Power equipment fault diagnosis method and system based on AI knowledge base

Abstract

The embodiment of the invention relates to the technical field of intelligent operation and maintenance and fault diagnosis of a power system, and discloses a power equipment fault diagnosis method and system based on an AI knowledge base, and the method comprises the steps: obtaining the current equipment data of each power equipment in the power system; mapping the current equipment data of each power equipment to defined entities and attributes in a basic knowledge graph to obtain a target entity and a target attribute corresponding to the current equipment data, and a target causal relationship and a target correlativity between each target entity and the target attribute; according to the target entities, the target attributes, the target causal relationship between the target entities and the target attributes and the target correlativity, a target fault knowledge graph is constructed and obtained; and inputting the target fault knowledge graph into a causal relationship network for reasoning to obtain equipment fault information. Through the above mode, the embodiment of the invention realizes high-precision, strong-interpretability and extensible fault detection.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance and fault diagnosis technology for power systems, specifically to a method and system for diagnosing power equipment faults based on an AI knowledge base. Background Technology

[0002] Currently, with the development of smart grids, a large number of sensors and monitoring devices have been deployed in power systems, enabling real-time collection of operational status data from power equipment (such as temperature, current, voltage, partial discharge, and oil chromatography analysis data). However, these data are typically high-dimensional, heterogeneous, and strongly coupled. Traditional fault diagnosis methods based on rules or shallow machine learning suffer from the following problems: 1. Insufficient knowledge utilization: Failure to effectively integrate domain expert knowledge (such as equipment structure, failure mechanism, causal logic); 2. Lack of causal relationship: Existing methods are mostly based on correlation analysis, ignoring the causal mechanism in fault propagation; 3. Weak knowledge graph modeling ability 4. Poor interpretability: The diagnostic results lack transparency, which is not conducive to the understanding and trust of maintenance personnel.

[0003] In recent years, the application of knowledge graphs (KG) in power systems has gradually increased. However, existing research mostly uses knowledge graphs as static prior knowledge bases, lacking dynamic updates and causal reasoning capabilities. At the same time, standard GNN models (such as GCN and GAT) have problems such as a single message passing mechanism and inability to capture system-level fault evolution patterns when processing power system data with hierarchical causal structures.

[0004] Therefore, there is an urgent need for a novel fault diagnosis method that can integrate dynamic knowledge graphs and hierarchical causal graph neural networks to achieve high-precision, highly interpretable, and scalable intelligent diagnosis. Summary of the Invention

[0005] In view of the above problems, embodiments of the present invention provide a power equipment fault diagnosis method and system based on an AI knowledge base, which is used to solve the problems of insufficient fault diagnosis accuracy and weak interpretability in the prior art.

[0006] According to one aspect of the present invention, a method for fault diagnosis of power equipment based on an AI knowledge base is provided, the method comprising: Obtain current equipment data for each power device in the power system; The current equipment data of each power device is mapped to entities and attributes defined in the basic knowledge graph to obtain the target entity, target attribute, target causal relationship and target correlation relationship between each target entity and target attribute corresponding to the current equipment data; the basic knowledge graph pre-stores the entity corresponding to each power device, the attribute corresponding to the current equipment data of the power device, and the causal relationship and correlation relationship between each entity and various attributes determined according to the correlation and causal relationship. Based on the target entities, target attributes, and the target causal and correlation relationships between the target entities and target attributes, a target fault knowledge graph is constructed. The target fault knowledge graph is input into a causal relationship network for reasoning to obtain equipment fault information. The hierarchical causal graph neural network includes an improved graph attention network. The improved graph attention network includes two layers: the first layer is a node-level message passing layer, and the second layer is a graph-level representation layer. The hierarchical causal graph neural network is trained based on training samples and corresponding fault causal chain labels and counterfactual labels. The fault causal chain labels include root cause entities, intermediate phenomena, observable symptoms, and fault types.

[0007] In one alternative approach, obtaining the current device data of each power device in the power system includes: collecting the current device data of each power device in real time from the monitoring system of the power system, wherein the current device data includes at least sensor data and device status quantities.

[0008] In one alternative approach, the underlying knowledge graph is constructed as follows: Collect historical data, including historical power equipment fault data, equipment specification data, and maintenance record data; Entities and attributes are defined based on the collected historical data, and causal and correlation relationships are established between entities; Use a graph database to store basic knowledge graphs.

[0009] In one alternative approach, the training process of the hierarchical causal graph neural network includes: Prepare training samples, each training sample including a fault knowledge graph instance; Each training sample is labeled with a fault causal chain label and a counterfactual label. The fault causal chain label indicates the root cause entity, intermediate phenomena, observable symptoms, and fault type. The counterfactual label indicates the fault changes in a counterfactual scenario. The model is trained using a graph neural network model, and the model parameters are optimized by minimizing a total loss function, which includes classification loss, fault causal chain label loss, contrastive loss, and counterfactual label-based loss.

[0010] In one alternative approach, the total loss function is formulated as follows: ; in, Weight of primary classification tasks; To maintain weights for causal structures; To compare the learning weights; For counterfactual reasoning weights; The strength of causal path regularization; Labels for actual fault types; The model predicts the probability distribution of failures. For entity indexes in the causal chain; The set of entities on the causal chain of the current sample; Score the importance of entity e in the causal chain; The probability of entity e in the causal chain predicted by the model; For causal side, This means that u causes v; Let be the set of all causal edges; This represents the margin of causal asymmetry. The embedded representation of nodes u and v; This represents the similarity calculation function; This is the graph-level representation of the original graph; This represents the positive sample image; This represents the negative sample image; N is the set of negative samples. For temperature parameters; For the cause node index, For the set of all cause nodes, This is a representation of the diagram after intervention.

[0011] In one alternative approach, training the graph neural network model and optimizing the model parameters by minimizing the total loss function includes: Self-supervised pre-training is performed through contrastive learning. The data augmentation process of contrastive learning includes node masking and edge dropping of the knowledge graph of the training samples, while retaining all causal relationship edges when dropping edges. Based on pre-training, multi-task supervised fine-tuning is performed using the fault causal chain labels and counterfactual labels, where the main task is fault type classification based on graph embedding, and the auxiliary task is prediction of causal relationship edges in the graph.

[0012] In one alternative approach, the first layer of the graph attention network is a node-level message passing layer, and its message passing mechanism is specifically implemented as follows: Calculate the attention coefficient between node i and its neighbor node j. Among them, attention coefficient The calculation formula is: ; in, and They are nodes and nodes The feature vectors are denoted by , W is a learnable weight matrix, and || is the vector concatenation operation. For connecting nodes i and j Relationship Type embedding vector, This is a learnable attention vector; Among them, when the relationship When the relationship is causal, its corresponding type embedding vector It is assigned an initial weight higher than that of the correlation.

[0013] In one alternative approach, the device fault information includes: a predicted fault type, an identified root cause entity, and a complete causal chain path connecting the root cause entity and the fault type.

[0014] In one alternative approach, mapping the current device data of each power device to entities and attributes already defined in the underlying knowledge graph specifically includes: Using a pre-trained natural language processing model, current device data in text form is mapped to corresponding entities and attributes in the underlying knowledge graph; and / or, Using a pre-trained computer vision model, current device data in image form is mapped to corresponding entities and attributes in the underlying knowledge graph.

[0015] According to another aspect of the present invention, a power equipment fault diagnosis system based on an AI knowledge base is provided, comprising: The acquisition module is used to acquire the current equipment data of each power device in the power system; The mapping module is used to map the current device data of each power device to the entities and attributes defined in the basic knowledge graph, so as to obtain the target entity, target attribute, target causal relationship and target correlation relationship between each target entity and target attribute corresponding to the current device data; the basic knowledge graph pre-stores the entity corresponding to each power device, the attribute corresponding to the current device data of the power device, and the causal relationship and correlation relationship between each entity and various attributes determined according to the correlation and causal relationship; The construction module is used to construct a target fault knowledge graph based on the target entities, target attributes, target causal relationships and target correlation relationships between the target entities and target attributes; The reasoning module is used to input the target fault knowledge graph into a causal relationship network for reasoning to obtain equipment fault information. The hierarchical causal graph neural network includes an improved graph attention network. This improved graph attention network comprises two layers: a first layer is a node-level message passing layer, and a second layer is a graph-level representation layer. The hierarchical causal graph neural network is trained using training samples and corresponding fault causal chain labels and counterfactual labels. The fault causal chain labels include root cause entities, observable symptoms of intermediate phenomena, and fault types.

[0016] This invention embodiment obtains the current equipment data of each power device in a power system; maps the current equipment data of each power device to entities and attributes defined in a basic knowledge graph, and obtains the target entities, target attributes, target causal relationships, and target correlation relationships between the target entities and target attributes corresponding to the current equipment data; the basic knowledge graph pre-stores the entities corresponding to each power device, the attributes corresponding to the current equipment data of the power devices, and the causal relationships and correlation relationships between each entity and various attributes determined according to correlation and causal relationships; based on the target entities, target attributes, and the relationships between the target entities and target attributes... By establishing causal and correlation relationships between targets, a target fault knowledge graph is constructed. This knowledge graph is then input into a causal relationship network for inference to obtain equipment fault information. The hierarchical causal graph neural network includes an improved graph attention network. This improved graph attention network comprises two layers: a first layer (node-level message passing layer) and a second layer (graph-level representation layer). The hierarchical causal graph neural network is trained using training samples and corresponding fault causal chain labels and counterfactual labels. The fault causal chain labels include root cause entities, intermediate phenomena (observable symptoms), and fault types. By explicitly modeling and inferring causal relationships, the root cause of the fault can be located, not just superficial phenomena, enabling high-precision, highly interpretable, and scalable intelligent diagnosis.

[0017] The above description is merely an overview of the technical solutions of the embodiments of the present invention. In order to better understand the technical means of the embodiments of the present invention and to implement them in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the embodiments of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0018] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart illustrating the power equipment fault diagnosis method based on an AI knowledge base provided in an embodiment of the present invention is shown. Figure 2 This illustration shows a schematic diagram of the hierarchical causal graph neural network structure in the power equipment fault diagnosis method based on an AI knowledge base provided in an embodiment of the present invention. Figure 3 A schematic diagram of the structure of the power equipment fault diagnosis system based on an AI knowledge base provided in an embodiment of the present invention is shown. Detailed Implementation

[0019] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein.

[0020] Figure 1 The flowchart illustrates a power equipment fault diagnosis method based on an AI knowledge base provided in an embodiment of the present invention. This method is executed by a power equipment fault diagnosis device based on an AI knowledge base. This device can be a computer device, distributed device, intelligent terminal device, etc., connected to the power equipment system; the present invention does not impose specific limitations. Figure 1 As shown, the method includes the following steps: Step 110: Obtain the current equipment data of each power device in the power system.

[0021] This involves real-time acquisition of current equipment data from the power system's monitoring system. This current equipment data includes at least sensor data and equipment status parameters. Specifically, in this embodiment, real-time acquisition of current equipment data from the power system's monitoring system is possible. The current equipment data includes sensor data (such as current, voltage, temperature, vibration frequency, etc.) and equipment status parameters (such as switch open / closed status, protection action signals, etc.). It may also include unstructured data such as infrared thermal images acquired through image acquisition devices and equipment operating sounds acquired through audio acquisition devices.

[0022] In this embodiment of the invention, the sensing data may specifically include: current, voltage, power factor, oil temperature, winding temperature, vibration frequency, etc. The sampling frequency can be 1kHz, and the data precision is 16 bits. Equipment status quantities may specifically include circuit breaker open / close status, protection device action signals, grounding switch status, etc., using status coding (0-normal, 1-warning, 2-fault). Image data may specifically include infrared thermal images with a resolution of 640×480 and a temperature sensitivity of 0.05°C; and visible light images with a resolution of 1920×1080. Audio data may have a sampling rate of 44.1kHz, 16-bit depth, and a sampling duration of 10 seconds.

[0023] Step 120: Map the current equipment data of each power device to the entities and attributes defined in the basic knowledge graph to obtain the target entities, target attributes, target causal relationships and target correlation relationships between the target entities and target attributes corresponding to the current equipment data.

[0024] The basic knowledge graph pre-stores entities corresponding to each power device, attributes corresponding to the current equipment data of the power device, and causal and correlation relationships between each entity and various attributes, determined based on correlation and causal relationships.

[0025] In this embodiment of the invention, the basic knowledge graph is constructed in the following manner: First, historical data is collected, including historical power equipment failure data, equipment specification data, and maintenance record data. Then, entities and attributes are defined based on the collected historical data, and causal and correlation relationships are established between entities; finally, a graph database is used to store the basic knowledge graph.

[0026] In a specific implementation of this invention, the following are examples of constructing basic knowledge graphs: For example, a causal relationship can be represented as: (Cooling fan failure) --[leading to]--> (poor heat dissipation); (Poor heat dissipation) --[leading to]--> (excessively high oil temperature); (Loose winding) --[leading to]--> (abnormal vibration).

[0027] The correlation can be represented as: (Oil temperature too high) --[is a phenomenon]--> (transformer failure); (Abnormal vibration) --[accompanied by]--> (abnormal noise).

[0028] Step 130: Construct a target fault knowledge graph based on the target entities, target attributes, target causal relationships and target correlation relationships between the target entities and target attributes.

[0029] Specifically, this embodiment maps the current device data of each power device to entities and attributes already defined in the basic knowledge graph, including: Using pre-trained natural language processing models, such as BERT, the current device data in text form is mapped to corresponding entities and attributes in the underlying knowledge graph; and / or, Using pre-trained computer vision models, such as CNNs, current device data in image form is mapped to corresponding entities and attributes in the underlying knowledge graph.

[0030] For example, when transformer A's oil temperature is detected to be 105°C and abnormal noise is present, the constructed target fault knowledge graph includes: Fault type: Transformer overheating fault (confidence level: 92%). Root cause entity: Cooling fan. Causal chain path: Cooling fan failure → leading to → poor heat dissipation → leading to → excessively high oil temperature → is the phenomenon of → transformer overheating fault. It can be seen that in this embodiment, the endpoint of the causal chain path is the fault type.

[0031] Step 140: Input the target fault knowledge graph into the causal relationship network for reasoning to obtain equipment fault information.

[0032] The hierarchical causal graph neural network includes an improved graph attention network. This improved graph attention network comprises two layers: a first layer is a node-level message passing layer, and a second layer is a graph-level representation layer. The hierarchical causal graph neural network is trained using training samples and corresponding fault causal chain labels and counterfactual labels. The fault causal chain labels include root cause entities, observable symptoms of intermediate phenomena, and fault types. In causal inference, counterfactual refers to a situation contrary to fact, i.e., "what would have happened if certain conditions had been different?" In machine learning, counterfactual samples are obtained by modifying the original data to simulate this "what if" situation. Counterfactual labels are derived from counterfactual samples. For example, how would the fault outcome change if a cause node in the fault knowledge graph were changed (e.g., changing "cooling fan fault" to "cooling fan normal")? Through such samples, the model can be trained to understand causal relationships, not just correlations. Counterfactual samples are constructed as follows: 1. Identify the nodes to intervene in: Select one or more nodes in the knowledge graph for intervention; these nodes can be causal nodes. 2. Perform the intervention: Change the state of the selected node to its counterfactual state. For example, if a node represents "cooling fan failure," change it to "cooling fan normal." Simultaneously, adjust the states of related nodes according to the causal logic. 3. Relabel: Based on the intervened knowledge graph, determine the new failure type and causal chain. By setting up counterfactual samples and their corresponding counterfactual labels, and training the model together with the original samples, the model can learn to distinguish between genuine causal relationships and spurious correlations.

[0033] like Figure 2 As shown, the hierarchical causal graph neural network of this embodiment includes an improved graph attention network, specifically comprising two layers: the first layer is a node-level message passing layer, which performs causal awareness message passing; the second layer is a graph-level representation layer, which generates graph-level representations through global pooling.

[0034] The first layer of the graph attention network is a node-level message passing layer, and its message passing mechanism is specifically implemented as follows: Calculate the attention coefficient between node i and its neighbor node j. Among them, attention coefficient The calculation formula is: ; in, and They are nodes i and nodesj The feature vectors are denoted by , W is a learnable weight matrix, and || is the vector concatenation operation. For connecting nodes i and j Relationship Type embedding vector, a This is a learnable attention vector; Among them, when the relationship When the relationship is causal, its corresponding type embedding vector It is assigned an initial weight higher than that of the correlation.

[0035] In this embodiment of the invention, the second-layer graph-level representation layer performs attention pooling, Set2Set pooling, and feature fusion. Attention pooling calculates an importance weight for each node, with important nodes receiving higher weights and less important nodes receiving lower weights. All node features are then summed according to their weights. In power equipment diagnostics, this means that if the graph contains two nodes, "oil temperature too high" and "slight vibration," the attention mechanism might assign a weight of 0.8 to "oil temperature too high" and a weight of 0.2 to "slight vibration," thus making the final graph representation more focused on the critical signal of temperature anomaly. After attention pooling, a multi-step Set2Set pooling mechanism captures richer structural information. This allows the second layer to retain structural information, remember the relationship patterns between nodes in the graph, ensure permutation invariance (generating the same graph representation regardless of node arrangement), and enrich information by capturing structural features at different levels through multi-step processing. Feature fusion concatenates the two pooling results along the feature dimension and maps the fused features to the target dimension.

[0036] In this embodiment of the invention, before inputting the target fault knowledge graph into the causal relationship network for reasoning, a hierarchical causal graph neural network is also trained. The specific training process includes: Prepare training samples, each training sample including a fault knowledge graph instance; Each training sample is labeled with a fault causal chain label and a counterfactual label. The fault causal chain label indicates the root cause entity, intermediate phenomena, observable symptoms, and fault type. The counterfactual label indicates the fault changes in a counterfactual scenario. The model is trained using a graph neural network model, and the model parameters are optimized by minimizing a total loss function, which includes classification loss, fault causal chain label loss, contrastive loss, and counterfactual label-based loss.

[0037] The formula for the total loss function is: ; in, Weight of primary classification tasks; To maintain weights for causal structures; To compare the learning weights; For counterfactual reasoning weights; The strength of causal path regularization; Labels for actual fault types; The model predicts the probability distribution of failures. For entity indexes in the causal chain; The set of entities on the causal chain of the current sample; Score the importance of entity e in the causal chain; The probability of entity e in the causal chain predicted by the model; For causal side, This means that u causes v; Let be the set of all causal edges; This represents the margin of causal asymmetry. The embedded representation of nodes u and v; This represents the similarity calculation function; This is the graph-level representation of the original graph; This represents the positive sample image; This represents the negative sample image; N is the set of negative samples. For temperature parameters; For the cause node index, For the set of all cause nodes, This is a representation of the diagram after intervention.

[0038] The total loss function combines fault classification loss and causal relationship prediction loss, and introduces counterfactual comparison loss. By using counterfactual samples, the model's causal reasoning ability is enhanced. This not only optimizes diagnostic accuracy, but more importantly, forces the model to learn the real causal mechanism behind power equipment faults. This enables the model to perform reliable causal reasoning even when data is scarce and fault modes are complex. In this way, the accuracy of the model is effectively improved, which is something that traditional loss functions cannot achieve.

[0039] In this embodiment of the invention, the training of the graph neural network model and the optimization of model parameters by minimizing the total loss function include: self-supervised pre-training through contrastive learning, wherein the data augmentation process of contrastive learning includes node masking and edge dropping of the knowledge graph of the training samples, and retaining all causal relationship edges when dropping edges; based on the pre-training, multi-task supervised fine-tuning is performed using the fault causal chain label and counterfactual label, wherein the main task is fault type classification based on graph embedding, and the auxiliary task is prediction of causal relationship edges in the graph.

[0040] The equipment fault information includes: the predicted fault type, the identified root cause entity, and the complete causal chain path connecting the root cause entity and the fault type.

[0041] This invention embodiment obtains the current equipment data of each power device in a power system; maps the current equipment data of each power device to entities and attributes defined in a basic knowledge graph, and obtains the target entities, target attributes, target causal relationships, and target correlation relationships between the target entities and target attributes corresponding to the current equipment data; the basic knowledge graph pre-stores the entities corresponding to each power device, the attributes corresponding to the current equipment data of the power devices, and the causal relationships and correlation relationships between each entity and various attributes determined according to correlation and causal relationships; based on the target entities, target attributes, and the relationships between the target entities and target attributes... By establishing causal and correlation relationships between targets, a target fault knowledge graph is constructed. This knowledge graph is then input into a causal relationship network for inference to obtain equipment fault information. The hierarchical causal graph neural network includes an improved graph attention network. This improved graph attention network comprises two layers: a first layer (node-level message passing layer) and a second layer (graph-level representation layer). The hierarchical causal graph neural network is trained using training samples and corresponding fault causal chain labels and counterfactual labels. The fault causal chain labels include root cause entities, intermediate phenomena (observable symptoms), and fault types. By explicitly modeling and inferring causal relationships, the root cause of the fault can be located, not just superficial phenomena, enabling high-precision, highly interpretable, and scalable intelligent diagnosis.

[0042] Figure 3 A schematic diagram of the structure of a power equipment fault diagnosis system based on an AI knowledge base provided in an embodiment of the present invention is shown. Figure 3 As shown, the device 200 includes: The acquisition module 210 is used to acquire the current equipment data of each power device in the power system; The mapping module 220 is used to map the current device data of each power device to entities and attributes defined in the basic knowledge graph, so as to obtain the target entity, target attribute, target causal relationship and target correlation relationship between each target entity and target attribute corresponding to the current device data; the basic knowledge graph pre-stores the entity corresponding to each power device, the attribute corresponding to the current device data of the power device, and the causal relationship and correlation relationship between each entity and various attributes determined according to the correlation and causal relationship. Construction module 230 is used to construct a target fault knowledge graph based on the target entity, target attribute, target causal relationship and target correlation relationship between each target entity and target attribute; The reasoning module 240 is used to input the target fault knowledge graph into a causal relationship network for reasoning to obtain equipment fault information. The hierarchical causal graph neural network includes an improved graph attention network. The improved graph attention network comprises two layers: a first layer is a node-level message passing layer, and a second layer is a graph-level representation layer. The hierarchical causal graph neural network is trained using training samples and corresponding fault causal chain labels and counterfactual labels. The fault causal chain labels include root cause entities, observable symptoms of intermediate phenomena, and fault types.

[0043] This invention embodiment obtains the current equipment data of each power device in a power system; maps the current equipment data of each power device to entities and attributes defined in a basic knowledge graph, and obtains the target entities, target attributes, target causal relationships, and target correlation relationships between the target entities and target attributes corresponding to the current equipment data; the basic knowledge graph pre-stores the entities corresponding to each power device, the attributes corresponding to the current equipment data of the power devices, and the causal relationships and correlation relationships between each entity and various attributes determined according to correlation and causal relationships; based on the target entities, target attributes, and the relationships between the target entities and target attributes... By establishing causal and correlation relationships between targets, a target fault knowledge graph is constructed. This knowledge graph is then input into a causal relationship network for inference to obtain equipment fault information. The hierarchical causal graph neural network includes an improved graph attention network. This improved graph attention network comprises two layers: a first layer (node-level message passing layer) and a second layer (graph-level representation layer). The hierarchical causal graph neural network is trained using training samples and corresponding fault causal chain labels and counterfactual labels. The fault causal chain labels include root cause entities, intermediate phenomena (observable symptoms), and fault types. By explicitly modeling and inferring causal relationships, the root cause of the fault can be located, not just superficial phenomena, enabling high-precision, highly interpretable, and scalable intelligent diagnosis.

[0044] The algorithms or displays provided herein are not inherently related to any particular computer, virtual system, or other device. Various general-purpose systems can also be used in conjunction with the teachings herein. The required structure for constructing such systems is apparent from the above description. Furthermore, the embodiments of the present invention are not directed to any particular programming language. It should be understood that the content of the invention described herein can be implemented using various programming languages, and the above description of specific languages ​​is for the purpose of disclosing the best mode of implementation of the invention.

[0045] Numerous specific details are set forth in the specification provided herein. However, it will be understood that embodiments of the invention may be practiced without these specific details. In some instances, well-known methods, structures, and techniques have not been shown in detail so as not to obscure the understanding of this specification.

[0046] Similarly, it should be understood that, in order to streamline the invention and aid in understanding one or more of the various aspects of the invention, features of the embodiments of the invention are sometimes grouped together in a single embodiment, figure, or description thereof in the above description of exemplary embodiments of the invention. However, this disclosure should not be construed as reflecting an intention that the claimed invention requires more features than are expressly recited in each claim.

[0047] Those skilled in the art will understand that modules in the device of the embodiments can be adaptively changed and placed in one or more devices different from that embodiment. Modules, units, or components in the embodiments can be combined into a single module, unit, or component, and can be divided into multiple sub-modules, sub-units, or sub-components. Except where at least some of such features and / or processes or units are mutually exclusive, any combination can be used to combine all features disclosed in this specification (including the accompanying claims, abstract, and drawings) and all processes or units of any method or device so disclosed. Unless expressly stated otherwise, each feature disclosed in this specification (including the accompanying claims, abstract, and drawings) may be replaced by an alternative feature that serves the same, equivalent, or similar purpose.

[0048] It should be noted that the above embodiments are illustrative of the invention and not restrictive, and that those skilled in the art can devise alternative embodiments without departing from the scope of the appended claims. In the claims, any reference signs placed between parentheses should not be construed as limiting the claims. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The invention can be implemented by means of hardware comprising several different elements and by means of a suitably programmed computer. In the unit claims enumerating several means, several of these means may be embodied by the same item of hardware. The use of the words first, second, and third, etc., does not indicate any order. These words can be interpreted as names. The steps in the above embodiments, unless otherwise specified, should not be construed as limiting the order of execution.

Claims

1. An AI knowledge base-based power equipment fault diagnosis method, characterized in that, The method comprises: acquiring current device data of each power device in a power system; mapping the current device data of each power device to entities and attributes defined in a basic knowledge graph to obtain target entities, target attributes, target causal relationships and target correlation relationships between each of the target entities and the target attributes; the basic knowledge graph pre-stores entities corresponding to each power device, attributes corresponding to the current device data of the power device, and causal relationships and correlation relationships between each of the entities and various attributes determined according to the correlation relationships and the causal relationships; constructing a target fault knowledge graph according to the target entities, the target attributes, the target causal relationships and the target correlation relationships between each of the target entities and the target attributes; inputting the target fault knowledge graph into a causal relationship network for reasoning to obtain device fault information, wherein the hierarchical causal graph neural network comprises an improved graph attention network; the improved graph attention network comprises two layers of graph attention networks, the first layer of graph attention networks is a node-level message passing layer, and the second layer of graph attention networks is a graph table-level representation layer; the hierarchical causal graph neural network is obtained by training the graph neural network according to training samples and corresponding fault causal chain labels and counterfactual labels; the fault causal chain label comprises a root cause entity, an intermediate phenomenon observable symptom and a fault type.

2. The method of claim 1, wherein, The acquisition of the current device data of each power device in the power system comprises: real-time collection of the current device data of each power device from a monitoring system of the power system, wherein the current device data at least comprises sensing data and device state quantity.

3. The method of claim 1, wherein, The basic knowledge graph is constructed by: collecting historical data, wherein the historical data comprises historical power device fault data, device specification data and maintenance record data; defining entities and attributes based on the collected historical data, and establishing causal relationships and correlation relationships between the entities; storing the basic knowledge graph using a graph database.

4. The method of claim 1, wherein, The training process of the hierarchical causal graph neural network comprises: preparing training samples, each training sample comprising a fault knowledge graph instance; annotating fault causal chain labels and counterfactual labels for each training sample, wherein the fault causal chain label indicates a root cause entity, an intermediate phenomenon observable symptom and a fault type, and the counterfactual label indicates a fault change in a counterfactual scenario; training using a graph neural network model, optimizing model parameters by minimizing a total loss function, wherein the total loss function comprises a classification loss, a loss of the fault causal chain label, a contrastive loss and a loss based on the counterfactual label.

5. The method of claim 4, wherein, The total loss function formula is: ; wherein, is a main classification task weight; is a causal structure preservation weight; is a contrastive learning weight; is a counterfactual reasoning weight; is a causal path regularization strength; is a true fault type label; is a model predicted fault probability distribution; is an entity index in a causal chain; is a set of entities on the causal chain for the current sample; is an importance score of entity e in the causal chain; is a model predicted probability of entity e in the causal chain; is a causal edge, denotes that u causes v; is a set of all causal edges; is a causal asymmetry margin; denotes an embedding representation of nodes u and v; denotes a similarity computation function; is a graph-level representation of the original graph; is a representation of the positive sample graph; is a representation of the negative sample graph; N is a set of negative samples, is a temperature parameter; is a cause node index, is a set of all cause nodes, is a representation of the post-intervention graph.

6. The method of claim 4, wherein, The training using the graph neural network model, optimizing the model parameters by minimizing the total loss function, comprises: self-supervised pre-training by a contrastive learning method, wherein the data enhancement process of the contrastive learning comprises node masking and edge dropping of the training sample knowledge graph, and all causal relationship edges are retained when the edges are dropped; On the basis of pre-training, the failure causal chain label and counterfactual label are used for multi-task supervised fine-tuning, wherein a main task is graph embedding-based failure type classification, and an auxiliary task is prediction of causal relationship edges in a graph.

7. The method according to any one of claims 1 to 6, characterized in that, The first layer graph attention network is a node-level message passing layer, and a specific implementation of a message passing mechanism thereof is: The attention coefficient between the computing node i and the neighbor node j Wherein, the calculation formula of the attention coefficient is: ； wherein, and are feature vectors of nodes and nodes respectively, W is a learnable weight matrix, is a vector concatenation operation, is a type embedding vector of the relation connecting nodes and , is a learnable attention vector; wherein, when the relationship is a causal relationship, its corresponding type embedding vector is assigned a higher initial weight than the initial weight for a correlation relationship.

8. The method according to any one of claims 1 to 6, characterized in that, The device failure information includes a predicted failure type, an identified root cause entity, and a complete causal chain path connecting the root cause entity and the failure type.

9. The method according to any one of claims 1 to 6, characterized in that, The mapping of the current device data of each power device to the defined entities and attributes in the base knowledge graph specifically includes: The current device data in a text form is mapped to corresponding entities and attributes in the base knowledge graph by using a pre-trained natural language processing model; and / or The current device data in an image form is mapped to corresponding entities and attributes in the base knowledge graph by using a pre-trained computer vision model.

10. An AI knowledge base-based power equipment fault diagnosis system, characterized in that, The system comprises: An acquisition module configured to acquire current device data of each power device in a power system; A mapping module configured to map the current device data of each power device to defined entities and attributes in a base knowledge graph to obtain target entities, target attributes, target causal relationships and target correlation relationships between each of the target entities and the target attributes corresponding to the current device data; the base knowledge graph pre-stores entities corresponding to each power device, attributes corresponding to current device data of the power device, and causal relationships and correlation relationships between each of the entities and various attributes determined according to the correlation relationships and the causal relationships; A construction module configured to construct a target failure knowledge graph according to the target entities, the target attributes, the target causal relationships and the target correlation relationships between each of the target entities and the target attributes; An inference module configured to input the target failure knowledge graph into a causal relationship network for inference to obtain device failure information, wherein the hierarchical causal graph neural network comprises an improved graph attention network; the improved graph attention network comprises two layers of graph attention networks, a first layer of graph attention networks is a node-level message passing layer, and a second layer of graph attention networks is a graph table-level representation layer; the hierarchical causal graph neural network is obtained by training a graph neural network according to training samples and corresponding failure causal chain labels and counterfactual labels; the failure causal chain label includes a root cause entity, an intermediate phenomenon observable symptom and a failure type.

Images