Transformer substation fault diagnosis method, device and system based on multi-modal fusion

Abstract

The invention provides a transformer substation fault diagnosis method, device and system based on multi-modal fusion. The method comprises the steps that multi-modal fault information generated in the operation process of a transformer substation is acquired; based on a pre-trained fault diagnosis model, obtaining a fault diagnosis result of the transformer substation according to the multi-modal fault information; wherein the fault diagnosis model comprises a feature extraction network based on a graph attention network and a feature fusion network based on a self-attention mechanism, the feature extraction network is obtained by introducing a denoising self-encoding model for pre-training, and the feature fusion network is obtained by training and optimizing set loss functions under different fault classification tasks. According to the method, feature extraction and fusion of the multi-modal fault information are realized by comprehensively utilizing the multi-modal fault information and utilizing the graph attention network and the self-attention mechanism, the limitation that a single information source is difficult to comprehensively reflect the state change of a transformer substation system is overcome, and the accuracy and comprehensiveness of primary and secondary fault diagnosis of the transformer substation are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of substation fault diagnosis, and in particular to a substation fault diagnosis method, device and system based on multi-modal fusion. Background Art

[0002] As a key node in the power system, the safety and stability of substation operation directly affect the reliability of power supply. Therefore, it is crucial to quickly and accurately diagnose substation faults and take effective countermeasures in time to ensure the normal operation of the power system. As substations gradually move towards digitalization and intelligence, the number of secondary equipment has increased significantly, and the system structure has become increasingly complex, making fault diagnosis face unprecedented challenges. When a fault occurs, the substation system can collect a large amount of fault data in real time. How to fully mine and utilize this data to improve the accuracy and fault tolerance of diagnosis has become an important research topic in substation fault diagnosis.

[0003] Existing substation fault diagnosis methods are mainly divided into two categories: rule-based fault diagnosis methods and data-driven fault diagnosis methods. Rule-based fault diagnosis methods, such as expert systems, Bayesian networks, fuzzy set theory, etc., mainly rely on the action information of relay protection and circuit breakers for diagnosis. Despite continuous improvements, it is difficult to eliminate the influence of uncertain factors such as refusal or malfunction of these devices and missing or false alarm information. Single switch quantity information cannot meet the needs of high-accuracy diagnosis.

[0004] In recent years, researchers have begun to introduce electrical quantity information, such as current and voltage, as a basis for diagnosis, and proposed a multimodal fusion fault diagnosis method. This method makes up for the shortcomings of a single information source to a certain extent by leveraging the completeness of electrical quantity information and the applicability of switch quantity information, but it still has limitations. The rule-based multimodal fusion method relies on specific fault feature extraction and model assumptions, cannot fully adapt to complex or unknown fault modes, and the cost of model construction and optimization is high.

[0005] On the other hand, data-driven fault diagnosis methods, such as machine learning and deep learning, automatically identify fault modes by analyzing historical operating data. These methods have certain learning capabilities, but usually rely on a large number of fault samples, while fault events in actual power system operation are relatively scarce, especially some complex or rare fault modes. This makes it difficult for data-driven methods to adapt to real-world scenarios and give full play to their advantages.

[0006] In summary, the existing fault diagnosis methods have certain limitations and are difficult to meet the requirements of modern power systems for high accuracy, robustness and adaptability.

[0007] Therefore, how to solve the problem that the existing fault diagnosis methods are applicable to a single information source and cannot fully express the system state changes, and the fault diagnosis accuracy is not high, is an important issue that needs to be solved urgently in the field of substation fault diagnosis. Summary of the invention

[0008] The present invention provides a substation fault diagnosis method, device and system based on multimodal fusion, which are used to overcome the defects of existing fault diagnosis methods that a single information source is difficult to fully express system state changes and the fault diagnosis accuracy is not high, fully explore the potential correlation and complementarity of multi-source information, and improve the fault diagnosis reliability and fault tolerance of the substation system.

[0009] On the one hand, the present invention provides a substation fault diagnosis method based on multimodal fusion, including: obtaining multimodal fault information generated by the substation during operation; based on a pre-trained fault diagnosis model, obtaining the fault diagnosis result of the substation according to the multimodal fault information; wherein the fault diagnosis model includes a feature extraction network based on a graph attention network and a feature fusion network based on a self-attention mechanism, the feature extraction network is obtained by pre-training by introducing a denoising autoencoder model, and the feature fusion network is obtained by training and optimizing by setting loss functions under different fault classification tasks.

[0010] Furthermore, the multimodal fault information at least includes alarm information, fault recording information and SCADA measurement information; the number of the feature extraction networks is consistent with the number of categories of the multimodal fault information, and the number of the feature fusion networks is consistent with the number of fault classification tasks; wherein the feature extraction network is used to extract the fault characteristics of the multimodal fault information, and the feature fusion network is used to fuse the fault characteristics of the multimodal fault information; the fault classification task includes one or more combinations of a fault location diagnosis task, a fault type diagnosis task and a secondary equipment refusal to operate diagnosis task, and the importance of each category of multimodal fault information is different under different fault classification tasks.

[0011] Furthermore, the fault diagnosis model also includes a classifier; accordingly, the fault diagnosis model based on pre-training obtains the fault diagnosis result of the substation according to the multimodal fault information, including: based on a pre-trained feature extraction network, extracting the first fault feature of the alarm information, the second fault feature of the fault recording information and the third fault feature of the SCADA measurement information; splicing the first fault feature, the second fault feature and the third fault feature to obtain a fault feature matrix; based on a pre-trained feature fusion network, according to the fault feature matrix, obtaining the fused feature representation corresponding to different fault classification tasks; based on the pre-trained classifier, according to the fused feature representation corresponding to different fault classification tasks, predicting the fault diagnosis results corresponding to different fault classification tasks.

[0012] Furthermore, the denoising autoencoder model includes an encoder and a decoder based on a graph attention network; correspondingly, training and optimizing the feature extraction network specifically includes: collecting multi-modal fault information samples and obtaining the original feature representations corresponding to the multi-modal fault information samples; based on the encoder, encoding the randomly masked multi-modal fault information samples and splicing them to obtain a multi-source encoding representation; based on the decoder, reconstructing the data according to the multi-source encoding representation to obtain a reconstructed feature representation; minimizing the error between the original feature representation and the reconstructed feature representation through a mean square error loss function, iteratively optimizing the encoder and the decoder until convergence; and using the encoder trained to convergence as the feature extraction network.

[0013] Furthermore, training and optimizing the feature fusion network specifically includes: in the case where the fault classification task is a fault location diagnosis task and / or a fault type diagnosis task, training the feature fusion network using a supervised contrast loss function, and then training the classifier using a cross-entropy loss function; in the case where the fault classification task is a secondary equipment refusal-to-operate diagnosis task, training the feature fusion network and the classifier simultaneously using a binary cross-entropy loss function.

[0014] Furthermore, obtaining the multi-modal fault information generated during the operation of the substation, and then includes: encoding the multi-modal fault information through an embedding layer to complete preliminary feature extraction and standardization processing to obtain initial features; correspondingly, based on a pre-trained fault diagnosis model, obtaining the fault diagnosis result of the substation according to the multi-modal fault information, including: using the initial features as the input of the pre-trained fault diagnosis model to obtain the fault diagnosis result of the substation.

[0015] In a second aspect, the present invention also provides a substation fault diagnosis device based on multi-modal fusion, including a multi-modal fault information acquisition module for acquiring multi-modal fault information generated during the operation of the substation; a substation fault diagnosis module for obtaining the fault diagnosis result of the substation according to the multi-modal fault information based on a pre-trained fault diagnosis model; wherein, the fault diagnosis model includes a feature extraction network based on a graph attention network and a feature fusion network based on a self-attention mechanism, the feature extraction network is pre-trained by introducing a denoising autoencoder model, and the feature fusion network is trained and optimized through a set loss function under different fault classification tasks.

[0016] In a third aspect, the present invention further provides a substation fault diagnosis system based on multimodal fusion, including: an actual physical system and a virtual simulation system that establish a two-way connection, wherein the actual physical system is configured to collect multimodal fault information generated during the operation of the substation and forward the multimodal fault information to the virtual simulation system; the virtual simulation system is constructed based on digital twin technology and is configured to receive the multimodal fault information and execute the substation fault diagnosis method based on multimodal fusion as described in any one of the above.

[0017] In a fourth aspect, the present invention further provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it implements the substation fault diagnosis method based on multimodal fusion as described in any one of the above.

[0018] In a fifth aspect, the present invention further provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the substation fault diagnosis method based on multimodal fusion as described in any one of the above.

[0019] The substation fault diagnosis method based on multimodal fusion provided by the present invention obtains multimodal fault information generated during the operation of the substation and, based on a pre-trained fault diagnosis model, obtains a fault diagnosis result of the substation according to the multimodal fault information; wherein, the fault diagnosis model includes a feature extraction network based on a graph attention network and a feature fusion network based on a self-attention mechanism. The feature extraction network is pre-trained by introducing a denoising autoencoder model, and the feature fusion network is trained and optimized by a set loss function under different fault classification tasks. This method comprehensively utilizes multimodal fault information, uses the graph attention network and the self-attention mechanism to realize the feature extraction and fusion of multimodal fault information, overcomes the limitation that a single information source is difficult to comprehensively reflect the state change of the substation system, and improves the accuracy and comprehensiveness of the primary and secondary fault diagnosis of the substation. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0021] Figure 1 It is a schematic flowchart of the substation fault diagnosis method based on multimodal fusion provided by an embodiment of the present invention.

[0022] Figure 2It is a schematic diagram of the process of a substation failure provided by an embodiment of the present invention.

[0023] Figure 3 It is a schematic diagram of the overall process of a substation fault diagnosis method based on multi-modal fusion provided by an embodiment of the present invention.

[0024] Figure 4 It is a schematic diagram of the training and optimization of a feature extraction network provided by an embodiment of the present invention.

[0025] Figure 5 It is a schematic diagram of the structure of a substation fault diagnosis device based on multi-modal fusion provided by an embodiment of the present invention.

[0026] Figure 6 It is a schematic diagram of the overall architecture of a substation fault diagnosis system based on multi-modal fusion provided by an embodiment of the present invention.

[0027] Figure 7 It is a schematic diagram of the physical structure of an electronic device provided by an embodiment of the present invention. Detailed implementation manners

[0028] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions in the present invention will be clearly and completely described below with reference to the accompanying drawings in the present invention. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present invention without creative efforts shall fall within the protection scope of the present invention.

[0029] Figure 1 It shows a schematic diagram of the process of a substation fault diagnosis method based on multi-modal fusion provided by an embodiment of the present invention.

[0030] As Figure 1 shown, the method includes: S110, obtaining multi-modal fault information generated during the operation of the substation; S120, based on a pre-trained fault diagnosis model, obtaining a fault diagnosis result of the substation according to the multi-modal fault information; wherein, the fault diagnosis model includes a feature extraction network based on a graph attention network and a feature fusion network based on a self-attention mechanism, the feature extraction network is pre-trained by introducing a denoising auto-encoder model, and the feature fusion network is trained and optimized by a set loss function under different fault classification tasks.

[0031] The following will elaborate on steps S110 - S120 and related steps in detail.

[0032] S110, obtaining multi-modal fault information generated during the operation of the substation.

[0033] In a substation, a fault in the primary system is usually accompanied by changes in the physical characteristics of the primary system and the control and protection responses of the secondary system, while generating a large amount of relevant fault data, including alarm information, fault recording information, and SCADA measurement information, etc. These fault data are important bases for analyzing the cause of the fault, locating the fault position, and evaluating the operating state of the system.

[0034] It is easy to understand that in this embodiment, multi-modal fault information generated when a substation fails during operation is obtained, specifically including obtaining alarm information, fault recording information, and SCADA measurement information generated during the operation of the substation.

[0035] Specifically, Figure 2 Fig. shows a schematic diagram of the process of a substation failing provided by an embodiment of the present invention. When a fault occurs in the primary system of the substation, the physical characteristics of the primary equipment will first change significantly. For example, electrical quantities such as current and voltage will mutate. These abnormal changes are sensed by the secondary protection equipment, and the relay protection acts according to the preset protection logic, while sending a trip signal to control the circuit breaker to isolate the faulty equipment or line. If the main protection equipment fails to operate normally due to its own fault or other reasons, the backup protection equipment will take over the protection function.

[0036] During this process, the action information of the protection equipment and the circuit breaker will be recorded as alarm information, including the time stamp of the equipment action, the equipment name, and the action status, etc.; the fault recording equipment records the change of electrical quantities during the fault at a high sampling frequency (i.e., fault recording information, such as the waveform change and dynamic characteristics of current and voltage), which can reflect in detail the electrical quantity characteristics and transient process of the primary equipment. This type of data can accurately capture the change of the physical characteristics of the primary system when the fault occurs; the measurement data from the SCADA (Supervisory Control and Data Acquisition) system (i.e., SCADA measurement information) has a relatively low sampling frequency and mainly records the real-time operating conditions of the power grid (such as steady-state operating information such as current, voltage, frequency, and power), which can reflect the overall operating state of the system and the steady-state characteristics before and after the fault.

[0037] It can be seen that the multi-modal fault information comprehensively reflects the whole process of the fault occurrence and handling from multiple angles, including the mutation of electrical quantities, the response behavior of the protection equipment, the action of the circuit breaker, and the detailed process of fault propagation and isolation, providing multi-dimensional and multi-level support for fault diagnosis.

[0038] It is worth mentioning that a single information source has significant limitations in fault diagnosis and cannot comprehensively characterize the complexity and diversity of fault scenarios. Therefore, in this embodiment, in order to accurately diagnose the fault location, type, and the operation status of secondary protection equipment in the substation system, alarm information, fault recording, measurement data, and the topological structure of the primary system are comprehensively utilized. These information complement each other and jointly describe the occurrence mechanism and propagation law of faults.

[0039] S120, based on a pre-trained fault diagnosis model, obtain the fault diagnosis result of the substation according to the multi-modal fault information; wherein, the fault diagnosis model includes a feature extraction network based on a graph attention network and a feature fusion network based on a self-attention mechanism. The feature extraction network is pre-trained by introducing a denoising auto-encoder model, and the feature fusion network is trained and optimized through a set loss function under different fault classification tasks.

[0040] In this embodiment, a fault diagnosis model is pre-trained, which includes a feature extraction network and a feature fusion network. Among them, the feature extraction network is constructed based on a graph attention network and is used to extract feature expressions with global relevance and discriminability in multi-modal fault information; the feature fusion network is constructed based on a self-attention mechanism and is used to extract fault features from different sources output by the fusion feature extraction network. Subsequently, the fused feature representation passes through a classifier to obtain the final fault diagnosis result.

[0041] The Graph Attention Network (GAT) is a graph neural network model based on the attention mechanism. Its core is to integrate the attention mechanism into the message passing process of the graph neural network. Different from traditional graph neural networks that use fixed adjacency matrix weights, GAT dynamically adjusts the weight of information transmission according to the importance of neighbor nodes to the target node by learning attention parameters, thereby enhancing the modeling ability for graph data.

[0042] By training to learn attention parameters, it can automatically adjust according to the characteristics of graph data during the message passing process, thereby realizing the effective aggregation of adjacent node information. This method not only enhances the ability to capture graph structure information but also improves the applicability of the model to various graph data. When the topological structure of the graph changes, GAT can quickly adapt to the new graph structure by recalculating the attention weights. This adaptive adjustment ability enables GAT to efficiently handle dynamic graphs or complex and diverse graph data. Whether it is the addition or deletion of nodes or the change of edges, GAT can dynamically adjust its model parameters to continue to perform efficient message passing and feature learning.

[0043] The graph attention network in this embodiment is constructed based on the primary topology of the substation. The primary topology of the substation can be represented by an undirected graph It is represented that, where V is a set of nodes, which are the positions where faults may occur in the embodiments, including each transmission line, busbar, and the high- and low-voltage sides of each main transformer; E is a set of edges, representing the electrical connection relationship between nodes; A is an adjacency matrix, describing the topological structure of the graph.

[0044] It is worth mentioning that, in order to better extract the characteristics of each source fault information and reduce the impact of data loss on the fault diagnosis performance, a denoising autoencoder (DAE) pre-trained feature extraction network is introduced in this embodiment. The denoising autoencoder is an unsupervised learning model, mainly used for feature extraction and data reconstruction, and can learn the robust feature representation of data in the case of noise or missing in the input data.

[0045] The self-attention mechanism is a mechanism for capturing the relationships between different elements inside the data. It was first widely used in the field of natural language processing and performed particularly well in the Transformer model. The self-attention mechanism dynamically adjusts the importance of each element by calculating its correlation (i.e., attention weight) with all other elements in the input.

[0046] In this embodiment, the self-attention mechanism can not only model the global associations between multi-source fault features, but also adaptively adjust the importance of each source fault information according to the requirements of different classification tasks, so as to achieve high-precision diagnosis of fault classification tasks such as fault location, fault type, and protection refusal to operate. This mechanism gives full play to the advantages of each source fault information and ensures the flexibility and effectiveness of fault feature fusion.

[0047] It should be noted that the number of feature extraction networks in this embodiment is the same as the number of categories of multi-modal fault information. For example, the multi-modal fault information includes three categories: alarm information, fault recording information, and SCADA measurement information. Correspondingly, three feature extraction networks are set up to extract the fault features in the alarm information, fault recording information, and SCADA measurement information respectively.

[0048] The number of feature fusion networks in this embodiment is the same as the number of fault classification tasks. For example, the fault classification tasks include three tasks: fault location diagnosis task, fault type diagnosis task, and secondary equipment refusal to operate diagnosis task. Correspondingly, three feature fusion networks are set up to fuse and obtain the fused feature representations corresponding to the fault location diagnosis task, fault type diagnosis task, and secondary equipment refusal to operate diagnosis task respectively.

[0049] Among them, each feature extraction network is trained based on each source fault information sample respectively, and each feature fusion network is trained based on the samples corresponding to each fault classification task respectively. This means that the network parameters of each feature extraction network are different, and the network parameters of each feature fusion network are also different. However, the network structures of each feature extraction network can be the same. For example, they are all constructed based on a multi-layer graph attention network; and the network structures of each feature fusion network can also be the same. For example, they are all constructed based on a self-attention mechanism.

[0050] In this embodiment, by obtaining multi-modal fault information generated during the operation of a substation and based on a pre-trained fault diagnosis model, a fault diagnosis result of the substation is obtained according to the multi-modal fault information; among them, the fault diagnosis model includes a feature extraction network based on a graph attention network and a feature fusion network based on a self-attention mechanism. The feature extraction network is pre-trained by introducing a denoising auto-encoder model, and the feature fusion network is trained and optimized by a set loss function under different fault classification tasks. This method comprehensively utilizes multi-modal fault information, and uses a graph attention network and a self-attention mechanism to realize the feature extraction and fusion of multi-modal fault information, overcoming the limitation that a single information source is difficult to comprehensively reflect the state change of the substation system, and improving the accuracy and comprehensiveness of the primary and secondary fault diagnosis of the substation.

[0051] On the basis of the above embodiment, further, the process of the fault diagnosis model predicting the fault diagnosis result will be described in detail below.

[0052] Obtaining a fault diagnosis result of a substation according to multi-modal fault information based on a pre-trained fault diagnosis model includes: extracting a first fault feature of alarm information, a second fault feature of fault recording information, and a third fault feature of SCADA measurement information based on a pre-trained feature extraction network; splicing the first fault feature, the second fault feature, and the third fault feature to obtain a fault feature matrix; obtaining a fusion feature representation corresponding to different fault classification tasks according to the fault feature matrix based on a pre-trained feature fusion network; predicting a fault diagnosis result corresponding to different fault classification tasks according to the fusion feature representation corresponding to different fault classification tasks based on a pre-trained classifier.

[0053] Figure 3 The overall flow diagram of the substation fault diagnosis method based on multi-modal fusion provided by the embodiment of the present invention is shown.

[0054] As Figure 3As shown, for the acquired multimodal fault information, which at least includes alarm information, fault recording information, and SCADA measurement information, they are respectively encoded through an embedding layer to complete preliminary feature extraction and normalization processing, and at the same time map data of different dimensions to a unified feature space for subsequent model processing. Specifically, the alarm information is encoded through the first embedding layer Embedding1 to obtain the first initial feature , the fault recording information is encoded through the second embedding layer Embedding2 to obtain the second initial feature , and the SCADA measurement information is encoded through the third embedding layer Embedding3 to obtain the third initial feature .

[0055] Subsequently, the initial features obtained by encoding are used as the input of the feature extraction network. Through the neighborhood aggregation and adaptive weight allocation mechanism, the fault patterns and implicit relationships between nodes in the primary system topology are deeply mined, so as to extract feature expressions with global relevance and discriminability.

[0056] Specifically, the first initial feature is input into the pre-trained first feature extraction network GAT1 to obtain the first fault feature of the alarm information ; the second initial feature is input into the pre-trained second feature extraction network GAT2 to obtain the second fault feature of the fault recording information ; the third initial feature is input into the pre-trained third feature extraction network GAT3 to obtain the third fault feature of the alarm information .

[0057] The first fault feature , the second fault feature , and the third fault feature are flattened and concatenated to obtain the fault feature matrix .

[0058] Immediately afterwards, the fault feature matrix is feature-fused through the feature fusion network. The self-attention mechanism dynamically assigns weights according to the importance of the feature of each source fault information, which can not only eliminate the interference of redundant information, but also enhance the attention to key features, and fully realize the complementarity and integration of information.

[0059] Specifically, the fault feature matrix is input into the pre-trained first feature fusion network (i.e., self-attention layer 1) to obtain the fused feature representation for the fault location diagnosis task ; the fault feature matrix is input into the pre-trained second feature fusion network (i.e., self-attention layer 2) to obtain the feature fusion representation for the fault type diagnosis task ; The fault feature matrix is input into a pre-trained third feature fusion network (i.e., self-attention layer 3) to obtain a feature fusion representation for the secondary equipment refusal-to-operate diagnosis task .

[0060] Finally, the fusion representations corresponding to each fault classification task are passed through the corresponding classifier to obtain the fault diagnosis results corresponding to each fault classification task.

[0061] Specifically, the fault location diagnosis task is used to determine the specific location where a fault occurs in the system. This task belongs to a multi-classification problem. Its classifier is constructed using a multi-layer perceptron (MLP1), and a Softmax activation function is added to the output layer to convert the network output into a probability distribution. The final output represents the probability values of each possible fault location, and the fault location with the highest probability is selected as the fault diagnosis result.

[0062] The fault type diagnosis task is used to identify the specific type and phase of the fault, including 10 types such as single-phase grounding of A / B / C, two-phase short circuit of AB / BC / AC, two-phase short circuit grounding of AB / BC / AC, and three-phase short circuit. This task is also a multi-classification task, and its classifier is also composed of MLP2 and a Softmax layer. The final model output represents the probability values of each possible fault type, and the fault type with the highest probability is selected as the fault diagnosis result.

[0063] The secondary equipment refusal-to-operate diagnosis task is used to detect possible refusal-to-operate situations of protection equipment in the system. Since multiple secondary protection equipment may refuse to operate simultaneously, this task belongs to a multi-label classification problem. The classifier is also composed of MLP3, but the Sigmoid activation function is used in the output layer to generate independent refusal probability prediction values for each protection equipment. By setting a threshold, the refusal state of specific equipment can be judged to obtain the fault diagnosis result.

[0064] In this embodiment, by using a pre-trained feature extraction network, the first fault feature of the alarm information, the second fault feature of the fault recording information, and the third fault feature of the SCADA measurement information are extracted, and the first fault feature, the second fault feature, and the third fault feature are concatenated to obtain a fault feature matrix. Furthermore, based on the pre-trained feature fusion network, according to the fault feature matrix, the fusion feature representations corresponding to different fault classification tasks are obtained. Thus, based on the pre-trained classifier, according to the fusion feature representations corresponding to different fault classification tasks, the fault diagnosis results corresponding to different fault classification tasks are predicted. This method comprehensively utilizes multi-modal fault information, and uses the graph attention network and self-attention mechanism to realize the feature extraction and fusion of multi-modal fault information, overcomes the limitation that a single information source is difficult to comprehensively reflect the state change of the substation system, and improves the accuracy and comprehensiveness of the primary and secondary fault diagnosis of the substation.

[0065] Based on the above embodiments, further, the training and optimization process of the feature extraction network will be described in detail below.

[0066] Training and optimizing the feature extraction network specifically includes: collecting multi-modal fault information samples and obtaining the original feature representations corresponding to the multi-modal fault information samples; encoding the randomly masked multi-modal fault information samples based on the encoder, and concatenating them to obtain a multi-source encoded representation; based on the decoder, reconstructing the data according to the multi-source encoded representation to obtain a reconstructed feature representation; minimizing the error between the original feature representation and the reconstructed feature representation through the mean square error loss function, and iteratively optimizing the encoder and decoder until convergence; using the encoder trained to convergence as the feature extraction network.

[0067] It is easy to understand that, in order to better extract the features of each source of fault information and reduce the impact of data loss on the fault diagnosis performance, this embodiment introduces a denoising autoencoder (DAE) to pre-train the feature extraction network. The denoising autoencoder is an unsupervised learning model mainly used for feature extraction and data reconstruction, and can learn robust feature representations of data in the case of noisy or missing input data.

[0068] Figure 4 The training and optimization schematic diagram of the feature extraction network provided by the embodiment of the present invention is shown. As Figure 4 shown, the multi-modal fault information samples include but are not limited to alarm information samples , fault recording information samples and SCADA measurement information samples . The multi-modal fault information samples are encoded through an embedding layer to complete preliminary feature extraction and standardization processing to obtain feature representations , and thus, the feature representations are concatenated to obtain the original feature representation .

[0069] The DAE consists of two main parts: an encoder and a decoder. The encoder includes the first feature extraction network (GAT1), the second feature extraction network (GAT2), and the third feature extraction network (GAT3) described above, which respectively extract features from the alarm information samples, fault recording information samples, and SCADA measurement information samples to obtain encoded representations , and In the input stage, the model introduces a data missing mechanism, that is, by randomly shielding some original alarm information samples, fault recording information samples and SCADA measurement information samples (setting them to 0), simulating the situation of incomplete data in actual scenarios, so as to enhance the model's feature extraction and reconstruction capabilities when facing missing data. The high-dimensional encoding representations output by the three feature extraction networks are spliced ​​to form an overall multi-source encoding representation. .

[0070] The decoder also uses the GAT architecture based on a primary topology structure, and its input is a multi-source coding representation , through the multi-layer feature aggregation and update of the decoder, the reconstructed feature representation is generated , aims to reconstruct and approximate the original feature representation This decoding method can not only preserve the influence of the system topology on the data characteristics, but also effectively compensate for the shielded or missing information through feature reconstruction, making the decoded data as close to the original input characteristics as possible, thereby restoring the integrity of the original data to a certain extent.

[0071] Subsequently, the mean square error (MSE) is used as the loss function to optimize the parameters of the encoder and decoder to obtain the trained encoder and decoder. The trained encoder is the feature extraction network.

[0072] This denoising autoencoder model based on the graph attention network makes full use of the topological structure of the primary system. It minimizes the error between the original complete input data and the reconstructed data through unsupervised learning, and can automatically learn the potential laws of data under different topological structures and mine the effective features of multi-source data. When the topological structure changes, the model can adaptively capture the fault feature representation under different topologies; when the data is missing or incomplete, the model reconstructs the missing information through the denoising mechanism and extracts robust fault features, ensuring that the model maintains stable performance in complex and incomplete data environments.

[0073] Compared with traditional end-to-end training methods, the denoising autoencoder model has significant advantages in multi-source data feature extraction. Traditional methods usually directly input alarm information, SCADA measurement data, and fault recording data into the model to complete feature extraction and fault diagnosis at one time. It is difficult to effectively capture the potential correlation between data, especially when data is missing or there is a lot of noise, which can easily lead to insufficient feature expression and affect diagnostic performance. The denoising autoencoder model denoises and extracts features from the input data through a pre-training step, and learns more representative and robust high-dimensional features in the latent space, effectively reducing the negative impact of missing data and noise, so that the model still shows good stability and adaptability when the data is limited or of poor quality, effectively improving the accuracy and robustness of fault diagnosis.

[0074] Based on the above embodiments, further, the following will elaborate on the training optimization process of the feature fusion network in detail.

[0075] Training and optimizing the feature fusion network specifically includes: when the fault classification task is a fault location diagnosis task and / or a fault type diagnosis task, training the feature fusion network using a supervised contrast loss function, and then training the classifier using a cross-entropy loss function; when the fault classification task is a secondary equipment refusal-to-operate diagnosis task, training both the feature fusion network and the classifier using a binary cross-entropy loss function.

[0076] It is easy to understand that the training strategies used for the feature fusion network in different fault classification tasks are different.

[0077] For the fault location diagnosis task and the fault type diagnosis task, since this embodiment only considers single-point faults, the diagnosis of the primary fault location and the primary fault type belongs to a multi-classification problem, and the goal is to select the correct category from all possible fault device sets and fault type sets.

[0078] This embodiment designs a phased training strategy: First, train the feature fusion network including the self-attention layer using a supervised contrast loss function to enhance the discrimination ability of the feature representation; then, train the classifier using a cross-entropy loss function to optimize the final classification performance.

[0079] This phased training strategy makes full use of the advantages of supervised contrast learning in feature extraction, and at the same time combines the optimization ability of the cross-entropy loss for classification tasks, which can better improve the diagnostic accuracy and robustness of the feature fusion network and the classifier.

[0080] For the secondary equipment refusal-to-operate diagnosis task, the refusal-to-operate diagnosis of secondary protection equipment is a multi-label classification problem, which is different from the fault location and fault type classification tasks and is not suitable for using supervised contrast learning. First, the core goal of supervised contrast learning is to enhance the discrimination ability of features by pulling closer the feature representations of similar samples and pushing away the feature representations of dissimilar samples. However, in a multi-label classification task, the "similar" or "dissimilar" relationship between samples is not a simple binary definition but depends on the distribution of specific labels. For example, two samples may be the same in some labels but different in other labels, and this label overlap leads to ambiguity in the division of positive and negative samples, making it difficult to directly apply the positive and negative sample pair construction strategy of contrast learning. Second, in this multi-label task, there are complex dependencies between labels. Supervised contrast learning only focuses on the feature similarity between samples and cannot effectively model the correlations between labels in a multi-label task, thus limiting its ability to capture global diagnostic features.

[0081] Therefore, for the secondary equipment protection refusal diagnosis task, in this embodiment, the binary cross-entropy loss function is directly used to train the feature fusion network and classifier including the self-attention layer, which can more directly optimize the network performance.

[0082] Based on the above embodiments, further, the preprocessing process after obtaining multi-modal fault information will be described in detail below.

[0083] The multi-modal fault information is encoded through the embedding layer to complete preliminary feature extraction and normalization processing, and the initial features are obtained.

[0084] For the alarm information in the multi-modal fault information. When a fault occurs in the substation system, a large number of alarm information will be generated, including the action alarms of protection devices and the action alarms of circuit breakers. In different fault scenarios, the lengths of the alarm information sequences are often inconsistent, and due to false alarms, missed alarms, and the existence of irrelevant alarms, these information may introduce greater uncertainty. Therefore, before the alarm information is input into the fault diagnosis model, it is necessary to uniformly encode it, extract effective information, and process its potential uncertainty.

[0085] The fuzzy time-sequence Petri net is a modeling tool suitable for processing uncertain information and the time-sequence relationship of complex systems, and has important application value in power system fault diagnosis. This method can construct a model describing the control relationship between relay protection devices and circuit breakers according to the protection configuration of the substation system, and accurately reflect the logical and time-sequence relationship between devices. At the same time, by introducing fuzzy theory, the fuzzy time-sequence Petri net can effectively deal with the noise, false alarms, and uncertainty problems in the alarm information, thereby significantly improving the robustness of the diagnosis process. Combining the action alarms of protection devices, the action alarms of circuit breakers, and their time constraints, the fuzzy time-sequence Petri net calculates the confidence of each protection device action through fuzzy reasoning, realizes the unified encoding and key feature extraction of the alarm sequence, and lays a foundation for subsequent intelligent diagnosis methods.

[0086] Thus, the initial features corresponding to the alarm information can be obtained, that is, the confidence vector of each relay protection action inferred by the fuzzy time-sequence Petri net.

[0087] For the fault recording information. In order to comprehensively reflect the dynamic characteristics before and after the fault, the three-phase current and zero-sequence current data of each node source-side line in the 5 cycles before the fault and 20 cycles after the fault can be selected, and the effective value of each cycle is calculated to effectively extract the amplitude change characteristics of the signal, reduce noise interference, and strengthen key information. At the same time, the obtained data is normalized to eliminate the difference in feature magnitudes between different nodes, and ensure the stability of feature input and the convergence of the model. The normalized data is used as the initial feature of the fault recording information of each node.

[0088] For the SCADA measurement information, the data of the three-phase current and zero-sequence current at the source side of each primary node at 1 measurement point before the fault and 19 measurement points after the fault occurrence can be selected. After normalizing the obtained data, it is used as the initial feature of the SCADA measurement information of each node.

[0089] Subsequently, the initial features obtained from the preprocessed alarm information, fault recorder information, and SCADA measurement information are respectively used as the inputs of the corresponding feature extraction networks to obtain the corresponding first fault feature, second fault feature, and third fault feature.

[0090] The fault feature matrix obtained by splicing the first fault feature, second fault feature, and third fault feature is used as the input of the corresponding feature fusion network to obtain the fusion feature representations for different fault classification tasks.

[0091] Finally, the fusion feature representations for different fault classification tasks are respectively passed through the corresponding classifiers to obtain the fault diagnosis results for different fault classification tasks.

[0092] Corresponding to the substation fault diagnosis method based on multi-modal fusion described in the above embodiments, the present invention also provides a substation fault diagnosis device based on multi-modal fusion.

[0093] Figure 5 The structural schematic diagram of the substation fault diagnosis device based on multi-modal fusion provided by the embodiment of the present invention is shown.

[0094] As Figure 5 shown, the device includes: a multi-modal fault information acquisition module 510, configured to acquire multi-modal fault information generated during the operation of the substation; a substation fault diagnosis module 520, configured to obtain the fault diagnosis result of the substation based on a pre-trained fault diagnosis model according to the multi-modal fault information; wherein, the fault diagnosis model includes a feature extraction network based on a graph attention network and a feature fusion network based on a self-attention mechanism, the feature extraction network is pre-trained by introducing a denoising auto-encoder model, and the feature fusion network is trained and optimized by a set loss function under different fault classification tasks.

[0095] In this embodiment, the multimodal fault information acquisition module 510 is used to acquire the multimodal fault information generated during the operation of the substation, and the substation fault diagnosis module 520 is used to obtain the fault diagnosis result of the substation based on the pre-trained fault diagnosis model according to the multimodal fault information. Among them, the fault diagnosis model includes a feature extraction network based on a graph attention network and a feature fusion network based on a self-attention mechanism. The feature extraction network is pre-trained by introducing a denoising autoencoder model, and the feature fusion network is trained and optimized by a set loss function under different fault classification tasks. This device comprehensively utilizes multimodal fault information, and uses the graph attention network and the self-attention mechanism to realize the feature extraction and fusion of multimodal fault information, overcoming the limitation that a single information source is difficult to comprehensively reflect the state change of the substation system, and improving the accuracy and comprehensiveness of the primary and secondary fault diagnosis of the substation.

[0096] It should be noted that the substation fault diagnosis device based on multimodal fusion provided in the embodiment of the present invention can be correspondingly referred to the multimodal fusion-based substation fault diagnosis method described in the above embodiments, and will not be elaborated here.

[0097] In addition, the present invention also provides a substation fault diagnosis system based on multimodal fusion. The system includes an actual physical system and a virtual simulation system that are connected bidirectionally. Among them, the actual physical system is used to collect the multimodal fault information generated during the operation of the substation and forward the multimodal fault information to the virtual simulation system; the virtual simulation system is constructed based on digital twin technology and is used to receive the multimodal fault information and execute the multimodal fusion-based substation fault diagnosis method described in any of the above embodiments.

[0098] Figure 6 The overall architecture diagram of the substation fault diagnosis system based on multimodal fusion provided in the embodiment of the present invention is shown.

[0099] As Figure 6 shown, this architecture is divided into two main links: the model construction stage and the model application stage. In the model construction stage, a virtual simulation system is built relying on digital twin technology. Through the virtual space, the design scheme (such as system topology and protection equipment configuration) is simulated, tested and verified to identify potential problems and optimize the design. With the high flexibility of the virtual environment, repeated debugging is carried out to ensure the reliability and feasibility of the design scheme. Subsequently, a large number of samples are generated through batch fault simulation to provide data support for the training of the fault diagnosis model.

[0100] When the virtual simulation system, the fault diagnosis model, and their diagnostic performance reach the expectations, it enters the model application stage. In this stage, the actual physical system is built and a two-way link is established with the virtual simulation system to form a virtual-real integrated collaborative system. The actual fault data (including multi-modal fault information) generated during the operation of the actual physical system is transmitted in real-time to the trained fault diagnosis model, which infers the cause of the fault and provides decision-making assistance for the operation and maintenance personnel. At the same time, the fault diagnosis model is fine-tuned based on the actual fault data to enable it to quickly adapt to the actual operating conditions. Compared with traditional fault diagnosis methods, this system makes full use of the large-scale data generated by virtual simulation, solves the problem of limited diagnostic ability due to insufficient data in the initial operation stage, and greatly improves the accuracy and adaptability of fault prediction.

[0101] It is also worth mentioning that the virtual simulation system is also used to generate a large number of samples through batch fault simulation to provide data support for the training of the fault diagnosis model.

[0102] Figure 7 An example of the physical structure diagram of an electronic device is shown as Figure 7 shown. The electronic device may include: a processor 710, a communication interface 720, a memory 730, and a communication bus 740. Among them, the processor 710, the communication interface 720, and the memory 730 complete communication with each other through the communication bus 740. The processor 710 can call the logical instructions in the memory 730 to execute a substation fault diagnosis method based on multi-modal fusion. The method includes: obtaining multi-modal fault information generated during the operation of the substation; based on a pre-trained fault diagnosis model, obtaining a fault diagnosis result of the substation according to the multi-modal fault information; wherein, the fault diagnosis model includes a feature extraction network based on a graph attention network and a feature fusion network based on a self-attention mechanism. The feature extraction network is pre-trained by introducing a denoising autoencoder model, and the feature fusion network is trained and optimized through a set loss function under different fault classification tasks.

[0103] In addition, when the logical instructions in the above-mentioned memory 730 are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, 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. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical discs that can store program codes.

[0104] On the other hand, the present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it is used to implement the substation fault diagnosis method based on multimodal fusion provided by the above-mentioned various methods. The method includes: obtaining multimodal fault information generated during the operation of the substation; based on a pre-trained fault diagnosis model, obtaining a fault diagnosis result of the substation according to the multimodal fault information; wherein, the fault diagnosis model includes a feature extraction network based on a graph attention network and a feature fusion network based on a self-attention mechanism. The feature extraction network is pre-trained by introducing a denoising autoencoder model, and the feature fusion network is trained and optimized by a set loss function under different fault classification tasks.

[0105] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. A person of ordinary skill in the art can understand and implement it without creative labor.

[0106] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on such an understanding, the essence of the above technical solution, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0107] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A substation fault diagnosis method based on multi-modal fusion, characterized in that: include: Obtain multi-modal fault information generated during substation operation; Based on a pre-trained fault diagnosis model, the fault diagnosis result of the substation is obtained according to the multimodal fault information; wherein, the fault diagnosis model includes a feature extraction network based on a graph attention network and a feature fusion network based on a self-attention mechanism, the feature extraction network is obtained by pre-training by introducing a denoising autoencoder model, and the feature fusion network is obtained by training and optimizing by setting loss functions under different fault classification tasks.

2. The substation fault diagnosis method based on multimodal fusion according to claim 1 is characterized in that: The multimodal fault information at least includes alarm information, fault recording information and SCADA measurement information; the number of the feature extraction networks is consistent with the number of categories of the multimodal fault information, and the number of the feature fusion networks is consistent with the number of fault classification tasks; Wherein, the feature extraction network is used to extract fault features of multi-modal fault information, and the feature fusion network is used to fuse fault features of multi-modal fault information; The fault classification task includes one or more combinations of a fault location diagnosis task, a fault type diagnosis task, and a secondary equipment failure diagnosis task, and the importance of each category of multimodal fault information under different fault classification tasks is different.

3. The substation fault diagnosis method based on multimodal fusion according to claim 2 is characterized in that: The fault diagnosis model also includes a classifier; Accordingly, the fault diagnosis model based on the pre-training obtains the fault diagnosis result of the substation according to the multi-modal fault information, including: Based on a pre-trained feature extraction network, extract a first fault feature of the alarm information, a second fault feature of the fault recording information, and a third fault feature of the SCADA measurement information; Concatenate the first fault feature, the second fault feature, and the third fault feature to obtain a fault feature matrix; Based on the pre-trained feature fusion network, according to the fault feature matrix, fusion feature representations corresponding to different fault classification tasks are obtained; Based on the pre-trained classifier, the fault diagnosis results corresponding to the different fault classification tasks are predicted according to the fusion feature representations corresponding to the different fault classification tasks.

4. The substation fault diagnosis method based on multimodal fusion according to claim 1 is characterized in that: The denoising autoencoder model includes an encoder and a decoder based on a graph attention network; Accordingly, training and optimizing the feature extraction network specifically includes: Collecting multimodal fault information samples and obtaining original feature representations corresponding to the multimodal fault information samples; Based on the encoder, randomly masked multi-modal fault information samples are encoded and concatenated to obtain multi-source coding representation; Based on the decoder, reconstruct data according to the multi-source coding representation to obtain a reconstructed feature representation; Minimizing the error between the original feature representation and the reconstructed feature representation by a mean square error loss function, iteratively optimizing the encoder and the decoder until convergence; The encoder trained to convergence is used as the feature extraction network.

5. The substation fault diagnosis method based on multi-modal fusion according to claim 3 is characterized in that: Training and optimizing the feature fusion network specifically includes: In the case where the fault classification task is a fault location diagnosis task and / or a fault type diagnosis task, the feature fusion network is trained using a supervised contrast loss function, and then the classifier is trained using a cross entropy loss function; In the case where the fault classification task is a secondary equipment failure diagnosis task, a binary cross entropy loss function is used to simultaneously train the feature fusion network and the classifier.

6. The substation fault diagnosis method based on multi-modal fusion according to claim 1 is characterized in that: Obtain multi-modal fault information generated during substation operation, including: The multi-modal fault information is encoded through an embedding layer to complete preliminary feature extraction and standardization processing to obtain initial features; Accordingly, based on the pre-trained fault diagnosis model, the fault diagnosis result of the substation is obtained according to the multi-modal fault information, including: The initial features are used as inputs of a pre-trained fault diagnosis model to obtain a fault diagnosis result of the substation.

7. A substation fault diagnosis device based on multi-modal fusion, characterized in that: include: A multi-modal fault information acquisition module is used to acquire multi-modal fault information generated during the operation of the substation; A substation fault diagnosis module, used for obtaining a substation fault diagnosis result according to the multimodal fault information based on a pre-trained fault diagnosis model; Among them, the fault diagnosis model includes a feature extraction network based on a graph attention network and a feature fusion network based on a self-attention mechanism. The feature extraction network is obtained by pre-training by introducing a denoising autoencoder model, and the feature fusion network is obtained by training and optimizing by setting loss functions under different fault classification tasks.

8. A substation fault diagnosis system based on multi-modal fusion, characterized in that: include: Establish a two-way connection between the actual physical system and the virtual simulation system, where The actual physical system is used to collect multi-modal fault information generated during the operation of the substation and forward the multi-modal fault information to the virtual simulation system; The virtual simulation system is constructed based on digital twin technology, and is used to receive the multimodal fault information and execute the substation fault diagnosis method based on multimodal fusion as described in any one of claims 1 to 6.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the computer program, the substation fault diagnosis method based on multimodal fusion according to any one of claims 1 to 6 is implemented.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the substation fault diagnosis method based on multimodal fusion according to any one of claims 1 to 6 is implemented.

Images