Semantic segmentation technology-based atlas processing method and device

Abstract

The invention discloses an atlas processing method and device based on a semantic segmentation technology, and the method comprises the steps: obtaining a to-be-recognized atlas which is a phase resolution partial discharge atlas; processing the to-be-identified atlas according to a pre-constructed semantic segmentation model to obtain a dichotomy result of the to-be-identified atlas, the dichotomy result being used for indicating partial discharge pixels and background pixels of the to-be-identified atlas, the partial discharge pixels being pixels corresponding to partial discharge signals in the to-be-identified atlas, and the background pixels being pixels corresponding to the partial discharge signals in the to-be-identified atlas; the background pixels are pixels except the partial discharge pixels in the atlas to be identified; according to the dichotomy result, background pixels and partial discharge pixels of the to-be-recognized atlas are processed according to background colors and partial discharge pixel rendering modes of standard atlas contained in a standard recognition database to obtain processed to-be-recognized atlas, the processed to-be-recognized atlas and the standard atlas have the same atlas format, and the processed to-be-recognized atlas and the standard atlas have the same atlas format. The processed to-be-identified map is used for identifying the partial discharge type. According to the scheme, the domain offset problem of the PRPD atlas to be recognized can be solved by utilizing the semantic segmentation model.

Description

Technical Field

[0001] This application relates to the technical field of atlas processing, and particularly relates to an atlas processing method and device based on semantic segmentation technology. Background Art

[0002] With the long-term operation of Gas Insulated Switcher (GIS) equipment in operation, local discharge defects sometimes occur inside the equipment. Phase Resolved Partial Discharge (PRPD) atlas data contains statistical information of partial discharge pulse signals during the ultra-high frequency detection process, and can counter the influence of accidental interference in the detection signals. Therefore, at present, in order to ensure the normal operation of GIS equipment and the stable operation and production of the power system, the discharge type of local discharge defects is usually initially determined based on the ultra-high frequency PRPD atlas, so as to take appropriate countermeasures according to the discharge type.

[0003] The discharge types identified based on the PRPD atlas are usually divided into five categories: tip corona, metal particles, solid insulation, floating electrode, and interference.

[0004] Currently, one method for identifying the discharge type based on the PRPD atlas is to collect a large amount of ultra-high frequency partial discharge data through experimental tests, construct a PRPD atlas dataset based on these data to train a model, and use this model to achieve the identification of the GIS equipment atlas. This method has a domain shift problem. The domain shift problem means that when the PRPD atlas to be identified and the constructed dataset belong to the same probability distribution, the trained model can have good accuracy. However, when there is a domain shift between the PRPD atlas to be identified and the dataset, the accuracy of the model is relatively poor.

[0005] During the generation process of the PRPD atlas, due to the use of different ultra-high frequency detection products, there are also significant differences in the background color of the generated atlas, the rendering method of partial discharge pixels, etc., so it is easy to cause a domain shift between the PRPD atlas to be identified and the dataset.

[0006] In the related art, generally, methods such as sample adaptation, feature adaptation, and model adaptation can be used to solve the domain shift problem. However, these methods all require providing a target domain dataset, that is, a PRPD atlas dataset with the same format as the PRPD atlas to be identified. In the case of lacking a dataset with the same format, there is currently no effective method to solve the domain shift problem. Summary of the Invention

[0007] Therefore, the present application discloses the following technical solutions:

[0008] The first aspect of the present application provides a map processing method based on semantic segmentation technology, including:

[0009] Obtain a map to be recognized, where the map to be recognized is a phase-resolved partial discharge map;

[0010] Process the map to be recognized according to a pre-constructed semantic segmentation model to obtain a binary classification result of the map to be recognized. The binary classification result is used to indicate the partial discharge pixels and background pixels of the map to be recognized. The partial discharge pixels are the pixels corresponding to the partial discharge signals in the map to be recognized, and the background pixels are the pixels other than the partial discharge pixels in the map to be recognized;

[0011] According to the binary classification result, process the background pixels and partial discharge pixels of the map to be recognized respectively according to the background color of the standard map contained in the standard recognition database and the rendering method of the partial discharge pixels, so as to obtain the processed map to be recognized. The processed map to be recognized has the same map format as the standard map, and the processed map to be recognized is used to identify the type of partial discharge.

[0012] Optionally, the method for constructing the semantic segmentation model includes:

[0013] Obtain a map background library and a feature template library. The map background library includes multiple phase-resolved partial discharge map backgrounds, and the feature template library includes multiple feature templates. Each feature template includes partial discharge pixels obtained from a phase-resolved partial discharge map;

[0014] Overlay the feature templates contained in the feature template library on the phase-resolved partial discharge map background of the map background library to obtain an overlaid map;

[0015] Generate annotation information for the overlaid map, where the annotation information is used to indicate partial discharge pixels and background pixels in the corresponding overlaid map;

[0016] Train an initial model according to multiple overlaid maps and the annotation information of the overlaid maps to obtain the semantic segmentation model.

[0017] Optionally, the method for obtaining the feature template library includes:

[0018] Extract partial discharge pixels from a phase-resolved partial discharge map to obtain a first feature template composed of the extracted local pixels;

[0019] Perform size transformation processing on the first feature template according to a preset proportional stretching factor to obtain a second feature template;

[0020] Add or delete multiple partial discharge pixels in the first feature template to obtain a third feature template;

[0021] Among them, the first feature template, the second feature template, and the third feature template constitute the feature template library.

[0022] Optionally, the method for obtaining the atlas background library includes:

[0023] Obtain the background of the first phase-resolved partial discharge atlas;

[0024] Perform size transformation on the background of the first phase-resolved partial discharge atlas based on multiple randomly determined target sizes to obtain the background of the second phase-resolved partial discharge atlas;

[0025] Among them, the background of the first phase-resolved partial discharge atlas and the background of the second phase-resolved partial discharge atlas constitute the atlas background library.

[0026] Optionally, it further includes:

[0027] Determine multiple gray value intervals;

[0028] For each of the gray value intervals, count the number of background pixels corresponding to the gray value interval according to the binary classification result, where the number of background pixels is the number of background pixels whose gray values are within the gray value interval;

[0029] Determine the target gray value interval from the multiple gray value intervals according to the number of background pixels;

[0030] Change the partial discharge pixels indicated by the binary classification result and whose gray values are within the target gray value interval to background pixels to obtain the updated binary classification result;

[0031] The processing of the background pixels and partial discharge pixels of the to-be-identified atlas according to the binary classification result, respectively, according to the background color of the standard atlas and the rendering method of partial discharge pixels contained in the standard recognition database includes:

[0032] According to the updated binary classification result, process the background pixels and partial discharge pixels of the to-be-identified atlas, respectively, according to the background color of the standard atlas and the rendering method of partial discharge pixels contained in the standard recognition database.

[0033] The second aspect of this application provides an atlas processing device based on semantic segmentation technology, including:

[0034] An obtaining unit, configured to obtain a to-be-identified atlas, where the to-be-identified atlas is a phase-resolved partial discharge atlas;

[0035] A segmentation unit, configured to process the to-be-recognized atlas according to a pre-constructed semantic segmentation model, and obtain a binary classification result of the to-be-recognized atlas, where the binary classification result is used to indicate the partial discharge pixels and background pixels of the to-be-recognized atlas, the partial discharge pixels are the pixels corresponding to the partial discharge signals in the to-be-recognized atlas, and the background pixels are the pixels other than the partial discharge pixels in the to-be-recognized atlas;

[0036] A processing unit, configured to process the background pixels and partial discharge pixels of the to-be-recognized atlas respectively according to the background color of the standard atlas and the rendering method of the partial discharge pixels included in the standard recognition database according to the binary classification result, so as to obtain a processed to-be-recognized atlas, where the processed to-be-recognized atlas and the standard atlas have the same atlas format, and the processed to-be-recognized atlas is used to identify the partial discharge type.

[0037] Optionally, it further includes a construction unit, configured to:

[0038] Obtain an atlas background library and a feature template library, where the atlas background library includes multiple phase-resolved partial discharge atlas backgrounds, and the feature template library includes multiple feature templates, and each feature template includes partial discharge pixels obtained from a phase-resolved partial discharge atlas;

[0039] Overlay the feature templates included in the feature template library on the phase-resolved partial discharge atlas background of the atlas background library to obtain an overlaid atlas;

[0040] Generate annotation information for the overlaid atlas, where the annotation information is used to indicate partial discharge pixels and background pixels in the corresponding overlaid atlas;

[0041] Train an initial model according to multiple overlaid atlases and the annotation information of the overlaid atlases to obtain the semantic segmentation model.

[0042] Optionally, when obtaining the feature template library, the construction unit is configured to:

[0043] Extract partial discharge pixels from a phase-resolved partial discharge atlas to obtain a first feature template composed of the extracted local pixels;

[0044] Perform size transformation processing on the first feature template according to a preset scaling factor to obtain a second feature template;

[0045] Add or delete multiple partial discharge pixels in the first feature template to obtain a third feature template;

[0046] Wherein, the first feature template, the second feature template and the third feature template constitute the feature template library.

[0047] Optionally, when obtaining the atlas background library, the construction unit is configured to:

[0048] Obtain the background of the first phase-resolved partial discharge pattern;

[0049] Perform size transformation on the background of the first phase-resolved partial discharge pattern based on multiple randomly determined target sizes to obtain the background of the second phase-resolved partial discharge pattern;

[0050] Wherein, the background of the first phase-resolved partial discharge pattern and the background of the second phase-resolved partial discharge pattern constitute the pattern background library.

[0051] Optionally, the segmentation unit is further configured to:

[0052] Determine multiple gray value intervals;

[0053] For each of the gray value intervals, count the number of background pixels corresponding to the gray value interval according to the binary classification result, where the number of background pixels is the number of background pixels whose gray values are within the gray value interval;

[0054] Determine a target gray value interval from the multiple gray value intervals according to the number of background pixels;

[0055] Change the partial discharge pixels indicated by the binary classification result and whose gray values are within the target gray value interval to background pixels to obtain an updated binary classification result;

[0056] When the processing unit processes the background pixels and partial discharge pixels of the to-be-recognized pattern according to the background color and the rendering method of the partial discharge pixels of the standard pattern contained in the standard recognition database based on the binary classification result, it is configured to:

[0057] Process the background pixels and partial discharge pixels of the to-be-recognized pattern according to the updated binary classification result, respectively, according to the background color and the rendering method of the partial discharge pixels of the standard pattern contained in the standard recognition database.

[0058] The beneficial effect of this solution is as follows:

[0059] Before identifying the partial discharge type based on the phase-resolved partial discharge pattern, this solution can use a semantic segmentation model to segment the partial discharge pixels and background pixels in the to-be-recognized pattern, and then process the background pixels and partial discharge pixels of the to-be-recognized pattern according to the background color and the rendering method of the partial discharge pixels of the standard pattern, respectively, to obtain a processed to-be-recognized pattern with the same pattern format as the standard pattern. Thus, this solution can make the pattern used for recognition have the same pattern format as the standard pattern, thereby solving the problem of domain shift. Description of the Drawings

[0060] To more clearly illustrate the technical solutions in the embodiments of the present application 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 only the embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on the provided drawings.

[0061] Figure 1 It is a flowchart of a graph processing method based on semantic segmentation technology provided by an embodiment of the present application;

[0062] Figure 2 It is a schematic structural diagram of a semantic segmentation model provided by an embodiment of the present application;

[0063] Figure 3 It is a schematic diagram of a graph to be recognized and a graph to be recognized after processing provided by an embodiment of the present application;

[0064] Figure 4 It is a schematic diagram of graphs in a graph background library and a feature template library provided by an embodiment of the present application;

[0065] Figure 5 It is a schematic diagram of a method for obtaining a superimposed graph provided by an embodiment of the present application;

[0066] Figure 6 It is a schematic diagram of multiple superimposed graphs provided by an embodiment of the present application;

[0067] Figure 7 It is a schematic diagram of annotation information of multiple superimposed graphs provided by an embodiment of the present application;

[0068] Figure 8 It is a schematic structural diagram of a graph processing device based on semantic segmentation technology provided by an embodiment of the present application. Detailed implementation manners

[0069] The following will clearly and completely describe the technical solutions in the embodiments of the present application with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present application.

[0070] This embodiment provides a graph processing method based on semantic segmentation technology. Please refer to Figure 1 , which is a flowchart of this method. This method may include the following steps.

[0071] S101, obtain a graph to be recognized, and the graph to be recognized is a phase-resolved partial discharge graph.

[0072] The spectrum to be identified can be obtained based on a gas insulated switch (GIS) device that needs to perform partial discharge detection. For example, ultra-high frequency detection can be performed on the GIS device that needs to perform partial discharge detection, and the pulse signal obtained by the detection can be converted into a phase-resolved partial discharge (PRPD) spectrum, which is used as the spectrum to be identified in step S101.

[0073] In this embodiment, the spectrum to be identified can be represented by PRPD_d.

[0074] Generally, the obtained image to be identified may be a grayscale image, in which each pixel corresponds to a pulse signal obtained by ultra-high frequency detection, and the grayscale value of the pixel is related to the intensity or other properties of the pulse signal.

[0075] S102, processing the to-be-identified graph according to a pre-built semantic segmentation model to obtain a binary classification result of the to-be-identified graph, wherein the binary classification result is used to indicate the partial discharge pixels and background pixels of the to-be-identified graph, wherein the partial discharge pixels are pixels corresponding to the local discharge signals in the to-be-identified graph, and the background pixels are pixels other than the partial discharge pixels in the to-be-identified graph.

[0076] The semantic segmentation model can be trained using a large number of graphs that are pre-labeled with local discharge pixels and background pixels. The semantic segmentation model can adopt any existing neural network model structure, which is not limited in this embodiment. As an example, the structure of the semantic segmentation model can be a pyramid scene parsing network (Pyramid Scene Parsing Network, Pspnet) structure. The schematic diagram of the structure can be seen in Figure 2 .

[0077] like Figure 2 As shown, the semantic segmentation model implemented based on the Pspnet network may include a mobile network (Mobilenet) structure as a backbone network, and the backbone network is used to extract graph features of an input RPD graph (eg, a graph to be identified).

[0078] The obtained graph features can be input into the pyramid scene parsing block (pspblock) of the semantic segmentation model. The pspblock can include a pyramid pooling structure. The pyramid pooling structure uses kernel modules with sizes of 1*1, 2*2, 3*3, and 6*6 to perform average pooling operations on the input graph features. Finally, the outputs corresponding to the above multiple kernels are fused to obtain fused pooling features.

[0079] Then, the fused pooling features can be upsampled, and the upsampled fused pooling features and the atlas features are concatenated to obtain concatenated features. Finally, the concatenated features are processed by the convolutional network module in the semantic segmentation model, and the binary classification result corresponding to the input PRPD atlas can be obtained.

[0080] The PRPD atlas that the above semantic segmentation model can process can be a PRPD atlas with a size of [473, 473], or the semantic segmentation model can be adjusted according to actual needs so that the model can process PRPD atlases of other sizes.

[0081] The training parameters used when training the semantic segmentation model can be set as needed without limitation. As an example, the number of samples batch_size used for one training can be set to 4, the upper limit of the number of training epochs can be set to 115, and the learning rate lr can be set to 0.0001.

[0082] During the process of training the semantic segmentation model, the loss function used can be an additional loss function or other existing loss functions in the field of neural network technology without limitation.

[0083] In this embodiment, the trained semantic segmentation model can be saved as a best.pth file in the electronic device. When S102 needs to be executed, the processor of the electronic device loads the semantic segmentation model from this file, and then inputs the above-mentioned atlas to be recognized into the loaded semantic segmentation model for processing, and the binary classification result of the atlas to be recognized (denoted as PRPD_c) can be obtained.

[0084] Since the partial discharge pixel distribution in the PRPD atlas has the characteristics of coexistence of discreteness and aggregation, the semantic segmentation result needs to contain independent partial discharge pixels, rather than a closed connected region. A network model for pixel-level segmentation needs to be adopted. The traditional fully convolutional neural network (FCN) belongs to a pixel-level segmentation network, but only uses global information and does not fully combine context information. In this embodiment, a semantic segmentation model based on the Pspnet network structure is adopted, which can combine context information during the processing process, so as to obtain a more accurate binary classification result.

[0085] In the PRPD atlas, each pixel corresponds to a detected pulse signal. Among the detected pulse signals, some pulse signals are related to the partial discharge phenomenon of the GIS device (for example, the pulse signals generated by the partial discharge phenomenon). Such signals are partial discharge signals, and the pixels in the PRPD atlas corresponding to these partial discharge signals are partial discharge pixels.

[0086] S103. According to the binary classification result, process the background pixels and partial discharge pixels of the to-be-recognized map respectively according to the background color of the standard maps contained in the standard recognition database and the rendering method of partial discharge pixels, so as to obtain the processed to-be-recognized map. The processed to-be-recognized map and the standard map have the same map format, and the processed to-be-recognized map is used to identify the type of partial discharge.

[0087] Step S103 aims to convert the domain shift problem existing between the to-be-recognized map and the standard map into a map format standardization operation. Use the semantic segmentation result to obtain the positions of the background pixels and the partial discharge map positions in the map, extract the map format in the standard PRPD recognition data set, and redefine the gray value of the background pixels and the color rendering method of the partial discharge map of the to-be-recognized map.

[0088] Specifically, in step S103, the color value of the background pixels of each standard map in the standard recognition database, that is, the above-mentioned background color, can be determined first, and the rendering method of the partial discharge pixels of each standard map can be determined.

[0089] Then, on the one hand, based on the binary classification result of the to-be-recognized map, find the background pixels in the to-be-recognized map, and assign the above-mentioned background color to the background pixels of the to-be-recognized map, so that the color of the background pixels of the to-be-recognized map is the same as the color of the background pixels of the standard map.

[0090] On the other hand, based on the binary classification result of the to-be-recognized map, find the partial discharge pixels in the to-be-recognized map, and render the partial discharge pixels of the to-be-recognized map according to the rendering method of the partial discharge pixels of the standard map, so that the rendering method of the partial discharge pixels in the to-be-recognized map is the same as the rendering method of the partial discharge pixels in the standard map.

[0091] The map obtained after processing the partial discharge pixels and background pixels of the to-be-recognized map in the above manner is the processed to-be-recognized map of S103 (denoted as PRPD_d’).

[0092] When obtaining the PRPD map, the color of the partial discharge pixel corresponding to the partial discharge signal in the PRPD map can be determined according to the intensity of the detected partial discharge signal. In the PRPD map, the mapping relationship between the intensity of the partial discharge signal and the color of the corresponding partial discharge pixel is equivalent to the rendering method of the partial discharge pixels of the PRPD map. Therefore, the process of rendering the partial discharge pixels of the to-be-recognized map in the second aspect above is equivalent to re-determining the color of each partial discharge pixel in the to-be-recognized map according to the mapping relationship between the intensity of the partial discharge signal and the color of the corresponding partial discharge pixel in the standard map, so that the mapping relationship of the processed to-be-recognized map is the same as that of the standard map.

[0093] As some examples, please refer to Figure 3 , Figure 3(1), (2), and (3) are the to-be-recognized spectrograms before processing. Among them, the pixels in the grayish-black part are the background pixels indicated by the binary classification result, and the pixels in the red part are the partial discharge pixels indicated by the binary classification result. Figure 3 (4) is the to-be-recognized spectrogram after processing obtained by processing a certain to-be-recognized spectrogram in the manner of S103. Among them, the pixels in the dark blue part are the background pixels, and the pixels in the colored part are the partial discharge pixels.

[0094] The standard recognition database of this embodiment can be understood as a database for training a partial discharge classification model. The multiple standard spectrograms included in the standard recognition database can be the sample spectrograms used when training the partial discharge classification model. The partial discharge classification model is a pre-trained neural network model for identifying the corresponding partial discharge type according to the PRPD spectrogram. After obtaining the to-be-recognized spectrogram after processing in S103, the to-be-recognized spectrogram after processing can be processed by the partial discharge classification model to determine the partial discharge type of the GIS device that needs to be detected for partial discharge in S101.

[0095] The spectrogram format of the PRPD spectrogram can include: the color of the background pixels in the PRPD spectrogram, the color of the partial discharge pixels in the PRPD spectrogram, and the mapping relationship of the intensity of the partial discharge signal. If the colors of the background pixels of two PRPD spectrograms are the same and the mapping relationships are the same, it can be considered that these two PRPD spectrograms have the same spectrogram format.

[0096] It can be seen from this that through the processing method of this embodiment, the to-be-recognized spectrogram after processing for identifying the partial discharge type can have the same background color as the standard spectrogram, and the mapping relationship between the color of the partial discharge pixels and the intensity of the partial discharge signal in the to-be-recognized spectrogram after processing is the same as the mapping relationship between the color of the partial discharge pixels and the intensity of the partial discharge signal in the standard spectrogram. In other words, the to-be-recognized spectrogram after processing and the standard spectrogram have the same spectrogram format, thereby correcting the domain shift problem caused by the inconsistent spectrogram format between the to-be-recognized spectrogram and the standard spectrogram.

[0097] The beneficial effects of this embodiment are as follows:

[0098] In the related art, there is a domain shift problem between the standard spectrogram and the to-be-recognized spectrogram. The domain shift problem refers to that the signal intensity represented by the color of the partial discharge pixels in the standard spectrogram is different from the signal intensity represented by the color of the partial discharge pixels in the to-be-recognized spectrogram. For example, the same color represents the signal intensity X1 in the standard spectrogram but represents the signal intensity X2 in the to-be-recognized spectrogram. Due to the existence of this difference, the relationship between the color distribution of the partial discharge pixels learned by the partial discharge classification model from the standard spectrogram and the partial discharge type may not be applicable to the color distribution of the partial discharge pixels in the to-be-recognized spectrogram, resulting in inaccurate identification of the partial discharge type.

[0099] Before identifying the partial discharge type based on the phase-resolved partial discharge pattern, this solution can use a semantic segmentation model to segment the partial discharge pixels and background pixels in the pattern to be identified, and then process the background pixels and partial discharge pixels of the pattern to be identified according to the background color of the standard pattern and the rendering method of the partial discharge pixels respectively, thereby correcting the above-mentioned domain shift problem and obtaining the processed pattern to be identified with the same pattern format as the standard pattern. Therefore, this solution can make the pattern used for identification have the same pattern format as the standard pattern, thus solving the problem of domain shift.

[0100] In some alternative embodiments, the method for constructing the semantic segmentation model may include:

[0101] A1. Obtain a pattern background library and a feature template library. The pattern background library includes multiple phase-resolved partial discharge pattern backgrounds, and the feature template library includes multiple feature templates. Each feature template includes the partial discharge pixels obtained from a phase-resolved partial discharge pattern;

[0102] A2. Superimpose the feature templates contained in the feature template library on the phase-resolved partial discharge pattern backgrounds in the pattern background library to obtain a superimposed pattern;

[0103] A3. Generate the annotation information of the superimposed pattern, and the annotation information is used to indicate the partial discharge pixels and background pixels in the corresponding superimposed pattern;

[0104] A4. Train an initial model according to multiple superimposed patterns and the annotation information of the superimposed patterns to obtain the semantic segmentation model.

[0105] In step A1, the method for obtaining the pattern background library may include:

[0106] Obtain the first phase-resolved partial discharge pattern background;

[0107] Perform size transformation on the first phase-resolved partial discharge pattern background based on multiple randomly determined target sizes to obtain the second phase-resolved partial discharge pattern background;

[0108] Among them, the first phase-resolved partial discharge pattern background and the second phase-resolved partial discharge pattern background constitute the pattern background library.

[0109] The first phase-resolved partial discharge pattern background may include the original background of the pattern provided by mainstream UHF partial discharge manufacturers, or may include the picture background obtained by performing UHF detection on GIS equipment without partial discharge.

[0110] When performing size transformation, a transformation size range can be set first, for example, defined as 56*224, and then several target sizes within this range are randomly generated, and each target size is smaller than the set transformation size range;

[0111] After determining the target size, the principle of iconography can be used to change the size of the background of any one or more first phase-resolved partial discharge patterns, so as to convert the background of the first phase-resolved partial discharge pattern to be processed into the background of the second phase-resolved partial discharge pattern with the corresponding target size.

[0112] Exemplarily, a randomly determined target size can be 45*200. After performing size transformation on the background of the first phase-resolved partial discharge pattern, a background of the second phase-resolved partial discharge pattern with a size of 45*200 can be obtained.

[0113] The unit of the above size can be pixels, centimeters, inches or other units applicable to two-dimensional patterns, which is not limited.

[0114] After performing size transformation in the above manner, the set composed of multiple backgrounds of the first phase-resolved partial discharge patterns and multiple backgrounds of the second phase-resolved partial discharge patterns obtained can be used as the pattern background library, denoted as Dataset_B.

[0115] The advantage of obtaining the pattern background library in the above manner is that the pixel resolutions of the PRPD patterns to be recognized are different, and the distribution form of the partial discharge pixels in the patterns is uncertain. By the above method, background patterns of various sizes can be obtained, improving the size diversity of the pattern background library, so that the semantic segmentation model trained using the pattern background library can learn the features of patterns of various sizes, and thus output more accurate binary classification results.

[0116] Optionally, the method for obtaining the feature template library in step A1 includes:

[0117] Extract partial discharge pixels from the phase-resolved partial discharge pattern to obtain a first feature template composed of the extracted partial pixels;

[0118] Perform size transformation processing on the first feature template according to a preset scaling factor to obtain a second feature template;

[0119] Add or delete multiple partial discharge pixels in the first feature template to obtain a third feature template;

[0120] Among them, the first feature template, the second feature template and the third feature template constitute the feature template library.

[0121] In this embodiment, the UHF detection can be first performed on the GIS device with partial discharge phenomenon to obtain n PRPD patterns with partial discharge phenomenon, and then the partial discharge pixels of each PRPD pattern are extracted to obtain n first feature templates, denoted as model1, model2... modeln in sequence. Among them, for each PRPD pattern, all the partial discharge pixels of the pattern constitute the first feature template corresponding to the pattern.

[0122] For each first feature template, the coordinates of the central pixel point of the first feature template can be defined as (0, 0), and a plane rectangular coordinate system of the first feature template can be defined with this central pixel point as the origin. The coordinates of any other pixel point i in the first feature template can be represented by (x i , y i ), where x i represents the distance from pixel point i to the central pixel point in the horizontal axis (x-axis) direction, and y i represents the distance from pixel point i to the central pixel point in the vertical axis (y-axis) direction.

[0123] After obtaining the first feature template, size transformation and sparsity transformation can be performed on the first feature template respectively.

[0124] Size transformation means performing size transformation processing on the first feature template according to a preset proportional stretching factor to obtain a second feature template.

[0125] Specifically, the proportional stretching factors θ and r can be predefined, and the specific values of these two factors can be set as needed and are not limited in this embodiment.

[0126] For any first feature template, the proportional stretching factor can be used to perform size transformation processing on each pixel point of the first feature template according to the following formulas (1) and (2) to obtain multiple corresponding processed pixel points. The new map composed of these processed pixel points is the second feature template obtained by size transformation processing.

[0127] x inew = r * cos(θ) * x i , (1).

[0128] y inew = r * sin(θ) * y i , (2).

[0129] In the above formulas, (x i , y i ) are the coordinates of the i-th pixel point of the first feature template, (x inew , y inew ) are the coordinates of the processed pixel point, cos is the cosine function, and sin is the sine function.

[0130] Optionally, when performing size transformation, different values of the proportional stretching factor can be set as needed, and the size transformation processing can be performed on the first feature template based on different values of the proportional stretching factor to obtain different second feature templates.

[0131] Sparsity transformation means adding or deleting multiple partial discharge pixels in the first feature template to obtain a third feature template.

[0132] Specifically, for any first feature template, the total number of all partial discharge pixels in the first feature template can be defined as N. Correspondingly, the coordinates of all partial discharge pixels in the first feature template can be sequentially recorded as (x 1 , y 1 ), (x 2 , y 2 )... (x N , y N );

[0133] On the one hand, for the first feature template that needs to be sparsely transformed, a subset can be randomly selected from the N partial discharge pixels of the first feature template, that is, a part of the partial discharge pixels are randomly selected, and the feature template formed by the selected partial discharge pixels is used as the third feature template obtained after deleting some partial discharge pixels from the original first feature template. This process reduces the pixel sparsity of the original template.

[0134] On the other hand, a part of the partial discharge pixels can be randomly selected from any first feature template and added to another first feature template, thereby obtaining a third feature template with additional partial discharge pixels on the basis of the first feature template. For example, 2 first feature templates modeli and modelj can be randomly selected, and some or all of the partial discharge pixels of modelj are selected, and the set formed by the selected partial discharge pixels and the partial discharge pixels of modeli is used as the third feature template. This process can change the pixel density within the first feature template, thereby changing the sparsity of the first feature template.

[0135] After performing size transformation and sparse transformation in the above manner, the set composed of the obtained multiple first feature templates, multiple second feature templates, and multiple third feature templates can be used as the feature template library, denoted as Dataset_F.

[0136] Obtaining the feature template library according to the above method can improve the data diversity of the feature template library, and further improve the accuracy and versatility of the semantic segmentation model trained using the feature template library to obtain more accurate binary classification results.

[0137] As some examples, please refer to Figure 4 , Figure 4 (1) and (2) of Figure 4 are schematic diagrams of two phase-resolved partial discharge pattern backgrounds in the atlas background library,

[0138] (3), (4), and (5) of

[0139] Randomly select a phase-resolved partial discharge pattern background (denoted as PRPD_B) and a feature template (denoted as PRPD_F) from the feature template library and the pattern background library respectively. Among them, the size of PRPD_B can be expressed as xb * yb, and the size of PRPD_F can be expressed as xf * yf;

[0140] Compare the sizes of the two. If xb is greater than xf and yb is greater than yf, directly execute the superposition step. If xb is less than or equal to xf, or yb is less than or equal to yf, perform the size conversion step on PRPD_F;

[0141] The size conversion step is specifically as follows: Set the proportional stretching factor, specifically set as: r is greater than 0 and less than, θ is equal to 45 degrees. Based on this proportional stretching factor, use the aforementioned formulas (1) and (2) to perform size conversion processing on each pixel in PRPD_F to obtain the corresponding converted pixel. When converting, (x i , y i ) is the coordinate of the pixel in PRPD_F, (x inew , y inew ) is the coordinate of the converted pixel. The pattern composed of the converted pixels can be used as the converted feature template. Through the size conversion step, the size of the feature template can be reduced to less than PRPD_B. That is to say, the sizes of the converted feature template in the x direction and the y direction are both less than PRPD_B;

[0142] Superposition step: First, determine the random value range of the superposition position (x0, y0) in PRPD_B. Specifically, if the central pixel point of PRPD_B is used as the origin of the coordinate system and its coordinate is defined as (0, 0), then the value range of x0 in the superposition position (x0, y0) is [- (xb - xf) / 2 + 1, (xb - xf) / 2 - 1], and the value range of y0 is [- (yb - yf) / 2 + 1, (yb - yf) / 2 - 1]. Then, a value of x0 and a value of y0 can be randomly determined within this random value range, thereby determining the superposition position (x0, y0);

[0143] After determining the superposition position, the central pixel point of the feature template can be aligned with the center of PRPD_B, so as to superpose the feature template PRPD_F and PRPD_B together to obtain a superposed pattern;

[0144] Exemplarily, please refer to Figure 5 , Figure 5 where (1) is the pattern background PRPD_B, Figure 5 and (2) is the feature template PRPD_F. After determining the superposition position in PRPD_B, align the superposition position of (1) and the center of (2) and then superpose them to obtain Figure 5 the superposed result of (1) and (2) inFigure 5 the superimposed spectrogram shown in (3) thereof (denoted as PRPD_gen);

[0145] Wherein, if the size conversion step is performed, PRPD_F used in the superimposing step should be replaced with the converted feature template, and xf and yf involved in the determined value range in the superimposing step should be correspondingly replaced with the size of the converted feature template in the x direction and the size in the y direction.

[0146] Performing the above steps on the spectrogram background library and the feature template library multiple times can obtain multiple superimposed spectrograms.

[0147] As some examples, please refer to Figure 6 which is a schematic diagram of multiple superimposed spectrograms obtained by the method of this embodiment.

[0148] In step A3, the form of the annotation information of the superimposed spectrogram is not limited. As some examples, for each superimposed spectrogram, a mask corresponding to the superimposed spectrogram can be generated as its annotation information (denoted as PRPD_map). Specifically, the color of each background pixel in the superimposed spectrogram can be set to black, and the color of each partial discharge pixel can be set to white. The image composed of these black background pixels and white partial discharge pixels is the mask of the superimposed spectrogram.

[0149] As some examples, please refer to Figure 7 which is a schematic diagram of the annotation information PRPD_map corresponding to multiple superimposed spectrograms generated by the method of this embodiment.

[0150] The above multiple superimposed spectrograms and the corresponding annotation information constitute a semantic segmentation dataset (Dataset_Y). In step A4, the aforementioned semantic segmentation model can be trained according to Dataset_Y.

[0151] In step A4, an initial model for training can be obtained first. This initial model has the same structure as the semantic segmentation model, and the model parameters included in the initial model can be generated by random initialization;

[0152] Secondly, the above multiple superimposed spectrograms can be processed with the initial model to obtain the binary classification result corresponding to each superimposed spectrogram. The binary classification result of the superimposed spectrogram indicates which pixels in the superimposed spectrogram are partial discharge pixels and which pixels are background pixels;

[0153] Then, compare the binary classification result of the superimposed spectrogram with the annotation information, and determine the model loss of the initial model based on the difference between the two. The size of the difference is positively correlated with the size of the model loss. The specific method for determining the model loss according to the difference can refer to relevant existing technologies and will not be elaborated;

[0154] Next, it is determined whether the model loss is less than a preset loss threshold. If it is less than the loss threshold, the current initial model is output as the trained semantic segmentation model. If it is greater than or equal to the loss threshold, the model parameters are updated according to the current model loss, and after the update is completed, the step of processing the above-mentioned multiple superimposed maps with the initial model is returned for execution until the model loss obtained in a certain iteration is less than the loss threshold;

[0155] Among them, the method of updating the model parameters according to the model loss can refer to the relevant prior art and will not be elaborated.

[0156] In some optional embodiments, before executing step S103, the binary classification result of the map to be recognized can also be updated in the following manner:

[0157] B1. Determine multiple gray value intervals;

[0158] B2. For each gray value interval, according to the binary classification result, count the number of background pixels corresponding to the gray value interval, where the number of background pixels is the number of background pixels whose gray values are within the gray value interval;

[0159] B3. Determine the target gray value interval from the multiple gray value intervals according to the number of background pixels;

[0160] B4. Change the partial discharge pixels indicated by the binary classification result and whose gray values are within the target gray value interval to background pixels to obtain the updated binary classification result.

[0161] In step B1, the value range of the gray values in the image can be evenly divided according to a certain number of intervals to obtain multiple gray value intervals. For example, the value range [0, 255] of the gray values is evenly divided into 15 gray value intervals, and the width of each gray value interval is 17, that is, divided into 15 gray value intervals such as [0, 17], [18, 35]... [238, 255].

[0162] In step B2, the set Set_b of all background pixels in the map to be recognized can be obtained according to the binary classification result, and then the gray histogram gray of Set_b is statistically calculated based on the above-mentioned multiple gray value intervals. The statistical method is to count the number of gray values of the pixels in Set_b in 15 gray value intervals, so as to draw gray with the horizontal axis being the gray value interval and the vertical axis being the number of pixels whose gray values are within the gray value interval;

[0163] Exemplarily, in this gray, the number of background pixels corresponding to [0, 17] can be 100, indicating that there are 100 pixels in Set_b whose gray values are within the interval [0, 17];

[0164] In step B3, the gray value interval with the largest number of corresponding background pixels can be determined according to the statistical gray, and this gray value interval is determined as the target gray value interval, which is recorded as [Gb, Gb+17];

[0165] As an example, suppose that after statistics, it is found that the number of background pixels corresponding to the interval [36, 53] is the largest, indicating that the number of pixels in Set_b that are located in this interval is the largest, so [36, 53] is determined as the target gray value interval. In this case, Gb is equal to 36;

[0166] In step B4, all partial discharge pixels in the to-be-identified graph can be obtained based on the binary classification result of the to-be-identified graph. The set formed by these partial discharge pixels is recorded as Set_f. Then, it is determined whether the grayscale value of each pixel in Set_f is within the target grayscale value interval [Gb, Gb+17]. If the grayscale value of a pixel in Set_f is within the target grayscale value interval, the pixel is removed from Set_f and added to Set_b, that is, the pixel is changed from a partial discharge pixel to a background pixel. In this way, the new binary classification result obtained after traversing each partial discharge pixel indicated by the binary classification result obtained in S102 is the updated binary classification result of step B4.

[0167] The above 15 grayscale value intervals are only examples. In other embodiments, other numbers of grayscale value intervals may be equally divided, not limited to 15.

[0168] If the binary classification result output by the semantic segmentation model is updated in the above manner, step S103 can be replaced by:

[0169] According to the updated binary classification results, the background pixels and partial discharge pixels of the to-be-identified atlas are processed respectively according to the background color and partial discharge pixel rendering method of the standard atlas contained in the standard recognition database.

[0170] In some examples, a background image library containing 400 phase-resolved partial discharge background images and a feature template library containing 400 feature templates can be obtained according to the aforementioned method. Then, 1000 superimposed images and 1000 annotation information corresponding to the 1000 superimposed images can be generated based on the above-mentioned background image library and feature template library, and the semantic segmentation model can be trained using the semantic segmentation dataset (Dataset_Y) composed of the 1000 superimposed images and their annotation information.

[0171] During training, Dataset_Y can be divided into a training set and a validation set in a ratio of 8:2, that is, 80% of the data in Dataset_Y is used for training, and the remaining 20% ​​of the data is used for validation.

[0172] Some of the superimposed maps in Dataset_Y can be found inFigure 6 , some of the annotation information can be found in Figure 7 .

[0173] It can be seen that the superimposed spectrogram generated in the above manner contains diverse background variations and formats, and the distribution patterns of the partial discharge pixels in the spectrogram are different. There are both partial discharge characteristic patterns aggregated into connected regions and scattered point and cluster-like partial discharge spatial patterns. Therefore, the semantic segmentation dataset constructed by the above method has rich background diversity and diversity in the distribution of partial discharge characteristics. The semantic segmentation model trained using this dataset can meet the semantic segmentation requirements of real PRPD spectrograms.

[0174] After training a semantic segmentation model with the Pspnet structure using the above dataset, when using this semantic segmentation model to process several spectrograms to be recognized, binary classification results as shown in (1) to (3) of Figure 3 can be obtained. It can be seen that Figure 3 the binary classification results of the spectrograms to be recognized in (1) to (3) of

[0175] include various spectrogram formats. For different spectrograms to be recognized with diverse partial discharge characteristics and background formats, the semantic segmentation model of this embodiment can effectively identify the partial discharge pixels and background pixels of the spectrograms to be recognized. Therefore, this semantic segmentation model can meet the semantic segmentation requirements of real PRPD spectrograms. Figure 3 Furthermore, the binary classification result as shown in (1) of Figure 3 can be updated according to the method of updating the binary classification result described above. Then, based on the updated binary classification result, the corresponding spectrogram to be recognized is processed to obtain

[0176] the processed spectrogram to be recognized as shown in (4) of Figure 3

[0177]

[0177] In summary, to solve the problem that the domain shift between the multi-format spectra to be recognized and the standard spectra of the standard recognition dataset leads to deviations in the recognition results obtained based on the spectra to be recognized, in this embodiment, in the case of a lack of a spectrum segmentation detection dataset, a semantic segmentation model based on the Pspnet network structure or other neural network structures is designed. Using this semantic segmentation model, the spectrum format conversion of the original spectra to be recognized into the spectrum format of the standard spectra is realized, and the domain shift problem between the spectra to be recognized and the standard spectra of the standard recognition database is solved during the process of recognizing the partial discharge type based on the PRPD spectra. Relying on a small number of accumulated PRPD spectra, the recognition and detection of the partial discharge type of the PRPD spectra driven by image data are realized.

[0178] This embodiment also provides a spectrum processing device based on semantic segmentation technology. Please refer to Figure 8 , which is a schematic structural diagram of the device. The device may include the following units.

[0179] An obtaining unit 801, configured to obtain a spectrum to be recognized, where the spectrum to be recognized is a phase-resolved partial discharge spectrum;

[0180] A segmentation unit 802, configured to process the spectrum to be recognized according to a pre-constructed semantic segmentation model to obtain a binary classification result of the spectrum to be recognized. The binary classification result is used to indicate the partial discharge pixels and background pixels of the spectrum to be recognized. The partial discharge pixels are the pixels corresponding to the partial discharge signals in the spectrum to be recognized, and the background pixels are the pixels other than the partial discharge pixels in the spectrum to be recognized;

[0181] A processing unit 803, configured to process the background pixels and partial discharge pixels of the spectrum to be recognized respectively according to the binary classification result and the background color and partial discharge pixel rendering method of the standard spectra included in the standard recognition database, so as to obtain the processed spectrum to be recognized. The processed spectrum to be recognized and the standard spectrum have the same spectrum format, and the processed spectrum to be recognized is used to recognize the partial discharge type.

[0182] Optionally, it further includes a construction unit 804, configured to:

[0183] Obtain a spectrum background library and a feature template library. The spectrum background library includes multiple phase-resolved partial discharge spectrum backgrounds, and the feature template library includes multiple feature templates. Each feature template includes partial discharge pixels obtained from a phase-resolved partial discharge spectrum;

[0184] Overlay the feature templates included in the feature template library on the phase-resolved partial discharge spectrum background of the spectrum background library to obtain an overlaid spectrum;

[0185] Generate annotation information for the overlaid spectrum, where the annotation information is used to indicate partial discharge pixels and background pixels in the corresponding overlaid spectrum;

[0186] Train an initial model based on multiple superimposed maps and the annotation information of the superimposed maps to obtain a semantic segmentation model.

[0187] Optionally, when the construction unit 804 obtains the feature template library, it is used for:

[0188] Extract partial discharge pixels from the phase-resolved partial discharge map to obtain a first feature template composed of the extracted partial pixels;

[0189] Perform size transformation processing on the first feature template according to a preset proportional stretching factor to obtain a second feature template;

[0190] Add or delete multiple partial discharge pixels in the first feature template to obtain a third feature template;

[0191] Among them, the first feature template, the second feature template, and the third feature template constitute the feature template library.

[0192] Optionally, when the construction unit 804 obtains the map background library, it is used for:

[0193] Obtain the background of the first phase-resolved partial discharge map;

[0194] Perform size transformation on the background of the first phase-resolved partial discharge map based on multiple randomly determined target sizes to obtain the background of the second phase-resolved partial discharge map;

[0195] Among them, the background of the first phase-resolved partial discharge map and the background of the second phase-resolved partial discharge map constitute the map background library.

[0196] Optionally, the segmentation unit 802 is further used for:

[0197] Determine multiple gray value intervals;

[0198] For each gray value interval, according to the binary classification result, count the number of background pixels corresponding to the gray value interval, and the number of background pixels is the number of background pixels whose gray value is within the gray value interval;

[0199] Determine the target gray value interval from multiple gray value intervals according to the number of background pixels;

[0200] Change the partial discharge pixels indicated by the binary classification result and whose gray value is within the target gray value interval to background pixels to obtain the updated binary classification result;

[0201] When the processing unit 803 processes the background pixels and partial discharge pixels of the map to be recognized according to the binary classification result, respectively, according to the background color of the standard map contained in the standard recognition database and the rendering method of the partial discharge pixels, it is used for:

[0202] According to the updated binary classification results, the background pixels and partial discharge pixels of the to-be-recognized map are processed according to the background color of the standard map contained in the standard recognition database and the rendering method of the partial discharge pixels respectively.

[0203] For the map processing device based on semantic segmentation technology in this embodiment, its working principle can refer to the relevant steps in the map processing method based on semantic segmentation technology provided in any embodiment, which will not be elaborated here.

[0204] It should be noted that the embodiments in this specification are all described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The same or similar parts among the embodiments can be referred to each other.

[0205] For the convenience of description, when describing the above system or device, it is divided into various modules or units according to functions for separate description. Of course, when implementing this application, the functions of each unit can be implemented in the same or multiple software and / or hardware.

[0206] From the description of the above embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus a necessary general hardware platform. Based on this understanding, the technical solution of this application, in essence, 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 storage medium, such as ROM / RAM, magnetic disk, optical disc, 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 various embodiments or some parts of the embodiments of this application.

[0207] Finally, it should also be noted that in this article, relational terms such as first, second, third, and fourth are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "including a..." does not exclude the existence of another identical element in the process, method, article or device including the said element.

[0208] The above are only the preferred embodiments of this application. It should be pointed out that for those of ordinary skill in the art, without departing from the principle of this application, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of this application.

Claims

1. A graph processing method based on semantic segmentation technology, characterized in that: include: Obtaining a to-be-identified spectrum, wherein the to-be-identified spectrum is a phase-resolved partial discharge spectrum; Processing the to-be-recognized graph according to a pre-built semantic segmentation model to obtain a binary classification result of the to-be-recognized graph, wherein the binary classification result is used to indicate a partial discharge pixel and a background pixel of the to-be-recognized graph, wherein the partial discharge pixel is a pixel corresponding to a partial discharge signal in the to-be-recognized graph, and the background pixel is a pixel other than the partial discharge pixel in the to-be-recognized graph; According to the binary classification result, the background pixels and partial discharge pixels of the to-be-identified spectrum are respectively processed according to the background color and partial discharge pixel rendering method of the standard spectrum contained in the standard recognition database to obtain a processed spectrum to be identified. The processed spectrum to be identified and the standard spectrum have the same spectrum format, and the processed spectrum to be identified is used to identify the type of partial discharge.

2. The method according to claim 1, characterized in that The method for constructing the semantic segmentation model includes: Obtaining a spectrum background library and a feature template library, wherein the spectrum background library includes a plurality of phase-resolved partial discharge spectrum backgrounds, and the feature template library includes a plurality of feature templates, each of which includes a partial discharge pixel obtained from a phase-resolved partial discharge spectrum; Superimposing the characteristic templates contained in the characteristic template library onto the phase-resolved partial discharge spectrum background of the spectrum background library to obtain a superimposed spectrum; Generating annotation information of the overlay map, wherein the annotation information is used to indicate the local discharge pixels and the background pixels in the corresponding overlay map; An initial model is trained according to the plurality of superimposed maps and the annotation information of the superimposed maps to obtain the semantic segmentation model.

3. The method according to claim 2, characterized in that The method for obtaining the feature template library includes: Extracting partial discharge pixels from the phase-resolved partial discharge map to obtain a first feature template composed of the extracted local pixels; Performing a size transformation process on the first feature template according to a preset proportional stretching factor to obtain a second feature template; Adding or deleting a plurality of partial discharge pixels in the first feature template to obtain a third feature template; The first feature template, the second feature template and the third feature template constitute the feature template library.

4. The method according to claim 2, characterized in that: The method for obtaining the atlas background library comprises: Obtaining a first phase-resolved partial discharge spectrum background; Performing a size transformation on the first phase-resolved partial discharge spectrum background based on a plurality of randomly determined target sizes to obtain a second phase-resolved partial discharge spectrum background; The first phase-resolved partial discharge spectrum background and the second phase-resolved partial discharge spectrum background constitute the spectrum background library.

5. The method according to claim 1, characterized in that Also includes: determining a plurality of gray value intervals; For each gray value interval, counting the number of background pixels corresponding to the gray value interval according to the binary classification result, the number of background pixels being the number of background pixels whose gray values ​​are within the gray value interval; Determining a target grayscale value interval from the plurality of grayscale value intervals according to the number of background pixels; Changing the PD pixels indicated by the binary classification result and having grayscale values ​​within the target grayscale value interval to background pixels to obtain an updated binary classification result; According to the binary classification result, the background pixels and partial discharge pixels of the to-be-identified atlas are processed respectively according to the background color and partial discharge pixel rendering mode of the standard atlas contained in the standard identification database, including: According to the updated binary classification result, the background pixels and partial discharge pixels of the to-be-identified atlas are processed respectively according to the background color and partial discharge pixel rendering method of the standard atlas contained in the standard recognition database.

6. A graph processing device based on semantic segmentation technology, characterized in that: include: An obtaining unit, used for obtaining a to-be-identified spectrum, wherein the to-be-identified spectrum is a phase-resolved partial discharge spectrum; A segmentation unit, used for processing the to-be-identified graph according to a pre-built semantic segmentation model to obtain a binary classification result of the to-be-identified graph, wherein the binary classification result is used to indicate a partial discharge pixel and a background pixel of the to-be-identified graph, wherein the partial discharge pixel is a pixel corresponding to a partial discharge signal in the to-be-identified graph, and the background pixel is a pixel other than the partial discharge pixel in the to-be-identified graph; A processing unit is used to process the background pixels and partial discharge pixels of the to-be-identified spectrum according to the background color and partial discharge pixel rendering method of the standard spectrum contained in the standard recognition database according to the binary classification result, so as to obtain a processed spectrum to be identified, wherein the processed spectrum to be identified and the standard spectrum have the same spectrum format, and the processed spectrum to be identified is used to identify the type of partial discharge.

7. The device according to claim 6, characterized in that Also included are building blocks for: Obtaining a spectrum background library and a feature template library, wherein the spectrum background library includes a plurality of phase-resolved partial discharge spectrum backgrounds, and the feature template library includes a plurality of feature templates, each of which includes a partial discharge pixel obtained from a phase-resolved partial discharge spectrum; Superimposing the characteristic templates contained in the characteristic template library onto the phase-resolved partial discharge spectrum background of the spectrum background library to obtain a superimposed spectrum; Generating annotation information of the overlay map, wherein the annotation information is used to indicate the local discharge pixels and the background pixels in the corresponding overlay map; An initial model is trained according to the plurality of superimposed maps and the annotation information of the superimposed maps to obtain the semantic segmentation model.

8. The device according to claim 7, characterized in that When the construction unit obtains the feature template library, it is used to: Extracting partial discharge pixels from the phase-resolved partial discharge map to obtain a first feature template composed of the extracted local pixels; Performing a size transformation process on the first feature template according to a preset proportional stretching factor to obtain a second feature template; Adding or deleting a plurality of partial discharge pixels in the first feature template to obtain a third feature template; The first feature template, the second feature template and the third feature template constitute the feature template library.

9. The device according to claim 7, characterized in that When the construction unit obtains the atlas background library, it is used to: Obtaining a first phase-resolved partial discharge spectrum background; Performing a size transformation on the first phase-resolved partial discharge spectrum background based on a plurality of randomly determined target sizes to obtain a second phase-resolved partial discharge spectrum background; The first phase-resolved partial discharge spectrum background and the second phase-resolved partial discharge spectrum background constitute the spectrum background library.

10. The device according to claim 6, characterized in that The segmentation unit is also used for: determining a plurality of gray value intervals; For each gray value interval, counting the number of background pixels corresponding to the gray value interval according to the binary classification result, the number of background pixels being the number of background pixels whose gray values ​​are within the gray value interval; Determining a target grayscale value interval from the plurality of grayscale value intervals according to the number of background pixels; Changing the PD pixels indicated by the binary classification result and having grayscale values ​​within the target grayscale value interval to background pixels to obtain an updated binary classification result; When the processing unit processes the background pixels and partial discharge pixels of the to-be-identified atlas respectively according to the background color and partial discharge pixel rendering mode of the standard atlas contained in the standard identification database based on the binary classification result, it is used to: According to the updated binary classification result, the background pixels and partial discharge pixels of the to-be-identified atlas are processed respectively according to the background color and partial discharge pixel rendering method of the standard atlas contained in the standard recognition database.

Images