Substation equipment infrared spectrum target identification method based on small samples

Abstract

The invention discloses a small sample-based power transformation equipment infrared spectrum target identification method, which comprises the following steps of: analyzing temperature data from a collected power transformation equipment infrared spectrum, and eliminating abnormal values in the temperature data; mapping the temperature data to an interval of [0, 255] in combination with mapping transformation to generate a grayscale image; performing target detection on the grayscale image based on a Yolov5s algorithm, generating an infrared spectrum rectangular frame and identifying the category of the power transformation equipment; and performing instance segmentation on the infrared spectrum rectangular frame by using a Yoract + + algorithm, separating the power transformation equipment from the image background, and marking different parts of the power transformation equipment. According to the method, a method of combining target detection and instance segmentation is adopted, on the basis of ensuring the accuracy, the real-time performance of target identification is improved, the time and energy of detection personnel for post-processing the infrared spectrum are saved, and reliable operation of power transformation equipment is ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent identification of substation equipment, and in particular to a substation equipment infrared spectrum target identification method based on a small sample. Background Art

[0002] As the main carrier of power energy, substation equipment plays a relatively important role in daily production and life. The continuous operation of the equipment is an important guarantee for people's lives. However, with the influence of factors such as the service life of substation equipment and poor working environment conditions, there are certain hidden dangers of failure in the operation of substation equipment. If these problems are not discovered in time, they may cause the entire power system to be paralyzed.

[0003] An existing method for super-resolution fault identification of infrared images of electrical equipment based on a multi-scale collaborative model is proposed. A super-resolution reconstruction network for infrared images of electrical equipment based on a multi-scale collaborative model is constructed. The network is based on a generative adversarial network. By introducing a multi-scale collaborative model and a dual-channel structure, the adaptability of the super-resolution reconstruction network to infrared images is improved, and the image feature extraction effect is optimized. On the basis of realizing super-resolution reconstruction of infrared images, a super-resolution fault identification model for infrared images of electrical equipment is established in combination with a deep learning target detection method. However, this method has the problem that the fault identification accuracy of equipment such as large and medium-sized transformers and motor carbon brushes is poor, resulting in poor fault identification accuracy.

[0004] There is an existing robot-based substation switch status image recognition method. The developed robot can automatically identify the position status of circuit breakers and various forms of disconnectors. At the same time, based on the IEC 61850 standard, standard modeling is performed on the position status information of circuit breakers and disconnectors to achieve seamless communication with the substation monitoring background switching operation module, creating the prerequisite for the "one-click" sequential control switching operation of outdoor open-type smart substations and the realization of unmanned and remote operation of ultra-high voltage substations; however, this method is computationally complex, it is difficult to make full use of image information, and only targets specific categories for target recognition, so the recognition accuracy is low.

[0005] The existing power component inspection image recognition method based on the MASK LSTM-CNN model (Long Short-Term Memory Mask Area Convolutional Neural Network) combines the existing Mask R-CNN (Mask Area Convolutional Neural Network) method, uses the long short-term memory neural network, and constructs the MASK LSTM-CNN model by fusing context information. Then, combined with the specific characteristics of the power components, the optimization algorithm is further used to optimize the parameters of the model, so that the designed model can still accurately identify the power components in the field environment with more interference information; however, this method still has certain limitations. Under the premise of ensuring the detection speed, the accuracy is not high, and under the premise of improving the accuracy, the detection speed cannot be guaranteed. Summary of the invention

[0006] The purpose of this section is to summarize some aspects of the embodiments of the present invention and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the abstract of the specification and the title of the invention of this application to avoid blurring the purpose of this section, the abstract of the specification and the title of the invention, and such simplifications or omissions cannot be used to limit the scope of the present invention.

[0007] In view of the above existing problems, the present invention is proposed.

[0008] Therefore, the technical problem solved by the present invention is that the existing infrared spectrum target recognition method for substation equipment has the problems of low target recognition accuracy and slow target detection speed.

[0009] In order to solve the above technical problems, the present invention provides the following technical solutions: parsing temperature data from the collected infrared spectrum of the substation equipment, and removing abnormal values ​​in the temperature data;

[0010] The temperature data is mapped to the interval [0, 255] in combination with a mapping transformation to generate a grayscale image;

[0011] Performing target detection on the grayscale image based on the Yolov5s algorithm, generating an infrared spectrum rectangular frame and identifying the category of the substation equipment;

[0012] The infrared image rectangular frame is segmented by Yolact++ algorithm to separate the substation equipment from the image background and mark the different components of the substation equipment.

[0013] As a preferred solution of the infrared spectrum target recognition method for substation equipment based on small samples described in the present invention, the analysis includes:

[0014] De-noising and size standardizing the collected infrared spectrum of the substation equipment;

[0015] Using image processing technology, extracting temperature data reflecting the temperature distribution of the substation equipment from the infrared spectrum;

[0016] Performing statistical analysis on the temperature data to identify and remove outliers that are far from the average or have a large standard deviation;

[0017] Performing calibration processing on the temperature data;

[0018] The temperature data after calibration and removal of outliers are organized into a format that is easy to handle.

[0019] As a preferred solution of the infrared spectrum target recognition method for substation equipment based on small samples described in the present invention, the grayscale image is realized by spectrum mapping, and the temperature data is mapped to the [0, 255] interval (image RGB threshold) to generate a grayscale image, which can be expressed by the following mathematical expression formula:

[0020]

[0021] Among them, Lut(T) is the mapping result, T is the processed temperature matrix, T min is the minimum temperature of T, T max is the maximum temperature of T, This is a floor operation.

[0022] As a preferred solution of the target recognition method of substation equipment infrared spectrum based on small samples described in the present invention, target detection is performed on the grayscale image based on Yolov5s algorithm, including:

[0023] The input grayscale image is processed through the convolution layer and the pooling layer in the backbone network Backbone, the size of the feature map is gradually reduced while the number of channels is increased, and the features in the grayscale image are extracted;

[0024] The neck layer Neck performs feature fusion and upsampling operations based on the features extracted by the backbone network;

[0025] The prediction head uses the combined feature map to output the category probability, bounding box coordinates, and possible confidence scores of each potential target through pre-trained classifiers and regressors to complete target detection.

[0026] As a preferred solution of the infrared spectrum target recognition method for substation equipment based on small samples described in the present invention, generating the infrared spectrum rectangular frame includes:

[0027] A threshold is set according to the generated confidence score, and prediction results above the threshold are filtered out;

[0028] For each filtered object, draw a rectangular box on the image according to its bounding box coordinates;

[0029] The rectangular frame closely surrounds the target area, covers the target with the smallest area, reduces the overlap with the background or other targets, and removes duplicate rectangular frames through non-maximum suppression (NMS).

[0030] As a preferred solution of the infrared spectrum target recognition method for substation equipment based on small samples described in the present invention, the types of substation equipment include transformers, lightning arresters, and high-voltage cables.

[0031] As a preferred solution of the infrared spectrum target recognition method for substation equipment based on small samples described in the present invention, the Yolact++ algorithm is used for instance segmentation. The Yolact++ algorithm includes a Prediction Head branch and a mask re-evaluation branch, wherein:

[0032] Optimize the Prediction Head branch, adopt the shared parameter approach for feature maps of different sizes, and apply the non-maximum suppression filtering algorithm to remove redundant prediction boxes;

[0033] The mask re-evaluation branch uses the cropped prototype mask generated during Yolact training as input and outputs the IoU of the real mask of each category of instance segmentation, allowing the model to self-check before finalizing the mask, thereby adjusting and optimizing the shape and position of the mask to improve its match with the real target.

[0034] As a preferred solution of the infrared spectrum target recognition method for substation equipment based on small samples described in the present invention, deformable convolution is added to Backbone, and a convolution layer is added after different convolution layers to learn the position offset of each pixel in the input feature map.

[0035] As a preferred solution of the infrared spectrum target recognition method for substation equipment based on small samples described in the present invention, the target recognition result includes:

[0036] The precise location and category of different components of substation equipment and corresponding equipment.

[0037] Beneficial effects of the present invention:

[0038] 1. Use Yolov5s algorithm to identify the type of substation equipment, shield the background information in the infrared spectrum, reduce the number of training library spectrum, and improve the recognition accuracy;

[0039] 2. Based on the Yolact algorithm, deformable convolution is added, and the sampling position offset is used as a free-style sampling of learnable parameters. A convolution layer is added after different convolution layers to learn the position offset of each pixel in the input feature map, thereby improving the feature sampling ability of the backbone network for different shape instances;

[0040] 3. Based on the Yolact algorithm, the Prediction Head prediction head is optimized, and the shared parameters are adopted for feature maps of different sizes. A better aspect ratio of the anchor box is used to improve the recall of large objects. The inference speed is improved while reducing the convolutional network parameters, so that the network inference speed can achieve the effect of real-time segmentation.

[0041] 4. The method of the present invention gives priority to target detection on small samples of substation equipment, has strong scalability, and can dynamically adjust the labeled objects according to different types of substation equipment, which is conducive to the updating and iterative training of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative labor. Among them:

[0043] Figure 1 It is a flow chart of the infrared spectrum target recognition method of substation equipment based on small samples shown in the present invention;

[0044] Figure 2 Schematic diagram of the implementation method of deformable convolution shown in the present invention;

[0045] Figure 3 Schematic diagram of the network structure of the Prediction Head branch shown in the present invention. DETAILED DESCRIPTION

[0046] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, but not all of the embodiments.

[0047] Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without making any creative work should fall within the scope of protection of the present invention.

[0048] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0049] At present, the main problems of the infrared spectrum target recognition method for substation equipment are concentrated on the problems of low target recognition accuracy and slow target detection speed. In view of the above problems, an embodiment of the present invention provides a substation infrared spectrum target recognition method based on a small sample. It is based on the instance segmentation model Yolact++ and adopts a method combining target detection and instance segmentation. On the basis of ensuring the accuracy, the real-time performance of target recognition is improved, the time and energy of the detection personnel in the later processing of the infrared spectrum is saved, and the reliable operation of the substation equipment is ensured.

[0050] According to an embodiment of the present invention, Figure 1 The flowchart shown is a method for identifying infrared spectrum targets of substation equipment based on small samples, and provides the following technical solutions:

[0051] The input end of the Yolov5s model is used to process the input data and perform data inference enhancement. The Backbone part is used to extract features from the image. The Neck part combines features of different levels for processing, thereby improving the model's recognition accuracy of the target. The Prediction Head part is used to predict image features, generate bounding boxes and predict categories to achieve target detection.

[0052] Based on the Yolact algorithm, Backbone adds deformable convolution, and uses free-style sampling with sampling position offset as a learnable parameter instead of rigid grid sampling in traditional CNN, which assumes that the sampling position offset is zero as a priori assumption. A convolution layer is added after different convolution layers to learn the position offset of each pixel in the input feature map.

[0053] Based on the Yolact algorithm, the Prediction Head is improved, and the shared parameters are adopted for feature maps of different sizes. The inference speed is improved while reducing the parameters of the convolutional network, so that the network inference speed can achieve the effect of real-time segmentation. The pixels of the feature map generate a large number of anchor points. According to the calculation characteristics of IoU (IoU calculates the ratio of the intersection and union of the "predicted bounding box" and the "real bounding box", the larger the value, the closer the "predicted box" is to the "real box"), the Fast NMS screening algorithm is applied to improve the screening efficiency with a slight loss of accuracy.

[0054] Preferably, in order to address the problems of low recognition accuracy and slow target detection speed in the infrared spectrum target recognition method of substation equipment, the infrared spectrum target recognition method of substation equipment based on small samples proposed in the embodiment of the present invention can recognize substation equipment targets, contour segmentation and temperature extraction in real time.

[0055] Specifically, a method for identifying infrared spectrum targets of substation equipment based on a small sample includes the following steps:

[0056] S1. Analyze the temperature data from the collected infrared spectrum of the substation equipment and remove abnormal values ​​in the temperature data. It should be noted that the analysis process includes the following steps:

[0057] (1) Data preprocessing: De-noising the collected infrared images of substation equipment to reduce the image quality degradation caused by environmental factors (such as dust, water vapor, etc.). At the same time, the image size is standardized to ensure the consistency of input data and provide a high-quality image foundation for subsequent processing.

[0058] (2) Temperature data extraction: Using image processing techniques, such as threshold segmentation and edge detection, the temperature data reflecting the temperature distribution of the device can be accurately extracted from the infrared spectrum. This data is usually in pixels and records the temperature information of the device surface.

[0059] (3) Outlier detection and elimination: Statistical analysis is performed on the extracted temperature data to identify and eliminate outliers that are far from the mean or have a large standard deviation. These outliers may be caused by measurement errors, equipment failures, or external interference, and their existence will seriously affect the accuracy of subsequent analysis.

[0060] (4) Data calibration: In order to eliminate the deviation that may occur under different infrared sensors or different measurement conditions, the temperature data is calibrated to convert it into a unified measurement standard.

[0061] (5) Temperature data organization: The temperature data after calibration and elimination of outliers are organized into an easy-to-process format, such as a two-dimensional array or matrix, to prepare for subsequent temperature data mapping and image generation.

[0062] Preferably, in the embodiment of the present invention, the temperature data is selected for analysis instead of other data (such as spectral information, texture features, etc.), mainly because:

[0063] Direct relevance: In the monitoring of substation equipment, temperature is an important indicator of the equipment's operating status and health. Abnormal temperature distribution often indicates a fault or potential problem inside the equipment.

[0064] Easy to measure: Infrared thermal imagers can quickly measure the temperature distribution on the surface of equipment without contact, without stopping or disassembling the equipment. It has the advantages of easy operation and accurate measurement.

[0065] Intuitive: Temperature data is presented in the form of images, which is intuitive and easy to understand, allowing technicians to quickly determine the operating status of the equipment.

[0066] S2. Combine the mapping transformation to map the temperature data to the interval [0, 255] to generate a grayscale image. It should be noted that the grayscale image is realized by atlas mapping, and the temperature data is mapped to the interval [0, 255] (image RGB threshold) to generate a grayscale image, which can be expressed by the following mathematical expression formula:

[0067]

[0068] Among them, Lut(T) is the mapping result, T is the processed temperature matrix, T min is the minimum temperature of T, T max is the maximum temperature of T, This is a floor operation.

[0069] S3, based on the Yolov5s algorithm, perform target detection on the grayscale image, generate an infrared spectrum rectangular frame and identify the type of substation equipment. Among them, what needs to be explained in this step is:

[0070] The input grayscale image is processed through the convolution layer and pooling layer in the backbone network, gradually reducing the size of the feature map while increasing the number of channels to extract features from the grayscale image;

[0071] The neck layer Neck performs feature fusion and upsampling operations based on the features extracted by the backbone network to provide more advanced semantic information and the ability to adapt to images of different scales;

[0072] The prediction head uses the combined feature map to output the category probability, bounding box coordinates, and possible confidence scores of each potential target through pre-trained classifiers and regressors to complete target detection.

[0073] Furthermore, an infrared spectrum rectangular frame is generated, including:

[0074] After the model is trained, it is applied to the infrared images of the substation equipment to be identified;

[0075] The model predicts each potential target in the image and outputs the class probability, bounding box coordinates (usually the upper left and lower right corner coordinates of the rectangular box), and confidence score of the substation.

[0076] A threshold is set based on the generated confidence score to filter out prediction results that are higher than the threshold;

[0077] For each filtered object, draw a rectangular box on the image according to its bounding box coordinates;

[0078] The rectangular box closely surrounds the target area, covers the target with the smallest area, reduces the overlap with the background or other targets, and removes duplicate rectangular boxes through non-maximum suppression (NMS).

[0079] As an example, the category of power transformation equipment includes transformers, lightning arresters, and high-voltage cables.

[0080] It should be noted that the overall structure of the Yolov5s model is similar to that of Yolov4, and is divided into four parts: Input, Backbone, Neck, and Prediction Head. The Input part includes data enhancement, adaptive anchor box calculation, and adaptive image scaling; Backbone is the main structure of Yolov5s, which contains Focus structure and CSP structure; Neck is composed of FPN and PAN structures, and uses the information extracted in the previous part to improve the feature extraction capability; Prediction Head uses GIOU_Loss as the loss function.

[0081] Preferably, the embodiment of the present invention adopts Yolov5s algorithm to perform target detection, identify the type of substation equipment, shield useless background information in infrared spectra, reduce the number of training library spectra, and thus improve the accuracy of target detection.

[0082] S4. Use the Yolact++ algorithm to perform instance segmentation on the infrared image rectangle frame, separate the substation equipment from the image background, and mark the different components of the substation equipment. It should be noted that this step uses the Yolact++ algorithm for instance segmentation. The Yolact++ algorithm includes a Prediction Head branch and a mask re-evaluation branch, where:

[0083] Optimize Prediction Head and adopt a shared parameter approach for feature maps of different sizes. Feature maps of different levels can use the same convolution kernel or other parameters when processing, which reduces the convolution network parameters and improves the reasoning speed. Apply the Fast NMS (non-maximum suppression) screening algorithm to remove redundant prediction boxes, which significantly improves the screening efficiency while maintaining high accuracy, thereby speeding up the reasoning speed of the model.

[0084] The mask re-evaluation branch uses the cropped prototype mask (basic, instance-category-independent mask template) generated during Yolact training as input, and outputs the IoU (the ratio of the intersection and union of the "predicted bounding box" and the "real bounding box") of the GT-mask (real mask) for each category of instance segmentation, allowing the model to self-check before finalizing the mask, thereby possibly adjusting or optimizing the shape and position of the mask to improve its match with the real target.

[0085] As an example, the Mask Re-Scoring branch uses the cropped prototype mask generated by Yolact (without thresholding) as input and outputs the IoU of the GT-mask corresponding to each category. The structure is composed of 5 convolution layers of 69×69×8, 35×35×16, 18×18×32, 9×9×64 and 5×5×128 and 1 pooling layer; thus, target recognition based on the infrared spectrum of substation equipment with high accuracy based on small samples is achieved. On this basis, the temperature of the substation equipment area can be read through the interface of the infrared spectrum processing software or by using the fitting curve of the corresponding relationship between temperature data and temperature value, the equipment operating status can be judged and fault diagnosis can be performed.

[0086] Furthermore, the target recognition results include the precise location and category of substation equipment (such as transformers, switchgear, etc.) and corresponding components. The operating status of the equipment is determined based on the temperature data of the substation equipment area and the normal operating temperature range or threshold of the equipment. If the temperature exceeds the normal range, it may indicate that the equipment is overheating, faulty or other abnormal conditions.

[0087] In an optional implementation, a deformable convolution is added to Backbone, and a convolution layer is added after different convolution layers to learn the position offset of each pixel in the input feature map.

[0088] As an example, substation equipment (transformers, lightning arresters, high-voltage cables, etc.), the precise location and category of different components of the corresponding equipment.

[0089] In an optional embodiment, deformable convolution (DCNs) uses free-style sampling of sampling position offset as a learnable parameter, replacing the rigid grid sampling in traditional CNN with zero sampling position offset as a priori assumption, thereby improving the accuracy of the image segmentation model. The implementation of 3×3 deformable convolution is as follows: Figure 2 shown.

[0090] As an example, deformable convolution is to add a convolution layer after different convolution layers to learn the position offset ΔP of each pixel in the input feature map. In traditional CNN, the position offset of p k The calculation formula for the convolution result value is as follows:

[0091]

[0092] Among them, y(p k ) is the convolution result, R is the rigid grid receptive field position set, p i is an enumeration of the positions listed in R, w(i) is the information corresponding to the i-th pixel in the input feature map, p k is the location where the convolution is performed.

[0093] Deformable convolution on p k The value at is calculated as follows:

[0094]

[0095] Among them, y(p k ) is the convolution result, R is the rigid grid receptive field position set, p i is an enumeration of the positions listed in R, w(i) is the information corresponding to the i-th pixel in the input feature map, p k is the position where the convolution is performed, ΔP is the position offset of each pixel, Δp i ∈ΔP, due to Δp i It is usually not an integer, so (p k +p i +Δp i ) needs to be calculated through bilinear interpolation.

[0096] In an alternative embodiment, see Figure 3 , the specific implementation methods for optimizing Prediction Head include: Figure 3 Here, W×H×· represents the size of the anchor box. The structure has three branches: Class Layer, Box Layer, and Mask Layer. It uses shared parameters for feature maps of different sizes, which reduces the parameters of the convolutional network and improves the reasoning speed, so that the network reasoning speed can achieve the effect of real-time segmentation. Three anchor boxes are generated for each pixel of the feature map. Since a large number of anchor boxes are generated, the Fast NMS screening algorithm is used according to the calculation characteristics of IoU to improve the screening efficiency with a slight loss of accuracy. Since Yolact++ does not use the fully connected (FC) network and splicing operation, the reasoning speed of the Yolact++ network has been greatly improved compared with Mask-RCNN.

[0097] Preferably, the substation equipment infrared spectrum target recognition method based on small samples provided in the embodiment of the present invention can identify substation equipment targets, contour segmentation and temperature extraction in real time. The algorithm can be further applied to systems with high real-time requirements such as drones and substation patrol robots. On the basis of the mature algorithm, an artificial intelligence platform is established to realize the full process of automatic uploading, automatic recognition and intelligent diagnosis of on-site infrared spectra, which is conducive to improving the intelligence level of substation equipment status detection and reducing the workload of on-site personnel.

[0098] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for identifying infrared spectrum targets of substation equipment based on small samples, characterized in that: include: Analyze temperature data from the collected infrared spectrum of the substation equipment and remove abnormal values ​​in the temperature data; The temperature data is mapped to the interval [0, 255] in combination with a mapping transformation to generate a grayscale image; Performing target detection on the grayscale image based on the Yolov5s algorithm, generating an infrared spectrum rectangular frame and identifying the category of the substation equipment; The infrared image rectangular frame is segmented by Yolact++ algorithm to separate the substation equipment from the image background and mark the different components of the substation equipment.

2. The infrared spectrum target recognition method for substation equipment based on small samples according to claim 1 is characterized in that: The analysis includes: De-noising and size standardizing the collected infrared spectrum of the substation equipment; Using image processing technology, extracting temperature data reflecting the temperature distribution of the substation equipment from the infrared spectrum; Performing statistical analysis on the temperature data to identify and remove outliers that are far from the average or have a large standard deviation; Performing calibration processing on the temperature data; The temperature data after calibration and removal of outliers are organized into a format that is easy to handle.

3. The method for identifying substation equipment infrared spectrum targets based on small samples according to claim 1 is characterized in that: The grayscale image is realized by atlas mapping, mapping the temperature data to the [0, 255] interval (image RGB threshold) to generate a grayscale image, which can be expressed by the following mathematical expression formula: Among them, Lut(T) is the mapping result, T is the processed temperature matrix, T min is the minimum temperature of T, T max is the maximum temperature of T, This is a floor operation.

4. The method for identifying substation equipment infrared spectrum targets based on small samples according to claim 1 is characterized in that: Performing target detection on the grayscale image based on the Yolov5s algorithm includes: The input grayscale image is processed through the convolution layer and the pooling layer in the backbone network Backbone, the size of the feature map is gradually reduced while the number of channels is increased, and the features in the grayscale image are extracted; The neck layer Neck performs feature fusion and upsampling operations based on the features extracted by the backbone network; The prediction head uses the combined feature map to output the category probability, bounding box coordinates, and possible confidence scores of each potential target through pre-trained classifiers and regressors to complete target detection.

5. The method for identifying substation equipment infrared spectrum targets based on small samples according to claim 1 is characterized in that: Generating the infrared spectrum rectangular frame includes: A threshold is set according to the generated confidence score, and prediction results above the threshold are filtered out; For each filtered object, draw a rectangular box on the image according to its bounding box coordinates; The rectangular frame closely surrounds the target area, covers the target with the smallest area, reduces the overlap with the background or other targets, and removes duplicate rectangular frames through non-maximum suppression (NMS).

6. The method for identifying substation equipment infrared spectrum targets based on small samples according to claim 1 or 4, characterized in that: The types of substation equipment include transformers, lightning arresters, and high-voltage cables.

7. The method for identifying substation equipment infrared spectrum targets based on small samples according to claim 1 is characterized in that: The Yolact++ algorithm is used for instance segmentation. The Yolact++ algorithm includes a Prediction Head branch and a mask re-evaluation branch, wherein: Optimize the Prediction Head branch, adopt the shared parameter approach for feature maps of different sizes, and apply the non-maximum suppression filtering algorithm to remove redundant prediction boxes; The mask re-evaluation branch uses the cropped prototype mask generated during Yolact training as input and outputs the IoU of the real mask of each category of instance segmentation, allowing the model to self-check before finalizing the mask, thereby adjusting and optimizing the shape and position of the mask to improve its match with the real target.

8. The method for identifying substation equipment infrared spectrum targets based on small samples according to claim 7 is characterized in that: Add deformable convolution to Backbone, and add a convolution layer after different convolution layers to learn the position offset of each pixel in the input feature map.

9. The method for identifying substation equipment infrared spectrum targets based on small samples according to claim 1 is characterized in that: The target recognition result includes: The precise location and category of different components of substation equipment and corresponding equipment.

Images