Transformer detection method and apparatus based on artificial intelligence and computer vision

Abstract

The application provides a transformer detection method and device based on artificial intelligence and computer vision, the method comprising: acquiring a live photo of a transformer; detecting the live photo by using a pre-trained transformer detection model to obtain a detection result of the transformer; the transformer detection model comprising a residual network for extracting features of the live photo, and a polarized self-attention (PSA) module and a fully convolutional network (FCN) module for decoding the features extracted by the residual network; and outputting a transformer abnormality prompt information when the detection result indicates that the transformer is in an abnormal state. According to the scheme, the live photo of the transformer is automatically analyzed by the transformer detection model to obtain a corresponding detection result, the detection efficiency is improved, and real-time detection of the transformer is realized.

Description

Technical Field

[0001] This invention relates to the field of transformer testing technology, and in particular to a transformer testing method and apparatus based on artificial intelligence and computer vision. Background Technology

[0002] With the continuous development of the power industry and the increasing electricity consumption year by year, the demand for stable and economical operation of transformers from various power users is also increasing. The safe and sustainable operation of transformers is crucial to the stability of the power grid and the quality of electricity for residents. When oil leakage occurs in a transformer, it not only seriously affects its appearance but also causes economic losses due to the need to shut down the transformer to eliminate the leakage. If there is a lot of oil on the ground, it may also become a fire hazard. When oil leakage causes the oil level inside the transformer to drop, it may cause live joints, switches, etc. to operate in an uninsulated state, leading to breakdown, short circuit, burnout, or even equipment explosion.

[0003] Therefore, real-time online monitoring of transformer oil leakage is essential. Currently, however, detecting transformer oil leakage mainly relies on monitoring equipment taking photos and then manually analyzing them, which is inefficient and cannot achieve real-time detection. Summary of the Invention

[0004] To address the shortcomings of the existing technologies, this invention provides a transformer detection method and apparatus based on artificial intelligence and computer vision, thereby providing a detection solution that can efficiently and in real-time detect whether a transformer is abnormal.

[0005] The first aspect of this application provides a transformer detection method based on artificial intelligence and computer vision, comprising:

[0006] Obtain real-time photos of the transformer;

[0007] The live photo is detected using a pre-trained transformer detection model to obtain the transformer detection result; wherein, the transformer detection model includes a residual network for extracting features from the live photo, and a polarized self-attention (PSA) module and a fully convolutional network (FCN) module for decoding the features extracted by the residual network.

[0008] When the detection result indicates that the transformer is in an abnormal state, a transformer abnormality prompt message is output.

[0009] Optionally, the process of training the transformer detection model includes:

[0010] Determine the transformer detection model to be trained;

[0011] Multiple sample images are obtained; wherein, the multiple sample images include negative sample images obtained by photographing a transformer in an abnormal state and positive sample images obtained by photographing a transformer in a normal state; each of the sample images corresponds to a pre-labeled manual detection result;

[0012] Each sample image is processed using the transformer detection model to be trained to obtain the model detection result for each sample image;

[0013] The model loss of the transformer detection model to be trained is determined based on the deviation between the manual detection results and the model detection results of the multiple sample images.

[0014] If the model loss does not meet the preset convergence condition, the parameters of the transformer detection model to be trained are updated according to the model loss, and the process of processing each sample image using the transformer detection model to be trained to obtain the model detection result of each sample image is returned until the model loss meets the convergence condition.

[0015] Optionally, after obtaining the real-time photos of the transformer, the method further includes:

[0016] The live photo is cropped to obtain a cropped live photo with the target size;

[0017] The process of using a pre-trained transformer detection model to detect the real-time photograph and obtain the transformer detection result includes:

[0018] The cropped real-world photograph was detected using a pre-trained transformer detection model to obtain the detection results of the transformer.

[0019] Optionally, the residual network includes five stages;

[0020] The process of using a pre-trained transformer detection model to detect the real-time photograph and obtain the transformer detection result includes:

[0021] The live photo is processed one by one using the five stages of the residual network to obtain the first feature map output by the fourth stage and the second feature map output by the fifth stage of the residual network.

[0022] The first feature map is decoded using the fully convolutional network to obtain the first detection result;

[0023] The second feature map is decoded using the polarization self-attention module to obtain the second detection result;

[0024] The first detection result and the second detection result are combined to obtain the detection result of the transformer.

[0025] Optionally, the detection result includes a marker box used to mark areas on the live photo where abnormal conditions exist.

[0026] A second aspect of this application provides a transformer detection device based on artificial intelligence and computer vision, comprising:

[0027] The acquisition unit is used to acquire real-time photos of the transformer.

[0028] The detection unit is used to detect the real-time photo using a pre-trained transformer detection model to obtain the detection result of the transformer; wherein, the transformer detection model includes a residual network for extracting features from the real-time photo, and a polarized self-attention (PSA) module and a fully convolutional network (FCN) module for decoding the features extracted by the residual network.

[0029] The output unit is used to output a transformer abnormality prompt message when the detection result indicates that the transformer is in an abnormal state.

[0030] Optionally, the apparatus further includes a training unit for training the transformer detection model, the training unit being used for:

[0031] Determine the transformer detection model to be trained;

[0032] Multiple sample images are obtained; wherein, the multiple sample images include negative sample images obtained by photographing a transformer in an abnormal state and positive sample images obtained by photographing a transformer in a normal state; each of the sample images corresponds to a pre-labeled manual detection result;

[0033] Each sample image is processed using the transformer detection model to be trained to obtain the model detection result for each sample image;

[0034] The model loss of the transformer detection model to be trained is determined based on the deviation between the manual detection results and the model detection results of the multiple sample images.

[0035] If the model loss does not meet the preset convergence condition, the parameters of the transformer detection model to be trained are updated according to the model loss, and the process of processing each sample image using the transformer detection model to be trained to obtain the model detection result of each sample image is returned until the model loss meets the convergence condition.

[0036] Optionally, the acquisition unit is further configured to:

[0037] The live photo is cropped to obtain a cropped live photo with the target size;

[0038] When the detection unit uses a pre-trained transformer detection model to detect the real-time photograph and obtain the detection result of the transformer, it is specifically used for:

[0039] The cropped real-world photograph was detected using a pre-trained transformer detection model to obtain the detection results of the transformer.

[0040] Optionally, the residual network includes five stages;

[0041] When the detection unit uses a pre-trained transformer detection model to detect the real-time photograph and obtain the detection result of the transformer, it is specifically used for:

[0042] The live photo is processed one by one using the five stages of the residual network to obtain the first feature map output by the fourth stage and the second feature map output by the fifth stage of the residual network.

[0043] The first feature map is decoded using the fully convolutional network to obtain the first detection result;

[0044] The second feature map is decoded using the polarization self-attention module to obtain the second detection result;

[0045] The first detection result and the second detection result are combined to obtain the detection result of the transformer.

[0046] Optionally, the detection result includes a marker box used to mark areas on the live photo where abnormal conditions exist.

[0047] This application provides a transformer detection method and apparatus based on artificial intelligence and computer vision. The method includes: acquiring a real-time photograph of the transformer; detecting the real-time photograph using a pre-trained transformer detection model to obtain a transformer detection result; the transformer detection model includes a residual network for extracting features from the real-time photograph, and a polarized self-attention (PSA) module and a fully convolutional network (FCN) module for decoding the features extracted by the residual network; when the detection result indicates that the transformer is in an abnormal state, outputting a transformer abnormality prompt message. This solution automatically analyzes the real-time photograph of the transformer through the transformer detection model to obtain the corresponding detection result, improving detection efficiency and realizing real-time transformer detection. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0049] Figure 1 A flowchart illustrating a transformer detection method based on artificial intelligence and computer vision, provided for embodiments of this application;

[0050] Figure 2 This is a schematic diagram of the structure of a transformer detection model provided in an embodiment of this application;

[0051] Figure 3 This is a schematic diagram of the structure of a stage in a residual network provided in an embodiment of this application;

[0052] Figure 4 A schematic diagram of the structure of a polarization self-attention module provided in an embodiment of this application;

[0053] Figure 5 This is a schematic diagram of a transformer detection device based on artificial intelligence and computer vision, provided as an embodiment of this application. Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] This application provides a transformer detection method based on artificial intelligence and computer vision. Please refer to [link to relevant documentation]. Figure 1 Here is a flowchart of the method, which may include the following steps.

[0056] S101, obtain real-time photos of the transformer.

[0057] Photos of the transformer in real time can be taken by cameras placed around the transformer.

[0058] The detection method provided in this embodiment can perform real-time detection of the transformer according to a certain detection cycle. Correspondingly, the aforementioned camera can take a real-time photo every one detection cycle. For example, the detection cycle can be set to 1 minute, then the camera placed around the transformer can take a real-time photo every minute. Each real-time photo is analyzed according to the method provided in this embodiment to determine whether the transformer is in an abnormal state when the photo is taken.

[0059] S102, using a pre-trained transformer detection model to detect real-world photos and obtain transformer detection results.

[0060] The structure of the transformer detection model provided in this embodiment can be found in [reference]. Figure 2 The transformer detection model includes a residual network for extracting features from live photos, and a polarized self-attention (PSA) module and a fully convolutional network (FCN) module for decoding the features extracted by the residual network.

[0061] The residual network used in this embodiment is an existing ResNet-50 network.

[0062] It can be seen that the residual network in the transformer detection model consists of five stages, referred to as stage one to stage five, each with the same residual structure. The residual structure of any stage in the residual network is as follows: Figure 3 As shown.

[0063] Based on the structure of the residual network, the specific execution process of step S102 may include:

[0064] The live photos were processed one by one using the five stages of the residual network to obtain the first feature map of the fourth stage output and the second feature map of the fifth stage output of the residual network.

[0065] The first feature map is decoded using a fully convolutional network to obtain the first detection result;

[0066] The second feature map is decoded using a polarization self-attention module to obtain the second detection result;

[0067] The first and second test results are combined to obtain the test results for the transformer.

[0068] In this transformer detection model, the residual network can be considered as the encoding module, while the PSA and FCN modules can be considered as the decoding modules. In other words, the transformer detection model adopts an encoding-decoding network architecture. The residual network, as the backbone, is responsible for extracting relevant features from the input image to complete the encoding process, while the decoding module is responsible for decoding the encoded features to obtain the final detection result.

[0069] Specifically, the FCN module in the decoding module is used to decode the features extracted from the first to fourth stages of the residual network, i.e., the aforementioned first feature map, to obtain the corresponding first detection result. The FCN module has a relatively simple structure, consisting of convolutional layers, normalization layers, and activation function layers. The PSA module in the decoding module is used to decode the features of the live photo extracted from the residual network after processing through the first to fifth stages, i.e., the aforementioned second feature map, to obtain the corresponding second detection result.

[0070] The structure of the PSA module can be found in [reference]. Figure 4 .from Figure 4 As can be seen, the PSA module includes two branches: a collection branch (i.e., ...) Figure 4 The branch where attention batches are collected) and the distribution branch (i.e. Figure 4 The PSA module can capture different feature maps of live photos from both the collection and distribution perspectives (the branch where attention batches are located) to generate the final result.

[0071] The process by which the PSA module decodes the second feature map to obtain the second detection result is as follows:

[0072] The PSA module accepts a spatial feature map as input data; in this embodiment, this input data is the aforementioned second feature map. The input data is then processed... Figure 4 The two branches shown generate pixel-level global attention maps for each location in the input data through several convolutional layers. Then, the global attention maps output from these two branches are aggregated to generate a new feature representation containing long-term contextual information. Convolutional layers, batch normalization layers, and activation layers are then applied for dimensionality reduction and feature fusion to obtain the global contextual features corresponding to the input data. These global contextual features are then concatenated with the input data, and the resulting new features are processed through one or more convolutional layers, batch normalization layers, and activation layers to obtain the second detection result output by the PSA module (i.e.,...). Figure 4 (output data in the file).

[0073] It should be noted that all operations in the PSA module are differentiable and can be jointly trained with other parts of the network in an end-to-end manner. Therefore, the PSA module can be flexibly attached to any feature map in the network and adaptively aggregate appropriate contextual information by predicting the contextual dependencies at each location.

[0074] In some optional embodiments, to improve detection efficiency, the target size can be preset. In this case, before executing S102, the following steps can be performed first:

[0075] Cropping the live photo yields a cropped live photo with the target size;

[0076] Correspondingly, after performing the above-mentioned trimming steps, step S102 can be adjusted as follows:

[0077] The transformer detection results are obtained by using a pre-trained transformer detection model to detect cropped real-world photos.

[0078] Cropping the live photos to the target size before inputting them into the transformer detection model can prevent the processing speed from decreasing due to excessively large photo sizes, thus ensuring high detection efficiency.

[0079] In this embodiment, the detection results can take various forms. For example, the detection results may include a marker box used to mark areas in the live photograph where abnormal conditions exist. That is, when the transformer detection model in S102 detects that the transformer is in an abnormal state, such as detecting an oil leak in the transformer's tank, the area corresponding to the abnormal state can be drawn in the live photograph using a rectangle of a specific color (e.g., a red rectangle). For example, when the abnormal state is an oil leak in the tank, the leaking area can be drawn with a red rectangle, and then the live photograph with the above-mentioned rectangle drawn can be output as the detection result.

[0080] Correspondingly, if the transformer in the real-time photo is not detected to be in an abnormal state, a prompt message indicating that the transformer is not abnormal can be directly output as the detection result.

[0081] Optionally, when the transformer is in an abnormal state, the first detection result output by the aforementioned FCN module may include areas in the real-time photos that may have abnormal conditions, such as areas where oil leakage may occur. Similarly, the second detection result output by the PSA module may also include areas in the real-time photos that may have abnormal conditions, such as areas where oil leakage may occur.

[0082] Therefore, when merging the first and second detection results, the area output by the FCN module and the area output by the PSA module can be directly overlapped, and the overlapped area can be drawn on the real-world photo to obtain the final transformer detection result, such as the oil leakage area drawn by the red rectangle above.

[0083] When the transformer is not in an abnormal state, the first detection result output by the aforementioned FCN module can be a prompt message indicating no abnormal area, and the second detection result output by the PSA module can also be a prompt message indicating no abnormal area. When merging the first and second detection results, the prompt messages output by the FCN module and the PSA module can be combined to determine the final transformer detection result, that is, to determine that the transformer in the real-time photo is not in an abnormal state.

[0084] S103: When the detection result indicates that the transformer is in an abnormal state, output transformer abnormality prompt information.

[0085] The aforementioned transformer anomaly alert information can take several forms. For example, it could be an alarm bell, in which case the alarm bell can be activated when the transformer anomaly alert information is output. Alternatively, the transformer anomaly alert information could be a text message sent to the terminal devices of the relevant maintenance personnel of the transformer. In this case, the terminal devices of the relevant maintenance personnel of the transformer in an abnormal state can be located, and then a text message indicating the transformer anomaly can be sent to these terminal devices, such as "A certain transformer is currently leaking oil, please go and check it in time."

[0086] Optionally, the process of training the transformer detection model includes:

[0087] Determine the transformer detection model to be trained;

[0088] Multiple sample images are obtained; these include negative sample images obtained from photographing a transformer in an abnormal state and positive sample images obtained from photographing a transformer in a normal state; each sample image corresponds to a pre-labeled manual detection result.

[0089] Each sample image is processed using the transformer detection model to be trained, and the model detection result for each sample image is obtained.

[0090] The model loss of the transformer detection model to be trained is determined based on the deviation between the manual detection results and the model detection results of multiple sample images.

[0091] If the model loss does not meet the preset convergence condition, update the parameters of the transformer detection model to be trained according to the model loss, and return to the step of processing each sample image with the transformer detection model to be trained to obtain the model detection result of each sample image, until the model loss meets the convergence condition.

[0092] The convergence condition can be set according to actual needs. For example, it can be set to the model loss being less than or equal to a preset convergence loss threshold, or it can be set to the number of iterations being greater than or equal to a preset maximum number of iterations. The number of iterations can be defined as the number of times the step of processing each sample image using the transformer detection model to be trained and obtaining the model detection result for each sample image is executed.

[0093] When determining the transformer detection model to be trained, one can first... Figure 2 In the model structure shown, each parameter is randomly assigned an initial value. After the assignment is completed, a transformer detection model to be trained is obtained.

[0094] The sample images mentioned above can be real-time photos of multiple transformers taken within a recent period (e.g., the last month). These multiple samples are obtained by manually labeling whether the transformers in these real-time photos are in an abnormal state (e.g., whether they are leaking oil).

[0095] The specific method for determining the model loss of the transformer detection model to be trained based on the discrepancy between the manual detection results and the model detection results of multiple sample images is not limited. As an example, the proportion of samples in all samples where the manual detection results and the model detection results are inconsistent can be counted, and this proportion can be determined as the model loss.

[0096] When updating parameters, an existing gradient backpropagation algorithm can be used to calculate the update amount of each parameter in the model based on the model loss, and then update the corresponding parameters according to the update amount.

[0097] Based on the transformer detection method based on artificial intelligence and computer vision provided in the embodiments of this application, the embodiments of this application also provide a transformer detection device based on artificial intelligence and computer vision. Please refer to [link to relevant documentation]. Figure 5 This is a schematic diagram of the structure of the device, which may include the following units.

[0098] Acquisition unit 501 is used to acquire real-time photos of the transformer.

[0099] The detection unit 502 is used to detect real-world photos using a pre-trained transformer detection model to obtain the transformer detection results.

[0100] The transformer detection model includes a residual network for extracting features from real-time photos, and a polarized self-attention (PSA) module and a fully convolutional network (FCN) module for decoding the features extracted by the residual network.

[0101] Output unit 503 is used to output transformer abnormality prompt information when the detection result indicates that the transformer is in an abnormal state.

[0102] Optionally, the device further includes a training unit 504 for training the transformer detection model, the training unit 504 being used for:

[0103] Determine the transformer detection model to be trained;

[0104] Multiple sample images are obtained; these include negative sample images obtained from photographing a transformer in an abnormal state and positive sample images obtained from photographing a transformer in a normal state; each sample image corresponds to a pre-labeled manual detection result.

[0105] Each sample image is processed using the transformer detection model to be trained, and the model detection result for each sample image is obtained.

[0106] The model loss of the transformer detection model to be trained is determined based on the deviation between the manual detection results and the model detection results of multiple sample images.

[0107] If the model loss does not meet the preset convergence condition, update the parameters of the transformer detection model to be trained according to the model loss, and return to the step of processing each sample image with the transformer detection model to be trained to obtain the model detection result of each sample image, until the model loss meets the convergence condition.

[0108] Optionally, the acquisition unit 501 is also used for:

[0109] Cropping the live photo yields a cropped live photo with the target size;

[0110] When the detection unit 502 uses a pre-trained transformer detection model to detect real-time photos and obtain the transformer detection results, it is specifically used for:

[0111] The transformer detection results are obtained by using a pre-trained transformer detection model to detect cropped real-world photos.

[0112] Optionally, the residual network includes five stages;

[0113] When the detection unit 502 uses a pre-trained transformer detection model to detect real-time photos and obtain the transformer detection results, it is specifically used for:

[0114] The live photos were processed one by one using the five stages of the residual network to obtain the first feature map of the fourth stage output and the second feature map of the fifth stage output of the residual network.

[0115] The first feature map is decoded using a fully convolutional network to obtain the first detection result;

[0116] The second feature map is decoded using a polarization self-attention module to obtain the second detection result;

[0117] The first and second test results are combined to obtain the test results for the transformer.

[0118] Optionally, the detection results include bounding boxes used to mark areas in the live photo where abnormal conditions exist.

[0119] The transformer detection device based on artificial intelligence and computer vision provided in this application can be found in the relevant steps and beneficial effects of the transformer detection method based on artificial intelligence and computer vision provided in this application, and will not be repeated here.

[0120] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0121] It should be noted that the concepts of "first" and "second" mentioned in this invention are only used to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependencies.

[0122] Those skilled in the art will be able to implement or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An artificial intelligence and computer vision based transformer detection method, characterized by, The method comprises the following steps: obtaining a live photo of a transformer; detecting the live photo by using a pre-trained transformer detection model to obtain a detection result of the transformer; wherein the transformer detection model comprises a residual network for extracting features of the live photo, and a polarized self-attention (PSA) module and a fully convolutional network (FCN) module for decoding the features extracted by the residual network; the residual network comprises five stages; the step of detecting the live photo by using the pre-trained transformer detection model to obtain the detection result of the transformer comprises the following steps: processing the live photo by using the five stages of the residual network one by one to obtain a first feature map output by the fourth stage of the residual network and a second feature map output by the fifth stage of the residual network; decoding the first feature map by using the FCN to obtain a first detection result; decoding the second feature map by using the PSA module to obtain a second detection result; when the first detection result and the second detection result both comprise a region for marking a possible abnormal state in the live photo, the regions marked by the first detection result and the second detection result are overlapped, and the overlapped region is drawn on the live photo to obtain a region of the transformer in an abnormal state, and a transformer abnormality prompt information is output.

2. The method of claim 1, wherein, The process of training the transformer detection model comprises the following steps: determining a transformer detection model to be trained; obtaining a plurality of sample images; wherein the plurality of sample images comprise negative sample images obtained by photographing a transformer in an abnormal state and positive sample images obtained by photographing a transformer in a normal state; each sample image corresponds to a pre-labeled artificial detection result; processing each sample image by using the transformer detection model to be trained to obtain a model detection result of each sample image; determining a model loss of the transformer detection model to be trained according to the deviation between the artificial detection results and the model detection results of the plurality of sample images; if the model loss does not satisfy a preset convergence condition, updating the parameters of the transformer detection model to be trained according to the model loss, and returning to the step of processing each sample image by using the transformer detection model to be trained to obtain a model detection result of each sample image until the model loss satisfies the convergence condition.

3. The method of claim 1, wherein, After the step of obtaining the live photo of the transformer, the method further comprises the following steps: cropping the live photo to obtain a cropped live photo with a target size; the step of detecting the live photo by using the pre-trained transformer detection model to obtain the detection result of the transformer comprises the following step: detecting the cropped live photo by using the pre-trained transformer detection model to obtain the detection result of the transformer.

4. The method of claim 1, wherein, The detection result comprises a marking box for marking a region of the live photo in an abnormal state.

5. An artificial intelligence and computer vision based transformer detection device, characterized in that, The method comprises the following steps: an obtaining unit is configured to obtain a live photo of a transformer; The detection unit detects the live photo by using a pre-trained transformer detection model to obtain a detection result of the transformer; wherein the transformer detection model comprises a residual network for extracting features of the live photo, and a polarized self-attention (PSA) module and a fully convolutional network (FCN) module for decoding the features extracted by the residual network; the residual network comprises five stages; When the detection unit detects the live photo by using the pre-trained transformer detection model to obtain the detection result of the transformer, the detection unit is specifically configured to: process the live photo by using the five stages of the residual network one by one to obtain a first feature map output by a fourth stage of the residual network and a second feature map output by a fifth stage of the residual network; decode the first feature map by using the FCN to obtain a first detection result; and decode the second feature map by using the PSA module to obtain a second detection result. The output unit is configured to: when the first detection result and the second detection result both comprise a region for labeling a possible abnormal state in the live photo, overlap the regions labeled by the first detection result and the second detection result, and draw the overlapped region on the live photo to obtain a region of the transformer in an abnormal state, and output a transformer abnormality prompt information.

6. The apparatus of claim 5, wherein, The device further comprises a training unit for training the transformer detection model, and the training unit is configured to: determine a transformer detection model to be trained; obtain a plurality of sample images; wherein the plurality of sample images comprise negative sample images obtained by photographing transformers in an abnormal state and positive sample images obtained by photographing transformers in a normal state; each sample image corresponds to a pre-labeled artificial detection result; process each sample image by using the transformer detection model to be trained to obtain a model detection result of each sample image; determine a model loss of the transformer detection model to be trained according to deviations of the artificial detection results and the model detection results of the plurality of sample images; if the model loss does not satisfy a preset convergence condition, update parameters of the transformer detection model to be trained according to the model loss, and return to execute the step of processing each sample image by using the transformer detection model to be trained to obtain a model detection result of each sample image until the model loss satisfies the convergence condition.

7. The apparatus of claim 5, wherein, The acquisition unit is further configured to: crop the live photo to obtain a cropped live photo with a target size; When the detection unit detects the live photo by using the pre-trained transformer detection model to obtain the detection result of the transformer, the detection unit is specifically configured to: detect the cropped live photo by using the pre-trained transformer detection model to obtain the detection result of the transformer.

8. The apparatus of claim 5, wherein, The detection result comprises a marking box for labeling a region of the live photo in an abnormal state.

Images