the map f:G→R d , mapping the entities and relations of the knowledge graph into a low-dimensional vector space;

[0028] Assume that entities and relations satisfy the translation relationship in vector space:

[0029]

[0030] Loss function:

[0031]

[0032] Among them, h, t, r represent the embedding vectors of the head entity, tail entity and relation respectively; h′, t′ are the corresponding negative samples; ||.|| is the distance function in the vector space; γ is a hyperparameter; ||.|| + Represents the ReLU operation;

[0033] Using the GNN model, the entity neighborhood information is considered for embedding update:

[0034]

[0035] in is the embedding of node i in the kth layer of the GNN model; Aggregate represents the neighbor information aggregation function

[0036] Introducing enhanced embedding technology for different modal characteristics: Introducing attribute embedding for nodes with sensor data and image data:

[0037] h node =MLP(Concat[h graph ,h sensor ,h image ]);

[0038] Among them, h graphh is the graph structure information embedding; sensor is sensor data embedding; h image Embedding for image data;

[0039] Introduce modal identifiers for relations:

[0040] h rel =W rel ·E mbed (modal);

[0041] After embedding, each node v i ∈V will correspond to a low-dimensional semantic vector Each edge r j ∈R will correspond to a relation embedding

[0042] Represent G as an embedding matrix:

[0043] Node Matrix:

[0044] Relationship Matrix:

[0045] In this embodiment, S23 is specifically:

[0046] For each modality, a mapping function is constructed to embed the modal features into a unified semantic space;

[0047] h aligned =MLP(Concat[h structured ,h text ,h sensor ,h image ]);

[0048] Among them, h structured ,h text ,h sensor ,h image They are structured data embedding of nodes, text data embedding of nodes, sensor data embedding of nodes and image data embedding of nodes respectively; MLP is multi-layer perceptron;

[0049] Alignment training is achieved by maximizing cross-modal consistency and minimizing embedding differences: maximizing the similarity of embeddings of the same entity in different modalities:

[0050]

[0051] Among them, mi and mj represent two different modes of node v; are two modal embedding representations of node v;

[0052] Minimize the distance between embeddings of the same device instance in different modalities:

[0053]

[0054] Total loss:

[0055] L=λ1L similarity +λ2L difference ;

[0056] Among them, λ1 and λ2 are weight coefficients;

[0057] Ultimately, alignment maximizes consistency while reducing embedding differences: entity embeddings of all modalities are unified into a shared semantic space, and each device instance has a modality-independent global semantic representation: All multimodal information of the device is retained to facilitate subsequent analysis and inference.

[0058] In this embodiment, step S3 is specifically:

[0059] Step S31: By decomposing the fault occurrence and propagation mechanism, the specific fault mode is divided into different stages, and a fault cause chain table is constructed;

[0060] Step S32: By means of probability modeling, probability distribution information is assigned to the triple relationship in the knowledge graph to indicate the uncertainty of the entity / relationship;

[0061] Step S33: Use causal inference technology to analyze the cause of the fault and its extension pattern, and clarify the strength and direction of the causal relationship in the equipment failure chain

[0062] Step S34: Generate the final knowledge graph by integrating the extended fault mode representation, probability distribution semantics and causal inference results.

[0063] In this embodiment, step S32 is specifically:

[0064] For the triple (h, r, t)(h, r, t) in the knowledge graph, construct the following probability model:

[0065] Define the conditional probability of a triple:

[0066] P(t|h,r);

[0067] represents the probability of tail entity t under the condition of entity h and its relation rr

[0068] Define the confidence level that a relationship occurs:

[0069] P(h,r,t);

[0070] Indicates the confidence of the triple as a whole;

[0071] Using existing historical data, calculate the observed triple frequency:

[0072]

[0073] Among them, P(t|h,r) is the conditional probability, which indicates the possibility of the tail entity t under the condition of the head entity h and the relation r; freq(h,r,t) is the frequency of the triple (h,r,t) in the data; freq(h,r) is the joint frequency of the head entity h and the relation r;

[0074] Under the condition of limited observation data, the prior probability is introduced and the posterior is updated using the Bayesian formula:

[0075]

[0076] P(h|D)∝P(D|h)·P(h);

[0077] Where P(h) is the prior probability of the head entity h; D is the observed data; P(t) is the marginal probability of the tail entity t; P(h|D) is the posterior probability of the head entity hh under the data D; P(D|h) is the likelihood function

[0078] Use the conditional probabilities on the triple chain for combined calculations to model the entire chain:

[0079] Add a probability representation to each edge in the knowledge graph, specifically: (h, r, t, P(h, r, t)).

[0080] In this embodiment, S33 is specifically:

[0081] Use the cause-effect diagram model to construct a cause-effect diagram that reflects the equipment failure chain;

[0082] ATE is used to quantify the average causal impact of a factor (such as a contributing factor) on another state or fault:

[0083]

[0084] Among them, X=1 means the inducement exists; X=0 means the inducement does not exist; Y is the state of the target event;

[0085] When direct observation is not possible or there are confounding variables, instrumental variables Z are used to quantify causal strength;

[0086]

[0087] Causal quantification based on a potential outcomes framework:

[0088]

[0089] Among them, Y1 is the consequence of applying intervention (such as high temperature environment); Y0 is the consequence without applying intervention;

[0090] Using time series data to determine the directionality of causality:

[0091]

[0092] According to the causal graph, direct reasoning is performed by predicting the consequences through the intensity of intervention, and chain reasoning is performed along the causal chain transmission relationship; the causal inference results are integrated into the knowledge graph, which can be expanded into a causal enhanced knowledge graph.

[0093] In this embodiment, step S4 is specifically as follows:

[0094] A fault prediction model is built based on LSTM, and correlation characteristics, historical fault modes, and dynamic state data are extracted from the knowledge graph as input features of the fault prediction model;

[0095] Predict potential failures in real time based on the equipment status time series;

[0096] Use semantic associations and causal reasoning in knowledge graphs to help machine learning models explain prediction results.

[0097] The present invention has the following beneficial effects:

[0098] 1. The present invention combines the fusion of multi-source heterogeneous data and the modeling of fault chains, combined with fault prediction and intelligent optimization technology, which can not only reduce the unplanned downtime of equipment, but also significantly improve the operation and maintenance efficiency and safety of power grid equipment;

[0099] 2. The present invention improves data utilization by fusion of multimodal data and unified semantic space mapping, and uses graph embedding technology to map multimodal data to the same semantic space, thereby solving the problem of data heterogeneity, ensuring the relevance and consistency between data, facilitating subsequent analysis and reasoning, fully tapping the potential value of multimodal data, and avoiding the blind spots in fault identification that may be caused by a single data source;

[0100] 3. The knowledge graph generated by the present invention can not only represent the equipment status and fault chain, but also quantify the causal influence between the fault cause and consequence, providing efficient support for intelligent diagnosis and prediction. By combining causal inference technology, based on historical fault data and operation and maintenance records, the cause of equipment failure and its impact chain are analyzed, solving the problem of "correlation does not equal causality" in traditional machine learning methods; BRIEF DESCRIPTION OF THE DRAWINGS

[0101] Figure 1 The figure is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0103] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0104] refer to Figure 1 In this embodiment, a method for optimizing the operation and maintenance of main equipment of a power grid based on multimodal data and knowledge graph is provided, which is characterized by comprising the following steps:

[0105] Step S1: Acquire multimodal data related to the operation and maintenance of the main equipment of the power grid, including structured data, unstructured data, sensor data and image data, and pre-process them;

[0106] Structured data: Equipment operation data, maintenance history, and configuration lists from EMS / SCADA systems. Unstructured data: Technical manuals, fault reports, and maintenance logs. Sensor data: Voltage, current, temperature, vibration, and other data provided by condition monitoring sensors. Image / video data: Equipment visual information collected by surveillance cameras and drone inspections.

[0107] Step S2: Use the graph construction method to build a knowledge graph, integrate data in different forms into the knowledge graph model, and map data in different forms to the same semantic space, using graph embedding technology to ensure consistency in form;

[0108] Step S3: Based on the constructed knowledge graph, the fault mode representation and probability distribution semantic enhancement features are expanded, combined with causal inference, to characterize the fault cause and its impact chain, and obtain the final knowledge graph;

[0109] Step S4: Analyze the equipment status in real time by combining the final knowledge graph and the fault prediction model, and predict potential faults through historical faults and real-time operation and maintenance data;

[0110] Step S5: Based on the equipment failure probability and remaining life predicted by the machine learning model and the potential failure chain inferred by the knowledge graph, the maintenance plan is optimized to reduce downtime.

[0111] In this embodiment, the preprocessing is as follows:

[0112] Extract equipment status, fault location and related information from unstructured data and convert them into structured representation;

[0113] For structured data, remove noise (outliers, erroneous records), fill missing values, and complete feature normalization;

[0114] For sensor data, extract key time series features and reduce data redundancy through sampling, noise filtering, and feature extraction methods;

[0115] For image data, bilateral filtering is used to remove noise, the fault area in the image is located based on the target detection model, and the ResNet deep neural network is used for feature extraction to convert image features into feature vectors.

[0116] In this embodiment, step S2 is specifically as follows:

[0117] Step S21: Using a graph construction method, data in different forms are integrated into the knowledge graph model to construct a knowledge graph G;

[0118] Step S22: Map the constructed knowledge graph G to a shared low-dimensional semantic vector space through graph embedding technology;

[0119] Step S23: By learning the alignment function, the representations of different modalities are embedded into a shared semantic space, maximizing the embedding consistency of the same device instance in different modalities and minimizing the cross-modal embedding difference.

[0120] In this embodiment, step S21 is specifically as follows:

[0121] Define a graph G = (V, E) consisting of a set of nodes V = {v1, v2, ..., v i …,v n} and edge set E = {(v i ,v j ,r)}

[0122] Where: v i ,v j ∈V is a node, representing the core entity in multimodal data (such as equipment, fault, status, etc.); r∈R is an edge, representing the relationship between entities (such as "connection", "trigger", "type belongs to"); E is an edge;

[0123] Convert attribute fields in structured data into node attributes. For example, typical attributes of transformers in substations (model, rated voltage, etc.) are directly constructed as nodes or relationships.

[0124] Example: Node "Transformer A" and edge (Transformer A, belongs to, Substation B)(Transformer A, belongs to, Substation B).

[0125] Convert the time series anomaly points after sensor data processing into event nodes and connect them to the corresponding device nodes; Example: (Motor C sensor, detected anomaly, December 3, 2023) (Motor C sensor, detected anomaly, December 3, 2023)

[0126] The defect areas (such as cracks and hot spots) extracted after image recognition are converted into fault characteristic nodes; example: infrared thermal imaging technology identifies overheating faults, constructs the node overheating area 1 overheating area 1, and the relationship (transformer B, temperature rise, overheating area 1) (transformer B, temperature rise, overheating area 1).

[0127] Finally, a multimodal fusion knowledge graph is constructed, represented as Gmulti = (V, E, R), which contains entities and relations extracted from multimodal data.

[0128] In this embodiment, step S22 is specifically as follows:

[0129] The goal of graph embedding is to learn the map f:G→R d , mapping the entities and relations of the knowledge graph into a low-dimensional vector space;

[0130] Assume that entities and relations satisfy the translation relationship in vector space:

[0131]

[0132] Loss function:

[0133]

[0134] Among them, h, t, r represent the embedding vectors of the head entity, tail entity and relation respectively; h′, t′ are the corresponding negative samples; ||.|| is the distance function in the vector space; γ is a hyperparameter; ||.|| + Represents the ReLU operation;

[0135] Using the GNN model, the entity neighborhood information is considered for embedding update:

[0136]

[0137] in is the embedding of node i in the kth layer of the GNN model; Aggregate represents the neighbor information aggregation function

[0138] Introducing enhanced embedding technology for different modal characteristics: Introducing attribute embedding for nodes with sensor data and image data:

[0139] h node =MLP(Concat[h graph ,h sensor ,h image ]);

[0140] Among them, h graph h is the graph structure information embedding; sensor is sensor data embedding; h image Embedding for image data;

[0141] Introduce modality identifiers for relations (e.g. whether the fault relation comes from a sensor or text):

[0142] h rel =W rel ·E mbed (modal);

[0143] After embedding, each node v i ∈V will correspond to a low-dimensional semantic vector Each edge r j ∈R will correspond to a relation embedding

[0144] Represent G as an embedding matrix:

[0145] Node Matrix:

[0146] Relationship Matrix:

[0147] In this embodiment, S23 is specifically:

[0148] For each modality, a mapping function is constructed to embed the modal features into a unified semantic space;

[0149] Input embedding for different modalities:

[0150]

[0151] Structured data:

[0152] Text Modal:

[0153] h aligned =MLP(Concat[h structured ,h text ,h sensor ,h image ]);

[0154] Among them, h structured ,h text ,h sensor ,h image They are structured data embedding of nodes, text data embedding of nodes, sensor data embedding of nodes and image data embedding of nodes respectively; MLP is multi-layer perceptron;

[0155] Alignment training is achieved by maximizing cross-modal consistency and minimizing embedding differences: maximizing the similarity of embeddings of the same entity in different modalities:

[0156]

[0157] Among them, mi and mj represent two different modes of node v; are two modal embedding representations of node v;

[0158] Minimize the distance between embeddings of the same device instance in different modalities:

[0159]

[0160] Total loss:

[0161] L=λ1L similarity +λ2L difference ;

[0162] Among them, λ1 and λ2 are weight coefficients;

[0163] Ultimately, alignment maximizes consistency while reducing embedding differences: entity embeddings of all modalities are unified into a shared semantic space, and each device instance has a modality-independent global semantic representation: All multimodal information of the device is retained to facilitate subsequent analysis and inference.

[0164] In this embodiment, step S3 is specifically:

[0165] Step S31: By decomposing the fault occurrence and propagation mechanism, the specific fault mode is divided into different stages, and a fault cause chain table is constructed;

[0166] The failure modes are logically divided into three levels:

[0167] Root Cause Layer: Indicates the root external causes (such as environment, operating errors) or equipment internal causes (such as component aging, material defects) that lead to failure.

[0168] Example:

[0169] High temperature weather causes the cooling capacity of equipment to decrease (external factor).

[0170] The performance degradation of the aging cooler leads to incomplete heat dissipation (internal cause).

[0171] Failure Propagation Layer:

[0172] Represents the intermediate mechanism or link of fault propagation (such as linkage reaction between components).

[0173] Includes: Interactions between fault causes (e.g. overheating → insulation aging → short circuit) and causal relationships between components (e.g. winding overheating may affect insulation performance).

[0174] Outcome Layer:

[0175] Indicates the final impact on the device or system.

[0176] Typical consequences: shutdown failure (such as whole machine shutdown). Safety disaster (such as explosion, power outage). Specific examples are as follows:

[0177]

[0178] Step S32: By means of probability modeling, probability distribution information is assigned to the triple relationship in the knowledge graph to indicate the uncertainty of the entity / relationship;

[0179] Step S33: Use causal inference technology to analyze the cause of the fault and its extension pattern, and clarify the strength and direction of the causal relationship in the equipment failure chain

[0180] Step S34: Generate the final knowledge graph by integrating the extended fault mode representation, probability distribution semantics and causal inference results.

[0181] In this embodiment, step S32 is specifically:

[0182] For the triple (h, r, t)(h, r, t) in the knowledge graph, construct the following probability model:

[0183] Define the conditional probability of a triple:

[0184] P(t|h,r);

[0185] represents the probability of tail entity t under the condition of entity h and its relation rr

[0186] Define the confidence level that a relationship occurs:

[0187] P(h,r,t);

[0188] Indicates the confidence of the triple as a whole (such as the degree of data support).

[0189] Using existing historical data (sensor-fault logs, etc.), calculate the observed triple frequency:

[0190]

[0191] Among them, P(t|h,r) is the conditional probability, which indicates the possibility of the tail entity t under the condition of the head entity h and the relation r; freq(h,r,t) is the frequency of the triple (h,r,t) in the data; freq(h,r) is the joint frequency of the head entity h and the relation r;

[0192] Under the condition of limited observation data, the prior probability is introduced and the posterior is updated using the Bayesian formula:

[0193]

[0194] P(h|D)∝P(D|h)·P(h);

[0195] Where P(h) is the prior probability of the head entity h; D is the observed data; P(t) is the marginal probability of the tail entity t; P(h|D) is the posterior probability of the head entity hh under the data D; P(D|h) is the likelihood function

[0196] Use the conditional probabilities on the triple chain for combined calculations to model the entire chain:

[0197] Add a probability representation to each edge in the knowledge graph, specifically: (h, r, t, P(h, r, t)).

[0198] In this embodiment, S33 is specifically:

[0199] Use the cause-effect diagram model to construct a cause-effect diagram that reflects the equipment failure chain;

[0200] ATE is used to quantify the average causal impact of a factor (such as a contributing factor) on another state or fault:

[0201]

[0202] Where X=1 indicates that the cause exists (such as high ambient temperature); X=0 indicates that the cause does not exist; Y is the state of the target event (such as winding overheating);

[0203] When direct observation is not possible or there are confounding variables, instrumental variables Z are used to quantify causal strength;

[0204]

[0205] Causal quantification based on a potential outcomes framework:

[0206]

[0207] Among them, Y1 is the consequence of applying intervention (such as high temperature environment); Y0 is the consequence without applying intervention;

[0208] Using time series data to determine the directionality of causality:

[0209]

[0210] According to the causal graph, direct reasoning is performed by predicting the consequences through the intensity of intervention, and chain reasoning is performed along the causal chain transmission relationship; the causal inference results are integrated into the knowledge graph, which can be expanded into a causal enhanced knowledge graph.

[0211] In this embodiment, step S4 is specifically as follows:

[0212] A fault prediction model is built based on LSTM, and correlation characteristics, historical fault modes, and dynamic state data are extracted from the knowledge graph as input features of the fault prediction model;

[0213] Predict potential failures in real time based on the device’s status time series (sensor data, etc.);

[0214] Use semantic associations and causal reasoning in knowledge graphs to help machine learning models explain prediction results.

[0215] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0216] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0217] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0218] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0219] The above is only a preferred embodiment of the present invention, and does not limit the present invention in other forms. Any technician familiar with the profession may use the above disclosed technical content to change or modify it into an equivalent embodiment with equivalent changes. However, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the technical solution of the present invention still belongs to the protection scope of the technical solution of the present invention.

Claims

1. A method for optimizing the operation and maintenance of main equipment of a power grid based on multimodal data and knowledge graph, characterized in that: The following steps are involved: Step S1: Acquire multimodal data related to the operation and maintenance of the main equipment of the power grid, including structured data, unstructured data, sensor data and image data, and pre-process them; Step S2: Use the graph construction method to build a knowledge graph, integrate data in different forms into the knowledge graph model, and map data in different forms to the same semantic space, using graph embedding technology to ensure consistency in form; Step S3: Based on the constructed knowledge graph, the fault mode representation and probability distribution semantic enhancement features are expanded, combined with causal inference, to characterize the fault cause and its impact chain, and obtain the final knowledge graph; Step S4: Analyze the equipment status in real time by combining the final knowledge graph and the fault prediction model, and predict potential faults through historical faults and real-time operation and maintenance data; Step S5: Based on the equipment failure probability and remaining life predicted by the machine learning model and the potential failure chain inferred by the knowledge graph, the maintenance plan is optimized to reduce downtime.

2. The method for optimizing the operation and maintenance of main equipment of a power grid based on multimodal data and knowledge graph according to claim 1, characterized in that: The preprocessing is specifically as follows: Extract equipment status, fault location and related information from unstructured data and convert them into structured representation; For structured data, remove noise (outliers, erroneous records), fill missing values, and complete feature normalization; For sensor data, extract key time series features and reduce data redundancy through sampling, noise filtering, and feature extraction methods; For image data, bilateral filtering is used to remove noise, the fault area in the image is located based on the target detection model, and the ResNet deep neural network is used for feature extraction to convert image features into feature vectors.

3. The method for optimizing the operation and maintenance of main equipment of a power grid based on multimodal data and knowledge graph according to claim 1, characterized in that: The step S2 is specifically as follows: Step S21: Using a graph construction method, data in different forms are integrated into the knowledge graph model to construct a knowledge graph G; Step S22: Map the constructed knowledge graph G to a shared low-dimensional semantic vector space through graph embedding technology; Step S23: By learning the alignment function, the representations of different modalities are embedded into a shared semantic space, maximizing the embedding consistency of the same device instance in different modalities and minimizing the cross-modal embedding difference.

4. The method for optimizing operation and maintenance of main equipment of power grid based on multimodal data and knowledge graph according to claim 3 is characterized in that: The step S21 is specifically as follows: Define a graph G = (V, E) consisting of a set of nodes V = {v1, v2, ..., v i …,v n } and edge set E = {(v i ,v j ,r)} Where: v i ,v j ∈V is a node, representing the core entity in multimodal data (such as equipment, fault, status, etc.); r∈R is an edge, representing the relationship between entities; E is an edge; Convert attribute fields in structured data into node attributes; convert time series anomalies after sensor data processing into event nodes and connect them to corresponding device nodes; convert defect areas extracted after image recognition into fault feature nodes; Finally, a multimodal fusion knowledge graph is constructed, represented as Gmulti = (V, E, R), which contains entities and relations extracted from multimodal data.

5. The method for optimizing the operation and maintenance of main equipment of a power grid based on multimodal data and knowledge graph according to claim 3 is characterized in that: The step S22 is specifically as follows: The goal of graph embedding is to learn the map f:G→R d , mapping the entities and relations of the knowledge graph into a low-dimensional vector space; Assume that entities and relations satisfy the translation relationship in vector space: Loss function: Among them, h, t, r represent the embedding vectors of the head entity, tail entity and relation respectively; h′, t′ are the corresponding negative samples; ||·|| is the distance function in the vector space; γ is a hyperparameter; ||·|| + Represents the ReLU operation; Using the GNN model, the entity neighborhood information is considered for embedding update: in is the embedding of node i in the kth layer of the GNN model; Aggregate represents the neighbor information aggregation function Introducing enhanced embedding technology for different modal characteristics: Introducing attribute embedding for nodes with sensor data and image data: h node =MLP(Concat[h graph ,h sensor ,h image ]); Among them, h graph h is the graph structure information embedding; sensor is sensor data embedding; h image Embedding for image data; Introduce modal identifiers for relations: h rel =W rel ·E mbed (modal); After embedding, each node v i ∈V will correspond to a low-dimensional semantic vector Each edge r j ∈R will correspond to a relation embedding Represent G as an embedding matrix: Node Matrix: Relationship Matrix:

6. The method for optimizing operation and maintenance of main equipment of power grid based on multimodal data and knowledge graph according to claim 3 is characterized in that: The S23 is specifically: For each modality, a mapping function is constructed to embed the modal features into a unified semantic space; h aligned =MLP(Concat[h structured ,h text ,h sensor ,h image ]); Among them, h structured ,h text ,h sensor ,h image They are structured data embedding of nodes, text data embedding of nodes, sensor data embedding of nodes and image data embedding of nodes respectively; MLP is multi-layer perceptron; Alignment training is achieved by maximizing cross-modal consistency and minimizing embedding differences: maximizing the similarity of embeddings of the same entity in different modalities: Among them, m i 、m j Represents two different modes of node v; are two modal embedding representations of node v; Minimize the distance between embeddings of the same device instance in different modalities: Total loss: L=λ1L similarity +λ2L difference ; Among them, λ1 and λ2 are weight coefficients; Ultimately, alignment maximizes consistency while reducing embedding differences: entity embeddings of all modalities are unified into a shared semantic space, and each device instance has a modality-independent global semantic representation: All multimodal information of the device is retained to facilitate subsequent analysis and inference.

7. The method for optimizing the operation and maintenance of main equipment of a power grid based on multimodal data and knowledge graph according to claim 1, characterized in that: The step S3 is specifically: Step S31: By decomposing the fault occurrence and propagation mechanism, the specific fault mode is divided into different stages, and a fault cause chain table is constructed; Step S32: By means of probability modeling, probability distribution information is assigned to the triple relationship in the knowledge graph to indicate the uncertainty of the entity / relationship; Step S33: Use causal inference technology to analyze the cause of the fault and its extension pattern, and clarify the strength and direction of the causal relationship in the equipment failure chain Step S34: Generate the final knowledge graph by integrating the extended fault mode representation, probability distribution semantics and causal inference results.

8. The method for optimizing operation and maintenance of main equipment of power grid based on multimodal data and knowledge graph according to claim 7 is characterized in that: The step S32 is specifically as follows: For the triple (h, r, t)(h, r, t) in the knowledge graph, construct the following probability model: Define the conditional probability of a triple: P(t|h,r); represents the probability of tail entity t under the condition of entity h and its relation rr Define the confidence level that a relationship occurs: P(h,r,t); Indicates the confidence of the triple as a whole; Using existing historical data, calculate the observed triple frequency: Among them, P(t|h,r) is the conditional probability, which indicates the possibility of the tail entity t under the condition of the head entity h and the relation r; freq(h,r,t) is the frequency of the triple (h,r,t) in the data; freq(h,r) is the joint frequency of the head entity h and the relation r; Under the condition of limited observation data, the prior probability is introduced and the posterior is updated using the Bayesian formula: P(h|D)∝(D|h)·P(h); Where P(h) is the prior probability of the head entity h; D is the observed data; P(t) is the marginal probability of the tail entity t; P(h|D) is the posterior probability of the head entity hh under the data D; P(D|h) is the likelihood function Use the conditional probabilities on the triple chain for combined calculations to model the entire chain: Add a probability representation to each edge in the knowledge graph, specifically: (h, r, t, P(h, r, t)).

9. The method for optimizing operation and maintenance of main equipment of power grid based on multimodal data and knowledge graph according to claim 7, characterized in that: The S33 is specifically: Use the cause-effect diagram model to construct a cause-effect diagram that reflects the equipment failure chain; ATE is used to quantify the average causal impact of a factor (such as a contributing factor) on another state or fault: Among them, X=1 means the inducement exists; X=0 means the inducement does not exist; Y is the state of the target event; When direct observation is not possible or there are confounding variables, instrumental variables Z are used to quantify causal strength; Causal quantification based on a potential outcomes framework: Among them, Y1 is the consequence of intervention (such as high temperature environment); Y0 is the consequence without intervention; Using time series data to determine the directionality of causality: According to the causal graph, direct reasoning is performed by predicting the consequences through the intensity of intervention, and chain reasoning is performed along the causal chain transmission relationship; the causal inference results are integrated into the knowledge graph, which can be expanded into a causal enhanced knowledge graph.

10. The method for optimizing operation and maintenance of main equipment of power grid based on multimodal data and knowledge graph according to claim 1, characterized in that: The step S4 is specifically as follows: A fault prediction model is built based on LSTM, and correlation characteristics, historical fault modes, and dynamic state data are extracted from the knowledge graph as input features of the fault prediction model; Predict potential failures in real time based on the equipment status time series; Use semantic associations and causal reasoning in knowledge graphs to help machine learning models explain prediction results.

Images