Fault detection method and device for secondary equipment of transformer substation

Abstract

The invention discloses a fault detection method and device for secondary equipment of a transformer substation. The method comprises the following steps: acquiring real-time operation parameters of the secondary equipment of the transformer substation, inputting the real-time operation parameters into a pre-trained graph neural network model to obtain equipment operation characteristics, and determining an equipment state vector according to a plurality of equipment operation characteristics in a continuous time window and a message passing mechanism; determining an equipment health degree score according to the equipment state vector and a knowledge graph, and triggering the intelligent inspection robot to execute a deep detection task under the condition that the health degree is lower than a preset health degree threshold value so as to collect target test data; determining a target fault type according to the target test data, the knowledge graph and the inference rule base, obtaining a fault processing scheme corresponding to a historical fault case, and generating corresponding maintenance guide information according to the fault processing scheme, thereby achieving the real-time detection, precise fault diagnosis and automatic maintenance of the secondary equipment of the transformer substation. And the operation safety and the maintenance efficiency of the power system are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of substation operation management, and in particular to a fault detection method and device for substation secondary equipment. Background Art

[0002] In the power system, health management and fault diagnosis of substation secondary equipment are key links to ensure the safe and stable operation of the power grid.

[0003] In the related technologies, fault diagnosis methods mainly rely on manual experience and regular maintenance, which has the problems of low efficiency, long response time and high misjudgment rate. And with the continuous expansion of the scale and increase of complexity of power systems, related fault diagnosis methods have been unable to meet the requirements of modern power systems for high reliability and high efficiency.

[0004] Therefore, an intelligent fault diagnosis and health management solution is needed that can detect equipment status in real time, warn of potential faults in advance, and provide scientific fault handling suggestions. Summary of the invention

[0005] The present invention provides a fault detection method and device for substation secondary equipment to solve the problems that substation secondary equipment fault diagnosis relies on manual experience, has a long response time, a high misjudgment rate and low maintenance efficiency.

[0006] According to one aspect of the present invention, a method for detecting a fault of a secondary device in a substation is provided, the method comprising:

[0007] Collecting real-time operating parameters of substation secondary equipment, inputting the real-time operating parameters of the substation secondary equipment into a pre-trained graph neural network model to obtain equipment operating characteristics of the substation secondary equipment, and determining the equipment state vector according to a plurality of equipment operating characteristics in a continuous time window and a message transmission mechanism;

[0008] Determine the equipment health score according to the equipment state vector and the pre-built knowledge graph, and trigger the intelligent inspection robot to perform a deep detection task to collect target test data when the health score is lower than a preset health threshold, wherein the knowledge graph is built based on the historical test data, operating parameter data and fault maintenance records of the substation secondary equipment;

[0009] Determine the target fault type corresponding to the substation secondary equipment according to the target test data, the knowledge graph and the pre-established inference rule base, obtain the fault handling plan corresponding to the historical fault case corresponding to the target fault type, and generate maintenance guidance information corresponding to the substation secondary equipment according to the fault handling plan.

[0010] According to another aspect of the present invention, there is provided a fault detection device for a secondary device of a substation, the device comprising:

[0011] An equipment information acquisition module is used to collect real-time operating parameters of substation secondary equipment, input the real-time operating parameters of the substation secondary equipment into a pre-trained graph neural network model to obtain equipment operating characteristics of the substation secondary equipment, and determine the equipment state vector according to multiple equipment operating characteristics in a continuous time window and a message transmission mechanism;

[0012] A test data acquisition module, used to determine the equipment health score according to the equipment state vector and a pre-built knowledge graph, and trigger the intelligent inspection robot to perform a deep detection task to collect target test data when the health score is lower than a preset health threshold, wherein the knowledge graph is constructed based on historical test data, operating parameter data and fault maintenance records of the substation secondary equipment;

[0013] An equipment maintenance guidance module is used to determine the target fault type corresponding to the substation secondary equipment based on the target test data, the knowledge graph and a pre-established inference rule base, obtain a fault handling plan corresponding to the historical fault case corresponding to the target fault type, and generate maintenance guidance information corresponding to the substation secondary equipment based on the fault handling plan.

[0014] According to another aspect of the present invention, an electronic device is provided, the electronic device comprising:

[0015] at least one processor; and

[0016] a memory communicatively connected to the at least one processor; wherein,

[0017] The memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can execute the fault detection method for substation secondary equipment described in any embodiment of the present invention.

[0018] According to another aspect of the present invention, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a processor to implement the fault detection method for substation secondary equipment described in any embodiment of the present invention when executed.

[0019] The technical solution of the embodiment of the present invention is as follows: first, by collecting the real-time operating parameters of the secondary equipment of the substation, the real-time operating parameters of the secondary equipment of the substation are input into a pre-trained graph neural network model to obtain the equipment operating characteristics of the secondary equipment of the substation, and the equipment state vector is determined according to the multiple equipment operating characteristics and the message transmission mechanism in the continuous time window, so as to accurately detect and evaluate the equipment status and give early warning of potential faults; then, the equipment health score is determined according to the equipment state vector and the pre-constructed knowledge graph, and when the health score is lower than the preset health threshold, the intelligent inspection robot is triggered to perform a deep detection task to collect target test data. Since the knowledge graph is based on the historical test data and operating parameters of the secondary equipment of the substation The data and fault maintenance records can be constructed to intelligently evaluate the health of equipment, start in-depth detection in time, and improve maintenance efficiency; the target fault type corresponding to the substation secondary equipment is determined according to the target test data, the knowledge graph and the pre-established inference rule library, and the fault handling plan corresponding to the historical fault case corresponding to the target fault type is obtained. According to the fault handling plan, maintenance guidance information corresponding to the substation secondary equipment is generated, which can quickly locate the fault and provide accurate maintenance guidance, solving the problems of substation secondary equipment fault diagnosis relying on manual experience, long response time, high misjudgment rate and low maintenance efficiency in related technologies, realizing real-time detection, accurate fault diagnosis and automated maintenance of substation secondary equipment, and improving the safety and maintenance efficiency of power system operation.

[0020] It should be understood that the contents described in this section are not intended to identify the key or important features of the embodiments of the present invention, nor are they intended to limit the scope of the present invention. Other features of the present invention will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0022] Figure 1 is a flow chart of a fault detection method for a substation secondary device provided according to the first embodiment of the present invention;

[0023] Figure 2 is a flow chart of a fault detection method for a substation secondary device provided according to the first embodiment of the present invention;

[0024] Figure 3is a structural schematic diagram of a fault detection device for secondary equipment of a substation provided according to a third embodiment of the present invention;

[0025] Figure 4 It is a structural schematic diagram of an electronic device for implementing the fault detection method for secondary equipment of a substation according to an embodiment of the present invention. DETAILED DESCRIPTION

[0026] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0027] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0028] It should be noted that the modifications of "one" and "plurality" mentioned in the present disclosure are illustrative rather than restrictive, and those skilled in the art should understand that unless otherwise clearly indicated in the context, it should be understood as "one or more".

[0029] The names of the messages or information exchanged between multiple devices in the embodiments of the present disclosure are only used for illustrative purposes and are not used to limit the scope of these messages or information.

[0030] It is understandable that before using the technical solutions disclosed in the embodiments of the present disclosure, the types, scope of use, usage scenarios, etc. of the personal information involved in the present disclosure should be informed to the user and the user's authorization should be obtained in an appropriate manner in accordance with relevant laws and regulations.

[0031] For example, in response to receiving an active request from a user, a prompt message is sent to the user to clearly prompt the user that the operation requested to be performed will require obtaining and using the user's personal information. Thus, the user can autonomously choose whether to provide personal information to software or hardware such as an electronic device, application, server, or storage medium that performs the operation of the technical solution of the present disclosure according to the prompt message.

[0032] As an optional but non-limiting implementation, in response to receiving an active request from the user, the prompt information may be sent to the user in the form of a pop-up window, in which the prompt information may be presented in text form. In addition, the pop-up window may also carry a selection control for the user to choose "agree" or "disagree" to provide personal information to the electronic device.

[0033] It is understandable that the above notification and the process of obtaining user authorization are merely illustrative and do not constitute a limitation on the implementation of the present disclosure. Other methods that meet the relevant laws and regulations may also be applied to the implementation of the present disclosure.

[0034] It is understandable that the data involved in this technical solution (including but not limited to the data itself, the acquisition or use of the data) shall comply with the requirements of relevant laws, regulations and relevant provisions.

[0035] Embodiment 1

[0036] Figure 1 A flowchart of a method for fault detection of secondary equipment in a substation is provided for the first embodiment of the present invention. This embodiment can be applied to the situation of real-time detection and intelligent diagnosis of faults in secondary equipment in a substation. The method can be executed by a fault detection device for secondary equipment in a substation. The fault detection device for secondary equipment in a substation can be implemented in the form of hardware and / or software, and optionally, can be implemented by an electronic device, which can be a mobile terminal, a PC or a server, etc.

[0037] like Figure 1 As shown, the method may specifically include:

[0038] S110. Collect real-time operating parameters of substation secondary equipment, input the real-time operating parameters of the substation secondary equipment into a pre-trained graph neural network model to obtain equipment operating characteristics of the substation secondary equipment, and determine the equipment state vector based on multiple equipment operating characteristics in a continuous time window and a message passing mechanism.

[0039] Among them, the substation secondary equipment can be understood as equipment that assists the primary equipment in working and is used for functions such as control, protection, measurement and communication. The substation secondary equipment may include but is not limited to at least one of relays, protection devices, measurement and control devices and automatic control systems. The primary equipment may be the main electrical equipment directly involved in the production, transmission and distribution of electric energy, such as transformers or circuit breakers. The real-time operating parameters can be understood as data used to reflect the current working status of the substation secondary equipment. The real-time operating parameters may include but are not limited to at least one of current, voltage, frequency and temperature. The real-time operating parameters can be obtained by collecting substation secondary equipment through distributed sensors. Exemplarily, intelligent sensors are installed on protection devices, measurement and control devices and other equipment to collect operating status quantities such as voltage, current or temperature in real time. The sampling frequency can be set according to the importance of the real-time operating parameters, such as sampling voltage and current per second, and sampling temperature once per minute. The graph neural network model (Graph Neural Network, GNN) can be understood as a deep learning model that processes graph-structured data, which can learn the topological relationship and parameter correlation between secondary devices and can be used to analyze the connection between various secondary devices in the substation. The device operation characteristics can be understood as the device behavior pattern or attribute obtained by analyzing the real-time operation parameters. The device operation characteristics may include but are not limited to at least one of power factor, load rate and temperature fluctuation range. The message transmission mechanism can be understood as a mechanism that enables each node to aggregate the information of its neighboring nodes and learn in combination with its own characteristics by propagating and updating node information in the graph. The device state vector can be understood as a vector determined according to multiple device operation characteristics in a continuous time window, which is used to describe the current state of the device and can be obtained by comprehensive analysis of data at multiple time points.

[0040] On the basis of the above scheme, optionally, the graph neural network model includes a plurality of graph convolution modules connected in series, each of the graph convolution modules includes a spatial convolution unit, a feature transformation unit, a batch normalization layer and an activation function layer; the real-time operation parameters of the substation secondary equipment are input into the pre-trained graph neural network model to obtain the equipment operation characteristics of the substation secondary equipment, including: for the first graph convolution module in the graph neural network model, the real-time operation parameters are converted into initial operation characteristics through the spatial convolution unit of the graph convolution module, and the substation secondary equipment is taken as the central node, and the substation secondary equipment directly connected to the central node is taken as the neighboring node, and the neighborhood aggregation function based on the attention weight is used to dynamically adjust the initial node according to the relationship type corresponding to the node-to-node connection relationship between the central node and the neighboring nodes. The feature aggregation weight corresponding to the initial operation feature is adjusted, and the operation aggregation feature corresponding to the real-time operation parameter is determined according to the adjusted aggregation weight; and the initial operation feature of the central node is weightedly combined with the operation aggregation feature through the residual connection mechanism of the graph convolution module in the feature transformation unit to obtain the operation fusion feature; the operation fusion feature is processed into the module output feature of the graph convolution module through the batch normalization layer and the activation function layer in sequence; for the subsequent graph convolution module of the graph neural network model, the module output feature of the previous graph convolution module connected to the graph convolution module is obtained, and the module output feature is input into the graph convolution module to obtain the module output feature of the graph convolution module, and the module output feature of the last graph convolution module is determined as the equipment operation feature of the substation secondary equipment.

[0041] Among them, the graph convolution module can be understood as the basic building block in the graph neural network, which is responsible for processing the node information in the graph data. Each graph convolution module may include multiple subcomponents, such as at least one of a spatial convolution unit, a feature transformation unit, a batch normalization layer and an activation function layer. The spatial convolution unit can be understood as a convolution layer that updates the feature representation of the current node by aggregating the information of adjacent nodes, and is used to convert the input real-time operation parameters into initial operation features. The feature transformation unit can be understood as a unit responsible for transforming node features, which can include transformation methods such as linear transformation (such as matrix multiplication) and nonlinear transformation (such as activation function). The feature transformation unit combines the initial operation features with the operation aggregation features by introducing a residual connection mechanism to enhance the learning ability of the model. The batch normalization layer can standardize the features of the data to accelerate the model training process and reduce the problem of internal covariate shift. The activation function layer can increase the expression ability of the model by introducing nonlinear factors. The activation function may include but is not limited to at least one of PReLU, Sigmoid and Tanh. The initial operation feature can be understood as the node feature representation obtained after being processed by the spatial convolution unit of the first graph convolution module. The attention weight can be understood as the different weights that the attention mechanism allows the graph neural network model to dynamically assign according to the importance differences between nodes in the graph neural network, which helps the graph neural network model to better focus on information that is more meaningful to the current task. The neighborhood aggregation function can be understood as the part of the graph convolution operation responsible for collecting and updating the feature representation of the current node. The attention weight is used to dynamically adjust the weights contributed by different neighbor nodes. The running aggregation feature can be understood as the feature representation obtained by combining the attention weight of the relationship type between the central node and its neighbor nodes based on the neighborhood aggregation function. The running fusion feature can be understood as the feature representation of the initial running feature of the central node and the running aggregation feature weighted by the residual connection mechanism in the feature transformation unit. The residual connection mechanism can be understood as a technology for alleviating the problem of gradient disappearance or explosion during deep network training, which can keep the information flow smooth by directly adding input features to output features. The module output feature can be understood as the final node feature representation output after all processing steps of a graph convolution module (including spatial convolution, feature transformation, batch normalization, activation function processing, etc.).

[0042] In an optional implementation, multi-layer graph convolution operations are performed on device nodes to extract features. Each layer of graph convolution includes two steps: spatial convolution and feature transformation. In the spatial convolution step, the device node currently performing feature extraction is selected as the central node, and the device nodes directly connected to the central node are selected as neighbor nodes. Then, the neighborhood aggregation function based on attention weight is used to fuse the information of the neighbor nodes. The aggregation function dynamically adjusts the weight of feature aggregation according to the type of connection relationship between the central node and each neighbor node. For example, a higher weight is given to the neighbor nodes in the "upstream and downstream" relationship, while a lower weight is given to the neighbor nodes in the "parallel" relationship. In the feature transformation step, the original features of the central node are weightedly combined with the aggregated features through the residual connection mechanism. Residual connection can effectively alleviate the gradient vanishing problem in deep graph neural networks. Batch normalization layers and PReLU activation functions are added after each layer of graph convolution to accelerate network convergence and enhance nonlinear expression capabilities. Through multiple layers of such graph convolution operations, the receptive field of each device node can be gradually expanded to capture topological structure information in a larger range.

[0043] In the above scheme, firstly, by combining the spatial convolution unit and the neighborhood aggregation function based on attention weight, the graph neural network model can effectively capture the complex interaction relationship between devices, and then dynamically adjust the feature aggregation weight to make the feature representation more accurate; secondly, the residual connection mechanism in the feature transformation unit helps to alleviate the gradient vanishing problem in the deep network and ensure the stability of model training; then, the application of batch normalization layer and activation function layer further optimizes the feature expression and enhances the nonlinear modeling ability of the model; finally, through the layer-by-layer abstraction of the multi-layer graph convolution module, highly representative equipment operation characteristics can be extracted from the complex operation parameters, which is conducive to the realization of real-time detection of equipment status, fault warning and maintenance decision support, which is of great significance.

[0044] On the basis of the above scheme, optionally, the device state vector is determined according to the multiple device operation characteristics and message transmission mechanism within the continuous time window, including: taking the substation secondary device that sends information as the source node, and the substation secondary device that receives information as the target node, converting the device operation characteristics of the source node into an initial message through a message generation function; using a gating unit determined by the device characteristics of the source node and the target node to filter the initial message to obtain multiple target messages, and performing weighted summation on the multiple target messages to obtain the target device characteristics of the target node; inputting the multiple target device characteristics of the substation secondary device within the continuous time window into the temporal attention layer, so as to weightedly splice the multiple target device characteristics through a multi-head attention mechanism and generate a device state vector after linear transformation.

[0045] Among them, the source node can be understood as a substation secondary device that sends information during the message transmission process, and is responsible for transmitting its own information (such as equipment operation characteristics) to other nodes. The target node can be understood as a substation secondary device that receives information from the source node, and the target node will update its own feature representation according to the received message. The message generation function can be understood as a function for converting the equipment operation characteristics of the source node into an initial message that can be transmitted to the target node, and the message generation function is associated with the relationship type corresponding to the node connection relationship between the source node and the target node. The gating unit can be understood as a unit for controlling the flow of information during the message transmission process. The gating unit can determine which information should be retained and which should be discarded according to the characteristics of the source node and the target node, so as to filter the initial message and generate the target message. After the target message is filtered by the gating unit, the message transmitted from the source node to the target node can reflect the influence of the source node on the target node. The target device characteristics can be understood as the updated device characteristics of the target node by integrating all the target messages received, which can be determined by weighted summing of all the target messages. The temporal attention layer can be understood as a convolutional layer that uses an attention mechanism to weight the target device features at different time points when processing data in a continuous time window. The multi-head attention mechanism can be understood as being able to capture the dependencies between temporal features from multiple angles. By applying the attention mechanism in parallel in different subspaces to weightedly splice multiple target device features, the model can focus on more information dimensions at the same time, thereby improving the model's expressiveness.

[0046] The above scheme, firstly, can dynamically adjust the information transmission process through message generation functions and gating units to ensure that only highly relevant information is retained, thereby improving the accuracy and relevance of features; secondly, by utilizing the multi-head attention mechanism, the model can simultaneously focus on the device features at different time points, enhancing the ability to understand time series data; finally, the device state vector generated by linear transformation not only integrates information from multiple time points, but also highlights key features, providing more accurate data support for device health assessment and fault prediction, improving the accuracy and reliability of device status detection, and helping to achieve early fault warning and preventive maintenance, thereby ensuring the stable operation of the power system.

[0047] S120. Determine the equipment health score based on the equipment state vector and a pre-constructed knowledge graph. When the health score is lower than a preset health threshold, trigger the intelligent inspection robot to perform a deep inspection task to collect target test data, wherein the knowledge graph is constructed based on historical test data, operating parameter data and fault maintenance records of the substation secondary equipment.

[0048] Among them, the knowledge graph can be understood as a structured knowledge representation method that uses a graph to express entities and their relationships. The knowledge graph integrates the historical test data, operating parameter data and fault maintenance records of the substation secondary equipment, which can provide a basis for the secondary equipment status evaluation. The historical test data can be understood as a data set obtained by performing various tests on the substation secondary equipment in a certain period of time in the past, which can be used to identify the change trend of the secondary equipment over time. The historical test data may include but is not limited to at least one of the functional test results and performance indicators of the equipment. The operating parameter data can be understood as various parameter values ​​generated during the operation of the equipment, such as current, voltage, temperature and humidity, which are used to reflect the performance of the equipment under actual working conditions. The operating parameter data is the basis for detecting the operating status of the equipment and diagnosing equipment failures. By continuously collecting and analyzing the operating parameter data, abnormal conditions in the operation of the secondary equipment can be discovered in time, and preventive measures can be taken to avoid failures. The fault maintenance record can be understood as recording the fault conditions that have occurred in the history of secondary equipment and the maintenance measures and results taken for these problems. It can include at least one or more information such as the time, location, cause, maintenance method and effect evaluation after the fault occurs. It can be used to quickly locate the current problem and also as part of building a knowledge graph to provide experience and guidance for future fault prevention and handling. The health score can be understood as a score calculated based on the equipment state vector to quantify the overall health of the equipment. The score range can be 0-100, with 100 representing the healthiest state. The health score can be used as a criterion for determining whether further inspection or maintenance is required. The health threshold can be understood as a criterion for determining whether the health meets the health status setting. The intelligent inspection robot can be understood as a robot with autonomous navigation capabilities, which can inspect the inside of the substation without human intervention and collect target test data. The robot is usually equipped with a variety of sensors, such as cameras, infrared imagers and other electronic devices. The target test data can be understood as data collected by the intelligent inspection robot when performing deep detection tasks, especially for equipment with low health scores, for more accurate diagnosis of problems.

[0049] On the basis of the above scheme, optionally, determining the equipment health score based on the equipment state vector and a pre-constructed knowledge graph includes: determining the state category corresponding to the equipment state vector based on the equipment type and operating conditions of the substation secondary equipment, extracting the historical state sequence corresponding to the state category from the pre-constructed knowledge graph, and using a dynamic time normalization algorithm to calculate the dynamic time normalization distance between the current equipment state vector and the historical state sequence of the corresponding category; converting the dynamic time normalization distance into a similarity score through a soft maximization function, and performing weighted summation on the similarity scores based on the timeliness and completeness of the historical state sequence to obtain the equipment health score.

[0050] Among them, the equipment type can be understood as a specific type of secondary equipment in a substation, such as a relay protection device, an automatic control system, etc. It is understandable that different equipment types may have different operating characteristics and failure modes. The operating condition can be understood as the working environment and conditions of the secondary equipment in actual operation, which may include but is not limited to at least one of factors such as load level, ambient temperature and humidity. The operating condition will have an impact on the health status of the secondary equipment. The state category can be understood as classifying the equipment state vector into one of several predefined categories according to the equipment type and operating condition. Each state category represents a group of equipment states with similar characteristics. The historical state sequence can be understood as a sequence of equipment states related to a specific state category extracted from the knowledge graph that change over time. The historical state sequence contains state information of the same or similar equipment at different time points in the past. The dynamic time warping algorithm (Dynamic Time Warping, DTW) can be understood as finding the optimal alignment between two time series by dynamic programming, and can handle sequences of different lengths and time distortions. The dynamic time warping distance can be understood as the distance between the current equipment state vector and the historical state sequence calculated using the DTW algorithm. The smaller the time normalization distance value is, the higher the similarity between the two is. The softmax function can be understood as a function that converts a set of numerical values ​​into a probability distribution. It can map any real number into the interval (0,1) and ensure that the sum of all converted values ​​is 1. It is used to convert the dynamic time normalization distance into a similarity score, so that the smaller the distance value is, the higher the similarity score is. The similarity score can be understood as a numerical value indicating the degree of similarity between the current device state vector and the historical state sequence. A high similarity score reflects that the current state is very close to the historical state. The timeliness can be understood as the latest degree of the historical state sequence. The newer timeliness will be considered to be more valuable for reference and better reflect the operating environment and technical level of the current device. The integrity can be understood as the integrity and continuity of the data in the historical state sequence. A historical state sequence with high integrity can provide more comprehensive information and help improve the accuracy of detection and evaluation.

[0051] In an optional implementation, the DTW distance is converted into a similarity score through a softmax function. The similarity score is then weighted based on the timeliness and completeness of the historical state sequence. For example, a higher weight can be assigned to the historical state sequence within the last month, and a lower weight can be assigned to the historical state sequence of an earlier period; at the same time, a higher weight can be assigned to the historical state sequence with higher data integrity. Finally, the final device health score is calculated through weighted average.

[0052] The above scheme combines the device state vector with the knowledge graph, uses the dynamic time normalization algorithm and the soft maximization function to accurately calculate the device health score, can effectively identify the similarity between the current state of the device and the historical data, and comprehensively consider the timeliness and completeness of the historical data to improve the accuracy and reliability of the score, which helps to timely discover potential faults, realize preventive maintenance, and improve the operation safety and efficiency of the power system.

[0053] On the basis of the above scheme, before extracting the historical state sequence corresponding to the state category from the pre-constructed knowledge graph, it also includes: extracting the historical state sequence of each substation secondary equipment from the knowledge graph, and classifying the historical state sequence according to the equipment type and operating conditions of the substation secondary equipment to obtain the state category corresponding to the historical state sequence.

[0054] The above scheme ensures the accuracy and relevance of status categories by extracting and classifying historical status sequences from the knowledge graph, providing a reliable basis for subsequent equipment health assessment and fault diagnosis, and improving the accuracy and efficiency of the assessment.

[0055] On the basis of the above scheme, optionally, before determining the equipment health score according to the equipment state vector and the pre-constructed knowledge graph, it also includes: collecting historical test data, operating parameter data and fault maintenance records of substation secondary equipment, and constructing a multi-layer knowledge graph including equipment nodes, test parameter nodes and fault type nodes based on the historical test data, the operating parameter data and the fault maintenance records.

[0056] Among them, the historical test data can be understood as a data set obtained by performing various tests on the secondary equipment of the substation in a certain period of time in the past, which can be used to identify the change trend of the secondary equipment over time. The historical test data may include but is not limited to at least one of the functional test results and performance indicators of the equipment, and the operating parameter data can be understood as various parameter values ​​generated during the operation of the equipment, such as current, voltage, temperature and humidity, which are used to reflect the performance of the equipment under actual working conditions. The operating parameter data is the basis for detecting the operating status of the equipment and diagnosing equipment failures. By continuously collecting and analyzing the operating parameter data, abnormal conditions in the operation of the secondary equipment can be discovered in time, and preventive measures can be taken to avoid failures. The fault maintenance record can be understood as a record of the fault conditions that have occurred in the history of the secondary equipment and the maintenance measures and results taken for these problems. It can include at least one or more information such as the time, location, cause, maintenance method and effect evaluation after the maintenance, which can be used to quickly locate the current problems. It is also part of the knowledge graph to provide experience and guidance for future fault prevention and processing. The device node can be understood as a node of the secondary equipment of the substation, and each device node contains basic information about the equipment, including but not limited to at least one of the equipment type, model and installation location. The test parameter node can be understood as a node of parameter values ​​generated by the secondary device under different test conditions, which contains the performance data of the secondary device in a specific test, such as at least one of the current value or voltage value. The fault type node can be understood as a node of the possible fault type of the secondary device, which helps to understand the fault mode of the device. The fault type node may include but is not limited to at least one of the information such as the description, cause and impact of the fault.

[0057] The above solution, by collecting historical test data, operating parameter data and fault maintenance records and constructing a multi-layer knowledge graph, can comprehensively integrate the multi-dimensional information of the equipment. It not only enriches the data background and improves the accuracy and reliability of secondary equipment status assessment, but also provides a basis for fault diagnosis and preventive maintenance, which helps to detect potential problems in a timely manner and improve the operating safety and efficiency of the power system.

[0058] S130. Determine the target fault type corresponding to the substation secondary equipment according to the target test data, the knowledge graph and the pre-established inference rule base, obtain the fault handling plan corresponding to the historical fault case corresponding to the target fault type, and generate maintenance guidance information corresponding to the substation secondary equipment according to the fault handling plan.

[0059] Among them, the inference rule base can be understood as a predefined logical rule, which is used to infer a specific target fault type from the target test data and the knowledge graph. The rule can be formulated based on experience or statistical analysis of historical data. The target fault type can be understood as a specific fault category determined according to the target test data and the inference rule base. The historical fault case can be understood as a fault instance that has occurred and its fault handling solution, which is recorded and used to guide future fault handling. The fault handling solution can be understood as a series of countermeasures for a specific fault type, aiming to restore the normal operation of the secondary equipment. The fault handling solution can be obtained from historical fault cases, or customized by professionals according to actual conditions, and there is no limitation here. The maintenance guidance information can be understood as the final generated document or instruction set, which is provided to on-site technicians to guide them on how to implement specific maintenance work according to the fault handling solution. Specifically, the maintenance guidance information can be "It is recommended to replace the secondary terminal of the No. 2 main transformer CT and perform insulation treatment" and the like.

[0060] On the basis of the above scheme, optionally, before determining the target fault type corresponding to the substation secondary equipment according to the target test data, the knowledge graph and the pre-established inference rule base, it also includes: converting the data in the knowledge graph corresponding to the substation secondary equipment into a three-dimensional tensor, decomposing the three-dimensional tensor into the outer product form of three factor matrices by using a tensor decomposition method, and obtaining a low-dimensional embedding representation of the three-dimensional tensor; inputting the low-dimensional embedding representation into a transposed convolutional network, so as to map the low-dimensional embedding representation back to the original dimensional space of the knowledge graph through the multi-layer transposed convolutional layers contained in the transposed convolutional network, so as to obtain a target embedding identifier for knowledge completion, and updating the knowledge graph according to the target embedding identifier; extracting a device state change sequence from the updated knowledge graph, dividing the device state change sequence into a plurality of subsequence fragments containing at least one identical state node by a sliding window method, applying a sequence pattern mining algorithm to the subsequence fragments to identify a plurality of target state transition patterns, and establishing an inference rule base containing the device state evolution law according to the plurality of target state transition patterns; wherein the device state change sequence contains the state nodes and transition relationships of the substation secondary equipment at different time points.

[0061] The three-dimensional tensor can be understood as a three-dimensional array for representing data. The two dimensions of the three-dimensional tensor correspond to the head entity and the tail entity, respectively, and the third dimension represents the relationship type. The tensor decomposition method can be understood as a method for decomposing a high-dimensional tensor into multiple low-dimensional matrices, which is used to simplify the data structure and extract potential feature representations. The tensor decomposition method can be CANDECOMP / PARAFAC (CP) decomposition or Tucker decomposition, etc., which are not limited here. The low-dimensional embedding representation can be understood as a low-dimensional representation obtained by the tensor decomposition method, which can capture the key features of the data in the knowledge graph. Low-dimensional embedding representations are usually easier to process and analyze, while retaining important information of the data. The transposed convolutional network can be understood as a convolutional network for mapping the low-dimensional embedding representation back to the original high-dimensional space, that is, a deconvolutional network. The transposed convolutional network gradually restores the dimension of the data through multiple layers of transposed convolutional layers to match the dimension of the original data. The multi-layer transposed convolutional layer can be understood as a core component in the transposed convolutional network, which is used to gradually increase the dimension of the data by applying the transposed convolution operation multiple times. Each transposed convolution layer usually includes parameters such as convolution kernel, step size and padding. The knowledge completion can be understood as the process of supplementing and improving the knowledge graph by predicting missing entities or relationships through the model. In this embodiment, the low-dimensional embedding representation is mapped back to the original dimensional space through the transposed convolution network, and new entities or relationships can be generated, thereby updating the knowledge graph. The target embedding identifier can be understood as an embedding vector generated by the transposed convolution network to represent the new entity or relationship, which is used to update the knowledge graph and supplement the missing information. The device state change sequence can be understood as a sequence of state nodes and state transition relationships of the secondary device at different time points. Each state node represents a specific state of the secondary device at a certain time point, and the state transition relationship represents the change process of the state of the secondary device. The sliding window can be understood as a data processing method that divides the data into multiple subsequence fragments by moving a fixed-size window on the time series data, which is used to extract local patterns in the time series. The subsequence fragment can be understood as a small segment of data containing at least one identical state node cut from the device state change sequence by the sliding window method. Each subsequence fragment can reflect the state change of the device within a specific time period. The sequence pattern mining algorithm can be understood as being used to discover frequently occurring device state change patterns from subsequence fragments. The sequence pattern mining algorithm may include but is not limited to at least one of the Apriori algorithm and the PrefixSpan algorithm. The target state transition pattern can be understood as a frequently occurring device state change pattern mined from subsequence fragments, which is used to reflect the law of the device state change pattern and can be used to establish an inference rule base. The inference rule base can be understood as a rule set containing multiple target state transition patterns, which is used to indicate the evolution law of device state changes.The rules in the inference rule base can identify potential failure modes based on the regularity of equipment status changes and provide maintenance suggestions.

[0062] An optional implementation method is to represent the knowledge related to the secondary equipment of the substation as a knowledge graph and convert it into a three-dimensional tensor form, where the two dimensions of the three-dimensional tensor correspond to the head entity and the tail entity respectively, and the third dimension represents the relationship type. For example, for a piece of knowledge such as "protection device-protection-transformer", it can be represented as a non-zero element in the tensor, where (i, j, k) corresponds to the index position of the head entity "protection device", the tail entity "transformer" and the relationship "protection" respectively. Next, the three-dimensional tensor is decomposed into the outer product form of three low-dimensional factor matrices using the tensor decomposition method. Specifically, CP decomposition can be used to decompose the original tensor X into the outer product of three factor matrices A, B, and C: X≈[A, B, C]. The column vectors of A and B correspond to the low-dimensional embedding representations of the head entity and the tail entity respectively, and the column vector of C corresponds to the low-dimensional representation of the relationship type. In this way, the high-dimensional sparse knowledge graph can be compressed into a low-dimensional dense vector space. Then, the obtained low-dimensional embedding representation is input into a multi-layer transposed convolutional network for knowledge completion. The network contains multiple transposed convolutional layers, each equipped with batch normalization and ReLU activation function. The input of the network is a low-dimensional embedding matrix, which is finally output as a tensor with the same dimension as the original knowledge graph through layer-by-layer upsampling and feature reconstruction. Based on the completed knowledge graph, the device state change sequence is further extracted. These sequences contain the state nodes of the device at different time points and their transition relationships. For example, for a circuit breaker, its state sequence may be: "normal operation → fault → maintenance → return to normal". The sliding window method is used to divide these long sequences into multiple overlapping subsequence fragments, and the window size can be set to 5 and the step size is 1. Sequence pattern mining algorithms (such as PrefixSpan) are applied to these subsequence fragments to identify frequently occurring state transition patterns, such as the triple pattern of "fault → maintenance → return to normal". The frequent patterns mined are organized into an inference rule base to describe the evolution law of the device state.

[0063] The above scheme completes the knowledge graph by converting the knowledge graph data into a three-dimensional tensor and performing tensor decomposition to generate a low-dimensional embedding representation, and then uses a transposed convolutional network to map it back to the original dimensional space. It not only enriches the information of the knowledge graph, but also identifies the equipment state transition mode by extracting the equipment state change sequence and applying the sequence pattern mining algorithm, and establishes an inference rule library containing the equipment state evolution law, which improves the accuracy of equipment state assessment and fault diagnosis, provides a basis for preventive maintenance, and improves the operation safety and efficiency of the power system.

[0064] The technical solution of the embodiment of the present invention is as follows: first, by collecting the real-time operating parameters of the secondary equipment of the substation, the real-time operating parameters of the secondary equipment of the substation are input into a pre-trained graph neural network model to obtain the equipment operating characteristics of the secondary equipment of the substation, and the equipment state vector is determined according to the multiple equipment operating characteristics and the message transmission mechanism in the continuous time window, so as to accurately detect and evaluate the equipment status and give early warning of potential faults; then, the equipment health score is determined according to the equipment state vector and the pre-constructed knowledge graph, and when the health score is lower than the preset health threshold, the intelligent inspection robot is triggered to perform a deep detection task to collect target test data. Since the knowledge graph is based on the historical test data and operating parameters of the secondary equipment of the substation The data and fault maintenance records can be constructed to intelligently evaluate the health of equipment, start in-depth detection in time, and improve maintenance efficiency; the target fault type corresponding to the substation secondary equipment is determined according to the target test data, the knowledge graph and the pre-established inference rule library, and the fault handling plan corresponding to the historical fault case corresponding to the target fault type is obtained. According to the fault handling plan, maintenance guidance information corresponding to the substation secondary equipment is generated, which can quickly locate the fault and provide accurate maintenance guidance, solving the problems of substation secondary equipment fault diagnosis relying on manual experience, long response time, high misjudgment rate and low maintenance efficiency in related technologies, realizing real-time detection, accurate fault diagnosis and automated maintenance of substation secondary equipment, and improving the safety and maintenance efficiency of power system operation.

[0065] Embodiment 2

[0066] Figure 2 A flowchart of a fault detection method for a substation secondary device provided in the second embodiment of the present invention. This embodiment is based on the above embodiment, and further refines the target fault type corresponding to the substation secondary device according to the target test data, the knowledge graph and the pre-established inference rule base. For specific implementation methods, please refer to the description of this embodiment. Among them, the technical features that are the same or similar to the above embodiments are not repeated here.

[0067] like Figure 2 As shown, the method may specifically include:

[0068] S210. Collect real-time operating parameters of substation secondary equipment, input the real-time operating parameters of the substation secondary equipment into a pre-trained graph neural network model to obtain equipment operating characteristics of the substation secondary equipment, and determine the equipment state vector based on multiple equipment operating characteristics and a message passing mechanism within a continuous time window.

[0069] S220. Determine the equipment health score based on the equipment state vector and a pre-constructed knowledge graph. When the health score is lower than a preset health threshold, trigger the intelligent inspection robot to perform a deep inspection task to collect target test data, wherein the knowledge graph is constructed based on historical test data, operating parameter data and fault maintenance records of the substation secondary equipment.

[0070] S230. Determine multiple candidate fault propagation paths based on the target test data, the knowledge graph and a pre-established inference rule base, determine the target fault type corresponding to the substation secondary equipment based on the multiple candidate fault propagation paths, obtain a fault handling plan corresponding to the historical fault case corresponding to the target fault type, and generate maintenance guidance information corresponding to the substation secondary equipment based on the fault handling plan.

[0071] The candidate fault propagation path can be understood as a possible fault propagation path determined from the target test data and the knowledge graph. Each candidate fault propagation path determines the process of fault propagation from one node to another.

[0072] On the basis of the above scheme, optionally, the knowledge graph includes fault feature data corresponding to multiple historical fault types; the determining multiple candidate fault propagation paths according to the target test data, the knowledge graph and the pre-established reasoning rule base includes: normalizing the target test data, calculating the similarity between the target test data and the fault feature data corresponding to each historical fault type in the knowledge graph and the importance weight corresponding to the fault feature data through a multi-head attention network, and calculating a comprehensive similarity score based on the importance weight; based on the target test data and the fault feature data with the highest comprehensive similarity score and the fault propagation rules in the reasoning rule base, rule reasoning is performed by combining forward linking and backward linking to generate an initial fault propagation path; using equipment status data, historical fault records and preset identification standards as evidence sources, the initial fault propagation path is credibility evaluated based on a preset evidence theory, and evidence conflicts are processed through a dynamic weight adjustment mechanism to obtain the confidence of the fault propagation path; the fault propagation paths whose confidence is lower than a preset screening threshold are deleted, and the remaining fault propagation paths are used as candidate fault propagation paths.

[0073] Among them, the fault feature data can be understood as a set of feature data related to a specific historical fault type, which is used to reflect the state and performance of the device when the fault occurs, such as at least one of abnormal current, temperature rise or vibration change. The multi-head attention network can be understood as a neural network model that processes historical fault types in parallel through multiple attention heads. Each attention head of the multi-head attention network corresponds to a historical fault type. The similarity can be understood as an indicator used to measure the similarity between the target test data and the fault feature data. The importance weight can be understood as the importance of the relationship between the target test data and the fault feature data corresponding to a certain historical fault type. The importance weight can be calculated by the attention mechanism in the multi-head attention network. The comprehensive similarity score can be understood as a score calculated by combining the similarity and the importance weight, which is used to evaluate the matching degree between the target test data and a certain historical fault type. The fault propagation rule can be understood as a rule used in the inference rule base to describe the propagation of faults between devices. The fault propagation rule may include but is not limited to at least one of the physical connection relationship rules of the device, the functional dependency relationship rules and the information interaction relationship rules. The physical connection relationship rules of the device can be understood as the rules of the physical connection relationship between devices, which may include but are not limited to cable connection and mechanical connection. The physical connection relationship rules of the devices can be used to determine the physical path along which the fault may propagate. The functional dependency rules can be understood as the rules of the functional dependency between devices, such as the normal operation of one device needs to rely on the output of another device, etc., which are used to analyze the impact of the fault on the overall function of the system, etc. The information interaction relationship rules can be understood as the rules of the information interaction relationship between devices, such as data communication or signal transmission, etc., which are used to identify the propagation path of the fault at the information level, etc. The forward link can be understood as starting from known facts and gradually deriving new conclusions according to the rules in the rule base. In this embodiment, the forward link can be used to derive the method of the initial fault propagation path from the target test data and the fault feature data with the highest comprehensive similarity score. The backward link can be understood as starting from the target (such as a specific fault type), and gradually finding facts that support the target according to the rules in the rule base. In this embodiment, the backward link can be used in combination with the forward link to generate the initial fault propagation path. The preset evidence theory can be understood as a mathematical framework for processing uncertainty and conflicting information, which is used to evaluate the credibility of evidence. In an embodiment, the preset evidence theory is used to evaluate the credibility of the initial fault propagation path. The evidence conflict can be understood as the contradiction or inconsistency between the information provided by different evidence sources. The evidence conflict can be handled by a dynamic weight adjustment mechanism to improve the accuracy of the reasoning result. The dynamic weight adjustment mechanism can be understood as a method for adjusting the weights of different evidence sources to resolve the evidence conflict problem.The dynamic weight adjustment mechanism for handling evidence conflicts can enhance the influence of credible evidence to weaken the influence of unreliable evidence. The confidence level can be understood as an indicator for measuring the credibility of a fault propagation path. The confidence level can be calculated through a preset evidence theory and a dynamic weight adjustment mechanism, and is used to screen out the most likely fault propagation path. The preset screening threshold can be understood as a set threshold for screening fault propagation paths. Specifically, only fault propagation paths with a confidence level higher than a preset screening threshold will be retained, and the remaining paths will be deleted. The candidate fault propagation path can be understood as a fault propagation path retained after confidence evaluation and screening. The candidate fault propagation path can be considered as a possible fault propagation path for further fault diagnosis and processing.

[0074] An optional implementation method is to perform normalization preprocessing on the high-dimensional target test data. Specifically, the minimum-maximum normalization method is used to map the eigenvalues ​​of each dimension to the interval [0,1]. For example, for the original eigenvalue x, the normalized value is calculated by (x-min) / (max-min), where min and max are the minimum and maximum values ​​of the dimension, respectively. Then, a multi-head attention network is constructed to calculate the similarity between the target test data and the fault feature data of the historical fault type in the knowledge graph. The network includes multiple attention heads, each of which corresponds to a type of historical fault type. For example, for different types of features such as temperature, pressure, and vibration, corresponding attention heads are set respectively. Each attention head calculates the correlation between feature vectors through the self-attention mechanism to obtain the importance weight of the feature of this type. Finally, the outputs of each attention head are spliced ​​and linearly transformed to obtain a comprehensive feature weight vector. Based on the weight vector, the weighted cosine similarity between the test data and the fault mode in the knowledge graph is calculated as the final similarity score. Relevant fault propagation rules are extracted from the inference rule base. These rules include rules for physical connection relationships of devices (such as device A is directly connected to device B), rules for functional dependency relationships (such as function C depends on function D), and rules for information interaction relationships (such as device E sends data to device F). A combination of forward chaining and backward chaining is used to start from the initial fault point and gradually deduce possible fault propagation paths according to the rules. For example, if device A fails, it can be deduced to device B according to the physical connection rules, and then to function C according to the functional dependency rules. At the same time, reverse deduction is performed from the final affected function or device to find all possible fault sources. Through bidirectional reasoning, an initial set of fault propagation paths is generated. In order to evaluate the credibility of the fault propagation path, the Dempster-Shafer evidence theory is introduced. Device status data, historical fault records, and expert knowledge are used as different evidence sources. For each propagation path, the support and negation of each evidence source are calculated respectively, and a basic probability distribution function is constructed. Then, the Dempster combination rule is used to fuse the information of multiple evidence sources. In order to deal with evidence conflicts, a dynamic weight adjustment mechanism based on evidence distance is adopted. Specifically, the pairwise distance between evidence sources is calculated, and the weight of each evidence source is dynamically adjusted according to the distance. Finally, the comprehensive confidence of each propagation path is obtained. The confidence of the fault propagation path is compared with the preset screening threshold (such as 0.8). Paths with confidence higher than the threshold are deleted because these paths are considered to be propagation paths with high certainty. Paths with confidence lower than the threshold are retained as multi-dimensional fault propagation paths to reflect the uncertainty and diversity of fault propagation.

[0075] The above scheme calculates the similarity between the target test data and the historical fault characteristics through a multi-head attention network, generates the initial fault propagation path in combination with the inference rule library, and uses the preset evidence theory and dynamic weight adjustment mechanism to evaluate the credibility of the path, and screens out high-confidence candidate fault propagation paths. It not only improves the accuracy of fault diagnosis, but also can effectively identify and handle evidence conflicts, provide a basis for fault handling, and improve the operation safety and maintenance efficiency of the power system.

[0076] On the basis of the above scheme, optionally, determining the target fault type corresponding to the substation secondary equipment according to the multiple candidate fault propagation paths includes: extracting the features of each candidate fault propagation path at multiple scales through a feature extraction network, calculating the importance coefficient of the features at each scale using an attention weight network, adaptively weighting the features at multiple scales according to the importance coefficient to obtain scale-weighted features, and encoding the position information of the substation secondary equipment in the fault propagation path to obtain position encoding features, and combining the position encoding features with the scale-weighted features into enhanced features; predicting fault types based on the enhanced features corresponding to the multiple candidate fault propagation paths and a pre-built multi-task learning framework to obtain first candidate fault types corresponding to the multiple candidate fault propagation paths; and inputting the enhanced features corresponding to the multiple candidate fault propagation paths into a multi-layer graph attention network to obtain the probability distribution of multiple second candidate fault types corresponding to the multiple candidate fault propagation paths; determining the target fault type corresponding to the substation secondary equipment according to the probability distribution of the first candidate fault type and multiple second candidate fault types.

[0077] Among them, the feature extraction network can be understood as a neural network model for extracting useful features from input data. In an embodiment, the feature extraction network is used to extract features of each candidate fault propagation path at multiple scales. The features at multiple scales can be understood as features at different scales. The features at multiple scales can include at least one of local structural features and global structural features. Local structural features can be obtained by 1-hop neighbor node sampling, and global structural features can be obtained by random walk sampling. The attention weight network can be understood as a neural network model for calculating the importance coefficient of features at each scale. The attention weight network can help the model focus on the most critical information by learning the importance of different features. The importance coefficient can be understood as a numerical value that measures the importance of a certain feature in the model, which is used to weight features at multiple scales. The scale-weighted feature can be understood as a feature representation obtained by adaptively weighting features at multiple scales. The scale-weighted feature integrates information at different scales and improves the expressiveness of the feature. The location information can be understood as the specific location of the substation secondary equipment in the fault propagation path. The location information helps to analyze the propagation path of the fault between devices. The location encoding feature can be understood as a feature representation obtained by encoding the location information of the substation secondary equipment. Position coding features can be used to capture the spatial distribution and relationship of substation secondary equipment. The enhanced features can be understood as feature representations that combine position coding features with scale-weighted features. The enhanced features integrate multiple feature information and improve the prediction ability of the model. The multi-task learning framework can be understood as a machine learning method that improves the performance of the model by learning multiple related tasks at the same time. In this embodiment, the multi-task learning framework is used to predict fault types. The first candidate fault type can be understood as a fault type predicted based on enhanced features corresponding to multiple candidate fault propagation paths and a multi-task learning framework. The first candidate fault type is a preliminary fault diagnosis result. The multi-layer graph attention network can be understood as a neural network model that processes graph structure data through a multi-layer attention mechanism. In this embodiment, the multi-layer graph attention network is used to further refine the prediction of the fault type. The probability distribution of the second candidate fault type can be understood as the probability distribution output by the multi-layer graph attention network that represents the possibility of each fault type. The probability distribution of the second candidate fault type provides more detailed fault diagnosis information

[0078] In an optional implementation, local structural features and global structural features are obtained by neighborhood sampling. Local structural features are obtained by sampling 1-hop neighbor nodes, and global structural features are obtained by random walk sampling. Then, an attention weight network is used to calculate the importance coefficients of features of different scales. The attention weight network consists of a fully connected layer and a softmax layer, with feature vectors of different scales as input and corresponding importance coefficients as output. Features of different scales are adaptively weighted according to the importance coefficient to obtain a multi-scale feature vector. The structural position information of the node in the graph is encoded and combined with the multi-scale feature vector to generate enhanced features. The structural position information can be obtained by algorithms such as node2vec. Then, a multi-task learning framework is constructed based on the enhanced features. A fault category recognition task and a fault sub-category recognition task are set. The fault category recognition task divides faults into three categories: device-level faults, system-level faults, and cascading faults. The fault sub-category recognition task identifies specific fault types under the corresponding categories based on the fault category prediction results. For example, for a device-level fault, it is further identified as a hardware fault, a software fault, etc. The two tasks share the underlying feature extraction network, but use independent classifiers. Finally, the enhanced features are input into a multi-layer graph attention network. Each layer of the network contains multiple attention heads. Each attention head extracts fault features of a specific type based on the attention query matrix and generates a feature vector. The attention coefficient is calculated through the interaction of the node feature vector and the edge feature vector. Specifically, the node feature vector is concatenated with the edge feature vector, and the attention coefficient is obtained through the fully connected layer. The feature vector is weighted according to the attention coefficient, and the weighted feature vector is input into the softmax function to obtain the probability distribution of each fault type. The probability distribution is combined with the prediction results of the multi-task classification model to obtain the final fault classification result.

[0079] The above scheme calculates the importance coefficient of features through multi-scale feature extraction and attention weight network, generates scale-weighted features, and combines position encoding features to form enhanced features. Preliminary fault type prediction is performed based on enhanced features and a multi-task learning framework, and then the fault type probability distribution is further refined through a multi-layer graph attention network. Finally, the target fault type is determined by combining the probability distributions of the first candidate fault type and the second candidate fault type. This not only improves the accuracy and reliability of fault diagnosis, but also can effectively handle complex fault propagation paths, provide a basis for fault handling, and improve the operation safety and maintenance efficiency of the power system.

[0080] The technical solution of the embodiment of the present invention extracts the equipment operation characteristics of real-time operation parameters through a graph neural network, and generates an equipment state vector in combination with a message passing mechanism to evaluate the health status of the equipment in real time. When the health level is lower than the preset threshold, the intelligent inspection robot is triggered to perform in-depth detection and collect target test data. Next, the initial fault propagation path is determined using a multi-head attention network and a comprehensive similarity score, and rule reasoning is performed by combining forward links with backward links to generate multiple candidate fault propagation paths. The credibility of the path is evaluated through a preset evidence theory and a dynamic weight adjustment mechanism, and high-confidence candidate paths are screened out. Finally, through multi-scale feature extraction, position encoding, and a multi-task learning framework, combined with a multi-layer graph attention network, the target fault type is comprehensively determined, which not only improves the accuracy and timeliness of fault diagnosis, but also can effectively handle complex fault propagation paths, reduce misjudgments and missed judgments, and improve the operation safety and maintenance efficiency of the power system.

[0081] Embodiment 3

[0082] Figure 3 This is a schematic diagram of the structure of a fault detection device for a substation secondary device provided in the third embodiment of the present invention. Figure 3 As shown, the device includes: an equipment information acquisition module 310, a test data collection module 320, and an equipment maintenance guidance module 330.

[0083] Among them, the equipment information acquisition module 310 is used to collect the real-time operating parameters of the substation secondary equipment, input the real-time operating parameters of the substation secondary equipment into the pre-trained graph neural network model to obtain the equipment operating characteristics of the substation secondary equipment, and determine the equipment state vector according to multiple equipment operating characteristics and message passing mechanism in a continuous time window; the test data acquisition module 320 is used to determine the equipment health score according to the equipment state vector and the pre-constructed knowledge graph, and when the health score is lower than the preset health threshold, trigger the intelligent inspection robot to perform a deep detection task to collect target test data, wherein the knowledge graph is constructed based on the historical test data, operating parameter data and fault maintenance records of the substation secondary equipment; the equipment maintenance guidance module 330 is used to determine the target fault type corresponding to the substation secondary equipment according to the target test data, the knowledge graph and the pre-established inference rule base, obtain the fault handling plan corresponding to the historical fault case corresponding to the target fault type, and generate the maintenance guidance information corresponding to the substation secondary equipment according to the fault handling plan.

[0084] The technical solution of the embodiment of the present invention is as follows: first, the real-time operating parameters of the secondary equipment of the substation are collected through the equipment information acquisition module 310, and the real-time operating parameters of the secondary equipment of the substation are input into the pre-trained graph neural network model to obtain the equipment operating characteristics of the secondary equipment of the substation, and the equipment state vector is determined according to the multiple equipment operating characteristics and the message transmission mechanism in the continuous time window, so as to accurately detect and evaluate the equipment status and give early warning of potential faults; then, the equipment health score is determined according to the equipment state vector and the pre-built knowledge graph through the test data acquisition module 320, and when the health score is lower than the preset health threshold, the intelligent inspection robot is triggered to perform the deep detection task to collect the target test data. Since the knowledge graph is based on the historical test data of the secondary equipment of the substation, The construction of operating parameter data and fault maintenance records can intelligently evaluate the health of equipment, start deep detection in time, and improve maintenance efficiency; finally, through the equipment maintenance guidance module 330, the target fault type corresponding to the substation secondary equipment is determined according to the target test data, the knowledge graph and the pre-established inference rule library, and the fault handling plan corresponding to the historical fault case corresponding to the target fault type is obtained. According to the fault handling plan, maintenance guidance information corresponding to the substation secondary equipment is generated, which can quickly locate the fault and provide accurate maintenance guidance, solving the problems of substation secondary equipment fault diagnosis relying on manual experience, long response time, high misjudgment rate and low maintenance efficiency in related technologies, realizing real-time detection, accurate fault diagnosis and automated maintenance of substation secondary equipment, and improving the safety and maintenance efficiency of power system operation.

[0085] Based on the above scheme, optionally, the graph neural network model includes multiple graph convolution modules connected in series, each of which includes a spatial convolution unit, a feature transformation unit, a batch normalization layer and an activation function layer; the device information acquisition module includes: a module output feature determination submodule and a device operation feature determination submodule.

[0086] Among them, the module output feature determination submodule is used for converting the real-time operation parameters into initial operation features through the spatial convolution unit of the graph convolution module for the first graph convolution module in the graph neural network model, and taking the substation secondary equipment as the central node and the substation secondary equipment directly connected to the central node as the neighboring nodes, the neighborhood aggregation function based on the attention weight dynamically adjusts the feature aggregation weight corresponding to the initial operation feature according to the relationship type corresponding to the node connection relationship between the central node and the neighboring nodes, and determines the operation aggregation feature corresponding to the real-time operation parameter according to the adjusted aggregation weight; and, through the feature transformation unit of the graph convolution module, The residual connection mechanism of the element performs weighted combination of the initial operation characteristics of the central node and the operation aggregation characteristics to obtain the operation fusion characteristics; the operation fusion characteristics are processed into the module output characteristics of the graph convolution module through the batch normalization layer and the activation function layer in sequence; the equipment operation characteristic determination submodule is used to obtain the module output characteristics of the previous graph convolution module connected to the graph convolution module for the subsequent graph convolution module of the graph neural network model, input the module output characteristics into the graph convolution module to obtain the module output characteristics of the graph convolution module, and determine the module output characteristics of the last graph convolution module as the equipment operation characteristics of the substation secondary equipment.

[0087] On the basis of the above scheme, optionally, the device information acquisition module includes: an initial message conversion submodule, a target device feature determination submodule, and a device state vector generation submodule.

[0088] Among them, the initial message conversion submodule is used to take the substation secondary equipment that sends information as the source node and the substation secondary equipment that receives information as the target node, and convert the equipment operation characteristics of the source node into the initial message through the message generation function; wherein the message generation function is associated with the relationship type corresponding to the node-to-node connection relationship between the source node and the target node; the target device feature determination submodule is used to filter the initial message using a gating unit jointly determined by the equipment characteristics of the source node and the target node to obtain multiple target messages, and weighted sum the multiple target messages to obtain the target device characteristics of the target node; the device state vector generation submodule is used to input the multiple target device characteristics of the substation secondary equipment in a continuous time window into the temporal attention layer, so as to weightedly splice the multiple target device characteristics through a multi-head attention mechanism and generate a device state vector after linear transformation.

[0089] On the basis of the above scheme, optionally, the test data acquisition module includes: a dynamic time normalization distance calculation submodule and an equipment health score determination submodule.

[0090] Among them, the dynamic time normalization distance calculation submodule is used to determine the state category corresponding to the equipment state vector according to the equipment type and operating conditions of the substation secondary equipment, extract the historical state sequence corresponding to the state category from the pre-constructed knowledge graph, and use the dynamic time normalization algorithm to calculate the dynamic time normalization distance between the current equipment state vector and the historical state sequence of the corresponding category; the equipment health score determination submodule is used to convert the dynamic time normalization distance into a similarity score through a soft maximization function, and perform weighted summation on the similarity score based on the timeliness and completeness of the historical state sequence to obtain the equipment health score.

[0091] On the basis of the above solution, optionally, the equipment maintenance guidance module includes: a target fault type determination submodule.

[0092] Among them, the target fault type determination submodule is used to determine multiple candidate fault propagation paths based on the target test data, the knowledge graph and a pre-established inference rule base, and determine the target fault type corresponding to the substation secondary equipment based on the multiple candidate fault propagation paths.

[0093] Based on the above scheme, optionally, the knowledge graph includes fault feature data corresponding to multiple historical fault types, and the target fault type determination submodule includes: a comprehensive similarity score calculation unit, an initial fault propagation path generation unit, a confidence acquisition unit, and a candidate fault propagation path determination unit.

[0094] The comprehensive similarity score calculation unit is used to normalize and preprocess the target test data, calculate the similarity between the target test data and the fault feature data corresponding to each historical fault type in the knowledge graph through a multi-head attention network, and calculate the importance weight corresponding to the fault feature data, and calculate the comprehensive similarity score based on the importance weight; wherein each attention head of the multi-head attention network corresponds to a historical fault type; the initial fault propagation path generation unit is used to generate an initial fault propagation path based on the target test data and the fault feature data with the highest comprehensive similarity score and the fault propagation rules in the inference rule library, using a combination of forward links and backward links. The method comprises the following steps: performing rule reasoning in a combined manner to generate an initial fault propagation path; wherein the fault propagation rules include equipment physical connection relationship rules, functional dependency relationship rules and information interaction relationship rules; the confidence acquisition unit is used to use equipment status data, historical fault records and preset identification standards as evidence sources, perform credibility assessment on the initial fault propagation path based on preset evidence theory, and handle evidence conflicts through a dynamic weight adjustment mechanism to obtain the confidence of the fault propagation path; the candidate fault propagation path determination unit is used to delete the fault propagation paths whose confidence levels are lower than a preset screening threshold, and use the remaining fault propagation paths as candidate fault propagation paths.

[0095] On the basis of the above solution, optionally, the target fault type determination submodule includes: an enhanced feature combination unit, a probability distribution determination unit, and a target fault type determination unit.

[0096] An enhanced feature combination unit is used to extract the features of each candidate fault propagation path at multiple scales through a feature extraction network, calculate the importance coefficient of the features at each scale using an attention weight network, adaptively weight the features at multiple scales according to the importance coefficient to obtain scale-weighted features, and encode the position information of the substation secondary equipment in the fault propagation path to obtain position encoding features, and combine the position encoding features with the scale-weighted features into enhanced features; a probability distribution determination unit is used to predict the fault type based on the enhanced features corresponding to the multiple candidate fault propagation paths and a pre-built multi-task learning framework to obtain a first candidate fault type corresponding to the multiple candidate fault propagation paths; and the enhanced features corresponding to the multiple candidate fault propagation paths are input into a multi-layer graph attention network to obtain the probability distribution of multiple second candidate fault types corresponding to the multiple candidate fault propagation paths; a target fault type determination unit is used to determine the target fault type corresponding to the substation secondary equipment according to the probability distribution of the first candidate fault type and multiple second candidate fault types.

[0097] On the basis of the above scheme, optionally, the fault detection device for substation secondary equipment further includes: a multi-layer knowledge graph construction module.

[0098] Among them, the multi-layer knowledge graph construction module is used to collect historical test data, operating parameter data and fault maintenance records of substation secondary equipment before determining the equipment health score according to the equipment state vector and the pre-constructed knowledge graph, and construct a multi-layer knowledge graph containing equipment nodes, test parameter nodes and fault type nodes based on the historical test data, the operating parameter data and the fault maintenance records.

[0099] On the basis of the above scheme, optionally, the fault detection device for substation secondary equipment further includes: a low-dimensional embedding representation acquisition module, a knowledge graph updating module, and an inference rule base resume module.

[0100] The low-dimensional embedding representation acquisition module is used to convert the data in the knowledge graph corresponding to the substation secondary equipment into a three-dimensional tensor before determining the target fault type corresponding to the substation secondary equipment according to the target test data, the knowledge graph and the pre-established inference rule base, and use the tensor decomposition method to decompose the three-dimensional tensor into the outer product form of three factor matrices to obtain the low-dimensional embedding representation of the three-dimensional tensor; wherein the two dimensions of the three-dimensional tensor correspond to the head entity and the tail entity respectively, and the third dimension represents the relationship type, and the knowledge graph updating module is used to input the low-dimensional embedding representation into the transposed convolutional network to convert the low-dimensional embedding representation into the transposed convolutional network through the multi-layer transposed convolutional layer contained in the transposed convolutional network. The low-dimensional embedding representation is mapped back to the original dimensional space of the knowledge graph to obtain the target embedding identifier of knowledge completion, and the knowledge graph is updated according to the target embedding identifier; the inference rule base resume module is used to extract the equipment state change sequence from the updated knowledge graph, divide the equipment state change sequence into multiple subsequence fragments containing at least one identical state node by means of a sliding window, apply a sequence pattern mining algorithm to the subsequence fragments to identify multiple target state transfer patterns, and establish an inference rule base containing the equipment state evolution law according to the multiple target state transfer patterns; wherein, the equipment state change sequence contains the state nodes and transfer relationships of the substation secondary equipment at different time points.

[0101] The fault detection device for substation secondary equipment provided in the embodiment of the present invention can execute the fault detection method for substation secondary equipment provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0102] Embodiment 4

[0103] Figure 4A schematic diagram of the structure of an electronic device 10 that can be used to implement an embodiment of the present invention is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices (such as helmets, glasses, watches, etc.) and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present invention described and / or required herein.

[0104] like Figure 4 As shown, the electronic device 10 includes at least one processor 11, and a memory connected to the at least one processor 11, such as a read-only memory (ROM) 12, a random access memory (RAM) 13, etc., wherein the memory stores a computer program that can be executed by at least one processor, and the processor 11 can perform various appropriate actions and processes according to the computer program stored in the read-only memory (ROM) 12 or the computer program loaded from the storage unit 18 to the random access memory (RAM) 13. In the RAM 13, various programs and data required for the operation of the electronic device 10 can also be stored. The processor 11, the ROM 12, and the RAM 13 are connected to each other through a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0105] A number of components in the electronic device 10 are connected to the I / O interface 15, including: an input unit 16, such as a keyboard, a mouse, etc.; an output unit 17, such as various types of displays, speakers, etc.; a storage unit 18, such as a disk, an optical disk, etc.; and a communication unit 19, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 19 allows the electronic device 10 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks.

[0106] The processor 11 may be a variety of general and / or special processing components with processing and computing capabilities. Some examples of the processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, digital signal processors (DSPs), and any appropriate processors, controllers, microcontrollers, etc. The processor 11 executes the various methods and processes described above, such as a fault detection method for a secondary device in a substation.

[0107] In some embodiments, a fault detection method for a substation secondary device may be implemented as a computer program, which is tangibly contained in a computer-readable storage medium, such as a storage unit 18. In some embodiments, part or all of the computer program may be loaded and / or installed on the electronic device 10 via the ROM 12 and / or the communication unit 19. When the computer program is loaded into the RAM 13 and executed by the processor 11, one or more steps of the fault detection method for a substation secondary device described above may be performed. Alternatively, in other embodiments, the processor 11 may be configured to execute a fault detection method for a substation secondary device in any other appropriate manner (e.g., by means of firmware).

[0108] Various implementations of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), load programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include: being implemented in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0109] Computer programs for implementing the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that when the computer program is executed by the processor, the functions / operations specified in the flow chart and / or block diagram are implemented. The computer program may be executed entirely on the machine, partially on the machine, partially on the machine and partially on a remote machine as a stand-alone software package, or entirely on a remote machine or server.

[0110] In the context of the present invention, a computer-readable storage medium may be a tangible medium that may contain or store a computer program for use by or in combination with an instruction execution system, device or equipment. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices or equipment, or any suitable combination of the foregoing. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. A more specific example of a machine-readable storage medium may include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0111] To provide interaction with a user, the systems and techniques described herein may be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and a pointing device (e.g., a mouse or trackball) through which the user can provide input to the electronic device. Other types of devices may also be used to provide interaction with the user; for example, the feedback provided to the user may be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user may be received in any form (including acoustic input, voice input, or tactile input).

[0112] The systems and techniques described herein may be implemented in a computing system that includes backend components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes frontend components (e.g., a user computer with a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such backend components, middleware components, or frontend components. The components of the system may be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: a local area network (LAN), a wide area network (WAN), a blockchain network, and the Internet.

[0113] A computing system may include a client and a server. The client and the server are generally remote from each other and usually interact through a communication network. The client and server relationship is generated by computer programs running on the corresponding computers and having a client-server relationship with each other. The server may be a cloud server, also known as a cloud computing server or cloud host, which is a host product in the cloud computing service system to solve the defects of difficult management and weak business scalability in traditional physical hosts and VPS services.

[0114] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps described in the present invention can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solution of the present invention can be achieved, and this document does not limit this.

[0115] The above specific implementations do not constitute a limitation on the protection scope of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A fault detection method for secondary equipment of a substation, characterized in that: include: Collecting real-time operating parameters of substation secondary equipment, inputting the real-time operating parameters of the substation secondary equipment into a pre-trained graph neural network model to obtain equipment operating characteristics of the substation secondary equipment, and determining the equipment state vector according to a plurality of equipment operating characteristics in a continuous time window and a message transmission mechanism; Determine the equipment health score according to the equipment state vector and the pre-built knowledge graph, and trigger the intelligent inspection robot to perform a deep detection task to collect target test data when the health score is lower than a preset health threshold, wherein the knowledge graph is built based on the historical test data, operating parameter data and fault maintenance records of the substation secondary equipment; Determine the target fault type corresponding to the substation secondary equipment according to the target test data, the knowledge graph and the pre-established inference rule base, obtain the fault handling plan corresponding to the historical fault case corresponding to the target fault type, and generate maintenance guidance information corresponding to the substation secondary equipment according to the fault handling plan.

2. The method according to claim 1, characterized in that The graph neural network model includes a plurality of graph convolution modules connected in series, each of which includes a spatial convolution unit, a feature transformation unit, a batch normalization layer, and an activation function layer; the real-time operation parameters of the substation secondary equipment are input into the pre-trained graph neural network model to obtain the equipment operation characteristics of the substation secondary equipment, including: For the first graph convolution module in the graph neural network model, the real-time operation parameters are converted into initial operation features through the spatial convolution unit of the graph convolution module, and the substation secondary equipment is used as the central node, and the substation secondary equipment directly connected to the central node is used as the neighboring node. The neighborhood aggregation function based on the attention weight dynamically adjusts the feature aggregation weight corresponding to the initial operation feature according to the relationship type corresponding to the node connection relationship between the central node and the neighboring node, and determines the operation aggregation feature corresponding to the real-time operation parameter according to the adjusted aggregation weight; and the initial operation feature of the central node is weightedly combined with the operation aggregation feature through the residual connection mechanism of the graph convolution module in the feature transformation unit to obtain the operation fusion feature; the operation fusion feature is processed into the module output feature of the graph convolution module through the batch normalization layer and the activation function layer in sequence; For the subsequent graph convolution module of the graph neural network model, the module output features of the previous graph convolution module connected to the graph convolution module are obtained, and the module output features are input into the graph convolution module to obtain the module output features of the graph convolution module, and the module output features of the last graph convolution module are determined as the equipment operation features of the substation secondary equipment.

3. The method according to claim 1, characterized in that The determining of the device state vector according to the plurality of device operation characteristics and the message transmission mechanism within the continuous time window comprises: The substation secondary equipment sending information is taken as the source node, and the substation secondary equipment receiving information is taken as the target node, and the equipment operation characteristics of the source node are converted into an initial message through a message generation function; wherein the message generation function is associated with a relationship type corresponding to the node-to-node connection relationship between the source node and the target node; The initial message is filtered using a gating unit determined by the device characteristics of the source node and the target node to obtain a plurality of target messages, and a weighted sum of the plurality of target messages is performed to obtain a target device characteristic of the target node; The multiple target device features of the substation secondary equipment in the continuous time window are input into the temporal attention layer, so as to weightedly splice the multiple target device features through the multi-head attention mechanism and generate a device state vector after linear transformation.

4. The method according to claim 1, characterized in that: Determining the device health score according to the device state vector and a pre-built knowledge graph includes: Determine the state category corresponding to the device state vector according to the device type and operating condition of the substation secondary equipment, extract the historical state sequence corresponding to the state category from the pre-built knowledge graph, and use the dynamic time normalization algorithm to calculate the dynamic time normalization distance between the current device state vector and the historical state sequence of the corresponding category; The dynamic time normalization distance is converted into a similarity score through a soft maximization function, and the similarity score is weighted and summed based on the timeliness and completeness of the historical state sequence to obtain a device health score.

5. The method according to claim 1, characterized in that: The determining the target fault type corresponding to the substation secondary equipment according to the target test data, the knowledge graph and a pre-established inference rule base includes: A plurality of candidate fault propagation paths are determined according to the target test data, the knowledge graph and a pre-established inference rule base, and a target fault type corresponding to the substation secondary equipment is determined according to the plurality of candidate fault propagation paths.

6. The method according to claim 5, characterized in that The knowledge graph includes fault feature data corresponding to various historical fault types; The determining of multiple candidate fault propagation paths according to the target test data, the knowledge graph and a pre-established inference rule base includes: The target test data is normalized and preprocessed, and the similarity between the target test data and the fault feature data corresponding to each historical fault type in the knowledge graph and the importance weight corresponding to the fault feature data are calculated through a multi-head attention network, and a comprehensive similarity score is calculated based on the importance weight; wherein each attention head of the multi-head attention network corresponds to a historical fault type; Based on the target test data and the fault feature data with the highest comprehensive similarity score and the fault propagation rules in the inference rule library, rule reasoning is performed by combining forward linking and backward linking to generate an initial fault propagation path; wherein the fault propagation rules include equipment physical connection relationship rules, functional dependency relationship rules and information interaction relationship rules; Using equipment status data, historical fault records and preset identification standards as evidence sources, the credibility of the initial fault propagation path is evaluated based on the preset evidence theory, and evidence conflicts are handled through a dynamic weight adjustment mechanism to obtain the confidence of the fault propagation path; The fault propagation paths whose confidence levels are lower than a preset screening threshold are deleted, and the remaining fault propagation paths are used as candidate fault propagation paths.

7. The method according to claim 5, characterized in that The determining, according to the plurality of candidate fault propagation paths, a target fault type corresponding to the substation secondary equipment comprises: The features of each candidate fault propagation path at multiple scales are extracted by a feature extraction network, the importance coefficient of the feature at each scale is calculated by an attention weight network, the features at multiple scales are adaptively weighted according to the importance coefficient to obtain scale-weighted features, and the position information of the substation secondary equipment in the fault propagation path is encoded to obtain position coding features, and the position coding features are combined with the scale-weighted features to form enhanced features; Fault type prediction is performed based on the enhanced features corresponding to the multiple candidate fault propagation paths and a pre-built multi-task learning framework to obtain first candidate fault types corresponding to the multiple candidate fault propagation paths; and, the enhanced features corresponding to the multiple candidate fault propagation paths are input into a multi-layer graph attention network to obtain probability distributions of multiple second candidate fault types corresponding to the multiple candidate fault propagation paths; A target fault type corresponding to the substation secondary equipment is determined according to the probability distribution of the first candidate fault type and a plurality of the second candidate fault types.

8. The method according to claim 1, characterized in that: Before determining the device health score according to the device state vector and the pre-built knowledge graph, the method further includes: The historical test data, operating parameter data and fault maintenance records of the substation secondary equipment are collected, and a multi-layer knowledge graph including equipment nodes, test parameter nodes and fault type nodes is constructed based on the historical test data, the operating parameter data and the fault maintenance records.

9. The method according to claim 1, characterized in that: Before determining the target fault type corresponding to the substation secondary equipment according to the target test data, the knowledge graph and the pre-established inference rule base, the method further includes: The data in the knowledge graph corresponding to the secondary equipment of the substation is converted into a three-dimensional tensor, and the three-dimensional tensor is decomposed into the outer product form of three factor matrices using the tensor decomposition method to obtain a low-dimensional embedding representation of the three-dimensional tensor; wherein the two dimensions of the three-dimensional tensor correspond to the head entity and the tail entity respectively, and the third dimension represents the relationship type. Inputting the low-dimensional embedding representation into a transposed convolutional network to map the low-dimensional embedding representation back to the original dimensional space of the knowledge graph through multiple transposed convolutional layers contained in the transposed convolutional network to obtain a target embedding identifier for knowledge completion, and updating the knowledge graph according to the target embedding identifier; The equipment state change sequence is extracted from the updated knowledge graph, and the equipment state change sequence is divided into multiple subsequence fragments containing at least one identical state node through a sliding window method. A sequence pattern mining algorithm is applied to the subsequence fragments to identify multiple target state transfer patterns, and an inference rule library containing equipment state evolution laws is established based on the multiple target state transfer patterns; wherein the equipment state change sequence includes the state nodes and transfer relationships of the substation secondary equipment at different time points.

10. A fault detection device for secondary equipment of a substation, characterized in that: include: An equipment information acquisition module is used to collect real-time operating parameters of substation secondary equipment, input the real-time operating parameters of the substation secondary equipment into a pre-trained graph neural network model to obtain equipment operating characteristics of the substation secondary equipment, and determine the equipment state vector according to multiple equipment operating characteristics in a continuous time window and a message transmission mechanism; A test data acquisition module, used to determine the equipment health score according to the equipment state vector and a pre-built knowledge graph, and trigger the intelligent inspection robot to perform a deep detection task to collect target test data when the health score is lower than a preset health threshold, wherein the knowledge graph is constructed based on historical test data, operating parameter data and fault maintenance records of the substation secondary equipment; An equipment maintenance guidance module is used to determine the target fault type corresponding to the substation secondary equipment based on the target test data, the knowledge graph and a pre-established inference rule base, obtain a fault handling plan corresponding to the historical fault case corresponding to the target fault type, and generate maintenance guidance information corresponding to the substation secondary equipment based on the fault handling plan.

Images