GIS equipment partial discharge diagnosis method and system based on multi-modal graph convolutional network

Abstract

The invention relates to the technical field of transformer detection, in particular to a GIS equipment partial discharge diagnosis method and system based on a multi-modal graph convolutional network. Firstly, ultrasonic waves, UHF and optical signals of GIS equipment are obtained; secondly, constructing a space-time diagram structure, extracting partial discharge features by using a multi-modal diagram convolutional network, and fusing multi-modal information; and finally, outputting a partial discharge diagnosis result. According to the method, by innovating the network structure design, the dynamic graph construction mechanism and the multi-modal information fusion strategy, the problems existing in an existing method are effectively solved, the diagnosis accuracy and reliability are remarkably improved, and meanwhile good adaptability and expandability are achieved. The method provides powerful technical support for safe operation and maintenance of GIS equipment, is of great significance in improving the safety and stability of a power system, and has wide application prospects.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment diagnosis, and in particular to a method and system for diagnosing partial discharge of GIS equipment based on a multimodal graph convolutional network. Background Art

[0002] GIS (gas-insulated switchgear) is a critical component of modern power systems, and its safe and reliable operation is crucial to the stability of the entire power grid. Partial discharge (PD) is a common insulation failure in GIS equipment, potentially leading to performance degradation and even serious accidents. Therefore, accurately diagnosing PD in GIS equipment is crucial for preventing failures and ensuring power grid safety.

[0003] Traditional methods for diagnosing partial discharge (PD) in GIS equipment primarily rely on a single signal source, such as ultrasonic or UHF signals. While these methods can detect PD to a certain extent, they have numerous limitations. First, a single signal source cannot fully reflect the characteristics of PD and is easily affected by environmental noise and interference, resulting in inaccurate diagnostic results. Second, these methods typically employ traditional signal processing techniques, such as Fourier transforms or wavelet analysis, which struggle to fully capture the complex temporal and spatial characteristics of PD signals.

[0004] In recent years, with the development of artificial intelligence (AI) technology, some researchers have attempted to apply machine learning methods to partial discharge (PD) diagnosis in GIS equipment. While these methods have improved diagnostic accuracy to some extent, they still face challenges. For example, most methods rely on single-modality signals and fail to fully utilize the complementary information provided by multiple sensors. Furthermore, these methods often treat signal processing and feature extraction as separate steps, making end-to-end optimization difficult.

[0005] Meanwhile, some researchers have begun exploring multimodal information fusion methods. These approaches attempt to combine multiple signals, such as ultrasonic, UHF, and optical, to obtain more comprehensive partial discharge (PD) characteristics. However, existing multimodal fusion methods mostly rely on simple feature concatenation or weighted averaging, making it difficult to effectively capture the complex relationships between different modal signals. Furthermore, these methods often overlook the spatiotemporal structure of the signals and fail to fully exploit the temporal information and spatial correlations in the PD process.

[0006] Furthermore, most existing diagnostic methods use static model structures that struggle to adapt to the dynamic changes in GIS equipment operating conditions and environmental conditions. This results in insufficient generalization and robustness of diagnostic models in practical applications, making them prone to misjudgment or omission, particularly when faced with novel or rare partial discharge patterns. Summary of the Invention

[0007] In view of the above problems, there is an urgent need to develop a partial discharge diagnosis method for GIS equipment that can fully utilize multimodal information, effectively capture the spatiotemporal characteristics of partial discharge, and have dynamic adaptability.

[0008] The present invention proposes a method and system for GIS equipment partial discharge diagnosis based on a multimodal graph convolutional network, comprising: The acquisition steps include: Acquire ultrasonic, UHF and optical signals from GIS equipment; Processing steps include: constructing a space-time graph structure based on the ultrasonic signal, the UHF signal and the optical signal; Extracting partial discharge features through a multimodal graph convolutional network according to the spatiotemporal graph structure; Based on the partial discharge characteristics, fusing multimodal information; Output steps include: Based on the fused multimodal information, the partial discharge diagnosis results of the GIS equipment are output.

[0009] Preferably, the obtaining step specifically includes: Acquiring the ultrasonic signal through an ultrasonic sensor; Acquire the UHF signal through a UHF antenna; acquiring the optical signal by an optical sensor; The ultrasonic signal is a time domain signal, and the UHF signal and the optical signal are frequency domain signals.

[0010] Preferably, the step of constructing the spatiotemporal graph structure specifically includes: A sliding window method is used to extract the spatiotemporal characteristics of the ultrasonic signal, the UHF signal and the optical signal; Calculating the correlation between the ultrasonic signal and the time domain signal; Based on the correlation, dynamically adjust the graph structure to generate a spatiotemporal graph structure; Each node in the spatiotemporal graph structure corresponds to a time sampling point, and the edges between the nodes represent the measurement values ​​of different parts of the GIS equipment at different time points.

[0011] Preferably, the step of extracting partial discharge features through a multimodal graph convolutional network specifically includes: Inputting the spatiotemporal graph structure into the ultrasound-time domain graph convolution layer to extract the time domain features of the ultrasound signal; Inputting the spatiotemporal graph structure into an optical-temporal graph convolution layer to extract the temporal features of the optical signal; Inputting the spatiotemporal graph structure into the UHF-frequency domain graph convolution layer to extract the frequency domain features of the UHF signal; The ultrasonic-time domain graph convolution layer, the optical-time domain graph convolution layer and the UHF-frequency domain graph convolution layer share parameters.

[0012] Preferably, the calculation formula of the ultrasound-time domain image convolution layer is: , in, Indicates the The output features of the layer, represents the activation function, represents the normalized adjacency matrix, represents the input features of the lth layer, represents the learnable weight matrix.

[0013] Preferably, the step of fusing multimodal information specifically includes: Adopt feature-level fusion method to fuse features of different modalities; The model-level fusion method is used to further fuse the results of feature-level fusion; The feature-level fusion method is defined as: , , The model-level fusion method is defined as: , in, represents the time domain signal fusion feature, represents the frequency domain signal fusion feature, represents the activation function, and represents the learnable weight moment.

[0014] Preferably, the step of fusing multimodal information further comprises: The self-attention mechanism is used to weight the fused features, and the calculation formula is: , in, 、 、 represent the query matrix, key matrix and value matrix respectively, Indicates the dimension of the key vector.

[0015] As an advantage, it also includes: Introducing residual connections into the multimodal graph convolutional network to gain data input and output; Through the dynamic graph construction mechanism, the weights of the feature matrix are updated in real time.

[0016] Preferably, the step of outputting the partial discharge diagnosis result of the GIS equipment specifically includes: Input the fused multimodal information into the fully connected layer; Calculate the probability distribution of different partial discharge types through the softmax function; The type with the highest probability is selected as the final partial discharge diagnosis result.

[0017] The GIS equipment partial discharge diagnosis system based on multimodal graph convolutional network includes: Signal acquisition module, used to obtain ultrasonic signals, UHF signals and optical signals of GIS equipment; A graph structure construction module, configured to construct a spatiotemporal graph structure based on the ultrasonic signal, the UHF signal, and the optical signal; a feature extraction module, configured to extract partial discharge features from the spatiotemporal graph structure through a multimodal graph convolutional network; Information fusion module, used to fuse partial discharge features from different modalities; Diagnostic output module, used to output partial discharge diagnostic results of GIS equipment based on fused multimodal information; Among them, the multimodal graph convolution network includes an ultrasonic-time domain graph convolution layer, an optical-time domain graph convolution layer, a UHF-frequency domain graph convolution layer, an attention mechanism module and a cross-modal feature fusion module.

[0018] This paper aims to address the challenges of existing GIS equipment partial discharge diagnosis methods by proposing a diagnostic method and system based on a multimodal graph convolutional network. Through innovative network structure design, dynamic graph construction, and multimodal information fusion strategies, this method achieves high-precision diagnosis of partial discharge in GIS equipment.

[0019] The method of the present invention has the following significant beneficial effects: First, this invention significantly improves the comprehensiveness and accuracy of diagnosis by fusing signals from three modalities: ultrasonic, ultra-high-frequency (UHF), and optical. Signals from different modalities provide complementary information: ultrasonic signals reflect the mechanical vibration characteristics of the discharge, UHF signals capture high-frequency electromagnetic radiation, and optical signals detect photon emission during the discharge process. This multimodal fusion strategy significantly enhances the model's ability to identify various types of partial discharges and reduces the impact of environmental noise and interference.

[0020] Secondly, the spatiotemporal graph structure and dynamic graph construction mechanism introduced in this paper effectively capture the temporal characteristics and spatial correlations of partial discharge signals. This approach not only preserves the temporal dynamics of the signal but also reflects the interactions between different parts of the GIS equipment, thereby providing a richer and more discriminative feature representation. In particular, the dynamic graph construction mechanism enables the model to adjust the graph structure in real time based on the characteristics of the input data, greatly enhancing the model's adaptability and generalization capabilities.

[0021] Furthermore, the multimodal graph convolutional network designed in this paper achieves efficient feature extraction from signals of different modalities through specialized ultrasonic-time domain graph convolutional layers, optical-time domain graph convolutional layers, and UHF-frequency domain graph convolutional layers. This design fully considers the characteristics of various signals and can characterize partial discharges from multiple dimensions. Furthermore, the residual connections and parameter sharing mechanisms introduced into the network effectively alleviate the vanishing gradient problem in deep learning model training, improving the model's expressiveness and training stability.

[0022] Furthermore, the present invention employs a multi-level, multi-stage information fusion strategy, including feature-level fusion, model-level fusion, and a self-attention mechanism, to achieve deep integration of multimodal information. This fusion approach not only fully exploits the complementarity of different modal signals but also adaptively adjusts the importance of different features, resulting in a more comprehensive and robust feature representation.

[0023] Finally, the end-to-end learning framework proposed in this paper organically combines multiple steps, including signal processing, feature extraction, information fusion, and diagnostic decision-making, to achieve global optimization. This approach avoids the fragmented nature of traditional methods, enabling better extraction of useful information from the data and improving overall diagnostic performance.

[0024] In summary, the proposed method and system for GIS equipment partial discharge diagnosis based on a multimodal graph convolutional network effectively addresses the challenges of existing approaches through innovative network structure design, dynamic graph construction mechanisms, and multimodal information fusion strategies. This method not only significantly improves diagnostic accuracy and reliability but also exhibits excellent adaptability and scalability, providing strong technical support for the safe operation and maintenance of GIS equipment and possessing significant practical significance for improving the safety and stability of power systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 This is an overall flow chart of the partial discharge diagnosis method for GIS equipment of the present invention; Figure 2 A module workflow diagram for the graph structure construction of the present invention; Figure 3 This is a workflow diagram of the feature extraction module of the present invention; Figure 4 This is a workflow diagram of the information fusion module of the present invention; Figure 5 This is a workflow diagram of the diagnostic output module of the present invention. DETAILED DESCRIPTION

[0026] Please refer to the attached Figure 1-5 The present invention provides a method and system for diagnosing partial discharge of GIS equipment based on a multimodal graph convolutional network. The method includes an acquisition step, a processing step, and an output step.

[0027] During the acquisition step, the method of the present invention acquires ultrasonic, UHF, and optical signals from GIS equipment. These three signals each have unique characteristics, revealing the characteristics of partial discharge from different perspectives. Ultrasonic signals primarily reflect the mechanical vibrations during the discharge process, UHF signals capture high-frequency electromagnetic radiation, and optical signals detect photons generated during the discharge process. By simultaneously acquiring these three signals, the method of the present invention can comprehensively capture partial discharge information.

[0028] In the processing step, a spatiotemporal graph structure is first constructed based on the acquired ultrasonic, UHF, and optical signals. The spatiotemporal graph structure is a data structure that can simultaneously represent the temporal and spatial characteristics of a signal and is very suitable for processing multimodal signals. In a preferred embodiment of the present invention, the process of constructing the spatiotemporal graph structure includes extracting the spatiotemporal characteristics of the signal using a sliding window method, calculating the correlation between the ultrasonic signal and the time domain signal, and dynamically adjusting the graph structure based on this correlation. This dynamic adjustment mechanism enables the constructed graph structure to better adapt to changes in the signal and improve the accuracy of subsequent processing.

[0029] Next, the method of the present invention extracts local discharge features from the spatiotemporal graph structure through a multimodal graph convolutional network. The multimodal graph convolutional network is a deep learning model specifically designed to process multimodal graph structure data. It can effectively utilize the topological information in the graph structure and simultaneously process signal features of different modalities. In an embodiment of the present invention, the multimodal graph convolutional network includes an ultrasonic-time domain graph convolution layer, an optical-time domain graph convolution layer, and a UHF-frequency domain graph convolution layer. These different convolutional layers perform feature extraction for different types of signals, thereby fully utilizing the characteristics of various signals.

[0030] After feature extraction, the method of the present invention performs multimodal information fusion. This step aims to organically combine features from different modalities to form a comprehensive feature representation. Information fusion can improve feature richness and discriminative power, helping to improve the accuracy of the final diagnostic result.

[0031] In the output step, the method of the present invention outputs the partial discharge diagnosis results of the GIS equipment based on the fused multimodal information. This result may include information such as the type and severity of the partial discharge, providing an important basis for the maintenance and management of the GIS equipment.

[0032] Furthermore, in the acquisition step, the method of the present invention acquires different types of signals using specific sensors. Specifically, ultrasonic signals are acquired using ultrasonic sensors, UHF signals are acquired using UHF antennas, and optical signals are acquired using optical sensors. It is worth noting that ultrasonic signals are time-domain signals, while UHF and optical signals are frequency-domain signals. This difference in signal type is one of the key reasons why the present invention employs a multimodal processing approach.

[0033] In constructing the spatiotemporal graph structure, the method of the present invention utilizes a sliding window technique. This method is a commonly used signal processing technique that can extract local signal features while maintaining temporal resolution. In the present invention, the sliding window size can be adjusted based on actual needs. For example, for ultrasonic signals, a 10ms window size can be selected to capture the complete process of most partial discharge events. For UHF signals, due to their higher frequency, a smaller window size, such as 1ms, can be selected to capture rapidly changing features.

[0034] The method of the present invention also calculates the correlation between the ultrasonic signal and the time domain signal. The purpose of this step is to find the relationship between different signals and provide a basis for subsequent graph structure adjustment. The correlation can be calculated using the Pearson correlation coefficient, which is mathematically expressed as: , in, is the correlation coefficient, for and The covariance of and They are and The standard deviation of .

[0035] Based on the calculated correlations, the method of the present invention dynamically adjusts the graph structure. Specifically, if the correlation between two signals exceeds a certain threshold (e.g., 0.7), an edge between the corresponding nodes of the two signals is added to the graph structure. Choosing this threshold requires a balance between the sparsity of the graph structure and the completeness of the information. A threshold that is too high may cause important correlations to be overlooked, while a threshold that is too low may introduce noise, increasing the complexity of subsequent processing.

[0036] Through the above steps, the method of the present invention constructs a graph structure that effectively represents the spatiotemporal characteristics of multimodal signals. In this graph, each node corresponds to a time sampling point, and the edges between nodes represent the associations between measurements from different parts of the GIS equipment at different time points. This structure not only preserves the temporal information of the signals but also reflects the spatial relationships between them, providing an excellent foundation for subsequent feature extraction.

[0037] The method and system of the present invention achieve accurate diagnosis of partial discharge in GIS equipment by fusing multimodal information and leveraging the powerful feature extraction capabilities of graph convolutional networks. Compared to traditional single-signal processing methods, this approach more comprehensively captures the characteristics of partial discharge, improving diagnostic accuracy and reliability. Furthermore, the present invention's dynamic graph structure construction mechanism enables the system to better adapt to varying operating conditions and equipment states, demonstrating excellent adaptability and generalization.

[0038] The method presented in this paper employs an innovative network architecture to extract partial discharge features using a multimodal graph convolutional network. Specifically, the network comprises an ultrasonic-time domain graph convolution layer, an optical-time domain graph convolution layer, and a UHF-frequency domain graph convolution layer. This design fully considers the characteristics of different signal types and effectively extracts key features from various modal signals.

[0039] In a preferred embodiment of the present invention, the spatiotemporal graph structure is first input into the ultrasonic-temporal graph convolution layer. This layer is specifically designed to process the temporal characteristics of ultrasonic signals. Ultrasonic signals typically contain rich temporal information, which can reflect the mechanical vibration characteristics of partial discharge processes. Through graph convolution operations, this layer can effectively capture the temporal patterns and local structural features in ultrasonic signals.

[0040] Simultaneously, the spatiotemporal graph structure is also input into the optical-temporal graph convolution layer. Although optical signals are typically viewed as frequency-domain signals, in some cases, their time-domain features also contain valuable information. For example, the time-domain waveform of the optical signal may reflect the dynamic characteristics of photon release during the discharge process. Therefore, the method of the present invention extracts these potentially useful features through the optical-temporal graph convolution layer.

[0041] For UHF signals, the method of this invention utilizes a UHF-frequency-domain graph convolution layer. The frequency-domain characteristics of UHF signals are particularly important for identifying different types of partial discharges. By inputting the spatiotemporal graph structure into this layer, frequency-domain features of UHF signals, such as the signal's spectral distribution and primary frequency components, can be effectively extracted.

[0042] Notably, our method implements parameter sharing across these graph convolutional layers. Parameter sharing is an effective technique that can reduce the number of model parameters, prevent overfitting, and promote information exchange between different modalities. Through parameter sharing, our method not only improves the model's generalization capabilities but also enhances the learning of correlations between signals from different modalities.

[0043] In another embodiment of the present invention, the calculation process of the ultrasound-time domain image convolution layer can be expressed by the following formula: , Indicates the The output features of the layer, represents the activation function, is the normalized adjacency matrix, It is The input features of the layer, is a learnable weight matrix. This formula describes the core operation of graph convolution, which is to update node features by aggregating neighborhood information.

[0044] Activation Function The choice of has a significant impact on the performance of the model. In a preferred embodiment of the present invention, a ReLU (Rectified Linear Unit) function can be used as the activation function. The mathematical expression of the ReLU function is: , The ReLU function has the advantages of simple calculation and effective relief of the gradient vanishing problem, and is particularly suitable for deep neural networks.

[0045] Adjacency Matrix The normalization is to stabilize the numerical calculation and prevent the characteristic scale from changing dramatically during the propagation process. A commonly used normalization method is symmetric normalization, and its mathematical expression is: , in, is the original adjacency matrix, is the degree matrix (i.e., a diagonal matrix whose diagonal elements are node degrees).

[0046] After extracting partial discharge features, the method of the present invention fuses multimodal information. This step is crucial for fully leveraging the complementary information of different modal signals. Specifically, this method employs a strategy that combines feature-level fusion with model-level fusion.

[0047] In feature-level fusion, this method directly concatenates features from different modalities. For time-domain signals (including time-domain features of ultrasonic and optical signals), the fused features can be expressed as: , in, Indicates the The eigenvector of the time domain signal, is the number of time domain signals. Similarly, for frequency domain signals (mainly the frequency domain features of UHF signals), the fusion features can be expressed as: , in, Indicates the The eigenvector of the frequency domain signal, is the number of the frequency domain signal.

[0048] In model-level fusion, the method of the present invention further performs weighted combination on the results of feature-level fusion. This process can be expressed by the following formula: , In this formula, It also represents the activation function, and is a learnable weight matrix. In this way, the model can adaptively adjust the importance of different modal information, thereby obtaining a more comprehensive and effective feature representation.

[0049] It is worth mentioning that the method of the present invention also introduces a self-attention mechanism when fusing multimodal information. The self-attention mechanism can help the model focus on the most relevant and important features, further improving the quality of feature representation. In one embodiment of the present invention, the calculation formula of the self-attention mechanism is as follows: , In this formula, 、 and They represent the query matrix, key matrix and value matrix respectively, which are obtained from the input features through linear transformation. Represents the dimension of the key vector, introducing The purpose of is to alleviate the gradient problem caused by the increase in dimension. The softmax function is used to normalize the attention scores so that their sum is 1.

[0050] By introducing the self-attention mechanism, the method of the present invention can adaptively adjust the importance of different features, effectively extract and combine the most relevant information, and thus improve the accuracy of partial discharge diagnosis.

[0051] This multi-level, multi-stage fusion strategy fully utilizes the complementarity of different modal signals to obtain a more comprehensive and robust feature representation, providing a solid foundation for subsequent partial discharge diagnosis.

[0052] The method of the present invention introduces residual connections in multimodal graph convolutional networks, a technique that effectively improves the performance of deep neural networks. The core idea of ​​residual connections is to add direct connections across layers in the network, allowing information to flow more easily within the network. In a preferred embodiment of the present invention, residual connections are applied between data input and output, thereby achieving data amplification.

[0053] Specifically, the residual connection can be expressed as: , in, Indicates the The output of the layer, Indicates the Nonlinear transformation of layer output, The first In this way, the network can learn new feature representations while retaining the original input information, thereby improving the model's expressiveness and alleviating the gradient vanishing problem in deep network training.

[0054] Another key innovation is the introduction of a dynamic graph construction mechanism for real-time updating of the feature matrix weights. This mechanism enables the model to dynamically adjust the graph structure based on the characteristics of the input data, thereby better adapting to the time-varying and complex characteristics of partial discharge in GIS equipment.

[0055] In one embodiment of the present invention, the core of the dynamic graph construction mechanism is a learnable adjacency matrix. This adjacency matrix is ​​not pre-fixed, but is learned through network learning. Specifically, the following formula can be used to update the adjacency matrix: , In this formula, yes The adjacency matrix at time , are input features, and is a learnable parameter matrix. In this way, the model can dynamically adjust the connection relationship between nodes according to the characteristics of the input data, thereby constructing a graph structure that best suits the current input.

[0056] The method of the present invention employs a series of carefully designed steps to output partial discharge diagnostic results for GIS equipment. First, the fused multimodal information is input into a fully connected layer. This fully connected layer maps high-dimensional features into the space of diagnostic results. In a preferred embodiment of the present invention, a multi-layer fully connected network is used, with nonlinear activation functions added between each layer to enhance the model's expressive power.

[0057] Next, this method calculates the probability distribution of different partial discharge types through the softmax function. The mathematical expression of the softmax function is: , in, It is The raw scores of the categories, is the total number of categories. The softmax function converts the raw scores into a probability distribution so that the sum of the probabilities of all categories is 1.

[0058] Finally, the method of the present invention selects the type with the highest probability as the final partial discharge diagnosis result. This probability-based decision-making method not only provides a diagnosis result, but also provides the confidence level of the result, providing more information support for subsequent decision-making.

[0059] It's worth noting that in practical applications, a probability threshold can be set based on specific needs. For example, only when the probability of a certain type exceeds 0.8 will it be output as the final diagnosis result. If the probabilities of all types are below the threshold, a diagnosis result of an unknown type can be output to avoid making uncertain judgments.

[0060] The present invention also proposes a GIS equipment partial discharge diagnosis system based on a multimodal graph convolutional network, corresponding to the above method. The system includes multiple functional modules, each with its own specific function, which work together to complete the task of partial discharge diagnosis.

[0061] Signal acquisition module 1 is responsible for acquiring ultrasonic, UHF, and optical signals from GIS equipment. This module may include various sensors, such as ultrasonic sensors, UHF antennas, and optical sensors. The design of signal acquisition module 1 must consider factors such as the sampling rate and accuracy of the signal to ensure high-quality raw data.

[0062] Graph Structure Construction Module 2 is responsible for constructing a spatiotemporal graph structure based on the collected multimodal signals. This module implements the dynamic graph construction mechanism described above, enabling real-time adjustments to the graph structure based on signal characteristics. The output of Graph Structure Construction Module 2 provides a robust data foundation for subsequent feature extraction.

[0063] Feature Extraction Module 3 is one of the core components of the system. It extracts partial discharge features from the spatiotemporal graph structure using a multimodal graph convolutional network. This module includes the previously described ultrasonic-time domain graph convolution layer, optical-time domain graph convolution layer, and UHF-frequency domain graph convolution layer, as well as residual connections. The design of Feature Extraction Module 3 directly impacts system performance and diagnostic accuracy.

[0064] Information Fusion Module 4 is responsible for fusing PD features from different modalities. This module implements the feature-level fusion and model-level fusion described above, as well as the self-attention mechanism. The output of Information Fusion Module 4 is a high-level feature representation that integrates multimodal information.

[0065] Finally, the diagnostic output module 5 outputs the partial discharge diagnostic results of the GIS equipment based on the fused multimodal information. This module implements the fully connected layer, softmax function calculation, and final decision logic described above.

[0066] This modular system design offers excellent flexibility and scalability. Each module can be independently optimized and upgraded, while the interfaces between modules remain stable. This makes the system easy to maintain and further improve. For example, if new sensor technologies are developed in the future, only Signal Acquisition Module 1 needs to be updated, without changing the entire system structure.

[0067] In summary, the proposed method and system for GIS equipment partial discharge diagnosis based on a multimodal graph convolutional network (MGCN) achieves high-precision diagnosis of partial discharge in GIS equipment through innovative network structure design, dynamic graph construction, and multimodal information fusion strategies. This method and system not only improves diagnostic accuracy but also exhibits excellent adaptability and scalability, providing strong technical support for the safe operation and maintenance of GIS equipment.

[0068] It should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A partial discharge diagnosis method for GIS equipment based on a multimodal graph convolutional network, characterized in that: include: The acquisition steps include: Acquire ultrasonic, UHF and optical signals from GIS equipment; Processing steps include: constructing a space-time graph structure based on the ultrasonic signal, the UHF signal and the optical signal; Extracting partial discharge features through a multimodal graph convolutional network according to the spatiotemporal graph structure; Based on the partial discharge characteristics, fusing multimodal information; Output steps include: Based on the fused multimodal information, the partial discharge diagnosis results of the GIS equipment are output.

2. The method according to claim 1, characterized in that The obtaining step specifically includes: Acquiring the ultrasonic signal through an ultrasonic sensor; Acquire the UHF signal through a UHF antenna; acquiring the optical signal by an optical sensor; The ultrasonic signal is a time domain signal, and the UHF signal and the optical signal are frequency domain signals.

3. The method according to claim 1, characterized in that The steps of constructing the spatiotemporal graph structure specifically include: A sliding window method is used to extract the spatiotemporal characteristics of the ultrasonic signal, the UHF signal and the optical signal; Calculating the correlation between the ultrasonic signal and the time domain signal; Based on the correlation, dynamically adjust the graph structure to generate a spatiotemporal graph structure; Each node in the spatiotemporal graph structure corresponds to a time sampling point, and the edges between the nodes represent the measurement values ​​of different parts of the GIS equipment at different time points.

4. The method according to claim 1, wherein The step of extracting partial discharge features through a multimodal graph convolutional network specifically includes: Inputting the spatiotemporal graph structure into the ultrasound-time domain graph convolution layer to extract the time domain features of the ultrasound signal; Inputting the spatiotemporal graph structure into an optical-temporal graph convolution layer to extract the temporal features of the optical signal; Inputting the spatiotemporal graph structure into the UHF-frequency domain graph convolution layer to extract the frequency domain features of the UHF signal; The ultrasonic-time domain graph convolution layer, the optical-time domain graph convolution layer and the UHF-frequency domain graph convolution layer share parameters.

5. The method according to claim 4, characterized in that The calculation formula of the ultrasound-time domain image convolution layer is: in, Indicates the The output features of the layer, represents the activation function, represents the normalized adjacency matrix, represents the input features of the lth layer, represents the learnable weight matrix.

6. The method according to claim 1, characterized in that The step of fusing multimodal information specifically includes: Adopt feature-level fusion method to fuse features of different modalities; The model-level fusion method is used to further fuse the results of feature-level fusion; The feature-level fusion method is defined as: , , The model-level fusion method is defined as: , in, represents the time domain signal fusion feature, represents the frequency domain signal fusion feature, represents the activation function, and represents the learnable weight moment.

7. The method according to claim 6, characterized in that The step of fusing multimodal information further includes: The self-attention mechanism is used to weight the fused features, and the calculation formula is: , in, 、 、 represent the query matrix, key matrix and value matrix respectively, Indicates the dimension of the key vector.

8. The method according to claim 1, characterized in that Also includes: Introducing residual connections into the multimodal graph convolutional network to gain data input and output; Through the dynamic graph construction mechanism, the weights of the feature matrix are updated in real time.

9. The method according to claim 1, characterized in that The step of outputting the partial discharge diagnosis result of the GIS equipment specifically includes: Input the fused multimodal information into the fully connected layer; Calculate the probability distribution of different partial discharge types through the softmax function; The type with the highest probability is selected as the final partial discharge diagnosis result.

10. A GIS equipment partial discharge diagnosis system based on a multimodal graph convolutional network for executing the method according to any one of claims 1 to 9, characterized in that: include: Signal acquisition module, used to obtain ultrasonic signals, UHF signals and optical signals of GIS equipment; A graph structure construction module, configured to construct a spatiotemporal graph structure based on the ultrasonic signal, the UHF signal, and the optical signal; a feature extraction module, configured to extract partial discharge features from the spatiotemporal graph structure through a multimodal graph convolutional network; Information fusion module, used to fuse partial discharge features from different modalities; Diagnostic output module, used to output partial discharge diagnostic results of GIS equipment based on fused multimodal information; Among them, the multimodal graph convolution network includes an ultrasonic-time domain graph convolution layer, an optical-time domain graph convolution layer, a UHF-frequency domain graph convolution layer, an attention mechanism module and a cross-modal feature fusion module.

Images