Partial discharge fault type identification method
By constructing a comprehensive model for partial discharge fault type identification, and utilizing MobileNetV3 and Graph Neural Network (GNN) in conjunction with PRPD graph processing and comprehensive judgment, the problem of inaccurate partial discharge fault type identification in existing technologies has been solved. This achieves efficient and accurate partial discharge type identification, thereby improving the safety and stability of the power system.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- LIAOYUAN POWER SUPPLY COMPANY STATE GRID JILIN ELECTRIC POWER
- Filing Date
- 2025-12-30
- Publication Date
- 2026-05-19
AI Technical Summary
The identification performance of existing partial discharge fault type identification methods needs further improvement, especially in the application of deep learning models, where it is difficult to accurately identify multiple partial discharge types.
A comprehensive model for identifying partial discharge fault types was constructed. A partial discharge fault type identification sub-model based on MobileNetV3 and graph neural network GNN was adopted. Data was processed and trained using PRPD graphs. The output results of the two models were combined for comprehensive judgment, and four types of partial discharge were identified.
It improves the accuracy of partial discharge fault type identification, reduces computational complexity and time, and provides a guarantee for the safe and stable operation of the power system.
Smart Images
Figure CN122064941A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method for identifying partial discharge fault types, belonging to the field of power system fault diagnosis technology. Background Technology
[0002] Partial discharge is a non-penetrating discharge phenomenon caused by the concentration of local electric fields in the insulation system of electrical equipment. Its long-term presence can accelerate insulation degradation and even lead to equipment breakdown or explosion accidents. my country is accelerating the construction of a new power system, and accurate identification of partial discharge faults is one of the key technologies for ensuring the safe operation of the power system.
[0003] Currently, partial discharge fault identification mainly employs feature-based identification methods, machine learning-based identification methods, and deep learning-based identification methods. However, the identification effectiveness needs further improvement. With the continuous development of deep learning models, partial discharge fault identification based on multiple deep learning models is a technological trend. Therefore, reliable and effective identification of partial discharge faults is of great significance for the stable operation of power systems. Summary of the Invention
[0004] To solve the above-mentioned technical problems, this invention provides a method for identifying partial discharge fault types, which can identify four types of partial discharge based on a comprehensive model for identifying partial discharge fault types.
[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:
[0006] A method for identifying partial discharge fault types is proposed. A comprehensive model for partial discharge fault type identification is constructed, comprising two sub-models. The input of each sub-model is a partial discharge phase distribution (PRPD) map (hereinafter referred to as "PRPD map"), and the output of each sub-model is the corresponding partial discharge type. The comprehensive model takes the PRPD map as input and outputs the partial discharge type, thus identifying four types of partial discharge. The specific steps are as follows:
[0007] Step 1: Construct two partial discharge fault type identification sub-models in the comprehensive model for partial discharge fault type identification, and train them;
[0008] Step 2: Process the partial discharge ultrasonic signal to be identified to form a processed partial discharge signal;
[0009] Step 3: Input the processed partial discharge signal into the trained sub-models for identifying two types of partial discharge faults;
[0010] Step 4: Combine the output results of the two partial discharge fault type identification sub-models to make a comprehensive judgment and output the partial discharge fault type.
[0011] A further improvement to the technical solution of this invention lies in: the construction and training of two partial discharge fault type identification sub-models in the comprehensive model for partial discharge fault type identification described in step 1, specifically as follows:
[0012] Step 1.1: Generate partial discharge ultrasonic signals, including four types of partial discharge: tip discharge, bubble discharge, suspension discharge, and surface discharge. The partial discharge ultrasonic signals can be generated using actual or simulated partial discharge ultrasonic signals.
[0013] Step 1.2: Partial discharge ultrasonic signal processing, including four steps: noise reduction, conversion to PRPD spectrum, data enhancement, and data normalization. The necessity of these four steps depends on the specific circumstances. The specific sub-steps are as follows:
[0014] Step 1.2.1: The noise reduction process requires determining the source form of the partial discharge ultrasonic signal. If the actual partial discharge ultrasonic signal is used, noise reduction is required. Existing noise reduction methods such as wavelet methods can be used for noise reduction. If the simulated partial discharge ultrasonic signal is used, noise reduction is not required.
[0015] Step 1.2.2: Convert the partial discharge ultrasonic signal into a PRPD map;
[0016] Step 1.2.3: Data augmentation requires determining the amount of data in the PRPD graph. If the amount of data is insufficient for training, data augmentation is necessary; otherwise, it is not necessary.
[0017] Step 1.2.4: Normalize the PRPD map by using min-max normalization to linearly map the pixel values of the PRPD map to the [0, 1] interval;
[0018] Step 1.3: Divide the normalized PRPD map into training set, validation set and test set;
[0019] Step 1.4: Construct a comprehensive model for identifying partial discharge fault types, including two sub-models for identifying partial discharge fault types: sub-model A based on MobileNetV3 and sub-model B based on graph neural network (GNN).
[0020] The network structure of the partial discharge fault type identification sub-model A based on MobileNetV3 includes an input layer, an initial convolutional layer, inverted residual block 1, inverted residual block 2, inverted residual block 3, a global average pooling layer, a fully connected layer, and an output layer. The input layer takes a normalized PRPD map as input. The initial convolutional layer quickly extracts low-level features (such as edges and textures) from the PRPD map. Inverted residual block 1 aggregates the low-level features from the initial convolutional layer and extracts high-level feature maps (such as overall shape) from the PRPD map through a cyclic structure of dimensionality increase-depth separable convolution-dimensionality reduction. Inverted residual block 2 aggregates... The high-level feature map of inverted residual block 1 is used to extract the high-level feature map of the PRPD map again through a recurrent structure of dimensionality increase-depthseparable convolution-dimensionality reduction; inverted residual block 3 aggregates the high-level feature map of inverted residual block 2 and extracts the high-level feature map of the PRPD map again through a recurrent structure of dimensionality increase-depthseparable convolution-dimensionality reduction; a global average pooling layer performs global average pooling on the high-level feature map output by inverted residual block 3, converting it into a feature value; a fully connected layer maps the output of the global average pooling layer to obtain the prediction score for each partial discharge type; the output layer outputs the partial discharge type result;
[0021] The network structure of the partial discharge fault type identification sub-model B based on graph neural network (GNN) includes an input layer, graph convolutional layer 1, graph convolutional layer 2, graph convolutional layer 3, global average pooling layer, fully connected layer, and output layer. The input layer takes a normalized PRPD map as input. Graph convolutional layer 1 performs the first layer feature extraction of the PRPD map, dynamically assigning weights from neighboring nodes to the current node based on an attention mechanism. Graph convolutional layer 2 performs the second layer feature extraction of the PRPD map, dynamically assigning weights from neighboring nodes to the current node based on an attention mechanism. Graph convolutional layer 3 performs the third layer feature extraction of the PRPD map, dynamically assigning weights from neighboring nodes to the current node based on an attention mechanism. The global average pooling layer reduces the features of the PRPD map to a fixed-length vector. The fully connected layer maps the output of the global average pooling layer to obtain the prediction score for each partial discharge type. The output layer outputs the partial discharge type result.
[0022] Step 1.5: Train the partial discharge fault type identification sub-models A and B respectively. The training requires setting the number of training samples and the number of training rounds. The input of identification sub-models A and B is the PRPD map, and the output of identification sub-models A and B is the partial discharge fault type.
[0023] A further improvement to the technical solution of this invention lies in the fact that the data processing in step 2 includes three sub-steps: noise reduction, conversion to PRPD map, and normalization data processing.
[0024] Step 2.1: Noise reduction processing. Existing noise reduction methods, such as wavelet methods, are used for noise reduction processing.
[0025] Step 2.2: Convert the partial discharge ultrasonic signal into a PRPD map;
[0026] Step 2.3: Normalize the PRPD map by using min-max normalization to linearly map the pixel values of the PRPD map to the [0, 1] interval.
[0027] A further improvement of the technical solution of the present invention is that: in step 3, the partial discharge signal after data processing is input into the two trained partial discharge fault type identification sub-models respectively, and the output result r_A of fault type identification sub-model A and the output result r_B of fault type identification sub-model B are obtained respectively.
[0028] A further improvement of the technical solution of the present invention is that: in step 4, the output results of the two partial discharge fault type identification models are combined and judged. If r_A and r_B are the same, the partial discharge fault type is output as r_A; if r_A and r_B are different, a random number α between 0 and 1 is generated using a random function. If α≤0.5, the partial discharge fault type is output as r_A; if α>0.5, the partial discharge fault type is output as r_B.
[0029] The technological advancements achieved by this invention due to the adoption of the above technical solutions are as follows:
[0030] The partial discharge fault type identification method constructed in this invention uses MobileNetV3 and Graph Neural Network (GNN) to build partial discharge fault type identification sub-models. Both sub-models have the advantage of small training amount. The proposed comprehensive partial discharge fault type identification model avoids the problem of inaccurate identification results of a single identification model. It has the advantages of low computational complexity and short computation time. Furthermore, the proposed partial discharge fault type identification method has a high identification accuracy and can identify four types of partial discharge faults, providing technical support for improving the safe and stable operation of power systems. Attached Figure Description
[0031] Figure 1 This is a flowchart illustrating the construction of a comprehensive model for identifying partial discharge fault types in the embodiments of the present invention.
[0032] Figure 2 This is a flowchart of the method in an embodiment of the present invention. Detailed Implementation
[0033] It should be noted that the terms "comprising" and "having" and any variations thereof in the specification, claims and accompanying drawings of this invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such processes, methods, products or devices.
[0034] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments:
[0035] like Figure 1 As shown, a method for identifying partial discharge fault types is proposed. This method constructs a comprehensive model for partial discharge fault type identification, combining... Figure 1 and Figure 2 The process of identifying partial discharge fault types is explained below, with specific steps as follows:
[0036] Step 1: Construct two partial discharge fault type identification sub-models in the comprehensive model for partial discharge fault type identification, and train them;
[0037] Step 1.1: Generate partial discharge ultrasonic signals, including four types of partial discharge: tip discharge, bubble discharge, suspension discharge, and surface discharge. The partial discharge ultrasonic signals can be generated using actual partial discharge ultrasonic signals from the National Basic Science Public Data Center. The number of each type of partial discharge is as follows: 41 tip discharges, 37 bubble discharges, 40 suspension discharges, and 36 surface discharges.
[0038] Step 1.2, Partial Discharge Ultrasonic Signal Processing, includes four steps: noise reduction, conversion to PRPD spectrum, data enhancement, and data normalization. The specific sub-steps are as follows:
[0039] Step 1.2.1: The partial discharge ultrasonic wave is an actual signal and needs to be denoised. The wavelet method is used for denoising.
[0040] Step 1.2.2: Convert the partial discharge ultrasonic signal into a PRPD map with an output size of 224 pixels × 224 pixels;
[0041] Step 1.2.3: Because the number of actual partial discharge ultrasonic signals in the National Basic Science Public Data Center is small and does not meet the data volume required for training, data augmentation is required. Data augmentation is performed by adjusting brightness, sharpness, random rotation, horizontal translation (TranslateX), and vertical translation (TranslateY), resulting in 924 PRPD maps (246 for tip discharge, 222 for bubble discharge, 240 for suspension discharge, and 216 for surface discharge).
[0042] Step 1.2.4: Normalize the PRPD map by using min-max normalization to linearly map the pixel values of the PRPD map to the [0, 1] interval;
[0043] Step 1.3: Divide the normalized PRPD map into training set, validation set and test set, with proportions of 70%, 15% and 15%, respectively.
[0044] Step 1.4: Construct a comprehensive model for identifying partial discharge fault types, including two sub-models for identifying partial discharge fault types: sub-model A based on MobileNetV3 and sub-model B based on graph neural network (GNN).
[0045] The network structure of the partial discharge fault type identification sub-model A based on MobileNetV3 includes an input layer, an initial convolutional layer, inverted residual block 1, inverted residual block 2, inverted residual block 3, a global average pooling layer, a fully connected layer, and an output layer. The input layer takes the normalized PRPD map training data as input, consisting of 647 PRPD maps. The initial convolutional layer has the following parameters: 3×3 kernel, stride of 2, initial number of channels of 16, and h-swish activation function, which quickly extracts low-level features (such as edges and textures) from the PRPD maps. Inverted residual block 1 aggregates the low-level features from the initial convolutional layer and extracts the PRP through a cyclic structure of dimensionality increase-depth separable convolution-dimensionality reduction. The high-level feature maps of the D-map (e.g., overall shape) are extracted; Inverted residual block 2 aggregates the high-level feature maps of inverted residual block 1, and extracts the high-level feature maps of the PRPD map again through a recurrent structure of dimensionality increase-depth separable convolution-dimensionality reduction; Inverted residual block 3 aggregates the high-level feature maps of inverted residual block 2, and extracts the high-level feature maps of the PRPD map again through a recurrent structure of dimensionality increase-depth separable convolution-dimensionality reduction; A global average pooling layer performs global average pooling on the high-level feature maps output by inverted residual block 3, converting them into a single feature value with an output size of 1×1×4; A fully connected layer maps the output of the global average pooling layer to obtain the prediction score for each partial discharge type; The output layer outputs the partial discharge type results.
[0046] The network structure of the partial discharge fault type identification sub-model B based on graph neural network (GNN) includes an input layer, graph convolutional layer 1, graph convolutional layer 2, graph convolutional layer 3, global average pooling layer, fully connected layer, and output layer. The input layer takes the normalized PRPD map training data as input, consisting of 647 PRPD maps. Graph convolutional layer 1 performs the first layer feature extraction of the PRPD maps, dynamically assigning weights from neighboring nodes to the current node based on an attention mechanism. Graph convolutional layer 2 performs the second layer feature extraction of the PRPD maps, dynamically assigning weights from neighboring nodes to the current node based on an attention mechanism. Graph convolutional layer 3 performs the third layer feature extraction of the PRPD maps, dynamically assigning weights from neighboring nodes to the current node based on an attention mechanism. The global average pooling layer reduces the features of the PRPD maps to a vector of fixed length 4. The fully connected layer maps the output of the global average pooling layer to obtain the prediction score for each partial discharge type. The output layer outputs the partial discharge type result.
[0047] Step 1.5: Train the partial discharge fault type identification sub-models A and B respectively. The training requires setting the number of training samples and the number of training rounds. The required number of training samples is 647 and the number of training rounds is 50. The input of identification sub-models A and B is the PRPD map, and the output of identification sub-models A and B is the partial discharge fault type.
[0048] Step 2: Process the partial discharge ultrasonic signal to be identified to form a processed partial discharge signal;
[0049] Step 2.1: Noise reduction processing. The wavelet method is used to reduce the noise of the discharge ultrasonic signal to be identified.
[0050] Step 2.2: Convert the partial discharge ultrasonic signal (tip discharge) to be identified into a PRPD map with an output size of 224 pixels × 224 pixels;
[0051] Step 2.3: Normalize the PRPD map by using min-max normalization to linearly map the pixel values of the PRPD map to the [0, 1] interval.
[0052] Step 3: Input the processed partial discharge signal into the two trained partial discharge fault type identification sub-models to obtain the output result r_A (point discharge) of fault type identification sub-model A and the output result r_B (point discharge) of fault type identification sub-model B, respectively.
[0053] Step 4: Combine the output results of the two partial discharge fault type identification sub-models to make a comprehensive judgment and output the partial discharge fault type. Since r_A (point discharge) and r_B (point discharge) are the same, the output partial discharge fault type is r_A (point discharge).
[0054] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for identifying partial discharge fault types, characterized in that: A comprehensive model for identifying partial discharge fault types is constructed, comprising two sub-models for identifying partial discharge fault types. The input of each sub-model is a partial discharge phase distribution map, and the output of each sub-model is the corresponding partial discharge type. The comprehensive model for identifying partial discharge fault types takes the partial discharge phase distribution map as input and outputs the partial discharge type, thereby identifying four types of partial discharge. The specific steps are as follows: Step 1: Construct two partial discharge fault type identification sub-models in the comprehensive model for partial discharge fault type identification, and train them; Step 2: Process the partial discharge ultrasonic signal to be identified to form a processed partial discharge signal; Step 3: Input the processed partial discharge signal into the trained sub-models for identifying two types of partial discharge faults; Step 4: Combine the output results of the two partial discharge fault type identification sub-models to make a comprehensive judgment and output the partial discharge fault type.
2. The partial discharge fault type identification method according to claim 1, characterized in that: The specific sub-steps for constructing and training the two partial discharge fault type identification sub-models in the comprehensive partial discharge fault type identification model in step 1 are as follows: Step 1.1: Generate partial discharge ultrasonic signals, including four types of partial discharge: tip discharge, bubble discharge, suspension discharge, and surface discharge. The partial discharge ultrasonic signals can be generated using actual or simulated partial discharge ultrasonic signals. Step 1.2: Partial discharge ultrasonic signal processing, including four steps: noise reduction, conversion to partial discharge phase distribution map, data enhancement, and data normalization. The necessity of these four steps depends on the specific circumstances. The specific sub-steps are as follows: Step 1.2.1: The noise reduction process requires determining the source form of the partial discharge ultrasonic signal. If the actual partial discharge ultrasonic signal is used, noise reduction is required. Existing noise reduction methods such as wavelet methods can be used for noise reduction. If the simulated partial discharge ultrasonic signal is used, noise reduction is not required. Step 1.2.2: Convert the partial discharge ultrasonic signal into a partial discharge phase distribution map; Step 1.2.3: Data augmentation requires determining the amount of data in the partial discharge phase distribution spectrum. If the amount of data is insufficient for training, data augmentation is necessary; otherwise, it is not necessary. Step 1.2.4: Normalize the partial discharge phase distribution map by using minimum-maximum normalization to linearly map the pixel values of the partial discharge phase distribution map to the [0, 1] interval; Step 1.3: Divide the normalized partial discharge phase distribution map into a training set, a validation set, and a test set; Step 1.4: Construct a comprehensive model for identifying partial discharge fault types, including two sub-models for identifying partial discharge fault types: sub-model A based on MobileNetV3 and sub-model B based on graph neural network (GNN). The network structure of the partial discharge fault type identification sub-model A based on MobileNetV3 includes an input layer, an initial convolutional layer, inverted residual block 1, inverted residual block 2, inverted residual block 3, a global average pooling layer, a fully connected layer, and an output layer. The input layer takes a normalized partial discharge phase distribution map as its input. The initial convolutional layer quickly extracts low-level features (such as edges and textures) from the partial discharge phase distribution map. Inverted residual block 1 aggregates the low-level features from the initial convolutional layer and extracts high-level feature maps (such as overall shape) from the partial discharge phase distribution map through a cyclic structure of dimensionality increase-depth separable convolution-dimensionality reduction. Inverted residual block 2... The high-level feature map of inverted residual block 1 is aggregated, and a high-level feature map of the partial discharge phase distribution spectrum is extracted again through a recurrent structure of dimensionality increase-depth separable convolution-dimensionality reduction; the high-level feature map of inverted residual block 2 is aggregated, and a high-level feature map of the partial discharge phase distribution spectrum is extracted again through a recurrent structure of dimensionality increase-depth separable convolution-dimensionality reduction; a global average pooling layer performs global average pooling on the high-level feature map output by inverted residual block 3, converting it into a feature value; a fully connected layer maps the output of the global average pooling layer to obtain the prediction score for each partial discharge type; the output layer outputs the partial discharge type result; The network structure of the partial discharge fault type identification sub-model B based on graph neural network (GNN) includes an input layer, graph convolutional layer 1, graph convolutional layer 2, graph convolutional layer 3, global average pooling layer, fully connected layer, and output layer. The input layer takes a normalized partial discharge phase distribution map as input. Graph convolutional layer 1 performs the first layer feature extraction of the partial discharge phase distribution map, dynamically allocating weights from neighboring nodes to the current node based on an attention mechanism. Graph convolutional layer 2 performs the second layer feature extraction of the partial discharge phase distribution map, dynamically allocating weights from neighboring nodes to the current node based on an attention mechanism. Graph convolutional layer 3 performs the third layer feature extraction of the partial discharge phase distribution map, dynamically allocating weights from neighboring nodes to the current node based on an attention mechanism. The global average pooling layer reduces the features of the partial discharge phase distribution map to a fixed-length vector. The fully connected layer maps the output of the global average pooling layer to obtain a prediction score for each partial discharge type. The output layer outputs the partial discharge type result. Step 1.5: Train the partial discharge fault type identification sub-models A and B respectively. The training requires setting the number of training samples and the number of training rounds. The input of identification sub-models A and B is the partial discharge phase distribution map, and the output of identification sub-models A and B is the partial discharge fault type.
3. The partial discharge fault type identification method according to claim 1, characterized in that: Step 2, the data processing of the partial discharge ultrasonic signal to be identified, includes three sub-steps: noise reduction, conversion to a partial discharge phase distribution map, and normalization data processing. Step 2.1: Noise reduction processing. Existing noise reduction methods, such as wavelet methods, are used for noise reduction processing. Step 2.2: Convert the partial discharge ultrasonic signal into a partial discharge phase distribution map; Step 2.3: Normalize the partial discharge phase distribution map by using minimum-maximum normalization to linearly map the pixel values of the partial discharge phase distribution map to the [0, 1] interval.
4. The partial discharge fault type identification method according to claim 1, characterized in that: In step 4, the output results of the two partial discharge fault type identification sub-models are combined for judgment. The output results r_A and r_B of the fault type identification sub-model A and fault type identification sub-model B are obtained respectively. If r_A and r_B are the same, the partial discharge fault type is output as r_A. If r_A and r_B are different, a random number α between 0 and 1 is generated using a random function. If α≤0.5, the partial discharge fault type is output as r_A. If α>0.5, the partial discharge fault type is output as r_B.