Methods, devices, electronic equipment and storage media for defect identification of power equipment

Abstract

This application provides a method, apparatus, electronic device, and storage medium for defect identification of power equipment. The method first acquires an image of the power equipment to be detected. Then, the image is input into a pre-trained defect identification model to obtain the target defect type corresponding to the image. The defect identification model is trained on an autoencoder network based on training images, defect word vector features of the training images, visual embedding features of the training images, and reconstructed images. This technical solution processes image feature extraction by inputting the image of the power equipment to be detected into the pre-trained defect identification model, and then outputs the target defect type corresponding to the image. The defect identification model can not only learn the visual information of the image but also accurately understand the semantic information of the defect type. Using the defect identification model for defect identification of power equipment helps improve the efficiency and accuracy of defect identification.

Description

Technical Field

[0001] This application relates to the field of target detection technology, and in particular to a method, apparatus, electronic device and storage medium for identifying defects in power equipment. Background Technology

[0002] With the increasing number of power equipment and the growing complexity of its operating environment, the identification of power equipment defects faces unprecedented challenges. With the rapid development of intelligent technology, defect identification methods based on machine learning and artificial intelligence have gradually become a research hotspot.

[0003] In the existing technology, it is impossible to determine a high-accuracy defect identification method for power equipment based on existing defect images and corresponding category text labels.

[0004] Therefore, improving the accuracy of defect identification in power equipment has become an urgent technical problem to be solved. Summary of the Invention

[0005] This application provides a method, apparatus, electronic device, and storage medium for identifying defects in power equipment, in order to address existing technologies' challenges in improving the accuracy of defect identification in power equipment.

[0006] In a first aspect, embodiments of this application provide a method for identifying defects in power equipment, including:

[0007] Acquire the image of the power equipment to be inspected;

[0008] The image to be detected is input into a pre-trained defect recognition model to obtain the target defect type corresponding to the image to be detected. The defect recognition model is obtained by training an autoencoder network based on at least one training image, word vector features of the defect type corresponding to the at least one training image, visual embedding features corresponding to the at least one training image, and reconstructed images corresponding to the at least one image features.

[0009] In one possible implementation, before inputting the image to be detected into a pre-trained defect recognition model to obtain the target defect type corresponding to the image to be detected, the method further includes:

[0010] S1, input the at least one training image into the encoder in the autoencoder network to obtain the first image feature corresponding to the at least one training image;

[0011] S2, input the first image features corresponding to the at least one training image into the first feature recombination unit in the autoencoder network to obtain the visual embedding features corresponding to the at least one training image;

[0012] S3, input the visual embedding features corresponding to the at least one training image into the second feature recombination unit in the autoencoder network to obtain the second image features corresponding to the at least one training image;

[0013] S4, input the second image features corresponding to the at least one training image into the decoder in the autoencoder network to obtain the reconstructed image corresponding to the at least one training image;

[0014] S5, input the reconstructed image corresponding to the at least one training image, the at least one training image, the visual embedding feature corresponding to the at least one training image, and the word vector feature of the defect type corresponding to the at least one training image into the loss function in the autoencoder network to obtain the loss function value;

[0015] S6. If the loss function value is greater than or equal to a preset threshold, update the parameters in the autoencoder network and repeat S1-S5 until the loss function value is less than the preset threshold. Then, use the autoencoder network corresponding to the loss function value that is less than the preset threshold as the defect identification model.

[0016] In one possible implementation, inputting the image to be detected into a pre-trained defect recognition model to obtain the target defect type corresponding to the image to be detected includes:

[0017] The image to be detected is input into the encoder in the defect recognition model to obtain the target image features corresponding to the image to be detected.

[0018] The target image features are input into the first feature recombination unit in the defect recognition model to obtain the target visual embedding features corresponding to the target image features;

[0019] The target visual embedding features are matched with the word vector features of the defect type in the semantic representation space of the defect recognition model to obtain at least one matching result and the probability value corresponding to the at least one matching result;

[0020] The defect type corresponding to the word vector feature of the matching result with the highest probability value is taken as the target defect type.

[0021] In one possible implementation, the step of inputting the at least one training image into the encoder of the autoencoder network to obtain the first image feature corresponding to the at least one training image includes:

[0022] For each training image, the training image is sequentially input into the first convolutional layer unit, the first invertible neural network unit, the first inverted residual moving block unit, the recursive transformer encoder unit, and the second convolutional layer unit in the encoder to obtain the first image feature corresponding to the training image.

[0023] In one possible implementation, the step of inputting the second image features corresponding to the at least one training image into the decoder in the autoencoder network to obtain the reconstructed image corresponding to the at least one training image includes:

[0024] For each second image feature, the second image feature is sequentially input into the third convolutional layer unit, the recursive transformer decoder unit, the second inverted residual moving block unit, the second invertible neural network unit, and the fourth convolutional layer unit to obtain the reconstructed image corresponding to the training image.

[0025] In one possible implementation, before inputting the image to be detected into a pre-trained defect recognition model to obtain the target defect type corresponding to the image to be detected, the method further includes:

[0026] Obtain at least one training image, the defect type corresponding to the at least one training image, and the defect type that does not appear in the at least one training image;

[0027] The defect type corresponding to the at least one training image and the defect type that does not appear in the at least one training image are input into a preset algorithm model to obtain the word vector features corresponding to the at least one defect type.

[0028] Secondly, embodiments of this application provide a defect identification device for power equipment, comprising:

[0029] The acquisition module is used to acquire the image of the power equipment to be inspected.

[0030] The processing module is used to input the image to be detected into a pre-trained defect recognition model to obtain the target defect type corresponding to the image to be detected. The defect recognition model is obtained by training an autoencoder network based on at least one training image, word vector features of the defect type corresponding to the at least one training image, visual embedding features corresponding to the at least one training image, and reconstructed images corresponding to the at least one image features.

[0031] In one possible implementation, before inputting the image to be detected into a pre-trained defect recognition model to obtain the target defect type corresponding to the image to be detected, the processing module is further configured to:

[0032] S1 is executed, in which the at least one training image is input into the encoder in the autoencoder network to obtain the first image feature corresponding to the at least one training image;

[0033] S2 is executed, and the first image features corresponding to the at least one training image are input into the first feature recombination unit in the autoencoder network to obtain the visual embedding features corresponding to the at least one training image.

[0034] S3 is executed, and the visual embedding features corresponding to the at least one training image are input into the second feature recombination unit in the autoencoder network to obtain the second image features corresponding to the at least one training image.

[0035] Execute S4 to input the second image features corresponding to the at least one training image into the decoder in the autoencoder network to obtain the reconstructed image corresponding to the at least one training image;

[0036] Execute S5, inputting the reconstructed image corresponding to the at least one training image, the at least one training image, the visual embedding feature corresponding to the at least one training image, and the word vector feature of the defect type corresponding to the at least one training image into the loss function in the autoencoder network to obtain the loss function value;

[0037] Execute S6. If the loss function value is greater than or equal to a preset threshold, update the parameters in the autoencoder network and repeat S1-S5 until the loss function value is less than the preset threshold. Then, use the autoencoder network corresponding to the loss function value that is less than the preset threshold as the defect identification model.

[0038] In one possible implementation, the processing module is specifically used for:

[0039] The image to be detected is input into the encoder in the defect recognition model to obtain the target image features corresponding to the image to be detected.

[0040] The target image features are input into the first feature recombination unit in the defect recognition model to obtain the target visual embedding features corresponding to the target image features;

[0041] The target visual embedding features are matched with the word vector features of the defect type in the semantic representation space of the defect recognition model to obtain at least one matching result and the probability value corresponding to the at least one matching result;

[0042] The defect type corresponding to the word vector feature of the matching result with the highest probability value is taken as the target defect type.

[0043] In one possible implementation, the step of inputting the at least one training image into the encoder of the autoencoder network to obtain the first image feature corresponding to the at least one training image is specifically configured to:

[0044] For each training image, the training image is sequentially input into the first convolutional layer unit, the first invertible neural network unit, the first inverted residual moving block unit, the recursive transformer encoder unit, and the second convolutional layer unit in the encoder to obtain the first image feature corresponding to the training image.

[0045] In one possible implementation, the step of inputting the second image features corresponding to the at least one training image into the decoder of the autoencoder network to obtain the reconstructed image corresponding to the at least one training image is specifically configured to:

[0046] For each second image feature, the second image feature is sequentially input into the third convolutional layer unit, the recursive transformer decoder unit, the second inverted residual moving block unit, the second invertible neural network unit, and the fourth convolutional layer unit to obtain the reconstructed image corresponding to the training image.

[0047] In one possible implementation, before inputting the image to be detected into a pre-trained defect recognition model to obtain the target defect type corresponding to the image to be detected, the processing module is further configured to:

[0048] Obtain at least one training image, the defect type corresponding to the at least one training image, and the defect type that does not appear in the at least one training image;

[0049] The defect type corresponding to the at least one training image and the defect type that does not appear in the at least one training image are input into a preset algorithm model to obtain the word vector features corresponding to the at least one defect type.

[0050] Thirdly, embodiments of this application provide an electronic device, including: a processor, and a memory communicatively connected to the processor;

[0051] The memory stores computer-executed instructions;

[0052] The processor executes computer execution instructions stored in the memory to implement the method as described in the first aspect or any of the above methods.

[0053] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the method described in the first aspect or any of the above-described methods.

[0054] Fifthly, embodiments of this application provide a computer program, the computer program product including a computer program stored in a computer-readable storage medium, at least one processor can read the computer program from the computer-readable storage medium, and the at least one processor can implement the method described in the first aspect or any of the above methods when executing the computer program.

[0055] The present application provides a method, apparatus, electronic device, and storage medium for identifying defects in power equipment. The method first acquires an image of the power equipment to be detected. Then, the image is input into a pre-trained defect identification model to obtain the target defect type corresponding to the image. The defect identification model is trained on an autoencoder network based on at least one training image, word vector features of the defect type corresponding to at least one training image, visual embedding features corresponding to at least one training image, and a reconstructed image corresponding to at least one image feature. This technical solution involves inputting the image of the power equipment to be detected into the pre-trained defect identification model for image feature extraction. The defect identification model analyzes the extracted image features and then outputs the target defect type corresponding to the image. The defect identification model can not only learn the visual information of the image but also accurately understand the semantic information of the defect type. Using the defect identification model for power equipment defect identification helps improve the efficiency and accuracy of power equipment defect identification. Attached Figure Description

[0056] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0057] Figure 1 A schematic diagram of the architecture of a self-encoding network provided in an embodiment of this application;

[0058] Figure 2 A flowchart illustrating the defect identification method for power equipment provided in this application embodiment. Figure 1 ;

[0059] Figure 3 A flowchart illustrating the defect identification method for power equipment provided in this application embodiment. Figure 2 ;

[0060] Figure 4 A flowchart illustrating the defect identification method for power equipment provided in this application embodiment. Figure 3 ;

[0061] Figure 5 A flowchart illustrating the defect identification method for power equipment provided in this application embodiment. Figure 4 ;

[0062] Figure 6 This is a schematic diagram of the structure of the defect identification device for power equipment provided in the embodiments of this application;

[0063] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0064] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0065] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0066] Before introducing the embodiments of this application, the application background of the embodiments of this application will be explained first:

[0067] With the increasing number of power equipment and the growing complexity of its operating environment, the identification of power equipment defects faces unprecedented challenges. With the rapid development of intelligent technology, defect identification methods based on machine learning and artificial intelligence have gradually become a research hotspot. Among them, zero-shot learning (ZSL) has attracted widespread attention because of its ability to effectively use existing knowledge to infer unseen categories.

[0068] In the scenario of defect identification in power equipment, we often face situations where the types of equipment defects are unknown. For example, before some new power equipment is launched on the market, there is a lack of sufficient defect label data, which makes it impossible to directly apply traditional supervised learning methods. Zero-shot learning, by introducing semantic information, combines the visual features of visible equipment defects with their descriptive semantic information (such as defect type, defect feature attributes, etc.), overcoming the problem of scarce defect data and enabling the inference of defect types not seen during the training phase.

[0069] Existing defect identification methods still face many challenges in the application of power equipment. They cannot determine a highly accurate defect identification method for power equipment based on existing defect images and corresponding text labels.

[0070] Therefore, improving the accuracy of defect identification in power equipment has become an urgent technical problem to be solved.

[0071] To address the technical problems existing in the prior art, the inventors of this application have the following concept: Addressing the issue that existing technologies cannot determine a highly accurate method for identifying defects in power equipment, this application employs an autoencoder network to comprehensively acquire multiple features from the training image, the word vectors corresponding to the training image, the visual embedding features of the training image, and the reconstructed image corresponding to the training image. By fully utilizing the correlation between the visual and semantic features of the image, a defect identification model is obtained. Using this model to detect the defect type in the image to be detected can improve the accuracy of defect identification of power equipment in the image. In some possible implementations, the autoencoder network includes invertible neural network (INN) units, which can retain detailed information of the training image feature data and invert... residual The Inverted Residual Mobile Block (iRMB) can capture local features of the training image, achieving the goal of extracting richer image features. The Restoration Transformer (Restormer) unit enhances the multi-scale local-global learning ability on the training image. The above-mentioned units of the autoencoder network improve the accuracy of the model in defect recognition to varying degrees.

[0072] Specifically, Figure 1 This is a schematic diagram of the architecture of the autoencoder network provided in the embodiments of this application, as shown below. Figure 1 The following describes the autoencoder network involved in the embodiments of this application:

[0073] The architecture includes: an encoder, a visual semantic constraint module, a decoder, and a feature reshaping unit. Figure 1 (R in the text).

[0074] The encoder transforms the input image into a more information-dense and abstract representation by extracting important features from the original input data and mapping these features to a fixed-dimensional representation.

[0075] The visual semantic constraint module captures the semantic relationships between different input data (such as visual embedding features and word vector features) by comparing and matching them.

[0076] A decoder is a tool that decodes the low-dimensional representation generated by the encoder (usually an image feature vector or a context-rich feature representation) back into the form of the original image (i.e., training image) data.

[0077] The feature reorganization unit reorganizes or optimizes the extracted image features to better capture key information and obtain the visual embedding features of the image, thereby improving the performance and accuracy of the model.

[0078] For example, the terms used in the self-encoding network provided in the embodiments of this application are explained as follows:

[0079] Convolutional Layer (Conv): A commonly used neural network layer in deep learning, which extracts features from input data through convolution operations.

[0080] Reversible neural networks: neural network architectures designed based on the principle of invariance.

[0081] Inverted Residual Moving Block: A module that helps train networks to deeper structures by introducing shortcut connections. The design of the residual block allows the input signal to jump directly to the output, thereby solving the gradient vanishing problem and enabling the network to be trained better.

[0082] Recurrent Transformer Encoder (Resormer Encoder): This is an encoder architecture that combines recurrent neural networks and transformer architecture.

[0083] Recurrent Transformer Decoder (Resormer Decoder): This is a decoder architecture that combines a recurrent neural network and a transformer model.

[0084] The parts not described in detail are disclosed in the following embodiments.

[0085] The technical solution of this application will now be described in detail through specific embodiments. It should be noted that the following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0086] It is worth noting that the application fields of the defect identification methods, devices, electronic equipment and storage media for power equipment in this application are not limited.

[0087] The subject of this application is an electronic device, which may specifically be a server, terminal device, etc.

[0088] Figure 2 A flowchart illustrating the defect identification method for power equipment provided in this application embodiment. Figure 1 ,like Figure 2 As shown, the method may include the following steps:

[0089] Step 21: Obtain the image of the electrical equipment to be inspected;

[0090] In this step, in order to identify the type of defect in the electrical equipment to be inspected, image data containing the electrical equipment to be inspected can be obtained first, and denoted as the image to be inspected.

[0091] For example, the image to be detected may contain different components of the electrical equipment to be detected (such as transformers, cables, switches, etc.).

[0092] In one possible implementation, the image to be detected can be obtained by: using an image sensor, such as a camera or webcam, to capture images of the area where the power equipment to be detected is located; or by directly obtaining the image from a pre-set physical storage space, such as a service platform or memory card.

[0093] In addition, to ensure the accuracy of the detection results, the images to be detected can be preprocessed, such as image resizing, image noise reduction, and normalization.

[0094] Step 22: Input the image to be detected into the pre-trained defect recognition model to obtain the target defect type corresponding to the image to be detected.

[0095] The defect recognition model is obtained by training an autoencoder network based on at least one training image, word vector features of the defect type corresponding to at least one training image, visual embedding features corresponding to at least one training image, and reconstructed images corresponding to at least one image feature.

[0096] In this step, the image to be detected is input into a pre-trained defect recognition model to extract image features from the image to be detected, and the corresponding target defect type is determined based on the image features in the image to be detected.

[0097] The image to be detected may include defect types that did not appear in any of the training images (i.e., invisible defects).

[0098] For example, the defect recognition model is obtained by learning defect features and corresponding word vector features of defect types from a large number of training images, thereby ensuring that when a new image to be detected is input, accurate defect feature information can be extracted from the image to be detected, and then the corresponding defect type information can be matched from the word vector features of defect types in the training images.

[0099] In one possible implementation, the above defect recognition model can be a zero-shot learning-based model. In a zero-shot learning-based model, the training images are visible defect images (that is, the defects in the images are historically known defects, and the defect type can be represented by defect text labels). The images to be detected may include all defect types that have not appeared in the training images (that is, invisible defects). Invisible defects can only be described by text labels but do not have corresponding image representations (e.g., metal fatigue cracks).

[0100] The defect recognition model based on zero-shot learning can be trained on an autoencoder network using visible defect images, word vector features corresponding to visible defect images, visual embedding features corresponding to visible defect images, and reconstructed images corresponding to visible defect images.

[0101] For example, the image to be detected corresponding to the invisible type of defect is input into the above-mentioned defect recognition model based on zero-shot learning. The defect recognition model can analyze the visual embedding features of the image to be detected, the word vector features corresponding to the visible type of defect image, and the word vector features corresponding to the invisible type of defect image (which can be obtained through the implementation of steps 1 and 2 before step S1 below) to obtain the defect type corresponding to the image to be detected corresponding to the invisible type of defect.

[0102] It should be noted that the autoencoder network provided in this application embodiment can be used not only in zero-shot learning-based models, but also in other deep learning-based defect recognition models.

[0103] Optionally, step 22 may include the following implementation:

[0104] Step 1: Input the image to be detected into the encoder in the defect recognition model to obtain the target image features corresponding to the image to be detected.

[0105] In this implementation, the image to be detected is input into the encoder of the defect recognition model. The encoder will convert the image into a low-dimensional feature representation to obtain the target image features corresponding to the image to be detected.

[0106] The target image features include visual features such as shape, texture, and color distribution of defects in the image to be detected, which are used to describe the overall features of defects in the image to be detected.

[0107] Step 2: Input the target image features into the first feature recombination unit in the defect recognition model to obtain the target visual embedding features corresponding to the target image features.

[0108] In this implementation, the target image features are input into the first feature reconstructing unit in the defect recognition model. The feature reconstructing unit converts the low-dimensional feature representation in the image features into the target visual embedding feature representation.

[0109] The feature recombination process is equivalent to extracting a more refined structure or pattern from image features, so that the obtained target visual embedding features can be better matched with word vector features in the semantic representation space of the defect detection model.

[0110] Step 3: Match the target visual embedding features with the word vector features of the defect type in the semantic representation space of the defect recognition model to obtain at least one matching result and the probability value corresponding to at least one matching result.

[0111] In this implementation, the target visual embedding features corresponding to the target image are compared with the word vector features of the defect type in the semantic representation space of the defect recognition model. The matching degree between the target visual embedding features and the word vector features of each defect type is calculated, and the matching result and the corresponding matching probability value for each defect type can be obtained.

[0112] Among them, the word vector features of defect types in the semantic representation space can be the word vector features corresponding to visible defects in the training image, or the word vector features corresponding to new defect types (i.e., invisible defects) that have not appeared in the training image.

[0113] For example, the matching result is obtained by calculating the matching degree between the visual embedding features of the target image and the word vector features of each defect type, which is used to measure the similarity between the visual features of the target image and the features of each defect type.

[0114] In one possible implementation, the target image is matched as worn, with a matching probability of 80%.

[0115] Step 4: Select the defect type corresponding to the word vector feature of the matching result with the highest probability value as the target defect type.

[0116] In this implementation, the defect recognition model selects the matching result with the highest probability value based on the matching probability value. The defect type represented by the word vector feature corresponding to the matching result with the highest probability value is the target defect type in the image to be detected.

[0117] In one possible implementation, the first matching result of the target image is wear, with a matching probability of 80%, and the second matching result of the target image is fracture, with a matching probability of 95%. Therefore, the defect matching result of the target image is fracture, that is, the target defect type is fracture.

[0118] The defect identification method for power equipment provided in this application first acquires an image of the power equipment to be detected, then inputs the image into a pre-trained defect identification model to obtain the target defect type corresponding to the image. The defect identification model is trained on an autoencoder network based on at least one training image, word vector features of at least one defect type corresponding to at least one training image, visual embedding features corresponding to at least one training image, and a reconstructed image corresponding to at least one image feature. This technical solution involves inputting the image of the power equipment to be detected into the pre-trained defect identification model for image feature extraction. The defect identification model analyzes the extracted image features and then outputs the target defect type corresponding to the image. The defect identification model can not only learn the visual information of the image to be detected but also accurately understand the semantic information of the defect type. Using the defect identification model for power equipment defect identification helps improve the efficiency and accuracy of power equipment defect identification.

[0119] Based on the above embodiments, Figure 3 A flowchart illustrating the defect identification method for power equipment provided in this application embodiment. Figure 2 ,like Figure 3 As shown, before step 22, the following steps may be included:

[0120] S1, input at least one training image into the encoder in the autoencoder network to obtain the first image features corresponding to at least one training image;

[0121] In this step, the training image is input into the encoder part of the autoencoder network to perform feature extraction and dimensionality reduction on the input image, mapping the training image from a high-dimensional pixel space to a lower-dimensional latent feature space, thereby obtaining the first image feature corresponding to the training image.

[0122] The first image features include key feature information of the training image (such as defect shape, defect color, defect texture, etc.). These feature information are a concise representation of the content of the first image and can effectively represent the visual feature information of the training image.

[0123] Optionally, prior to step S1, the following implementation may also be included:

[0124] Step 1: Obtain at least one training image, at least one defect type corresponding to the training image, and at least one defect type that does not appear in the training image;

[0125] In this implementation, at least one training image, the corresponding defect type in the training image, and at least one defect type that did not appear in the training image are obtained from the historical database.

[0126] The training images cover various types of power equipment and defect situations to enhance the generalization ability of the subsequent defect recognition model. Each training image contains at least one power equipment defect (i.e., visible defects), and the defect types that do not appear in the training images are also defects that have not appeared in the training images (i.e., invisible defects).

[0127] For example, each training image has a corresponding text label for the defect type (i.e., the visible class of defects), and each invisible class of defect type also has a corresponding text label to indicate the defect type in the training image (such as cracks, damage, oil leaks, dirt, etc.).

[0128] Step 2: Input at least one defect type corresponding to a training image and at least one defect type not appearing in the training image into the preset algorithm model to obtain word vector features corresponding to at least one defect type.

[0129] In this implementation, the defect types corresponding to the training images (i.e., visible defect types) and the defect types that do not appear in the training images (i.e., invisible defect types) are input into the preset algorithm model, and the above visible and invisible defect types are converted into corresponding word vector features.

[0130] Word vector features are mathematical representations of the text labels corresponding to defect types, capturing the semantic relationships between individual words. By converting defect types into word vectors, the defect recognition model can understand the similarities and differences between different defect types and map them into a high-dimensional vector space. In this space, similar defect types are mapped to nearby positions, while different defect types are mapped to more distant positions.

[0131] For example, the preset algorithm model can be word embedding technology (Word to Vector, Word2Vec), bidirectional encoder representations from transformers (BERT), generative pre-trained transformers (GPT), etc.

[0132] S2, input the first image feature corresponding to at least one training image into the first feature recombination unit in the autoencoder network to obtain the visual embedding feature corresponding to at least one training image;

[0133] In this step, the first image features output by the encoder are input into the first feature reorganization unit in the autoencoder network to optimize and reorganize the input first image features, enhance their feature representation ability, and extract more refined and discriminative visual embedding features.

[0134] Among them, visual embedding features are refined representations of the first image features, containing key information of the first image features of the training image, and can be used for subsequent image reconstruction and defect recognition and classification tasks.

[0135] Optionally, step S2 can be implemented as follows:

[0136] For each training image, the training image is sequentially input into the first convolutional layer unit, the first invertible neural network unit, the first inverted residual moving block unit, the recursive transformer encoder unit, and the second convolutional layer unit in the encoder to obtain the first image feature corresponding to the training image.

[0137] In this implementation, each training image is sequentially input into the first convolutional layer unit, the first invertible neural network unit, the first inverted residual moving block unit, the recursive transformer encoder unit, and the second convolutional layer unit of the autoencoder network to progressively extract, transform, and enhance features until the high-level image features, i.e., the first image features corresponding to the training images, are finally obtained.

[0138] S3, input the visual embedding feature corresponding to at least one training image into the second feature recombination unit in the autoencoder network to obtain the second image feature corresponding to at least one training image;

[0139] In this step, at least one visual embedding feature corresponding to a training image is input into the second feature reorganization unit in the autoencoder network, and the input visual embedding feature is further processed to obtain a more refined and discriminative second image feature.

[0140] The second feature recombination unit can highly compress or optimize the features of the image and extract the most core visual information in the image, enabling the autoencoder network to better understand and represent the training image content and provide higher quality feature input for the subsequent decoding process.

[0141] S4, input the second image features corresponding to at least one training image into the decoder in the autoencoder network to obtain the reconstructed image corresponding to at least one training image;

[0142] In this step, the second image features are input into the decoder part of the autoencoder network, and the latent low-dimensional representation in the visual embedding features is transformed back into the high-dimensional space of the training image to obtain the reconstructed image corresponding to the training image.

[0143] Through the above process, the decoder gradually restores the spatial structure of the image through a series of operations, and restores the reconstructed image as close as possible to the training image.

[0144] Optionally, step S4 can be implemented as follows:

[0145] For each second image feature, the second image feature is sequentially input into the third convolutional layer unit, the recursive transformer decoder unit, the second inverted residual moving block unit, the second invertible neural network unit, and the fourth convolutional layer unit to obtain the reconstructed image corresponding to the training image.

[0146] In this implementation, each second image feature is sequentially input into the third convolutional layer unit, the recursive transformer decoder unit, the second inverted residual moving block unit, the second invertible neural network unit, and the fourth convolutional layer unit to perform feature restoration and reconstruction, thereby obtaining the reconstructed image corresponding to the training image.

[0147] S5, input the reconstructed image corresponding to at least one training image, at least one training image, the visual embedding feature corresponding to at least one training image, and the word vector feature of the defect type corresponding to at least one training image into the loss function in the autoencoder network to obtain the loss function value;

[0148] In this step, the reconstruction error (e.g., mean squared error) between the input training image and the reconstructed image is calculated. The mean squared errors of the computational visual embedding features corresponding to the training image and the word vector features of the defect type corresponding to the training image are also calculated. The reconstruction error and mean squared error are then input into the loss function in the autoencoder network to calculate the loss function value, which serves as the basis for evaluating the model performance.

[0149] In one possible implementation, regularization is introduced into the visual semantic constraint module between the encoder and decoder in the autoencoder network, and the final loss function formula is as follows:

[0150]

[0151] In the formula, X represents the sum of the two sides. Let H and S be the training image and the reconstructed image, respectively. H and S are the visual embedding features corresponding to the training image and the word vector features corresponding to the defect type of the training image, respectively. and It is a hyperparameter greater than 0. It refers to an exponential function that is similar in structure to the training and reconstructed images.

[0152] The first part of the loss function represents the constraints of image reconstruction (i.e., the reconstruction error), and the corresponding mathematical formula is:

[0153]

[0154] The above formula uses the basic formula of mean squared error, which ensures that the model retains the pixel intensity and detailed texture information of the training image in the reconstructed image during the training process, while minimizing the difference between the training image data and the reconstructed image data.

[0155] To further describe the internal structural information of the training and reconstructed images and preserve it in the visual embedding features, a structural similarity constraint is introduced based on the above image reconstruction constraints. This constraint is used to measure the similarity between the training and reconstructed images, and the new image reconstruction constraint can be expressed as:

[0156]

[0157] The visual embedding features in the new image reconstruction constraint formula are learned only from the training images (all of which are images corresponding to visible defects) and the word vectors corresponding to the training images. This may cause differences in defect recognition and classification for the images to be detected (i.e., invisible images).

[0158] In other words, due to potential discrepancies between the training images and the images to be detected, the attribute classifiers for the training and detection images may exhibit some bias in feature learning (for example, both crack defects and fracture defects involve cracks, but the degree of the cracks differs, resulting in significantly different image features). This can lead to matching errors when the visual embedding features of the image to be detected are projected into the semantic representation space and then matched with the semantic representation features (i.e., word vector features) of all training images. If the same method is used to project the visual embedding features into the semantic representation space and then match them with the word vector features for all categories, the recognition performance on the image to be detected will significantly decrease.

[0159] Therefore, in the visual semantic constraint module (such as...) Figure 1 (As shown) Semantic regularization is introduced to optimize the matching effect of the visual semantic constraint module shared between the encoder and decoder. In the visual semantic constraint module, the encoder projects the visual embedding feature representation of the image onto the semantic representation space to obtain the semantic representation (i.e., word vector) corresponding to the visual embedding feature.

[0160] However, it is worth noting that the visual embedding feature representation projection is used as input to the decoder to reconstruct the visual embedding features, resulting in a reconstructed image corresponding to the visual embedding features. To obtain a reconstructed image with the same features as the training image, all information contained in the visual features of the training image must be preserved during the visual embedding feature representation projection process. This ensures that this information can be effectively recovered by the decoder. Preserving all information contained in the visual features of the training image not only improves the accuracy of the projection but also enables the encoder to learn a more robust feature representation, better adapting to new unseen classes. The visual feature mapping generated by the encoder through the above process is reversible, allowing the autoencoder network to more effectively identify and infer defect categories not seen during the training phase in the prediction stage (i.e., the defect identification stage of the image to be detected). This improves the overall performance of the autoencoder network based on zero-shot learning in different application scenarios.

[0161] In summary, by introducing semantic regularization into the visual semantic constraint module, projecting visual embedding features onto the semantic representation space, minimizing the Euclidean distance between the visual embedding features and the semantic representation space through backpropagation, and combining this with the reconstructed image to constrain the projection process of the visual embedding features, the generalization inference ability of the autoencoder network on unknown categories can be significantly improved. The formula corresponding to the introduction of regularization is expressed as:

[0162]

[0163] Combining the new image reconstruction constraint formula and the formula corresponding to the introduction of regularization, we can obtain the overall loss function formula for the autoencoder network.

[0164] S6. If the loss function value is greater than or equal to the preset threshold, update the parameters in the autoencoder network and repeat S1-S5 until the loss function value is less than the preset threshold. Use the autoencoder network corresponding to the loss function value that is less than the preset threshold as the defect identification model.

[0165] In this step, if the loss function value is greater than or equal to the preset threshold, it means that the current defect identification model has not yet converged and the identification accuracy still needs to be improved. The parameters of the autoencoder network (such as weights and biases) need to be updated through backpropagation to reduce the loss function value. Repeat steps S1 to S5 iteratively, adjusting the autoencoder network parameters each time according to the loss function value, until the loss function value is less than the preset threshold. This means that the autoencoder network corresponding to the loss function value less than the preset threshold has converged. At this time, the obtained autoencoder network is the final defect identification model, which can efficiently and accurately identify and classify defects.

[0166] The method for defect identification of power equipment provided in this application first inputs at least one training image into the encoder of an autoencoder network to obtain a first image feature corresponding to at least one training image. The first image feature is then input into a first feature reconstructing unit of the autoencoder network to obtain a visual embedding feature corresponding to at least one training image. The visual embedding feature is then input into a second feature reconstructing unit of the autoencoder network to obtain a second image feature corresponding to at least one training image. The second image feature is then input into a decoder of the autoencoder network to obtain a reconstructed image corresponding to at least one training image. Next, the reconstructed image, the at least one training image, the visual embedding feature, and the word vector feature of the defect type are input into a loss function of the autoencoder network to obtain a loss function value. If the loss function value is greater than or equal to a preset threshold, the parameters in the autoencoder network are updated. This training process is repeated until the loss function value is less than the preset threshold. The autoencoder network with a loss function value less than the preset threshold is used as the defect identification model. This technical solution improves the robustness and accuracy of the defect identification model by continuously optimizing the parameters in the autoencoder network.

[0167] Based on the above embodiments, Figure 4 A flowchart illustrating the defect identification method for power equipment provided in this application embodiment. Figure 3 ,like Figure 4 As shown, in one possible implementation of S2, "for each training image, the training image is sequentially input into the first convolutional layer unit, the first invertible neural network unit, the first inverted residual moving block unit, the recursive transformer encoder unit, and the second convolutional layer unit in the encoder to obtain the first image feature corresponding to the training image" may include the following steps:

[0168] Step 41: Input the training image into the first convolutional layer unit of the encoder in the autoencoder network to obtain the feature map after convolution.

[0169] Step 42: Input the convolutional feature map into the first reversible neural network unit, perform specific processing and feature transformation, and output the feature map after feature processing and transformation.

[0170] The aforementioned processing and transformation can preserve the detailed information of the input data, that is, it can extract local information in the image data, thereby improving the feature extraction capability of the autoencoder network.

[0171] For example, the first reversible neural network unit is an INN module with an affine coupling layer.

[0172] Step 43: Input the feature map that has undergone feature processing and transformation into the first inverted residual moving block unit for processing. By first expanding the number of channels and then performing feature extraction, a deeper feature map is obtained.

[0173] The image feature processing described above can capture more feature information. Compared with the traditional residual block that first compresses the number of channels (conventional residual block), the first inverted residual moving block unit shows better performance in some lightweight networks, and performs feature extraction through low-dimensional mapping and layer-by-layer synthesis of features.

[0174] Step 44: Input the deeper feature map into the recursive transformer encoder unit to obtain a feature map with global contextual information representation.

[0175] The aforementioned recursive transformer encoder unit improves the Transformer module by using a gated depthwise separable convolutional network and a multidimensional depthwise separable convolutional network head attention transpose module, thereby enhancing multi-scale local-global representation learning on the image, capturing long-distance dependencies between input features through a self-attention mechanism, and enhancing the relevant features of the image.

[0176] Step 45: Input the feature map with global contextual information representation into the second convolutional layer unit for further processing and compression of features, extract higher-level image semantic feature information, and then perform channel adjustment to obtain the first image feature corresponding to the training image.

[0177] The channel adjustment can be achieved by using the activation function of a Leaky Rectified Linear Unit (Leaky ReLU).

[0178] The method for defect identification of power equipment provided in this application involves sequentially inputting each training image into a first convolutional layer unit, a first invertible neural network unit, a first inverted residual moving block unit, a recursive transformer encoder unit, and a second convolutional layer unit in an encoder to obtain image features corresponding to the training image. This technical solution sequentially inputs each training image into multiple encoder units to extract image features, enabling the autoencoder network to capture both detailed and global information in the training image. The resulting image features represent high-level information of the training image and can be effectively used in the image reconstruction process within the autoencoder network, improving the adaptability and performance of the autoencoder network in different application scenarios.

[0179] Based on the above embodiments, Figure 5 A flowchart illustrating the defect identification method for power equipment provided in this application embodiment. Figure 4 ,like Figure 5As shown, in one possible implementation of S3, "for each second image feature, the second image feature is sequentially input into the third convolutional layer unit, the recursive transformer decoder unit, the second inverted residual moving block unit, the second invertible neural network unit, and the fourth convolutional layer unit to obtain the reconstructed image corresponding to the training image" can include the following steps:

[0180] Step 51: Input the second image features corresponding to the training image into the third convolutional layer unit of the decoder in the autoencoder network to perform convolution operation, further extract details in the image or transform the features to obtain the feature map after convolution processing.

[0181] Step 52: Input the convolutional feature map into the recursive transformer decoder unit to decode the image features, recover more detailed features from the high-level features, and output the decoded feature map.

[0182] Step 53: Input the decoded feature map into the second inverted residual moving block unit for feature compression and expansion, adjust the dimension of the image features, and obtain the adjusted feature map.

[0183] Step 54: Input the adjusted feature map into the second reversible neural network unit, perform further nonlinear transformation on the feature map, and finely adjust the features through reversible operations to output the nonlinearly transformed feature map.

[0184] The second reversible neural network unit is an INN module with an affine coupling layer.

[0185] Step 55: Input the feature map after nonlinear transformation into the fourth convolutional layer unit, perform final feature extraction or reconstruction adjustment on the feature map, and output the reconstructed image corresponding to the training image.

[0186] The method for defect identification of power equipment provided in this application first inputs the second image features corresponding to the training image into the third convolutional layer unit of the decoder in the autoencoder network to perform convolution operations, further extracting details in the image or transforming the features to obtain a convolutionally processed feature map. The convolutionally processed feature map is then input into the recursive transformer decoder unit to decode the image features and recover more detailed features from high-level features, outputting a decoded feature map. The decoded feature map is then input into the second inverted residual moving block unit to compress and expand the features, adjusting the dimension of the image features to obtain an adjusted feature map. The adjusted feature map is then input into the second reversible neural network unit to perform further nonlinear transformations on the feature map, and finely adjusts the features through reversible operations, outputting a nonlinearly transformed feature map. Finally, the nonlinearly transformed feature map is input into the fourth convolutional layer unit to perform final feature extraction or reconstruction adjustments on the feature map, outputting a reconstructed image corresponding to the training image. This technical solution sequentially inputs each visual embedding feature into multiple decoder units to gradually recover the details and global information of the training image, ultimately generating a high-quality reconstructed image corresponding to the training image. The combination of various decoder units enables the autoencoder network to accurately recover the training image, capture the feature dependencies and detail changes of the training image, thereby improving the accuracy and adaptability of image reconstruction.

[0187] The following are embodiments of the apparatus described in this application, which can be used to execute the embodiments of the method described in this application. For details not disclosed in the apparatus embodiments of this application, please refer to the embodiments of the method described in this application.

[0188] Figure 6 This is a schematic diagram of the structure of a defect identification device for power equipment provided in an embodiment of this application. Figure 6 As shown, the device includes:

[0189] The acquisition module 61 is used to acquire the image of the power equipment to be detected.

[0190] The processing module 62 is used to input the image to be detected into a pre-trained defect recognition model to obtain the target defect type corresponding to the image to be detected. The defect recognition model is obtained by training an autoencoder network based on at least one training image, word vector features of at least one defect type corresponding to at least one training image, visual embedding features corresponding to at least one training image, and reconstructed image corresponding to at least one image feature.

[0191] In one possible implementation, before inputting the image to be detected into a pre-trained defect recognition model to obtain the target defect type corresponding to the image to be detected, the processing module 62 is further configured to:

[0192] Execute S1, inputting at least one training image into the encoder in the autoencoder network to obtain the first image features corresponding to at least one training image;

[0193] S2 is executed, where the first image feature corresponding to at least one training image is input into the first feature recombination unit in the autoencoder network to obtain the visual embedding feature corresponding to at least one training image.

[0194] Execute S3 to input the visual embedding features corresponding to at least one training image into the second feature recombination unit in the autoencoder network to obtain the second image features corresponding to at least one training image.

[0195] Execute S4 to input the second image features corresponding to at least one training image into the decoder in the autoencoder network to obtain the reconstructed image corresponding to at least one training image;

[0196] Execute S5, inputting the reconstructed image corresponding to at least one training image, at least one training image, the visual embedding feature corresponding to at least one training image, and the word vector feature of the defect type corresponding to at least one training image into the loss function in the autoencoder network to obtain the loss function value;

[0197] Execute S6. If the loss function value is greater than or equal to the preset threshold, update the parameters in the autoencoder network and repeat S1-S5 until the loss function value is less than the preset threshold. Use the autoencoder network corresponding to the loss function value that is less than the preset threshold as the defect identification model.

[0198] In one possible implementation, the processing module 62 is specifically used for:

[0199] The image to be detected is input into the encoder in the defect recognition model to obtain the target image features corresponding to the image to be detected.

[0200] The target image features are input into the first feature recombination unit in the defect recognition model to obtain the target visual embedding features corresponding to the target image features;

[0201] The target visual embedding features are matched with the word vector features of the defect type in the semantic representation space of the defect recognition model to obtain at least one matching result and the probability value corresponding to at least one matching result;

[0202] The defect type corresponding to the word vector feature of the matching result with the highest probability value is taken as the target defect type.

[0203] In one possible implementation, at least one training image is input into the encoder in the autoencoder network to obtain first image features corresponding to at least one training image. The processing module 62 is specifically used for:

[0204] For each training image, the training image is sequentially input into the first convolutional layer unit, the first invertible neural network unit, the first inverted residual moving block unit, the recursive transformer encoder unit, and the second convolutional layer unit in the encoder to obtain the first image feature corresponding to the training image.

[0205] In one possible implementation, the second image features corresponding to at least one training image are input into the decoder in the autoencoder network to obtain a reconstructed image corresponding to at least one training image. The processing module 62 is specifically used for:

[0206] For each second image feature, the second image feature is sequentially input into the third convolutional layer unit, the recursive transformer decoder unit, the second inverted residual moving block unit, the second invertible neural network unit, and the fourth convolutional layer unit to obtain the reconstructed image corresponding to the training image.

[0207] In one possible implementation, before inputting the image to be detected into a pre-trained defect recognition model to obtain the target defect type corresponding to the image to be detected, the processing module 62 is further configured to:

[0208] Obtain at least one training image, at least one defect type corresponding to the training image, and at least one defect type that does not appear in the training image;

[0209] Input at least one defect type corresponding to a training image and at least one defect type not appearing in the training image into a preset algorithm model to obtain word vector features corresponding to at least one defect type.

[0210] The apparatus provided in this application embodiment can be used to execute the determination method in any of the above embodiments. Its implementation principle and technical effect are similar, and will not be described again here.

[0211] It should be noted that the division of the various modules in the above device is merely a logical functional division. In actual implementation, they can be fully or partially integrated into a single physical entity, or they can be physically separated. Furthermore, these modules can be implemented entirely in software via processing element calls; they can be fully implemented in hardware; or some modules can be implemented in software via processing element calls, while others are implemented in hardware. Additionally, these modules can be fully or partially integrated together, or implemented independently. The processing element here can be an integrated circuit with signal processing capabilities. During implementation, each step of the above method or each of the above modules can be completed through the integrated logic circuits in the hardware of the processor element or through software instructions.

[0212] Figure 7 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application, such as... Figure 7As shown, the electronic device may include: a processor 71, a memory 72, and computer program instructions stored in the memory 72 and executable on the processor 71. When the processor 71 executes the computer program instructions, it implements the method provided in any of the foregoing embodiments.

[0213] Optionally, the various components of the electronic device can be connected via a system bus.

[0214] The memory 72 can be a separate memory unit or a memory unit integrated into the processor 71. The number of processors 71 can be one or more.

[0215] It should be understood that the processor 71 can be a Central Processing Unit (CPU), or other general-purpose processors 71, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor 71 can be a microprocessor 71, or any conventional processor 71. The steps of the method disclosed in this application can be directly manifested as being executed by the hardware processor 71, or being executed by a combination of hardware and software modules within the processor 71.

[0216] The system bus can be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus, etc. The system bus can be divided into address bus, data bus, control bus, etc. For ease of representation, only one thick line is used in the diagram, but this does not indicate that there is only one bus or one type of bus. Memory 72 may include random access memory (RAM) 72, and may also include non-volatile memory (NVM) 72, such as at least one disk storage device 72.

[0217] All or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a readable memory 72. When the program is executed, it performs the steps of the above method embodiments; and the aforementioned memory 72 (storage medium) includes: read-only memory 72 (ROM), RAM, flash memory 72, hard disk, solid-state hard disk, magnetic tape, floppy disk, optical disk, and any combination thereof.

[0218] The electronic device provided in this application embodiment can be used to execute the method provided in any of the above method embodiments. Its implementation principle and technical effect are similar, and will not be repeated here.

[0219] This application provides a computer-readable storage medium storing computer instructions that, when executed on a computer, cause the computer to perform the above-described method.

[0220] The aforementioned computer-readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory, electrically erasable programmable read-only memory, erasable programmable read-only memory, programmable read-only memory, read-only memory, magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0221] Optionally, a readable storage medium can be coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Alternatively, the readable storage medium can be an integral part of the processor. Both the processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components within the device.

[0222] This application also provides a computer program product, which includes a computer program stored in a computer-readable storage medium. At least one processor can read the computer program from the computer-readable storage medium, and the at least one processor can implement the above-described method when executing the computer program.

[0223] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A method for defect identification of power equipment, characterized in that, include: Acquire the image of the power equipment to be inspected; The image to be detected is input into a pre-trained defect recognition model to obtain the target defect type corresponding to the image to be detected. The defect recognition model is obtained by training an autoencoder network based on at least one training image, word vector features of the defect type corresponding to the at least one training image, visual embedding features corresponding to the at least one training image, and reconstructed images corresponding to the at least one image features.

2. The method according to claim 1, characterized in that, Before inputting the image to be detected into a pre-trained defect recognition model to obtain the target defect type corresponding to the image to be detected, the method further includes: S1, input the at least one training image into the encoder in the autoencoder network to obtain the first image feature corresponding to the at least one training image; S2, input the first image features corresponding to the at least one training image into the first feature recombination unit in the autoencoder network to obtain the visual embedding features corresponding to the at least one training image; S3, input the visual embedding features corresponding to the at least one training image into the second feature recombination unit in the autoencoder network to obtain the second image features corresponding to the at least one training image; S4, input the second image features corresponding to the at least one training image into the decoder in the autoencoder network to obtain the reconstructed image corresponding to the at least one training image; S5, input the reconstructed image corresponding to the at least one training image, the at least one training image, the visual embedding feature corresponding to the at least one training image, and the word vector feature of the defect type corresponding to the at least one training image into the loss function in the autoencoder network to obtain the loss function value; S6. If the loss function value is greater than or equal to a preset threshold, update the parameters in the autoencoder network and repeat S1-S5 until the loss function value is less than the preset threshold. Then, use the autoencoder network corresponding to the loss function value that is less than the preset threshold as the defect identification model.

3. The method according to claim 2, characterized in that, The step of inputting the image to be detected into a pre-trained defect recognition model to obtain the target defect type corresponding to the image to be detected includes: The image to be detected is input into the encoder in the defect recognition model to obtain the target image features corresponding to the image to be detected. The target image features are input into the first feature recombination unit in the defect recognition model to obtain the target visual embedding features corresponding to the target image features; The target visual embedding features are matched with the word vector features of the defect type in the semantic representation space of the defect recognition model to obtain at least one matching result and the probability value corresponding to the at least one matching result; The defect type corresponding to the word vector feature of the matching result with the highest probability value is taken as the target defect type.

4. The method according to claim 2 or 3, characterized in that, The step of inputting the at least one training image into the encoder of the autoencoder network to obtain the first image feature corresponding to the at least one training image includes: For each training image, the training image is sequentially input into the first convolutional layer unit, the first invertible neural network unit, the first inverted residual moving block unit, the recursive transformer encoder unit, and the second convolutional layer unit in the encoder to obtain the first image feature corresponding to the training image.

5. The method according to claim 2 or 3, characterized in that, The step of inputting the second image features corresponding to the at least one training image into the decoder of the autoencoder network to obtain the reconstructed image corresponding to the at least one training image includes: For each second image feature, the second image feature is sequentially input into the third convolutional layer unit, the recursive transformer decoder unit, the second inverted residual moving block unit, the second invertible neural network unit, and the fourth convolutional layer unit to obtain the reconstructed image corresponding to the training image.

6. The method according to any one of claims 1-3, characterized in that, Before inputting the image to be detected into a pre-trained defect recognition model to obtain the target defect type corresponding to the image to be detected, the method further includes: Obtain at least one training image, the defect type corresponding to the at least one training image, and the defect type that does not appear in the at least one training image; The defect type corresponding to the at least one training image and the defect type that does not appear in the at least one training image are input into a preset algorithm model to obtain the word vector features corresponding to the at least one defect type.

7. A defect identification device for power equipment, characterized in that, include: The acquisition module is used to acquire the image of the power equipment to be inspected. The processing module is used to input the image to be detected into a pre-trained defect recognition model to obtain the target defect type corresponding to the image to be detected. The defect recognition model is obtained by training an autoencoder network based on at least one training image, word vector features of the defect type corresponding to the at least one training image, visual embedding features corresponding to the at least one training image, and reconstructed images corresponding to the at least one image features.

8. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1 to 6.

10. A computer program, characterized in that, The computer program includes a computer program stored in a computer-readable storage medium, which at least one processor can read from the computer-readable storage medium, and which, when executing the computer program, can implement the method described in any one of claims 1 to 6.

Images