Power equipment rapid detection method and system based on infrared image and detr

Abstract

The application provides a power equipment rapid detection method and system based on an infrared image and a DETR, comprising the following steps: constructing an infrared image data set of power equipment and recording temperature data; firstly performing image enhancement processing and data enhancement processing on the data set to control the zoom ratio of the image; constructing an improved DETR target detection model, which combines a typical convolutional neural network and an improved transformer architecture based on temperature data, to extract feature information of the infrared image of the power equipment; training the target detection model by using the processed data set, and detecting the type and position of the power equipment in the infrared image by using the target detection model. In view of the real-time detection requirement in engineering application, the application proposes the DETR target detection model based on the improved transformer architecture, realizes the rapid detection of the infrared image of the power equipment, and assists maintenance personnel in screening a large number of operation and maintenance images.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power equipment detection, and particularly relates to a power equipment rapid detection method and system based on infrared images and DETR. BACKGROUND

[0002] With the development and progress of the country, the power system has gradually become the most important part of the energy system, and intelligentization and digitization are the keys to the construction of modern power systems. According to statistics, many electrical equipment failures in electrical equipment are caused by loose connections, poor contact, leakage current, magnetic leakage, etc. In order to discover substation faults in time and effectively eliminate potential dangers, it is necessary to regularly or irregularly inspect the substation equipment. At present, the main inspection methods of equipment are manual inspection, robot inspection and unmanned aerial vehicle inspection, and the detection technologies mainly include visible light detection, infrared detection, ultraviolet detection, etc. Among them, the infrared detection technology is widely used in the thermal fault diagnosis of substation equipment due to its high precision, non-electric operation and other advantages.

[0003] However, due to the influence of different equipment types, voltage levels and other factors, the number of images obtained by the infrared detection technology is huge, and the detection personnel need to find damaged or hidden safe equipment from these pictures and quickly eliminate equipment failures to ensure the normal operation of the substation. However, manual detection is not only time-consuming and labor-intensive, but also low in efficiency, inaccurate, and causes leakage or misjudgment. In recent years, with the development of artificial intelligence technology, intelligent classification and fault detection of substation equipment based on infrared image data have become possible. The target detection network model is used to extract the features of the infrared image visually, and the type of substation power equipment and its faults are intelligently identified and labeled. Therefore, it is necessary to detect the substation power equipment based on infrared images to assist unmanned aerial vehicles, robots and other equipment to realize automatic screening and processing, reduce the work pressure of the detection personnel, and reduce the consumption of human and material resources.

[0004] However, most of the existing infrared image detection algorithms of power equipment need to combine prior boxes for target positioning, which increases the calculation amount of the detection model and reduces the detection real-time performance, and cannot adapt to the needs of rapid detection of power equipment infrared images. SUMMARY

[0005] Therefore, the present application aims to solve the problem that the existing infrared image detection algorithm of power equipment has large calculation amount and low detection real-time performance, and cannot adapt to the needs of rapid detection of power equipment infrared images.

[0006] In order to solve the above technical problems, the present application provides the following technical solutions:

[0007] In a first aspect, the present application provides a power equipment rapid detection method based on infrared images and DETR, comprising the following steps:

[0008] An infrared image dataset of power equipment is constructed, and the average temperature of the infrared image images in the dataset is recorded, which is the average of the highest temperature and the lowest temperature of each image;

[0009] The dataset is first subjected to image enhancement processing to enhance the local features of the infrared images of the power equipment;

[0010] The dataset is further subjected to data enhancement processing to control the scaling ratio of the images;

[0011] An improved DETR target detection model is constructed, which combines a typical convolutional neural network and an improved transformer architecture to extract feature information of the infrared images of the power equipment, and the improved transformer architecture adopts a multi-head self-attention mechanism based on the average temperature to process the features;

[0012] The improved DETR target detection model is trained using the processed dataset, and the trained improved DETR target detection model is used to detect the type and position of the power equipment in the infrared images.

[0013] Further, the structure of the improved DETR target detection model includes a backbone network, a position encoding, an encoder, a decoder and a prediction head, wherein the backbone network adopts a ResNet50 network.

[0014] Further, in the improved DETR target detection model, the most important feature processing adopts an improved multi-head self-attention mechanism, and the calculation formula is as follows:

[0015]

[0016]

[0017]

[0018] In the formula, F is the feature vector extracted by the ResNet50 network, W Q , W K and W V are weight matrices, Q, K and V represent query vectors, key vectors and value vectors respectively, d is the dimension of the self-attention output matrix Z, T e is a temperature parameter corresponding to the temperature data in the infrared image of the power equipment, i is the index position in the feature vector D, and when T e =1, the formula is a standard softmax function.

[0019] Further, the image enhancement processing is performed by using the improved MSRCR algorithm, the improved MSRCR algorithm processes three channels of the infrared image color RGB respectively, and then superimposes to reconstitute a color image, and the calculation formula is as follows:

[0020]

[0021]

[0022]

[0023] wherein, R MSRCRi is the MSRCR algorithm calculation process, R MSRi is the MSR algorithm calculation process, I i (x,y) and I j (x,y) are the input images of the i and j channels respectively, N is the total number of channels 3, I' i (x,y) is the proportion of the i channel input image in the total channel image, f is a color space mapping function, C i represents a color recovery factor of the channel, which is used to adjust the proportion of the three channel colors, β is a gain constant, and α is a controlled nonlinear intensity.

[0024] Further, in the improved MSRCR algorithm, the color recovery factor C i is modified by the following formula:

[0025]

[0026] Further, the data enhancement processing adopts the MS-Mosaic algorithm, and in the MS-Mosaic algorithm, a multi-scale scaling factor is used to control the scaling ratio of the image, so as to reduce the gray area in the picture.

[0027] Further, when the improved DETR target detection model is trained, the MS-Mosaic algorithm is used for data enhancement processing on the data set in the set training round according to the set probability, and no data enhancement processing is performed in other training rounds.

[0028] In a second aspect, the present application provides a power equipment rapid detection system based on infrared images and DETR, comprising:

[0029] An image acquisition unit is configured to acquire and construct an infrared image data set of power equipment, and record the average temperature of the infrared image of the power equipment in the data set, wherein the average temperature is the average value of the highest temperature and the lowest temperature of each image.

[0030] A first preprocessing unit is configured to perform image enhancement processing on the data set first, so as to improve the local features of the infrared image of the power equipment.

[0031] a second pre-processing unit, configured to perform data augmentation on the data set again to control a zoom ratio of the image;

[0032] a model construction unit, configured to construct an improved DETR target detection model, the DETR target detection model being combined with a typical convolutional neural network and a transformer architecture to extract feature information of the infrared image of the power equipment, and the improved transformer architecture adopting a multi-head self-attention mechanism based on average temperature improvement to perform feature processing;

[0033] a target detection unit, configured to train the improved DETR target detection model by using the processed data set, and detect the type and position of the power equipment in the infrared image by using the trained improved DETR target detection model.

[0034] Further, in the model construction unit, the structure of the improved DETR target detection model includes a backbone network, a position encoding, an encoder, a decoder and a prediction head, wherein the backbone network adopts a ResNet50 network.

[0035] Further, in the improved DETR target detection model, the most important feature processing adopts an improved multi-head self-attention mechanism, and the calculation formula is as follows:

[0036]

[0037]

[0038]

[0039] In the formula, F is a feature vector extracted by the ResNet50 network, W Q , W K and W V are weight matrices, Q, K and V respectively represent a query vector, a key vector and a value vector, d is a dimension of a self-attention output matrix Z, T e is a temperature parameter corresponding to temperature data in the infrared image of the power equipment, and i is an index position in the feature vector D. When T e = 1, the formula is a standard softmax function.

[0040] In summary, the present application provides a power equipment rapid detection method and system based on infrared images and DETR, which includes constructing a power equipment infrared image dataset and recording the average temperature; the dataset is first subjected to image enhancement processing to improve the local features of the power equipment infrared image; the dataset is then subjected to data enhancement processing to control the scaling ratio of the image; an improved DETR target detection model is constructed, which combines a typical convolutional neural network and a transformer architecture improved based on the average temperature to extract feature information of the power equipment infrared image; the improved DETR target detection model is trained using the processed dataset, and the trained improved DETR target detection model is used to detect the type and position of the power equipment in the infrared image. The present application proposes a DETR target detection model based on a transformer architecture improved based on the average temperature to meet the real-time detection requirements in engineering applications, realizes rapid detection of power equipment infrared images, and assists maintenance personnel in screening massive operation and maintenance images. BRIEF DESCRIPTION OF DRAWINGS

[0041] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor based on these drawings.

[0042] Figure 1 The overall flowchart of the power equipment rapid detection method based on infrared images and DETR provided by the embodiment of the present application is shown in the figure.

[0043] Figure 2 The MS-Mosaic data enhancement effect diagram provided by the embodiment of the present application is shown in the figure.

[0044] Figure 3 The structure diagram of the improved DETR target detection model provided by the embodiment of the present application is shown in the figure.

[0045] Figure 4 The training and testing flowchart of the improved DETR target detection model provided by the embodiment of the present application is shown in the figure.

[0046] Figure 5 The test detection effect diagram of the improved DETR target detection model provided by the embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0047] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0048] This embodiment provides a rapid detection method for power equipment based on infrared images and DETR, including the following steps:

[0049] S1: Construct an infrared image dataset of power equipment, and record the average of the highest and lowest temperatures for each image.

[0050] It should be noted that infrared images of power equipment were collected by inspection personnel at different times, angles, and distances. These images, containing various types of power equipment, were then filtered to construct a power equipment infrared image dataset for training and testing the DETR target detection model. The LabelImg tool was then used to label the equipment type and location in the infrared images. Simultaneously, the average of the highest and lowest temperatures from each infrared image in the dataset was recorded as temperature data to improve the DETR target detection model.

[0051] In an optional embodiment, the labeled power equipment types include insulator strings (IS), voltage transformers (VT), current transformers (CT), arresters (AR), circuit breakers (CB), transformer bushings (TB), and oil pillows (OP).

[0052] S2: Perform image enhancement preprocessing on the infrared images of power equipment.

[0053] In an optional embodiment, the MSRCR algorithm is used for image enhancement. MSRCR introduces a color restoration factor C based on MSR (an industry-standard algorithm, which will not be described in detail here). i This is used to adjust the contrast of local areas in an infrared image. Since MSRCR is only a preprocessing algorithm for a single channel (grayscale image), it is necessary to process the three color channels (RGB) of the infrared image separately, and then superimpose them to reconstruct the color image. The specific calculation principle is as follows:

[0054]

[0055]

[0056]

[0057] In the formula, R MSRCRi For the calculation process of the MSRCR algorithm, R MSRi For the MSR algorithm calculation process, I i (x,y) and I j (x, y) represent the input images for the i-th and j-th channels, respectively, and N is the total number of channels (3). i (x,y) represents the proportion of the i-th channel input image to the total number of channels, f is the color space mapping function, and C i The color restoration factor (C) represents the channel's color ratio, used to adjust the proportions of the three channels' colors. β is the gain constant, and α is the controlled nonlinear intensity. Since the image processed by MSRCR is generally grayish, this embodiment uses the following formula to determine the color restoration factor C. i Correction:

[0058]

[0059] The improved MSRCR algorithm utilizes the modified color restoration factor C i Adjusting the proportional relationship between the three color channels in the infrared image of power equipment highlights pixels in relatively dark areas, thereby improving the local contrast of the processed infrared image of power equipment.

[0060] S3: Perform data enhancement preprocessing on infrared images of power equipment.

[0061] It should be noted that, assuming an input 416×416 image A, the scaling ratio of image A is first calculated, and then the cubic convolution interpolation (BICUBIC) algorithm is used to obtain image B. Pasting image B onto a 416×416 grayscale image results in a scaled-down image C1. This image scaling method reduces image resolution without distortion, but if the area of ​​the gray pixels filling C1 is too large (exceeding half the area of ​​the gray pixels), it will cause redundant computation in the model, increasing the inference time.

[0062] Therefore, in an optional embodiment, by introducing a multi-scale scaling factor (MS) to control the scaling ratio of the input image, excessive gray areas can be avoided as much as possible. The multi-scale scaling factor can be set to any number within [1,2], with the following effect: Figure 2As shown, the MS-Mosaic data augmentation algorithm is used to process four input infrared images of power equipment simultaneously using four different scaling factors. The four processed images are then pasted into a single image to construct a multi-scale infrared image of the power equipment, thus achieving data augmentation.

[0063] During model training, the total number of input images for MS-Mosaic data augmentation was increased from four to six, allowing six infrared images of power equipment to be embedded in a single image. Although MS-Mosaic data augmentation can enhance multi-scale target infrared images, the processed images differ significantly from the real images. Therefore, MS-Mosaic data augmentation was used for the first 30 epochs with a probability of 0.5, and then automatically disabled after 30 epochs. The number of training epochs and the probability can be set according to actual needs.

[0064] S4: Construct an improved DRTR target detection model based on average temperature

[0065] The improved DETR object detection model combines a typical convolutional neural network (CNN) with a temperature-parameter-based improved transformer architecture to directly predict the final detection result end-to-end. Its structure is as follows: Figure 3 As shown, it mainly includes a backbone network, an encoder, a decoder, and prediction heads. The backbone network can use a typical convolutional neural network (CNN) model to extract features. The encoder mainly processes the superposition of positional encoding features and CNN-extracted features. The decoder involves parallel decoding of N object queries, and the prediction heads predict bounding boxes and their categories through the FFN module.

[0066] In an optional embodiment, a ResNet50 is used as the backbone CNN network to extract infrared image feature information F of power equipment. Simultaneously, location encoding is used to add corresponding location feature information P to all features, enabling the network to distinguish power equipment in different areas.

[0067] The F and P are superimposed and input into the improved transformer architecture (encoder-decoder) of the DETR network. The most important feature processing module is the multi-head self-attention mechanism, which combines infrared image temperature data. The calculation principle is as follows:

[0068]

[0069]

[0070]

[0071] In the formula, F is the feature vector extracted by the ResNet50 network, and W Q W K and W V T is the weight matrix, optimized during random initialization training. Q, K, and V represent the query vector, key vector, and value vector, respectively. d is the dimension of the self-attention output matrix Z, where each element represents the corresponding level of attention. e Let i be the temperature parameter corresponding to the temperature data in the infrared image of the power equipment, and let i be the index position in the feature vector D. When T e When T = 1, this formula is the standard softmax function; as the temperature parameter T increases... e As the temperature increases, the range of the softmax function's output also expands. Introducing the temperature parameter can balance the model's output probability distribution, reduce overconfidence, and make the improved DETR model more reasonable and stable.

[0072] This invention designs eight self-attention modules in a multi-head self-attention module, each corresponding to a learnable weight matrix Z = [W Q W K W V Finally, these are concatenated to form [Z1, Z2, ..., Z8]. By using a decoder and a prediction head to decode and predict the enhanced features G extracted by multi-head attention, the total target category of power equipment (7 categories in this invention) and the bounding box (location of power equipment) or no bounding box (no power equipment) are predicted.

[0073] S5: Train and test the improved DETR object detection model.

[0074] The improved DETR target detection model was constructed according to... Figure 4 The process shown is for model training and testing. The specific steps are as follows:

[0075] (1) The constructed infrared image dataset of power equipment was processed by improving MSRCR to enhance the image contrast of the image samples, and was divided into training set and test set in an 8:2 ratio;

[0076] (2) Input the infrared images of power equipment in the training set into the MS-Mosaic data augmentation module to construct multi-scale infrared images of power equipment, and then input them into the improved DETR target detection model for training. Then input the infrared images in the test set into the trained improved DETR target detection model for testing to obtain its detection results.

[0077] (3) Based on the test set detection results, the accuracy and real-time performance of the method of the present invention for detecting power equipment based on infrared images are evaluated by using the detection accuracy (mean average precision, mAP) and detection speed (i.e. the total number of images detected in one second, FPS). The results are shown in Table 1.

[0078] (4) The additional infrared images of the power equipment were input into the improved DETR target detection model, which combines MS-Mosaic and improved MSRCR infrared image enhancement, to detect the type and location of the power equipment in the infrared images. The results are as follows: Figure 5 As shown.

[0079] Table 1 Test set detection results

[0080]

[0081] This embodiment provides a rapid detection method for power equipment based on infrared images and DETR. This method is designed to meet the requirements of real-time detection in engineering applications. It proposes an improved multi-head attention mechanism in the Transformer architecture based on temperature parameters, thereby constructing an improved DETR target detection model to achieve rapid detection of power equipment using infrared images and assist maintenance personnel in screening massive amounts of operation and maintenance images.

[0082] Furthermore, one embodiment of the present invention addresses the problems of low contrast and blurriness in infrared images of power equipment by proposing an improved MSRCR algorithm to enhance the original infrared image, thereby improving the quality of the infrared image itself and indirectly improving the accuracy of the improved DETR target detection model. Another embodiment of the present invention addresses the problem of decreased accuracy and robustness of the DETR target detection model due to the multi-scale nature of power equipment in infrared images by proposing a multi-scale mosaic (MS-Mosaic) data augmentation method to construct multi-scale infrared images of power equipment, thereby improving the model's robustness and multi-scale target detection capability.

[0083] Due to the large number of infrared images of power equipment in practical applications, and the low contrast, blurriness, irregular shapes, and diverse sizes of power equipment targets, this invention proposes to first enhance the original infrared images using an improved MSRCR algorithm to improve the image quality. Then, a multi-scale mosaic (MS-Mosaic) data augmentation method is employed to construct multi-scale infrared images of power equipment, directly improving the model's robustness and multi-scale target detection capability. Finally, combined with an improved multi-head attention mechanism in the Transformer architecture based on temperature parameters, an improved DETR target detection model is constructed, enabling rapid detection of power equipment infrared images and meeting the real-time detection requirements of engineering applications. The combined steps of this invention constitute a rapid detection method for power equipment based on infrared images, which helps inspection personnel process and judge massive amounts of infrared images, reducing workload and minimizing waste of human and material resources.

[0084] The above is a detailed description of an embodiment of a rapid detection method for power equipment based on infrared images and DETR according to the present invention. The following will provide a detailed description of an embodiment of a rapid detection system for power equipment based on infrared images and DETR according to the present invention.

[0085] This embodiment provides a rapid detection system for power equipment based on infrared imaging and DETR, including:

[0086] The image acquisition unit is used to acquire and construct an infrared image dataset of power equipment, and at the same time record the average temperature of the infrared images of power equipment in the dataset. The average temperature is the average of the highest and lowest temperatures of each image.

[0087] The first preprocessing unit is used to perform image enhancement processing on the dataset to improve the local features of the infrared images of power equipment.

[0088] The second preprocessing unit is used to perform data augmentation on the dataset to control the scaling ratio of the images;

[0089] The model building unit is used to build an improved DETR target detection model. The DETR target detection model combines a typical convolutional neural network and a transformer architecture to extract feature information from infrared images of power equipment. The improved transformer architecture uses a multi-head self-attention mechanism based on average temperature for feature processing.

[0090] The target detection unit is used to train an improved DETR target detection model using the processed dataset, and then uses the trained improved DETR target detection model to detect the type and location of power equipment in infrared images.

[0091] Furthermore, in the model building unit, the improved DETR object detection model structure includes: a backbone network, a positional encoder, an encoder, a decoder, and a prediction head, wherein the backbone network adopts a ResNet50 network.

[0092] Furthermore, in the improved DETR object detection model, the most important feature processing adopts an improved multi-head self-attention mechanism, and the calculation formula is as follows:

[0093]

[0094]

[0095]

[0096] In the formula, F is the feature vector extracted by the ResNet50 network, and W Q W K and W V Let T be the weight matrix, where Q, K, and V represent the query vector, key vector, and value vector, respectively, d is the dimension of the self-attention output matrix Z, and T is the weight matrix. e Let i be the temperature parameter corresponding to the temperature data in the infrared image of the power equipment, and let i be the index position in the feature vector D. When T e When =1, this formula is the standard softmax function.

[0097] It should be noted that the rapid detection system provided in this embodiment is used to implement the rapid detection method provided in the foregoing embodiments. The specific settings of each unit are based on the complete implementation of the method, and will not be repeated here.

[0098] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A rapid detection method for power equipment based on infrared images and DETR, characterized in that, Includes the following steps: A dataset of infrared images of power equipment is constructed, and the average temperature of the infrared images of power equipment in the dataset is recorded. The average temperature is the average of the highest and lowest temperatures of each image. The dataset is first subjected to image enhancement processing to improve the local features of the infrared images of power equipment; The dataset is then subjected to further data augmentation processing to control the scaling ratio of the images; An improved DETR target detection model is constructed, which combines a typical convolutional neural network and an improved transformer architecture to extract feature information from infrared images of power equipment. The improved transformer architecture uses a multi-head self-attention mechanism based on the average temperature for feature processing. The improved DETR target detection model is trained using the processed dataset, and the trained improved DETR target detection model is used to detect the type and location of power equipment in infrared images. The structure of the improved DETR target detection model includes: a backbone network, a position encoder, an encoder, a decoder, and a prediction head, wherein the backbone network adopts a ResNet50 network; In the improved DETR object detection model, the most important feature processing employs an improved multi-head self-attention mechanism, calculated as follows: ； ； ； In the formula, F is the feature vector extracted by the ResNet50 network, and W Q W K and W V Let T be the weight matrix, where Q, K, and V represent the query vector, key vector, and value vector, respectively, d is the dimension of the self-attention output matrix Z, and T is the weight matrix. e Let i be the temperature parameter corresponding to the temperature data in the infrared image of the power equipment, and let i be the index position in the feature vector D. When T e When =1, this formula is the standard softmax function.

2. The rapid detection method for power equipment based on infrared images and DETR according to claim 1, characterized in that, The image enhancement processing employs an improved MSRCR algorithm. This improved MSRCR algorithm processes the three color channels (RGB) of the infrared image separately, and then superimposes them to reconstruct the color image. The calculation formula is as follows: ； ； ； In the formula, R MSRCRi For the calculation process of the MSRCR algorithm, R MSRi This refers to the calculation process of the MSR algorithm. and Let be the input images for the i-th and j-th channels, respectively, and N be the total number of channels (3). The proportion of the i-th channel input image to the total number of channels, where f is the color space mapping function, and C... i The color recovery factor represents the channel and is used to adjust the proportion of colors in the three channels. β is the gain constant and α is the controlled nonlinear intensity.

3. The rapid detection method for power equipment based on infrared images and DETR according to claim 2, characterized in that, In the improved MSRCR algorithm, the color restoration factor C is calculated using the following formula. i Correction: 。 4. The rapid detection method for power equipment based on infrared images and DETR according to claim 1, characterized in that, The data augmentation process employs the MS-Mosaic algorithm, which uses a multi-scale scaling factor to control the scaling ratio of the image in order to reduce the gray area in the image.

5. The rapid detection method for power equipment based on infrared images and DETR according to claim 4, characterized in that, When training the improved DETR object detection model, the MS-Mosaic algorithm is used to perform data augmentation on the dataset at a set probability in a set training round, and no data augmentation is performed in other training rounds.

6. A rapid detection system for power equipment based on infrared imaging and DETR, characterized in that, include: The image acquisition unit is used to acquire and construct an infrared image dataset of power equipment, and at the same time record the average temperature of the infrared images of power equipment in the dataset. The average temperature is the average of the highest and lowest temperatures of each image. The first preprocessing unit is used to perform image enhancement processing on the dataset to improve the local features of the infrared image of the power equipment. The second preprocessing unit is used to perform data augmentation processing on the dataset to control the scaling ratio of the image; The model building unit is used to build an improved DETR target detection model. The DETR target detection model combines a typical convolutional neural network and a transformer architecture to extract feature information from infrared images of power equipment. The improved transformer architecture uses a multi-head self-attention mechanism based on the average temperature for feature processing. The target detection unit is used to train the improved DETR target detection model using the processed dataset, and to use the trained improved DETR target detection model to detect the type and location of power equipment in the infrared image. In the model building unit, the structure of the improved DETR target detection model includes: a backbone network, a position encoder, an encoder, a decoder, and a prediction head, wherein the backbone network adopts a ResNet50 network; In the improved DETR object detection model, the most important feature processing employs an improved multi-head self-attention mechanism, calculated as follows: ； ； ； In the formula, F is the feature vector extracted by the ResNet50 network, and W Q W K and W V Let T be the weight matrix, where Q, K, and V represent the query vector, key vector, and value vector, respectively, d is the dimension of the self-attention output matrix Z, and T is the weight matrix. e Let i be the temperature parameter corresponding to the temperature data in the infrared image of the power equipment, and let i be the index position in the feature vector D. When T e When =1, this formula is the standard softmax function.

Images