Transformer insulation creeping surface discharge defect detection method and device

Abstract

The invention provides a transformer insulation creeping surface discharge defect detection method and device which can be applied to the technical field of high-voltage test, and the method comprises the steps: obtaining an initial image of an oil paper insulation structure in response to a defect detection request for the to-be-detected oil paper insulation structure; the initial image is input into an encoder module, encoding features output by a plurality of encoders included in the encoder module are obtained, and the encoder module comprises k encoders; coding features output by the multiple encoders are input into a hop connection module based on the attention mechanism, multiple hop connection features are obtained, and the hop connection module based on the attention mechanism comprises k hop connection units; the multiple hop connection features are input into a decoder module, a target image is obtained, and the decoder module comprises k decoders; and based on the edge features of the target image, obtaining a creeping discharge defect detection result of the oil paper insulation structure.

Description

Technical Field

[0001] The present invention relates to the technical field of high voltage testing, and more specifically, to a method and device for detecting transformer insulation surface discharge defects. Background Art

[0002] Large oil-immersed transformers are the core equipment of the power system. Once a fault occurs, it will have a significant impact on the power supply stability of the power system. Surface discharge is one of the most serious types of discharge faults in the oil-paper insulation structure of the transformer. Accurately monitoring the occurrence and degree of surface discharge is an important prerequisite for carrying out transformer maintenance and ensuring the stable operation of the power system.

[0003] Oil-paper insulation surface discharge refers to the local electric field distortion caused by moisture and impurity erosion, insulation material aging, and component deformation inside the transformer, which leads to discharge along the oil-paper interface, causing carbonization damage to the oil-paper composite insulation, resulting in a decrease in transformer insulation performance, leakage, and even explosion. Discharge carbon marks are important visual features of oil-paper insulation surface discharge. The area, morphology and other features of carbon marks are of great reference value for judging and analyzing the causes, degree and development trends of discharge. The use of robotics and machine vision to achieve autonomous detection of discharge carbon marks has great development potential. Accurate segmentation of discharge carbon marks from the collected internal transformer image data is an important prerequisite and key technical requirement for robots to complete autonomous detection of discharge carbon marks, judgment of discharge degree and evaluation of transformer insulation status.

[0004] However, due to the spatial randomness of the arc erosion site during the surface discharge of oil-paper insulation, and the event randomness of the discharge severity, the edges of the dendritic carbon marks are extremely complex and the sizes of the carbon marks vary greatly, making it difficult to accurately segment the discharge carbon marks. Summary of the invention

[0005] In view of this, the present invention provides a transformer insulation surface discharge defect detection method, device, equipment, storage medium and program product.

[0006] One aspect of the present invention provides a transformer insulation creeping discharge defect detection method, comprising: in response to a defect detection request for an oil-paper insulation structure to be detected, obtaining an initial image of the oil-paper insulation structure; inputting the initial image into an encoder module to obtain encoding features output by each of a plurality of encoders included in the encoder module, wherein the encoder module includes k encoders; inputting the encoding features output by each of the plurality of encoders into a skip connection module based on an attention mechanism to obtain a plurality of skip connection features, wherein the skip connection module based on the attention mechanism includes k skip connection units; inputting the plurality of skip connection features into a decoder module to obtain a target image, wherein the decoder module includes k decoders; and obtaining a creeping discharge defect detection result of the oil-paper insulation structure based on edge features of the target image.

[0007] Another aspect of the present invention provides a transformer insulation creeping discharge defect detection device, comprising: an image acquisition module, used to respond to a defect detection request for the oil-paper insulation structure to be detected, and obtain an initial image of the oil-paper insulation structure; a first input module, used to input the initial image into an encoder module, and obtain encoding features output by each of the multiple encoders included in the encoder module, and the encoder module includes k encoders; a second input module, used to input the encoding features output by each of the multiple encoders into a skip connection module based on an attention mechanism, and obtain multiple skip connection features, and the skip connection module based on the attention mechanism includes k skip connection units; a third input module, used to input the multiple skip connection features into a decoder module, and obtain a target image, and the decoder module includes k decoders; and a result determination module, used to obtain a creeping discharge defect detection result of the oil-paper insulation structure based on the edge features of the target image.

[0008] Another aspect of the present invention provides an electronic device, comprising: one or more processors; and a memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the above-mentioned method.

[0009] Another aspect of the present invention provides a computer-readable storage medium storing computer-executable instructions, which are used to implement the above method when executed.

[0010] Another aspect of the present invention provides a computer program product, the computer program product comprising computer executable instructions, and the instructions are used to implement the method as described above when executed.

[0011] According to an embodiment of the present invention, by adopting a skip connection module based on the attention mechanism between the encoder module and the decoder module, the semantic gap between the encoder and the decoder can be alleviated, and accurate segmentation of discharge carbon marks can be achieved, so that the obtained target image can accurately characterize the characteristics of discharge carbon marks, and then a more accurate surface discharge defect detection result can be obtained based on the segmented target image. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] The above and other objects, features and advantages of the present invention will become more apparent through the following description of the embodiments of the present invention with reference to the accompanying drawings, in which:

[0013] Figure 1 An exemplary system architecture to which the transformer insulation creeping discharge defect detection method according to an embodiment of the present invention can be applied is shown.

[0014] Figure 2 A flow chart of a transformer insulation creeping discharge defect detection method according to an embodiment of the present invention is shown.

[0015] Figure 3 A structural schematic diagram of an MHAC module according to a specific embodiment of the present invention is shown.

[0016] Figure 4 A network structure diagram for generating a target image according to a specific embodiment of the present invention is shown.

[0017] Figure 5 A schematic structural diagram of a jump connection unit according to a specific embodiment of the present invention is shown.

[0018] Figure 6 A block diagram of a transformer insulation creeping discharge defect detection device according to an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0019] Below, embodiments of the present invention will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the present invention. In the following detailed description, for ease of explanation, many specific details are set forth to provide a comprehensive understanding of embodiments of the present invention. However, it is apparent that one or more embodiments may also be implemented without these specific details. In addition, in the following description, descriptions of known structures and technologies are omitted to avoid unnecessary confusion of concepts of the present invention.

[0020] The terms used herein are only for describing specific embodiments and are not intended to limit the present invention. The terms "comprise", "include", etc. used herein indicate the existence of features, steps, operations and / or components, but do not exclude the existence or addition of one or more other features, steps, operations or components.

[0021] All terms (including technical and scientific terms) used herein have the meanings commonly understood by those skilled in the art unless otherwise defined. It should be noted that the terms used herein should be interpreted as having a meaning consistent with the context of this specification and should not be interpreted in an idealized or overly rigid manner.

[0022] When using expressions such as "at least one of A, B, and C, etc.", they should generally be interpreted according to the meaning of the expression commonly understood by those skilled in the art (for example, "a system having at least one of A, B, and C" should include but is not limited to a system having A alone, B alone, C alone, A and B, A and C, B and C, and / or A, B, C, etc.).

[0023] In the embodiments of the present invention, the collection, updating, analysis, processing, use, transmission, provision, disclosure, storage and other aspects of the data involved (for example, including but not limited to user personal information) are in compliance with the provisions of relevant laws and regulations, are used for legitimate purposes, and do not violate public order and good morals. In particular, necessary measures are taken for user personal information to prevent illegal access to user personal information data and maintain the security of user personal information, network security and national security.

[0024] In the embodiment of the present invention, before obtaining or collecting the user's personal information, the user's authorization or consent is obtained.

[0025] Since the arc erosion site is spatially random during the surface discharge of oil-paper insulation, and the severity of the discharge is event-random, the edges of the dendritic carbon marks are abnormally complex and the sizes of the carbon marks vary greatly. This places high demands on the global information perception and local detail feature extraction capabilities of the semantic segmentation model, making it difficult to use the semantic segmentation model to achieve accurate segmentation of the discharge carbon marks.

[0026] An embodiment of the present invention provides a transformer insulation surface discharge defect detection method, comprising: in response to a defect detection request for an oil-paper insulation structure to be detected, obtaining an initial image of the oil-paper insulation structure; inputting the initial image into an encoder module to obtain encoding features output by each of a plurality of encoders included in the encoder module, wherein the encoder module includes k encoders; inputting the encoding features output by each of the plurality of encoders into a skip connection module based on an attention mechanism to obtain a plurality of skip connection features, wherein the skip connection module based on the attention mechanism includes k skip connection units; inputting the plurality of skip connection features into a decoder module to obtain a target image, wherein the decoder module includes k decoders; and obtaining a surface discharge defect detection result of the oil-paper insulation structure based on edge features of the target image.

[0027] The embodiment of the present invention can alleviate the semantic gap between the encoder module and the decoder module by adopting a skip connection module based on the attention mechanism between the encoder module and the decoder module, thereby realizing accurate segmentation of discharge carbon marks, so that the obtained target image can accurately characterize the characteristics of discharge carbon marks, and then obtain more accurate surface discharge defect detection results based on the segmented target image.

[0028] Figure 1 The exemplary system architecture to which the transformer insulation creeping discharge defect detection method according to the embodiment of the present invention can be applied is shown. It should be noted that: Figure 1 The examples shown are merely examples of system architectures to which the embodiments of the present invention may be applied, to help those skilled in the art understand the technical content of the present invention, but do not mean that the embodiments of the present invention may not be used in other devices, systems, environments or scenarios.

[0029] like Figure 1 As shown, the system architecture 100 according to this embodiment may include a first terminal device 101, a second terminal device 102, a third terminal device 103, a network 104, and a server 105. The network 104 is used to provide a medium for a communication link between the first terminal device 101, the second terminal device 102, the third terminal device 103, and the server 105. The network 104 may include various connection types, such as wired and / or wireless communication links, etc.

[0030] The user may use the first terminal device 101, the second terminal device 102, and the third terminal device 103 to interact with the server 105 through the network 104 to receive or send messages, etc. Various communication client applications may be installed on the first terminal device 101, the second terminal device 102, and the third terminal device 103, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, and / or social platform software, etc. (only for example).

[0031] The first terminal device 101, the second terminal device 102, and the third terminal device 103 may be various electronic devices having display screens and supporting web browsing, including but not limited to smart phones, tablet computers, laptop computers, desktop computers, and the like.

[0032] The server 105 may be a server that provides various services, such as a background management server (only as an example) that provides support for websites browsed by users using the first terminal device 101, the second terminal device 102, and the third terminal device 103. The background management server may analyze and process the received data such as user requests, and feed back the processing results (such as web pages, information, or data obtained or generated according to user requests) to the terminal device.

[0033] It should be noted that the transformer insulation surface discharge defect detection method provided in the embodiment of the present invention can generally be executed by the server 105. Accordingly, the transformer insulation surface discharge defect detection device provided in the embodiment of the present invention can generally be set in the server 105. The transformer insulation surface discharge defect detection method provided in the embodiment of the present invention can also be executed by a server or server cluster that is different from the server 105 and can communicate with the first terminal device 101, the second terminal device 102, the third terminal device 103 and / or the server 105. Correspondingly, the transformer insulation surface discharge defect detection device provided in the embodiment of the present invention can also be set in a server or server cluster that is different from the server 105 and can communicate with the first terminal device 101, the second terminal device 102, the third terminal device 103 and / or the server 105. Alternatively, the transformer insulation surface discharge defect detection method provided in the embodiment of the present invention can also be performed by the first terminal device 101, the second terminal device 102 or the third terminal device 103, or by other terminal devices different from the first terminal device 101, the second terminal device 102 or the third terminal device 103. Correspondingly, the transformer insulation surface discharge defect detection device provided in the embodiment of the present invention can also be arranged in the first terminal device 101, the second terminal device 102 or the third terminal device 103, or in other terminal devices different from the first terminal device 101, the second terminal device 102 or the third terminal device 103.

[0034] For example, the initial image may be originally stored in any one of the first terminal device 101, the second terminal device 102 or the third terminal device 103 (for example, the first terminal device 101, but not limited thereto), or stored in an external storage device and may be imported into the first terminal device 101. Then, in response to a defect detection request for an oil-paper insulation structure to be detected, the first terminal device 101 may locally execute the transformer insulation surface discharge defect detection method provided by the embodiment of the present invention, or send the initial image to other terminal devices, servers, or server clusters, and the other terminal devices, servers, or server clusters that receive the initial image may execute the transformer insulation surface discharge defect detection method provided by the embodiment of the present invention.

[0035] It should be understood that Figure 1 The number of terminal devices, networks and servers in the embodiment is only . According to the implementation requirements, there can be any number of terminal devices, networks and servers.

[0036] Figure 2 A flow chart of a transformer insulation creeping discharge defect detection method according to an embodiment of the present invention is shown.

[0037] like Figure 2As shown, the method includes operations S210 to S250.

[0038] In operation S210 , in response to a defect detection request for an oil-paper insulation structure to be detected, an initial image of the oil-paper insulation structure is acquired.

[0039] According to an embodiment of the present invention, the initial image may be obtained by photographing the oil-paper insulation structure to be inspected using a camera. After the initial image is photographed, the initial image may be stored in a preset storage area.

[0040] According to an embodiment of the present invention, an initial image of the oil-paper insulation structure can be obtained from a preset storage area. Specifically, the defect detection request may include storage address information corresponding to the preset storage area, and the initial image of the oil-paper insulation structure can be obtained from the preset storage area according to the storage address information.

[0041] According to an embodiment of the present invention, after the transformer discharges along the interface of the oil-paper insulation structure, a discharge carbon mark will be generated on the oil-paper insulation structure, and the area, shape and other characteristics of the discharge carbon mark have important reference value for judging and analyzing the cause, degree and development trend of the discharge. Therefore, surface discharge defect detection can be performed based on the initial image of the oil-paper insulation structure.

[0042] In operation S220, the initial image is input into the encoder module to obtain encoding features output by each of the plurality of encoders included in the encoder module.

[0043] According to an embodiment of the present invention, the encoder module includes k encoders, where k≥1 and k is an integer. The input of each encoder may include the encoding features output by the previous encoder, and the input of the first encoder may be an initial image.

[0044] In operation S230, the encoding features output by each of the plurality of encoders are input into a skip connection module based on an attention mechanism to obtain a plurality of skip connection features.

[0045] According to an embodiment of the present invention, the skip connection module based on the attention mechanism includes k skip connection units. The skip connection units correspond to the encoders one by one. Specifically, the encoding features output by each encoder can be input into the skip connection unit corresponding to the encoder to obtain the skip connection features.

[0046] According to an embodiment of the present invention, a skip connection module based on an attention mechanism can be used to enhance the activation strength of the coding feature and reduce the semantic gap between the encoder and the decoder. Specifically, in each skip connection unit of the skip connection module based on the attention mechanism, a weight for the corresponding coding feature can be first determined, and then the coding feature is weighted using the weight to obtain a skip connection feature.

[0047] In operation S240, the plurality of skip-connection features are input into a decoder module to obtain a target image.

[0048] According to an embodiment of the present invention, the decoder module includes k decoders. The decoders correspond to the skip connection units one by one, and the input of each decoder may include the skip connection feature output by the skip connection unit corresponding to the decoder.

[0049] According to an embodiment of the present invention, the skip connection features enhanced by the skip connection unit are used for decoding, so that the obtained target image can accurately characterize the discharge carbon mark features on the oil-paper insulation structure.

[0050] In operation S250, a creeping discharge defect detection result of the oil-paper insulation structure is obtained based on the edge features of the target image.

[0051] According to an embodiment of the present invention, the edge feature can characterize the edge feature of the discharge carbon mark in the target image. The creeping discharge defect detection result can include the area where the creeping discharge defect exists in the oil-paper insulation structure and the severity of the discharge.

[0052] According to the embodiment of the present invention, the discharge carbon mark feature can be determined according to the edge feature of the target image, and then the surface discharge defect detection result of the oil-paper insulation structure can be determined according to the discharge carbon mark feature.

[0053] According to an embodiment of the present invention, by adopting a skip connection module based on the attention mechanism between the encoder module and the decoder module, the semantic gap between the encoder and the decoder can be alleviated, and accurate segmentation of discharge carbon marks can be achieved, so that the obtained target image can accurately characterize the characteristics of discharge carbon marks, and then a more accurate surface discharge defect detection result can be obtained based on the segmented target image.

[0054] According to an embodiment of the present invention, based on the edge features of a target image, a surface discharge defect detection result of an oil-paper insulation structure is obtained, including: performing image classification on the target image to obtain an edge type recognition result of the target image; when the edge type recognition result is represented as a dendritic edge, determining a discharge defect region from the target image; determining the area of ​​a defect region of the oil-paper insulation structure based on a mapping relationship between the discharge defect region and the target image and the camera coordinate system of the initial image and the world coordinate system; and obtaining a surface discharge defect detection result of the oil-paper insulation structure based on the defect region area.

[0055] According to an embodiment of the present invention, the target image can be classified according to the edge type corresponding to the edge feature of the target image to obtain the edge type recognition result of the target image. The edge type may include, for example, a smooth edge, a jagged edge, a branch-shaped edge, etc.

[0056] According to an embodiment of the present invention, the edge type recognition result can be used to characterize the shape of the edge feature. Specifically, when the edge recognition result is represented as a dendritic edge, it is characterized that the oil-paper insulation structure has a surface discharge defect, and when the edge recognition result is represented as other edges, it is characterized that the oil-paper insulation structure does not have a surface discharge defect.

[0057] According to an embodiment of the present invention, the discharge defect region may represent a region where the discharge defect exists in the target image of the oil-paper insulation structure. In the target image, the discharge defect region may be represented as a region surrounded by dendritic edges.

[0058] According to the embodiment of the present invention, since the target image is obtained based on the initial image which is captured by a camera, the area of ​​the discharge defect region in the target image is not the actual defect region area of ​​the oil-paper insulation structure.

[0059] According to an embodiment of the present invention, when determining the area of ​​a defective region of an oil-paper insulation structure, the coordinate position of the discharge defective region in the target image can be determined based on the target image and the camera coordinate system, and then the coordinate position of the discharge region in reality can be determined based on the mapping relationship between the camera coordinate system and the world coordinate system, thereby determining the area of ​​the defective region.

[0060] According to an embodiment of the present invention, the area of ​​the defect region can characterize the severity of the surface discharge defect of the oil-paper insulation structure. Specifically, the larger the area of ​​the defect region, the more serious the surface discharge defect, and the smaller the area of ​​the defect region, the milder the surface discharge defect.

[0061] According to an embodiment of the present invention, by determining a surface discharge defect detection result based on a target image that can accurately characterize a characteristic edge of a discharge carbon mark, the accuracy of the discharge defect detection result is improved.

[0062] According to an embodiment of the present invention, an initial image is input into an encoder module to obtain encoding features output by each of the multiple encoders included in the encoder module, including: inputting the initial image into m first encoders connected in series to obtain encoding features output by each of the m first encoders; and inputting the encoding features output by the mth first encoder into n second encoders connected in series to obtain encoding features output by each of the n second encoders.

[0063] According to an embodiment of the present invention, the k encoders include m first encoders connected in series and n second encoders connected in series, wherein m≥1 and m is an integer, n≥1 and n is an integer, and m+n=k.

[0064] According to an embodiment of the present invention, the input of the n second encoders is the encoding features output by the mth first encoder, that is, the n second encoders are connected in series to the m first encoders. Specifically, the encoding features output by the mth first encoder can be input to the first second encoder.

[0065] According to an embodiment of the present invention, the accuracy of the encoding process is improved by using two different encoders for encoding.

[0066] According to an embodiment of the present invention, an initial image is input into m first encoders connected in series to obtain coding features output by each of the m first encoders, including: performing convolution processing on the coding features output by the i-1th first encoder to obtain the first features; and performing maximum pooling processing on the first features to obtain the coding features output by the i-th first encoder.

[0067] According to an embodiment of the present invention, 0<i≤m, when i=1, the encoding feature output by the i-1th first encoder is represented as the initial image, that is, the input of the first first encoder is the initial image.

[0068] According to an embodiment of the present invention, the i-th first encoder may sequentially perform convolution processing and maximum pooling processing on the encoding features output by the i-1-th encoder.

[0069] According to an embodiment of the present invention, when the coded features output by the i-1th first encoder are convolved using the i-th first encoder, two-dimensional convolution processing can be performed on the coded features output by the i-1th first encoder to obtain the first features.

[0070] According to an embodiment of the present invention, after performing maximum pooling on the first feature, the encoding feature output by the i-th first encoder can be obtained.

[0071] According to an embodiment of the present invention, the encoding feature is obtained by performing maximum pooling on the first feature, thereby reducing the size of the feature and further reducing the complexity of the model.

[0072] According to an embodiment of the present invention, the encoding features output by the mth first encoder are input into n second encoders connected in series to obtain the encoding features output by each of the n second encoders, including: normalizing the encoding features output by the j-1th second encoder to obtain a second feature; splitting the second feature according to the channel dimension to obtain a plurality of first sub-features; performing an element-by-element product attention operation on the first sub-feature based on the combination dimension corresponding to the first sub-feature to obtain a second sub-feature; and sequentially concatenating, normalizing, performing depth-wise separable convolution and maximum pooling on the plurality of second sub-features to obtain the encoding features output by the jth second encoder.

[0073] According to an embodiment of the present invention, 0<j≤n, when j=1, the encoding feature output by the j-1th second encoder is expressed as the encoding feature output by the mth first encoder, that is, the input of the first second encoder is the encoding feature output by the mth first encoder.

[0074] According to an embodiment of the present invention, the second encoder may include a multi-head attention convolution (MHAC) module with linear complexity, wherein the MHAC module may perform normalization, splitting, element-wise product attention operation, splicing, normalization, and depth-separable convolution on the encoded features in sequence.

[0075] According to an embodiment of the present invention, the size of the coding feature can be C×H×W. After normalizing the coding feature, the size of the obtained second feature is also C×H×W, wherein C can be used to characterize the number of channels of the feature, H can be used to characterize the height of the feature, and W can be used to characterize the width of the feature.

[0076] According to an embodiment of the present invention, when the second feature is split according to the channel dimension, the second feature can be split into three first sub-features, and the size of each first sub-feature is C / 3×H×W.

[0077] According to an embodiment of the present invention, the three first sub-features may be respectively X 1 , X 2 and X 3 , the element-by-element product attention operation of the three first sub-features includes: X on the HW, CH and CW axes respectively 1 With P xy , X 2 With P zx , X 3 With P zy Perform Hadamard Product operation, where P xy (Size is 1×C / 3×H×W), P zx (Size is 1×1×C / 3×H), P zy (of size 1×1×C / 3×W) are 3 learnable tensors.

[0078] According to an embodiment of the present invention, when multiple second sub-features are spliced, the splicing can be performed according to the channel dimension. Specifically, a merge (concat) instruction can be used to splice the second sub-features corresponding to the three first sub-features according to the channel dimension. The size of the features obtained after splicing is C×H×W.

[0079] According to an embodiment of the present invention, when the features after splicing multiple second sub-features are normalized and depthwise separable convolution is performed in sequence, the normalization (LayerNorm, LN) instruction and the depthwise separable convolution (DepthwiseSeparable Convolution, DW) instruction can be used respectively, and the resulting feature size is C×H×W.

[0080] According to an embodiment of the present invention, after performing depth-wise separable convolution on the normalized features, it is also necessary to perform maximum pooling processing on the features obtained by the depth-wise separable convolution, and finally obtain the encoded features output by the j-th second encoder.

[0081] According to an embodiment of the present invention, by adopting the MHAC module, multi-scale features in the coding features can be extracted, the loss of shallow visual features can be slowed down, the effective fusion of features of different scales can be strengthened, and the model's perception ability of multi-scale carbon mark features in the coding features can be improved while reducing the model complexity.

[0082] Figure 3 A structural schematic diagram of an MHAC module according to a specific embodiment of the present invention is shown.

[0083] like Figure 3 As shown in Figure 2, the MHAC module normalizes the input features (Input) (LayerNorm, LN) and splits them into X according to the channel dimension. 1 , X 2 , X 3 Three sub-features, for X 1 , X 2 , X 3 After performing element-by-element product attention (HA) operations, they are concatenated according to the channel dimension (C). The concatenated features are normalized (LayerNorm, LN) and processed with depthwise separable convolution (DW) in turn to obtain the output features (Out) of the MHAC module.

[0084] According to an embodiment of the present invention, a plurality of skip connection features are input into a decoder module to obtain a target image, including: inputting the skip connection features outputted by each of the kth skip connection unit to the k+1-nth skip connection unit into n second decoders connected in series to obtain the decoding features outputted by the nth second decoder; and inputting the skip connection features outputted by each of the knth skip connection unit to the 1st skip connection unit and the decoding features outputted by the nth second decoder into m first decoders connected in series to obtain the target image.

[0085] According to an embodiment of the present invention, the k decoders include n serially connected second decoders and m serially connected first decoders, wherein the n second decoders correspond one-to-one to the n second encoders, and the m first decoders correspond one-to-one to the m first encoders.

[0086] According to the embodiment of the present invention, since in the encoding process, the n second encoders are connected in series after the m first encoders, correspondingly, in the decoding process, the m first decoders are connected in series after the n second decoders.

[0087] According to an embodiment of the present invention, the input of the n second decoders includes the jump connection features output by the kth jump connection unit to the k+1-nth jump connection unit, and the input of the m first decoders includes the jump connection features output by the knth jump connection unit to the 1st jump connection unit, and the decoding features output by the nth second decoder.

[0088] According to an embodiment of the present invention, the m first decoders obtain a target image after decoding the skip connection features outputted by the kn-th skip connection unit to the 1st skip connection unit and the decoding features outputted by the n-th second decoder.

[0089] According to an embodiment of the present invention, by decoding the jump connection features output by the jump connection unit and the decoding features output by the decoder during the decoding process, feature transfer between the encoder and the decoder is achieved, the semantic gap between the encoder and the decoder is alleviated, and the obtained target image can more accurately characterize the carbon mark features.

[0090] According to an embodiment of the present invention, the jump connection features output by each of the kth jump connection unit to the k+1-nth jump connection unit are input into n second decoders connected in series to obtain the decoding features output by the nth second decoder, including: adding the decoding features output by the j-1th second decoder and the jump connection features output by the k+1-jth jump connection unit to obtain a third feature; normalizing the third feature to obtain a fourth feature; splitting the fourth feature according to the channel dimension to obtain multiple third sub-features; performing an element-by-element product attention operation on the third sub-feature based on the combination dimension corresponding to the third sub-feature to obtain a fourth sub-feature; and sequentially concatenating, normalizing, depth-separable convolution and activation processing on the multiple fourth sub-features to obtain the decoding features output by the jth second decoder.

[0091] According to an embodiment of the present invention, 0<j≤n, when j=1, the decoding feature output by the j-1th second decoder is represented as 0, that is, the input of the first second encoder only includes the skip connection feature output by the kth skip connection unit.

[0092] According to an embodiment of the present invention, when j≠1, the input of the j-th second decoder includes the decoding features output by the j-1-th second decoder and the skip connection features output by the k+1-j-th skip connection unit.

[0093] According to an embodiment of the present invention, the second decoder may include an MHAC module, wherein the MHAC module is used to perform normalization, splitting, splicing, normalization and depth-separable convolution on the third feature output by the addition module in sequence.

[0094] According to an embodiment of the present invention, the size of the third feature may be C×H×W, and after the decoded feature is normalized, the size of the fourth feature obtained is also C×H×W.

[0095] According to an embodiment of the present invention, when the fourth feature is split according to the channel dimension, the fourth feature can be split into three third sub-features, and the size of each third sub-feature is C / 3×H×W.

[0096] According to an embodiment of the present invention, the three first sub-features may be respectively X 4 , X 5 and X 6 , the element-by-element product attention operation of the three first sub-features includes: X on the HW, CH and CW axes respectively 4 With P xy , X 5 With P zx , X 6 With P zy Perform Hadamard Product operation, where P xy (Size is 1×C / 3×H×W), P zx (Size is 1×1×C / 3×H), P zy (of size 1×1×C / 3×W) are 3 learnable tensors.

[0097] According to an embodiment of the present invention, when multiple fourth sub-features are spliced, the splicing can be performed according to the channel dimension. Specifically, the fourth sub-features corresponding to the three third sub-features can be spliced ​​according to the channel dimension, and the resulting feature size is C×H×W.

[0098] According to an embodiment of the present invention, after the features obtained by concatenating multiple fourth sub-features are normalized and subjected to depthwise separable convolution in sequence, the resulting feature size is C×H×W.

[0099] According to an embodiment of the present invention, after performing depthwise separable convolution on the normalized features, it is also necessary to perform activation processing on the features obtained by the depthwise separable convolution. Specifically, the activation function may be a Hardswish function.

[0100] According to an embodiment of the present invention, by adopting the MHAC module, multi-scale features in the decoding features can be extracted, the loss of shallow visual features can be slowed down, the effective fusion of features of different scales can be strengthened, and the model's perception ability of multi-scale carbon mark features in the encoding features can be improved while reducing the model complexity.

[0101] According to an embodiment of the present invention, by using the Hardswish function as an activation function, the computational cost of the Hardswish function is greatly reduced compared to the conventional swish function and the GeLU function. In addition, since the Hardswish function has better robustness and lightweight advantages, the activation strength of the carbon mark feature is enhanced, and the gradient propagation effect in the model training and reasoning process is improved. Specifically, the Hardswish function can be shown as the following formula (1):

[0102] (1)

[0103] Among them, ReLU6 can represent the ReLU activation function, and Hardwish(x) can represent the Hardswish function.

[0104] According to an embodiment of the present invention, the skip connection features outputted by the knth skip connection unit to the 1st skip connection unit and the decoding features outputted by the nth second decoder are inputted into m first decoders connected in series to obtain a target image, including: adding the decoding features outputted by the i-1th first decoder and the skip connection features outputted by the k+1-inth skip connection unit to obtain a fifth feature; performing convolution processing on the fifth feature to obtain a sixth feature; and performing activation processing on the sixth feature to obtain a decoding feature outputted by the i-th first decoder.

[0105] According to an embodiment of the present invention, 0<i≤m, when i=1, the decoding features output by the i-1 first decoders are represented as the decoding features output by the nth second decoder, that is, the input of the first first decoder includes the decoding features output by the nth second decoder and the jump connection features output by the knth jump connection unit.

[0106] According to an embodiment of the present invention, when i≠1, the input of the i-th decoder may include the decoding features output by the i-1-th first decoder and the skip connection features output by the k+1-in-th skip connection unit.

[0107] According to an embodiment of the present invention, the first decoder may perform convolution processing and activation processing on the fifth feature in sequence.

[0108] According to an embodiment of the present invention, when the fifth feature is convolved using the first decoder, a two-dimensional convolution may be performed on the fifth feature to obtain a sixth feature. Specifically, the convolution kernel size of the two-dimensional convolution may be 3.

[0109] According to an embodiment of the present invention, after the sixth feature is activated, a decoding feature output by the i-th first decoder can be obtained. Specifically, the activation function can be a Hardswish function.

[0110] According to an embodiment of the present invention, the decoded feature output by the mth first decoder is represented as the target image.

[0111] According to an embodiment of the present invention, the semantic gap between the encoder and the decoder is alleviated by decoding based on skip connection features.

[0112] Figure 4 A network structure diagram for generating a target image according to a specific embodiment of the present invention is shown.

[0113] As shown in Figure 4, the network for generating the target image includes an encoding module, an attention-based skip connection module, and a decoding module. The encoding module includes 3 serially connected first encoders and 3 serially connected second encoders, the decoding module includes 3 serially connected second decoders and 3 serially connected first decoders, and the attention-based skip connection module includes 6 skip connection units.

[0114] According to an embodiment of the present invention, the i-th first encoder sequentially performs convolution (Conv) processing and maximum pooling (MaxPooling, MP) processing on the encoding features output by the i-1-th first encoder to obtain the encoding features output by the i-th first encoder, and the first first encoder sequentially performs convolution processing and maximum pooling processing on the initial image, wherein the convolution processing can be a two-dimensional convolution (Conv2d) processing, and the number of convolution kernels can be 3.

[0115] According to an embodiment of the present invention, the j-th second encoder sequentially performs multi-head attention convolution (MCH) processing and maximum pooling (MP) processing with linear complexity on the encoding features output by the j-1-th second encoder to obtain the encoding features output by the j-th second encoder, and the first second encoder sequentially performs multi-head attention convolution processing and maximum pooling processing with linear complexity on the encoding features output by the third first encoder.

[0116] According to an embodiment of the present invention, the encoding features output by each encoder are respectively input into a skip connection (Coordinate Attention, CA) unit corresponding to the encoder to obtain skip connection features output by each of the multiple skip connection units.

[0117] According to an embodiment of the present invention, the skip connection features output by each skip connection unit are respectively input into a decoder corresponding to the skip connection unit.

[0118] According to an embodiment of the present invention, the j-th second decoder sequentially adds (Addition), performs multi-head attention convolution processing with linear complexity, and performs activation processing on the decoding features output by the j-1-th second decoder and the jump connection features output by the k+1-j-th jump connection unit to obtain the decoding features output by the j-th second decoder. The first second decoder sequentially performs multi-head attention convolution (MCH) processing with linear complexity and activation (HardSwish, HS) processing on the jump connection features output by the sixth jump connection unit.

[0119] According to an embodiment of the present invention, the i-th first decoder sequentially adds, performs two-dimensional convolution processing, and activates the decoding features output by the i-1-th first decoder and the jump connection features output by the k+1-in-th jump connection unit to obtain the decoding features output by the i-th first decoder. The first first decoder sequentially adds, performs two-dimensional convolution processing, and activates the decoding features output by the third second decoder and the jump connection features output by the third jump connection unit. The output of the third first decoder is the target image.

[0120] According to an embodiment of the present invention, the encoded features output by each of the multiple encoders are input into a skip connection module based on the attention mechanism to obtain multiple skip connection features, including: performing average pooling on the encoded features output by the h-th encoder in the height direction and the width direction, respectively, to obtain the seventh feature and the eighth feature; concatenating the seventh feature and the eighth feature, and sequentially performing convolution, batch normalization and non-linear activation processing to obtain the ninth feature; splitting the ninth feature into a first tensor and a second tensor along the channel direction; performing convolution and non-linear activation processing on the first tensor and the second tensor, respectively, to obtain a first attention weight in the height direction and a second attention weight in the width direction; and weighting the encoded features output by the h-th encoder based on the first attention weight and the second attention weight to obtain the skip connection features output by the h-th skip connection unit.

[0121] According to an embodiment of the present invention, 0<h≤k, the seventh feature can be obtained by averaging the encoding features output by the h-th encoder in the height direction, and the eighth feature can be obtained by averaging the encoding features output by the h-th encoder in the width direction. Specifically, the above process can be shown in the following formula (2):

[0122] (2)

[0123] Among them, Z h c(h) can be used to characterize the seventh feature, W can be used to characterize the width of the coding feature, H can be used to characterize the height of the coding feature, and Z w c (w) can characterize the eighth feature.

[0124] According to an embodiment of the present invention, when the seventh feature and the eighth feature are concatenated, the concat instruction can be used. When the concatenated features are sequentially subjected to convolution, batch normalization, and nonlinear activation processing, the two-dimensional convolution (Conv2d) instruction, the batch normalization (BatchNorm) instruction, and the nonlinear activation (Sigmoid) instruction can be sequentially executed to obtain a ninth feature of a size of 1×(W+H)×C / r. Specifically, the above process can be shown in the following formula (3):

[0125] (3)

[0126] Among them, z h The seventh characteristic, z w The eighth characteristic, F 1 It can represent a two-dimensional convolution, can represent nonlinear activation, and f can represent the ninth feature.

[0127] According to an embodiment of the present invention, when convolution and nonlinear activation processing are performed on the first tensor and the second tensor respectively, the convolution processing may be a two-dimensional convolution processing, and the nonlinear activation processing may utilize a sigmoid activation function.

[0128] According to an embodiment of the present invention, the process of obtaining the first attention weight and the second attention weight can be shown in the following formula (4):

[0129] (4)

[0130] Among them, f h We can characterize the first tensor, F h It can be represented that the convolution processing of the first tensor is performed. can characterize the nonlinear activation process, g h The first attention weight can be represented, f w We can represent the second tensor, F w It can be represented that the convolution process is performed on the second tensor, g w The second attention weight can be characterized.

[0131] According to an embodiment of the present invention, the process of weighting the coding features can be shown as the following formula (5):

[0132] (5)

[0133] Among them, x c (i,j) can represent the encoding features, g h c (i) can represent the first attention weight, g w c (j) can represent the second attention weight.

[0134] According to an embodiment of the present invention, feature transfer between the encoder and the decoder is achieved based on the attention mechanism, so as to alleviate the semantic difference between the encoder and the decoder and achieve accurate semantic segmentation of oil-paper insulation structure images with complex edge features and significant size differences.

[0135] Figure 5 A schematic structural diagram of a jump connection unit according to a specific embodiment of the present invention is shown.

[0136] like Figure 5 As shown in the figure, after the input encoding feature (Input) is subjected to residual processing, average pooling (X Avg Pool) in the height direction and average pooling (Y Avg Pool) in the width direction are performed. Pool), obtain the seventh feature of size C×H×1 and the eighth feature of size C×1×W; after concatenating the seventh feature and the eighth feature (Concat), perform convolution (Conv2d), batch normalization (BatchNorm) and non-linear activation (Non-linear) processing in sequence to obtain the ninth feature of size C / r×1×(W+H); split the ninth feature along the channel direction (Split) into a first tensor of size C×H×1 and a second tensor of size C×1×W; perform convolution (Conv2d) and non-linear activation (Sigmoid) processing on the first tensor and the second tensor respectively to obtain the first attention weight in the height direction and the second attention weight in the width direction; and based on the first attention weight and the second attention weight, perform weighting (Re-weight) processing on the encoded feature to obtain the skip connection feature (Output) of size C×H×W output by the skip connection unit.

[0137] Figure 6 A block diagram of a transformer insulation creeping discharge defect detection device according to an embodiment of the present invention is shown.

[0138] like Figure 6 As shown, the transformer insulation creeping discharge defect detection device 600 includes an image acquisition module 610 , a first input module 620 , a second input module 630 , a third input module 640 and a result determination module 650 .

[0139] The image acquisition module 610 is used to acquire an initial image of the oil-paper insulation structure in response to a defect detection request for the oil-paper insulation structure to be detected. In one embodiment, the image acquisition module 610 can be used to perform the operation S210 described above, which will not be described in detail here.

[0140] The first input module 620 is used to input the initial image into the encoder module to obtain the encoding features output by each of the multiple encoders included in the encoder module, and the encoder module includes k encoders. In one embodiment, the first input module 620 can be used to perform the operation S220 described above, which will not be repeated here.

[0141] The second input module 630 is used to input the encoding features output by each of the multiple encoders into a skip connection module based on the attention mechanism to obtain multiple skip connection features, and the skip connection module based on the attention mechanism includes k skip connection units. In one embodiment, the second input module 630 can be used to perform the operation S230 described above, which will not be repeated here.

[0142] The third input module 640 is used to input multiple skip connection features into the decoder module to obtain a target image, and the decoder module includes k decoders. In one embodiment, the third input module 640 can be used to perform the operation S240 described above, which will not be repeated here.

[0143] The result determination module 650 is used to obtain the surface discharge defect detection result of the oil-paper insulation structure based on the edge features of the target image. In one embodiment, the result determination module 650 can be used to perform the operation S260 described above, which will not be repeated here.

[0144] According to an embodiment of the present invention, the result determination module 650 includes an image classification submodule, a region determination submodule, an area determination submodule and a result determination submodule.

[0145] The image classification submodule is used to classify the target image and obtain the edge type recognition result of the target image.

[0146] The region determination submodule is used to determine the discharge defect region from the target image when the edge type recognition result is represented as a dendritic edge.

[0147] The area determination submodule is used to determine the defect area of ​​the oil-paper insulation structure based on the mapping relationship between the camera coordinate system and the world coordinate system of the discharge defect area and the target image and the initial image.

[0148] The result determination submodule is used to obtain the surface discharge defect detection result of the oil-paper insulation structure based on the defect area.

[0149] According to an embodiment of the present invention, the k encoders include m first encoders and n second encoders. The first input module 620 includes a first input submodule and a second input submodule.

[0150] The first input submodule is used to input the initial image into m first encoders connected in series to obtain encoding features output by each of the m first encoders.

[0151] The second input submodule is used to input the coding features output by the mth first encoder into the n second encoders connected in series to obtain the coding features output by each of the n second encoders.

[0152] According to an embodiment of the present invention, the first input submodule includes a first convolution unit and a pooling unit.

[0153] The first convolution unit is used to perform convolution processing on the encoding feature output by the i-1th first encoder to obtain a first feature, where 0<i≤m, and when i=1, the encoding feature output by the i-1th first encoder is represented as an initial image.

[0154] The pooling unit is used to perform maximum pooling processing on the first feature to obtain the encoding feature output by the i-th first encoder.

[0155] According to an embodiment of the present invention, the second input submodule includes a first normalization unit, a first splitting unit, a first multiplication unit, and a first processing unit.

[0156] The first normalization unit is used to normalize the coding features output by the j-1th second encoder to obtain the second features, where 0<j≤n, and when j=1, the coding features output by the j-1th second encoder are expressed as the coding features output by the mth first encoder.

[0157] The first splitting unit is used to split the second feature according to the channel dimension to obtain multiple first sub-features.

[0158] The first product unit is used to perform an element-by-element product attention operation on the first sub-feature based on the combination dimension corresponding to the first sub-feature to obtain the second sub-feature.

[0159] The first processing unit is used to sequentially concatenate, normalize, perform depth-separable convolution and maximum pooling processing on the multiple second sub-features to obtain the encoding features output by the j-th second encoder.

[0160] According to an embodiment of the present invention, the k decoders include n second decoders and m first decoders. The third input module 640 includes a third input submodule and a fourth input submodule.

[0161] The third input submodule is used to input the skip connection features output by each of the k-th skip connection unit to the k+1-n-th skip connection units into the n second decoders connected in series to obtain the decoding features output by the n-th second decoder.

[0162] The fourth input submodule is used to input the skip connection features output by the kn-th skip connection unit to the 1st skip connection unit and the decoding features output by the n-th second decoder into the m first decoders connected in series to obtain the target image.

[0163] According to an embodiment of the present invention, the third input submodule includes a first adding unit, a second normalizing unit, a second splitting unit, a second multiplying unit and a second processing unit.

[0164] The first adding unit is used to add the decoding feature output by the j-1th second decoder and the jump connection feature output by the k+1-jth jump connection unit to obtain a third feature, where 0<j≤n, and when j=1, the decoding feature output by the j-1th second decoder is represented as 0.

[0165] The second normalization unit is used to perform normalization processing on the third feature to obtain a fourth feature.

[0166] The second splitting unit is used to split the fourth feature according to the channel dimension to obtain multiple third sub-features.

[0167] The second product unit is used to perform an element-by-element product attention operation on the third sub-feature based on the combination dimension corresponding to the third sub-feature to obtain a fourth sub-feature.

[0168] The second processing unit is used to sequentially concatenate, normalize, perform depth-separable convolution and activation processing on the multiple fourth sub-features to obtain the decoding features output by the j-th second decoder.

[0169] According to an embodiment of the present invention, the fourth input submodule includes a second adding unit, a second convolution unit and an activation unit.

[0170] The second adding unit is used to add the decoding features output by the i-1th first decoder and the jump connection features output by the k+1-inth jump connection unit to obtain the fifth feature, where 0<i≤m, and when i=1, the decoding features output by the i-1th first decoder are represented as the decoding features output by the nth second decoder.

[0171] The second convolution unit is used to perform convolution processing on the fifth feature to obtain the sixth feature.

[0172] The activation unit is used to perform maximum pooling processing on the sixth feature to obtain a decoded feature output by the i-th first decoder, wherein the decoded feature output by the m-th first decoder is represented as a target image.

[0173] According to an embodiment of the present invention, the second input module 630 includes an average pooling submodule, a first processing submodule, a splitting submodule, a second processing submodule and a weighted processing submodule.

[0174] The average pooling submodule is used to average pool the encoded features output by the h-th encoder in the height direction and the width direction respectively to obtain the seventh feature and the eighth feature, where 0<h≤k.

[0175] The first processing submodule is used to concatenate the seventh feature and the eighth feature, and then perform convolution, batch normalization and nonlinear activation processing in sequence to obtain the ninth feature.

[0176] The splitting submodule is used to split the ninth feature into a first tensor and a second tensor along the channel direction.

[0177] The second processing submodule is used to perform convolution and nonlinear activation processing on the first tensor and the second tensor respectively to obtain a first attention weight in the height direction and a second attention weight in the width direction.

[0178] The weighted processing submodule is used to perform weighted processing on the encoding features output by the h-th encoder based on the first attention weight and the second attention weight to obtain the skip connection features output by the h-th skip connection unit.

[0179] According to the embodiments of the present invention, any of the modules, submodules, units, image acquisition module 610, first input module 620, second input module 630, third input module 640 and result determination module 650 subunits, or at least part of the functions of any of them can be implemented in one module. According to the embodiments of the present invention, any one or more of the modules, submodules, units, and subunits can be split into multiple modules for implementation. According to the embodiments of the present invention, any one or more of the modules, submodules, units, and subunits can be at least partially implemented as hardware circuits, such as field programmable gate arrays (FPGAs), programmable logic arrays (PLAs), systems on chips, systems on substrates, systems on packages, application specific integrated circuits (ASICs), or can be implemented by hardware or firmware in any other reasonable way of integrating or packaging circuits, or can be implemented in any one of the three implementation methods of software, hardware, and firmware, or in any appropriate combination of any of them. Alternatively, according to the embodiments of the present invention, one or more of the modules, submodules, units, and subunits can be at least partially implemented as computer program modules, and when the computer program modules are run, the corresponding functions can be executed.

[0180] For example, any of the image acquisition module 610, the first input module 620, the second input module 630, the third input module 640 and the result determination module 650 can be combined in one module / unit / sub-unit for implementation, or any of the modules / units / sub-units can be split into multiple modules / units / sub-units. Alternatively, at least part of the functions of one or more of these modules / units / sub-units can be combined with at least part of the functions of other modules / units / sub-units and implemented in one module / unit / sub-unit. According to an embodiment of the present invention, at least one of the image acquisition module 610, the first input module 620, the second input module 630, the third input module 640 and the result determination module 650 can be at least partially implemented as a hardware circuit, such as a field programmable gate array (FPGA), a programmable logic array (PLA), a system on a chip, a system on a substrate, a system on a package, an application specific integrated circuit (ASIC), or can be implemented by hardware or firmware such as any other reasonable way of integrating or packaging the circuit, or by any one of the three implementation methods of software, hardware and firmware or by a suitable combination of any of them. Alternatively, at least one of the image acquisition module 610, the first input module 620, the second input module 630, the third input module 640 and the result determination module 650 may be at least partially implemented as a computer program module, which may perform corresponding functions when executed.

[0181] It should be noted that the transformer insulation surface discharge defect detection device part in the embodiment of the present invention corresponds to the transformer insulation surface discharge defect detection method part in the embodiment of the present invention. The description of the transformer insulation surface discharge defect detection device part specifically refers to the transformer insulation surface discharge defect detection method part, which will not be repeated here.

[0182] The flowcharts and block diagrams in the accompanying drawings illustrate the possible architecture, functions and operations of the systems, methods and computer program products according to various embodiments of the present invention. In this regard, each box in the flowchart or block diagram may represent a module, a program segment, or a part of a code, and the above-mentioned module, program segment, or a part of the code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box may also occur in an order different from the order marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram or flowchart, and the combination of boxes in the block diagram or flowchart, can be implemented with a dedicated hardware-based system that performs the specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions. It can be understood by those skilled in the art that the features recorded in the various embodiments of the present invention can be combined and / or combined in various ways, even if such a combination or combination is not explicitly recorded in the present invention. In particular, without departing from the spirit and teachings of the present invention, the features recorded in the various embodiments of the present invention can be combined and / or combined in various ways. All these combinations and / or combinations fall within the scope of the present invention.

[0183] The embodiments of the present invention are described above. However, these embodiments are only for the purpose of illustration, and are not intended to limit the scope of the present invention. Although each embodiment is described above, it does not mean that the measures in each embodiment cannot be used in combination advantageously. Without departing from the scope of the present invention, those skilled in the art may make various substitutions and modifications, which should all fall within the scope of the present invention.

Claims

1. A method for detecting transformer insulation surface discharge defects, characterized in that: The method comprises: In response to a defect detection request for an oil-paper insulation structure to be detected, acquiring an initial image of the oil-paper insulation structure; Inputting the initial image into an encoder module to obtain encoding features output by each of a plurality of encoders included in the encoder module, wherein the encoder module includes k encoders; Inputting the encoding features output by each of the multiple encoders into a skip connection module based on an attention mechanism to obtain multiple skip connection features, wherein the skip connection module based on the attention mechanism includes k skip connection units; Inputting the plurality of skip connection features into a decoder module to obtain a target image, wherein the decoder module comprises k decoders; and Based on the edge features of the target image, a surface discharge defect detection result of the oil-paper insulation structure is obtained.

2. The method according to claim 1, characterized in that The step of obtaining the surface discharge defect detection result of the oil-paper insulation structure based on the edge features of the target image includes: Performing image classification on the target image to obtain an edge type recognition result of the target image; In the case where the edge type recognition result indicates a dendritic edge, determining a discharge defect area from the target image; Determine the defect area of ​​the oil-paper insulation structure based on the mapping relationship between the discharge defect area and the camera coordinate system of the target image and the initial image and the world coordinate system; and Based on the defect region area, a surface discharge defect detection result of the oil-paper insulation structure is obtained.

3. The method according to claim 1, characterized in that The k encoders include m first encoders and n second encoders; The method is characterized in that the initial image is input into an encoder module to obtain encoding features output by each of a plurality of encoders included in the encoder module, including: Inputting the initial image into the m first encoders connected in series to obtain encoding features output by each of the m first encoders; and The encoding feature output by the mth first encoder is input into the n second encoders connected in series to obtain the encoding features output by each of the n second encoders.

4. The method according to claim 3, characterized in that Inputting the initial image into the m first encoders connected in series to obtain encoding features output by each of the m first encoders comprises: Performing convolution processing on the encoding feature output by the i-1th first encoder to obtain a first feature, wherein 0<i≤m, when i=1, the encoding feature output by the i-1th first encoder is represented as the initial image; and Perform maximum pooling processing on the first feature to obtain the encoding feature output by the i-th first encoder.

5. The method according to claim 3, characterized in that: The step of inputting the coding feature output by the mth first encoder into the n second encoders connected in series to obtain the coding features output by each of the n second encoders comprises: Normalizing the encoding feature output by the j-1th second encoder to obtain a second feature, wherein 0<j≤n, when j=1, the encoding feature output by the j-1th second encoder is expressed as the encoding feature output by the mth first encoder; Splitting the second feature according to the channel dimension to obtain multiple first sub-features; Based on the combination dimension corresponding to the first sub-feature, performing an element-by-element product attention operation on the first sub-feature to obtain a second sub-feature; and Multiple second sub-features are sequentially concatenated, normalized, subjected to depthwise separable convolution and maximum pooling processing to obtain the encoding features output by the j-th second encoder.

6. The method according to claim 1, characterized in that The k decoders include n second decoders and m first decoders; Characterized in that the step of inputting the plurality of skip connection features into a decoder module to obtain a target image comprises: Inputting the skip connection features outputted by each of the k-th skip connection unit to the k+1-n-th skip connection unit into the n second decoders connected in series to obtain a decoding feature outputted by the n-th second decoder; and The skip connection features outputted by the kn-th skip connection unit to the 1-th skip connection unit and the decoded features outputted by the n-th second decoder are inputted into the m first decoders connected in series to obtain the target image.

7. The method according to claim 6, characterized in that The step of inputting the skip connection features outputted by the kth skip connection unit to the k+1-nth skip connection unit into the n second decoders connected in series to obtain the decoding features outputted by the nth second decoder comprises: Add the decoding feature output by the j-1th second decoder to the skip connection feature output by the k+1-jth skip connection unit to obtain a third feature, wherein 0<j≤n, and when j=1, the decoding feature output by the j-1th second decoder is represented as 0; Normalize the third feature to obtain the fourth feature; Splitting the fourth feature according to the channel dimension to obtain multiple third sub-features; Based on the combined dimension corresponding to the third sub-feature, performing an element-by-element product attention operation on the third sub-feature to obtain a fourth sub-feature; and Multiple fourth sub-features are sequentially concatenated, normalized, depth-separable convolved, and activated to obtain a decoding feature output by the j-th second decoder.

8. The method according to claim 6, characterized in that The step of inputting the skip connection features respectively outputted by the kn-th skip connection unit to the first skip connection unit and the decoded features outputted by the n-th second decoder into the m first decoders connected in series to obtain the target image comprises: Add the decoding feature output by the i-1th first decoder and the skip connection feature output by the k+1-inth skip connection unit to obtain a fifth feature, wherein 0<i≤m, and when i=1, the decoding feature output by the i-1th first decoder is represented as the decoding feature output by the nth second decoder; Performing convolution processing on the fifth feature to obtain a sixth feature; and The sixth feature is activated to obtain a decoded feature output by the i-th first decoder, wherein the decoded feature output by the m-th first decoder is represented as the target image.

9. The method according to claim 1, characterized in that: The step of inputting the encoding features outputted by each of the multiple encoders into a skip connection module based on an attention mechanism to obtain multiple skip connection features includes: The encoded features output by the h-th encoder are average-pooled in the height direction and the width direction respectively to obtain the seventh feature and the eighth feature, where 0<h≤k; After concatenating the seventh feature and the eighth feature, convolution, batch normalization and nonlinear activation processing are performed in sequence to obtain a ninth feature; Splitting the ninth feature into a first tensor and a second tensor along the channel direction; Performing convolution and nonlinear activation processing on the first tensor and the second tensor respectively to obtain a first attention weight in a height direction and a second attention weight in a width direction; and Based on the first attention weight and the second attention weight, weighted processing is performed on the encoded features output by the h-th encoder to obtain the skip connection features output by the h-th skip connection unit.

10. A transformer insulation surface discharge defect detection device, characterized in that: The device comprises: An image acquisition module, configured to acquire an initial image of the oil-paper insulation structure in response to a defect detection request for the oil-paper insulation structure to be detected; A first input module, used for inputting the initial image into an encoder module to obtain encoding features output by each of a plurality of encoders included in the encoder module, wherein the encoder module includes k encoders; A second input module is used to input the encoding features output by each of the multiple encoders into a skip connection module based on an attention mechanism to obtain multiple skip connection features, wherein the skip connection module based on the attention mechanism includes k skip connection units; A third input module, configured to input the plurality of skip connection features into a decoder module to obtain a target image, wherein the decoder module includes k decoders; and The result determination module is used to obtain the surface discharge defect detection result of the oil-paper insulation structure based on the edge features of the target image.

Images