Power transmission and transformation equipment fault repair method, system, equipment and medium

Abstract

The invention discloses a power transmission and transformation equipment fault repair method, system, equipment and medium, and relates to the technical field of power transmission and transformation equipment overhaul, and the method comprises the steps: obtaining multi-modal time sequence data, obtaining a multi-modal fusion feature vector through employing an encoder in combination with a cross attention mechanism, outputting a fault type probability after processing, and carrying out the fault type maintenance of the multi-modal fusion feature vector; fusing and splicing the multi-modal fusion feature vector and the collected state data of the power transmission and transformation equipment to obtain an initial feature vector, constructing a topological graph based on physical equipment, processing the topological graph and the initial feature vector through a graph neural network, obtaining a component fault contribution degree in combination with an attention mechanism, and obtaining a fault component list and a fault cause. And generating and outputting a repair strategy for the fault component list and the fault reason according to the large language model. The method solves the problems that a traditional fault repairing method is poor in real-time performance, high in false alarm and missing alarm rate, inaccurate in fault positioning and depends on artificial experience decision making.

Description

Technical Field

[0001] This invention relates to the field of power transmission and transformation equipment maintenance technology, specifically to a method, system, equipment, and medium for repairing power transmission and transformation equipment faults. Background Technology

[0002] The existing technical means for fault diagnosis and repair of power transmission and transformation equipment are mainly divided into three categories: First, regular inspection and offline testing, in which maintenance personnel conduct power outage tests and manual inspections of key equipment such as transformers, circuit breakers, and GIS at fixed intervals, and judge the health status of the equipment based on test data such as insulation resistance, dielectric loss, and SF6 composition; Second, online monitoring of single parameter alarms, in which online monitoring terminals for temperature, partial discharge, vibration, etc. are installed on the equipment, and an alarm is triggered when a single monitored quantity exceeds a set threshold; Third, expert systems or rule bases, which use "IF-THEN" rules to perform simple logical judgments on the monitored quantities and output qualitative conclusions.

[0003] However, these technologies have significant drawbacks: First, they lack real-time performance, as periodic inspections and offline tests cannot continuously acquire equipment status, making it easy for faults to occur and escalate within the interval between inspections. Second, they have high false alarm and false negative rates, as single-parameter alarms are easily affected by environmental interference, fixed thresholds lead to false alarms, and it is difficult to capture multi-parameter coupled faults, resulting in false negatives. Third, their root cause localization is coarse, as existing rule bases can only provide vague conclusions such as "suspected overheating" or "suspected discharge," failing to accurately locate the fault location and mechanism, leading to blind maintenance and prolonged power outages. Fourth, their repair strategies are static, as existing systems mostly remain at the alarm level, lacking root cause-based repair strategy recommendations, requiring maintenance personnel to make decisions based on experience, which poses security risks.

[0004] In summary, existing technologies have limitations in real-time performance, accuracy, positioning precision, and strategy generation, making it difficult to meet the needs of efficient operation and maintenance of power transmission and transformation equipment. There is an urgent need for a new technology that can integrate multi-source data in real time, accurately locate the root cause of faults, and automatically generate repair strategies. Summary of the Invention

[0005] In view of the above-mentioned problems, the present invention is proposed.

[0006] Therefore, the technical problem solved by this invention is: how to provide a fault repair method for power transmission and transformation equipment, which involves acquiring multimodal time-series data, constructing a multimodal fusion feature vector to output fault type probabilities, fusing and splicing this vector with collected power transmission and transformation equipment status data to obtain an initial feature vector, constructing a topology map based on physical equipment, obtaining a list of faulty components and fault causes through a graph neural network combined with an attention mechanism, and generating and outputting a repair strategy based on the list of faulty components and fault causes using a large language model. This effectively improves fault identification and location accuracy, reduces manual decision-making, increases maintenance efficiency, and lowers operation and maintenance costs.

[0007] To address the aforementioned technical problems, this invention provides the following technical solution: a method for repairing faults in power transmission and transformation equipment, comprising the following steps: acquiring multimodal time-series data; obtaining a multimodal fusion feature vector based on the multimodal time-series data using an encoder combined with a cross-attention mechanism; outputting a fault type probability after processing the multimodal fusion feature vector; fusing and concatenating the multimodal fusion feature vector with collected power transmission and transformation equipment status data to obtain an initial feature vector; constructing a topology graph based on physical equipment; processing the topology graph and the initial feature vector using a graph neural network; obtaining the component fault contribution rate using an attention mechanism; obtaining a list of faulty components and fault causes; and generating and outputting a repair strategy based on the list of faulty components and fault causes using a large language model.

[0008] In a preferred embodiment of the fault repair method for power transmission and transformation equipment described in this invention, a multimodal fusion feature vector is obtained based on the multimodal time-series data using an encoder combined with a cross-attention mechanism. This includes: using the time-series features extracted from the multimodal time-series data by the encoder, calculating the correlation weights of the time-series features using the cross-attention mechanism, fusing them, and obtaining the multimodal fusion feature vector. The beneficial effect of this preferred embodiment is that it can capture the correlation of multimodal data, highlight key features through weight allocation, and integrate complementary information into the fusion vector, thereby improving the accuracy of fault identification.

[0009] As a preferred embodiment of the fault repair method for power transmission and transformation equipment described in this invention, the multimodal fusion feature vector is processed to output the fault type probability, including: the multimodal fusion feature vector is processed by global average pooling and then passed through a fully connected layer and an activation function to output the fault type probability.

[0010] As a preferred embodiment of the fault repair method for power transmission and transformation equipment according to the present invention, the method involves constructing a topology graph based on physical equipment, processing the topology graph and the initial feature vector using a graph neural network, and obtaining the component fault contribution degree using an attention mechanism. This includes: constructing the topology graph with equipment components as nodes and physical connections as edges; processing the topology graph using the graph neural network layer in conjunction with the initial feature vector to obtain an aggregated feature vector; calculating the correlation between each component and the current fault using a node-level attention mechanism; and outputting a list of faulty components and the cause of the fault. The beneficial effects of this preferred embodiment are that constructing the topology graph with components as nodes and physical connections as edges closely matches the actual structure of the equipment, facilitating the aggregation of component association information by the graph neural network; and that the node-level attention mechanism calculates the correlation between each component and the fault, overcoming the problem of coarse localization in existing technologies and reducing costs and risks.

[0011] As a preferred embodiment of the fault repair method for power transmission and transformation equipment described in this invention, the method involves generating and outputting a repair strategy based on the list of faulty components and the causes of the fault according to a large language model. This includes: training the large language model by injecting domain knowledge of power transmission and transformation equipment through low-rank fine-tuning, outputting the repair strategy described in natural language, and generating a visual chart of the strategy for output. The beneficial effects of this preferred embodiment are that injecting domain knowledge through low-rank fine-tuning enhances the professionalism and relevance of the repair strategy of the large language model; the natural language description and visual chart facilitate understanding and execution, enabling maintenance personnel to quickly master the strategy and improving maintenance efficiency and standardization.

[0012] As a preferred embodiment of the fault repair method for power transmission and transformation equipment described in this invention, the multimodal fusion feature vector is processed by global average pooling and then outputs the probability of the fault type through an activation function. This includes: converting the multimodal fusion feature vector into fixed features through global average pooling, mapping the fixed features to various components through the fully connected layer, and calculating the probability of the fault type using the Softmax activation function.

[0013] As a preferred embodiment of the fault repair method for power transmission and transformation equipment described in this invention, the acquisition of multimodal time-series data includes: obtaining the multimodal time-series data by collecting equipment operation data and environmental data, and then performing noise reduction, normalization and time-series alignment processing.

[0014] This invention provides a fault repair system for power transmission and transformation equipment.

[0015] To address the aforementioned technical problems, the present invention further provides the following technical solution: a power transmission and transformation equipment fault repair system, comprising: a data acquisition and analysis module, which acquires multimodal time-series data, and obtains a multimodal fusion feature vector based on the multimodal time-series data using an encoder combined with a cross-attention mechanism, and outputs a fault type probability after processing the multimodal fusion feature vector; a fault feature fusion module, which fuses and splices the multimodal fusion feature vector with the acquired power transmission and transformation equipment status data to obtain an initial feature vector; a fault tracing module, which constructs a topology map based on physical equipment, and uses a graph neural network and an attention mechanism to locate faulty components and the fault type probability based on the initial feature vector, and obtains a list of faulty components and fault causes; and a strategy repair module, which generates and outputs a repair strategy based on the list of faulty components and fault causes using a large language model.

[0016] The present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method for repairing faults in power transmission and transformation equipment.

[0017] The present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the steps of the method for repairing faults in power transmission and transformation equipment.

[0018] The beneficial effects of this invention are as follows: By collecting data and utilizing an encoder combined with a cross-attention mechanism to deeply fuse multimodal temporal features, the invention overcomes the lag and single-parameter monitoring limitations of traditional periodic inspections, enabling early warning of faults. A topology graph is constructed using equipment components as nodes and physical connections as edges. This graph neural network aggregates component association information, and a node-level attention mechanism calculates the correlation between each component and the fault, clearly identifying the faulty component and its cause. This solves the problem of blind maintenance caused by the coarse positioning of traditional methods, providing a basis for targeted maintenance. Furthermore, by injecting knowledge from the power transmission and transformation field into the large language model through low-rank fine-tuning, the model can generate natural language repair strategies that include operational steps and visual charts, making it easier for maintenance personnel to understand and improving the standardization and efficiency of maintenance. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a general flowchart of a fault repair method for power transmission and transformation equipment provided in one embodiment of the present invention. Detailed Implementation

[0021] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0022] Example 1, referring to Figure 1 This is the first embodiment of the present invention, which provides a method for repairing faults in power transmission and transformation equipment, including:

[0023] S100: Acquire multimodal time series data. Based on the multimodal time series data, use an encoder combined with a cross-attention mechanism to obtain a multimodal fusion feature vector. After processing, the multimodal fusion feature vector outputs the probability of the fault type.

[0024] S200: The multimodal fusion feature vector is fused and spliced ​​with the collected power transmission and transformation equipment status data to obtain the initial feature vector.

[0025] S300: Constructs a topology map based on physical devices, processes the topology map and initial feature vectors through a graph neural network, and obtains the component failure contribution by combining an attention mechanism, thereby obtaining a list of faulty components and the cause of failure.

[0026] S400: Generates and outputs repair strategies based on the list of faulty components and the causes of faults according to the large language model.

[0027] It should be noted that traditional technologies rely on periodic inspections and offline testing, requiring power outages for testing at fixed intervals. This makes it impossible to continuously acquire equipment status data, and faults can easily occur and escalate between inspections, increasing the difficulty and losses of emergency repairs. Online monitoring single-parameter alarms trigger alarms based on fixed thresholds for a single monitored quantity, making them susceptible to environmental interference and false alarms. Complex faults caused by multi-parameter coupling are often overlooked because they exceed the scope of a single criterion, interfering with maintenance decisions. Existing rule bases can only provide qualitative conclusions such as "suspected overheating" or "suspected discharge," failing to accurately locate faulty components and causes, resulting in highly blind maintenance and requiring extensive investigations, increasing maintenance costs. Furthermore, existing technologies mostly remain at the alarm level, failing to generate specific repair solutions based on the root cause of the fault, relying on the experience of maintenance personnel for decision-making, reducing the effectiveness of maintenance.

[0028] Therefore, addressing the aforementioned issues of insufficient real-time performance, false alarms and missed alarms, ambiguous localization, and lack of targeted repair strategies, a fault repair method for power transmission and transformation equipment is constructed through steps S100-S400. First, multimodal time-series data is acquired, and an encoder combined with a cross-attention mechanism is used to obtain a multimodal fusion feature vector, which is then processed to output the fault type probability. Second, the multimodal fusion feature vector is fused and concatenated with the collected power transmission and transformation equipment status data to obtain an initial feature vector. Next, a topology map is constructed based on the physical equipment, and a graph neural network is used to process the topology map and the initial feature vector, combined with an attention mechanism to obtain the component fault contribution, resulting in a list of faulty components and fault causes. Finally, a repair strategy is generated and output based on the list of faulty components and fault causes using a large language model, achieving accurate fusion of multimodal time-series data and equipment data.

[0029] Example 2, refer to Figure 1 This is the second embodiment of the present invention, which provides a method for repairing faults in power transmission and transformation equipment.

[0030] In this embodiment of the invention, step S100 involves acquiring multimodal time-series data, obtaining a multimodal fusion feature vector based on the multimodal time-series data using an encoder combined with a cross-attention mechanism, and then outputting the fault type probability after processing the multimodal fusion feature vector. This includes the following steps A1-A3:

[0031] A1: Collect equipment operation data and environmental data, and perform noise reduction, normalization and time-series alignment processing to obtain the multimodal time-series data.

[0032] A2: Extracting temporal features from multimodal time series data using an encoder.

[0033] A3: Calculate the correlation weights of temporal features through a cross-attention mechanism and perform deep fusion.

[0034] In step A2, the encoder uses a 6-layer stack to achieve deep feature extraction, a 256-dimensional vector to balance expressive power and complexity, and an 8-head attention mechanism to enhance the ability to capture multi-dimensional information.

[0035] Specifically, in step A3, the fusion process includes the following steps A31-A34:

[0036] A31: After calculation using the cross-attention mechanism, a multimodal fusion feature vector with a dimension of 256 is obtained.

[0037] A32: Perform global average pooling on the temporal dimension to transform the multimodal fusion feature vector into a fixed feature.

[0038] A33: The 256-dimensional features are mapped to the number of fault types through a fully connected layer.

[0039] A34: The Softmax activation function is used to normalize the fully connected layer and output the probability distribution of fault types.

[0040] Specifically, the probability distribution of fault types in step A34 is represented as follows:

[0041] P_health∈[0,1]^K;

[0042] Where P_health is the probability of fault type and K is the number of fault types, this formula represents the probability distribution characteristics of fault types in power transmission and transformation equipment.

[0043] In one implementation, temporal features can be extracted from multimodal time-series data using an encoder, or through a temporal convolutional network. In this network, causal convolution follows the temporal sequence, preventing future information from interfering with the current feature extraction and ensuring the rigor of the temporal logic. Dilated convolution expands the receptive field without increasing parameters by setting different dilation coefficients, effectively capturing long-period temporal dependencies. After processing the multimodal time-series data of power transmission and transformation equipment in this way, cross-modal correlation features are fused using a cross-attention mechanism, and then processed by a fully connected layer mapping and activation function to output accurate fault probabilities.

[0044] In another possible implementation, the extraction of temporal features from multimodal time-series data using an encoder can also be achieved through a Long Short-Term Memory (LSTM) network. In its gating mechanism, the forget gate filters redundant historical data from equipment operation, the input gate selectively retains key state information, and the output gate integrates effective features to stably capture the long-term dependencies of multimodal time-series data. For the temporal features extracted from the long-term operation of power transmission and transformation equipment, multimodal information is fused using cross-attention, and then processed to output the fault probability.

[0045] In this embodiment of the invention, in step S200, the multimodal fusion feature vector is fused and spliced ​​with the collected power transmission and transformation equipment status data to obtain the initial feature vector.

[0046] It should be noted that the encoder and cross-attention mechanism in step S100 were reused to fuse and splice the multimodal fusion feature vector with the collected power transmission and transformation equipment status data.

[0047] Specifically, the multimodal fusion feature vector is a high-dimensional feature of multimodal time series data after processing. The status data of power transmission and transformation equipment includes equipment model and operating years information. The two are concatenated according to the feature dimension to form an initial feature vector that contains both multimodal time series features and equipment status information, providing input for fault tracing based on graph neural networks in S300.

[0048] In one implementation, the fusion and concatenation can also be achieved through feature mapping fusion. This involves mapping two types of features to the same dimension using two independent fully connected layers before concatenation. The fully connected layers can adaptively adjust their parameters based on the characteristics of the multimodal fusion features and device state data, reducing feature information loss during the mapping process. This preserves the dynamic changes in the multimodal data while fully transmitting the static attributes of the device. By mapping the multimodal fusion features and state data to the same dimension and then concatenating them into an initial feature vector, the problem of feature dimensionality differences is resolved, satisfying the dimensionality requirements of the subsequent graph neural network for the input features.

[0049] In another possible implementation, the fusion and splicing can also be achieved through feature-weighted fusion. Weights are assigned to the multimodal fusion feature vector and the power transmission and transformation equipment status data, and then summed. The weights are learned through model training. By learning the importance weights of the equipment status data and multimodal features, the data is fused according to the weight ratio, preserving temporal features while highlighting key state parameters, thus providing effective input for the graph neural network.

[0050] In this embodiment of the invention, step S300 involves constructing a topology graph based on physical devices, processing the topology graph and initial feature vectors using a graph neural network, and obtaining the component failure contribution rate using an attention mechanism to obtain a list of faulty components and their causes. This includes the following steps B1-B4:

[0051] B1: Construct an equipment topology graph using the core components of power transmission and transformation equipment as nodes and the physical connections between components as edges.

[0052] B2: Use the initial feature vector generated in step S200 as the node feature input of the graph neural network.

[0053] B3: A three-layer graph neural network model is used to perform graph convolution operations on the topological graph.

[0054] B4: Calculate the contribution of each component to the fault through a node-level attention mechanism.

[0055] For example, in step B2, the features of each node include multimodal monitoring data features, device static parameters and cross-modal correlation features, ensuring that the features of each component are comprehensive.

[0056] Specifically, in step B3, a three-layer graph neural network model is adopted, including the first layer, where each node initially captures local correlation information by aggregating its direct neighbors; the second and third layers gradually expand the aggregation range to realize the modeling of the indirect influence between components in the overall structure of the equipment.

[0057] It should be noted that the graph neural network model has a hidden dimension of 128 in each layer, and finally outputs a 4-dimensional node feature vector. Through multi-layer iteration, the local and global correlation information of the components is incorporated into the features.

[0058] Specifically, in step B4, the node-level attention mechanism assigns weights to each node. The weight calculation is based on the matching degree between node features and global fault features. The higher the weight, the stronger the correlation between the component and the fault. Finally, a sorted list of faulty components and the corresponding fault causes are output, achieving accurate localization from fault type to specific component and root cause.

[0059] In one implementation, the node-level attention mechanism can also calculate the contribution using a multilayer perceptron attention mechanism, which learns the importance weights of component features. This mechanism performs a nonlinear transformation on the initial feature vector through stacked hidden layers, enhancing the expressive power of key features. Simultaneously, it adjusts the weight allocation based on the physical correlation strength of components in the topological connection relationship, making the contribution calculation of adjacent highly correlated components more closely reflect the actual fault propagation pattern. By automatically learning the influence weights of each component on the fault based on the initial feature vector through the multilayer perceptron attention mechanism, and optimizing the weight allocation in conjunction with the topological connection relationship, the contribution calculation of key components is strengthened.

[0060] In another possible implementation, the node-level attention mechanism can also calculate the contribution through a graph attention network. By using a self-attention mechanism, weights are assigned to the neighbors of each node in the topology graph. The weight calculation is based on the similarity of node features, which can adapt to complex topologies such as series and parallel connections in power transmission and transformation equipment, capture the dependencies between components, and thus calculate the contribution of each component to the fault. This adapts to the topology of power transmission and transformation equipment and improves the accuracy of the list of faulty components and their causes.

[0061] In this embodiment of the invention, step S400 generates and outputs a repair strategy based on the list of faulty components and the causes of the fault according to the large language model, specifically including the following steps C1-C3:

[0062] C1: Collect multi-dimensional key information as input to the large language model, including fault location results, basic equipment information, real-time operating status, and historical maintenance records.

[0063] C2: The repair strategy is generated using the LLaMA-7B model based on low-rank adaptation fine-tuning.

[0064] C3: The generated repair strategy is output through the visualization module after format verification.

[0065] Specifically, in step C2, the repair strategy is generated using the LoRA-based fine-tuned LLaMA-7B model, including the following steps C21 and C22:

[0066] C21: Fine-tune the LLaMA-7B model using LoRA technology to adapt it to the professional terminology and operation and maintenance rules in the field of power transmission and transformation equipment.

[0067] C22: The model generates repair strategies based on input instructions, combining pre-trained power operation and maintenance knowledge with fine-tuned domain experience, and outputs visual charts.

[0068] It should be noted that in step C21, the LLaMA-7B model is a large language model; the rank parameter `rank` is set to 16, indicating that the dimension of the low-rank matrix in the model weight matrix decomposition during LoRA fine-tuning is 16; the scaling parameter `alpha` is set to 32. The scaling parameter dynamically adjusts the influence of the low-rank matrix on the model output. The larger the value, the weaker the contribution of the low-rank matrix, and the closer the model is to the original pre-trained weights; the smaller the value, the stronger the influence of domain fine-tuning. The original weight matrix of the large language model has extremely high dimensionality, and directly fine-tuning all weights would lead to excessive computation and a high risk of overfitting. LoRA can decompose the weight update into two low-rank matrices.

[0069] LoRA, or low-rank adaptation for large language models, is a commonly used method for efficient parameter fine-tuning. Based on the pre-trained model weights, trainable low-rank decomposition matrices are injected into each layer of the encoder architecture to reduce the training load for subsequent tasks.

[0070] In one implementation, the visualization chart generated in step C22 can also be generated through 3D topology visualization. Based on the topology diagram construction logic, the components of the power transmission and transformation equipment are presented as 3D models, supporting interactive operations. The contribution of faulty components is intuitively marked using gradients of red, orange, and yellow, while also annotating the electrical details between components. This method not only intuitively displays the three-dimensional structure of the equipment and the distribution of faults, but also dynamically associates with the list of faulty components, ensuring the need for combining the topology diagram with the list of faulty components, making it easier for maintenance personnel to understand component relationships and the scope of fault impact.

[0071] In another possible implementation, the visualization chart generated in step C22 can also be generated through a repair strategy flowchart. Based on the repair strategy generated by the large language model, the operation steps are transformed into a flowchart, annotating the locations of spare parts, safety points, and related components corresponding to each step. This method visualizes the natural language strategy, reduces the understanding cost for maintenance personnel, and improves the standardization of repair operations.

[0072] In summary, this invention, by collecting data and utilizing an encoder combined with a cross-attention mechanism to deeply fuse multimodal temporal features, overcomes the lag and limitations of traditional periodic inspections and single-parameter monitoring, enabling early warning of faults. It constructs a topology graph using equipment components as nodes and physical connections as edges, aggregates component association information using a graph neural network, and then calculates the correlation between each component and the fault through a node-level attention mechanism, clearly identifying the faulty component and its cause. This solves the problem of blind maintenance caused by the coarse positioning of traditional methods, providing a basis for targeted maintenance. Furthermore, by injecting knowledge from the power transmission and transformation domain into a large language model through low-rank fine-tuning, it enables the model to generate natural language repair strategies that include operational steps and visual charts, facilitating understanding by maintenance personnel and improving maintenance standardization and efficiency.

[0073] Example 3, referring to Figure 1This is the third embodiment of the present invention, which provides a fault repair system for power transmission and transformation equipment, including a data acquisition and analysis module that acquires multimodal time-series data, obtains a multimodal fusion feature vector based on the multimodal time-series data using an encoder combined with a cross-attention mechanism, and outputs the fault type probability after processing the multimodal fusion feature vector; a fault feature fusion module that fuses and splices the multimodal fusion feature vector with the acquired power transmission and transformation equipment status data to obtain an initial feature vector; a fault tracing module that constructs a topology map based on physical equipment, uses a graph neural network and an attention mechanism to locate faulty components and fault type probabilities based on the initial feature vector, and obtains a list of faulty components and fault causes; and a strategy repair module that generates and outputs repair strategies based on the list of faulty components and fault causes using a large language model.

[0074] Example 4, the fourth embodiment of the present invention, differs from the previous three embodiments in that: if the function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0075] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0076] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0077] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0078] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for repairing faults in power transmission and transformation equipment, characterized in that: include, Acquire multimodal time series data, and based on the multimodal time series data, use an encoder combined with a cross-attention mechanism to obtain a multimodal fusion feature vector. After processing, the multimodal fusion feature vector outputs the probability of fault type. The multimodal fusion feature vector is fused and spliced ​​with the collected power transmission and transformation equipment status data to obtain an initial feature vector; A topology graph is constructed based on physical devices. The topology graph and the initial feature vector are processed by a graph neural network. An attention mechanism is used to obtain the component failure contribution and to obtain a list of faulty components and the cause of failure. Based on the large language model, a repair strategy is generated and output for the list of faulty components and the causes of the fault.

2. The method for repairing faults in power transmission and transformation equipment as described in claim 1, characterized in that, Based on the multimodal temporal data, a multimodal fusion feature vector is obtained using an encoder combined with a cross-attention mechanism, including: The temporal features extracted from the multimodal temporal data by the encoder are combined with the cross-attention mechanism to calculate the correlation weight of the temporal features, and then fused to obtain a multimodal fusion feature vector.

3. The method for repairing faults in power transmission and transformation equipment as described in claim 2, characterized in that, The multimodal fusion feature vector, after processing, outputs the probability of fault type, including: The multimodal fusion feature vector is processed by global average pooling and then outputs the probability of the fault type through a fully connected layer and an activation function.

4. The method for repairing faults in power transmission and transformation equipment as described in claim 3, characterized in that, A topology graph is constructed based on physical devices. The topology graph and the initial feature vector are processed using a graph neural network. An attention mechanism is then used to obtain the component failure contribution, including: Using equipment components as nodes and physical connections as edges, a topology graph is constructed. The graph neural network layer is used to process the topology graph in conjunction with the initial feature vector to obtain an aggregated feature vector. A node-level attention mechanism is used to calculate the aggregated feature vector to obtain the correlation between each component and the current fault, and the list of faulty components and the cause of the fault are output.

5. A method for repairing faults in power transmission and transformation equipment as described in claim 4, characterized in that, Based on the large language model, a repair strategy is generated and output for the list of faulty components and the causes of the faults, including: By injecting domain knowledge of power transmission and transformation equipment through LoRA, the large language model is trained, outputting a natural language description of the repair strategy, and generating a visual chart of the strategy for output.

6. The method for repairing faults in power transmission and transformation equipment as described in claim 3, characterized in that, The multimodal fused feature vector, after global average pooling, outputs the probability of the fault type through an activation function, including: The multimodal fusion feature vector is transformed into fixed features through global average pooling, and then the fixed features are mapped to each component through the fully connected layer. The probability of the fault type is calculated using the Softmax activation function.

7. A method for repairing faults in power transmission and transformation equipment as described in claim 1, characterized in that, Acquire multimodal time series data, including: The multimodal time series data is obtained by collecting equipment operation data and environmental data, and then processing them through noise reduction, normalization and time alignment.

8. A fault repair system for power transmission and transformation equipment, employing a fault repair method for power transmission and transformation equipment as described in any one of claims 1 to 7, characterized in that, include: The data acquisition and analysis module acquires multimodal time-series data, and obtains a multimodal fusion feature vector based on the multimodal time-series data using an encoder combined with a cross-attention mechanism. The multimodal fusion feature vector is then processed to output the probability of the fault type. The fault feature fusion module fuses and splices the multimodal fusion feature vector with the collected power transmission and transformation equipment status data to obtain an initial feature vector; The fault tracing module constructs a topology map based on physical devices, uses graph neural networks and attention mechanisms to locate faulty components and the probability of the fault type according to the initial feature vector, and obtains a list of faulty components and fault causes. The strategy repair module generates and outputs repair strategies based on the list of faulty components and the causes of the faults, according to the large language model.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the fault repair method for power transmission and transformation equipment as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the fault repair method for power transmission and transformation equipment as described in any one of claims 1 to 7.

Images