Switch cabinet state identification method based on small target perception

Abstract

The invention belongs to the technical field of computer vision. The invention provides a switch cabinet state identification method based on small target perception. According to the embodiment of the invention, the SPD non-stepping convolution is introduced into the backbone network of the YOLOv8 to replace the stepping convolution in the original backbone network, so that the feature extraction capability of the model is improved, and the loss of small target features is reduced. An SPPF module in a backbone network of YOLOv8 is improved into a multi-scale feature extraction module MSFEAC, so that multi-scale perception of a target can be realized, a receptive field is expanded, and global information of a small target of a switch cabinet indicator lamp is better reserved. And a multi-scale feature extraction module and a small target detection layer are added, so that the network can pay attention to more detail information, thereby improving the recognition effect of a small target, namely the state of a transformer substation drainage wire bolt and each indicator lamp of a switch cabinet, and being more suitable for an intelligent transformer substation application scene.

Description

Technical Field

[0001] The disclosed embodiments relate to the field of computer vision technology, and in particular to a switch cabinet state recognition method based on small target perception. Background Art

[0002] With the rapid development of artificial intelligence and Internet of Things technologies, more and more power equipment and systems (such as switch cabinets, etc.) need to be efficiently monitored and managed. In these systems, the equipment status indicator light is a key source of information, usually indicating the operating status of the equipment through colors such as red, yellow, and green. However, due to the wide variety of status lights that need to be monitored on the substation switch cabinet and their dense layout, manual detection is inefficient and it is difficult to ensure the accuracy and real-time performance of the detection. Therefore, how to accurately capture the indicator light status and perform real-time analysis and recognition through algorithms is an important and challenging topic in the field of computer vision.

[0003] Early research methods mainly focused on extracting the features of indicator lights using grayscale images and color gamut conversion technology. These methods extract the color and on / off status of indicator lights by converting color images into grayscale images or specific color gamut spaces. These feature extraction methods have improved the accuracy of indicator light status recognition to a certain extent, but they still have certain limitations in complex environments.

[0004] With the rapid development of deep learning in the field of computer vision, deep learning technology has gradually been applied to the detection task of industrial targets. At present, the YOLO (You Only LookOnce) series of algorithms are widely used in the field of target detection. Since the YOLOv5 series, the algorithm can simultaneously complete the two major tasks of target positioning and recognition. Compared with traditional algorithms, YOLO has significant advantages in speed, accuracy, model size and ease of use, but there are still certain challenges in small target detection. The indicator lights of substation switch cabinets are small in size when they are identified. The newer YOLOv8 network has insufficient feature extraction, and the neural network cannot learn effective information, resulting in the inability to accurately identify the status of the substation switch cabinet.

[0005] Therefore, it is necessary to improve one or more problems existing in the above-mentioned related technical solutions.

[0006] It should be noted that this section is intended to provide background or context for the technical solutions of the present disclosure stated in the claims. The description herein is not admitted to be prior art by virtue of being included in this section. Summary of the invention

[0007] The purpose of the embodiments of the present disclosure is to provide a switch cabinet state recognition method based on small target perception, thereby overcoming one or more problems caused by the limitations and defects of related technologies at least to a certain extent.

[0008] According to an embodiment of the present disclosure, a switch cabinet state recognition method based on small target perception is provided, the method comprising: Acquire a substation switch cabinet image, and preprocess the substation switch cabinet image; Annotating the preprocessed substation switch cabinet image, and dividing the annotated substation switch cabinet image into a training set and a test set according to a preset ratio; Constructing a YOLOv8 model; wherein the YOLOv8 model includes an input layer, a backbone network, a neck network and a detection head; The YOLOv8 model is trained and tested using the training set and the test set respectively to obtain the trained YOLOv8 model; The YOLOv8 model is used to detect the state of the substation switch cabinet to obtain the switch cabinet state recognition result.

[0009] Furthermore, the step of acquiring the substation switch cabinet image and preprocessing the substation switch cabinet image includes: The substation switch cabinet images corresponding to various working states of the substation switch cabinet are obtained, and random scaling, random cropping, contrast adjustment and data enhancement processing are sequentially performed on the substation switch cabinet images.

[0010] Furthermore, the step of annotating the preprocessed substation switch cabinet image and dividing the annotated substation switch cabinet image into a training set and a training set according to a preset ratio includes: The pre-processed substation switch cabinet image is labeled using the Labelme tool, and the labeled information is saved as a txt file, and the data format is set to VOC format; The labeled substation switch cabinet images are divided into the training set and the test set in a ratio of 8:2.

[0011] Furthermore, the backbone network includes: 5 SPD-Conv modules, 4 C2f modules and 1 MSFE_AC module; among them, One of the SPD-Conv modules, one of the SPD-Conv modules, one of the C2f modules, one of the SPD-Conv modules, one of the C2f modules, the SPD-Conv module, one of the C2f modules, one of the SPD-Conv modules, one of the C2f modules, one of the SPD-Conv modules, one of the C2f modules and one of the MFSE_AC modules are connected in series in sequence.

[0012] Furthermore, the SPD-Conv module includes one SPD layer and one non-strided convolution layer connected in sequence; wherein, The SPD layer slices the initial feature map of the substation switch cabinet image using a scaling factor and performs 2-fold downsampling to obtain four sub-feature maps of the same shape; Connecting the four sub-feature maps along the channel dimension to obtain a feature map that is half the size in space and four times the size in dimension; The feature map is input into the non-stride convolution to retain the discriminative feature information of the substation switch cabinet image.

[0013] Furthermore, the MSFE_AC module includes: 3 dilated convolutions, 3 Conv modules, 1 pool layer, 1 concat module, 1 Upsample layer and 1 CBAM attention module; among them, The initial feature map is input in parallel into one of the Conv modules, three of the dilated convolutions of different scales, and one of the pool layers, wherein the Conv module compresses the initial feature map, the dilated convolutions expand the receptive field of the initial feature map, and the pool layer performs a single pooling layer operation on the initial feature map; The output of the pool layer is input to the Upsample layer for processing; Input the output of the Conv module, the three dilated convolutions and the Upsample layer into the concat module for feature fusion to obtain a fused feature map; Using one of the Conv modules to compress the fused feature map, and adding the compressed fused feature map to the initial feature map; The added fused feature map and the initial feature map are input into the CBAM attention module to obtain the final feature map.

[0014] Furthermore, the CBAM attention module includes a CAM submodule and a SAM submodule; wherein, The input fusion feature map F (H×W×C) is subjected to global maximum pooling and global average pooling based on width W and height H respectively to obtain two 1×1×C first feature maps; Sending the first feature maps to a two-layer MLP network for processing; wherein the number of neurons in the first layer of the MLP network is C / r, the activation function is Relu, and the number of neurons in the second layer of the MLP network is C; The features output by the MLP network are added based on element-wise operation, and then activated by sigmoid operation to generate the final Mc; Perform a multiplication operation on the Mc and the fused feature map F to generate a second feature map; Input the second feature into the input of the SAM submodule, perform a space-based global maximum pooling and a global average pooling, and obtain two H×W×1 third feature maps; The two third feature maps are concatenated based on the channel, and the dimension is reduced to one channel through a 7×7 convolution operation, and Ms is generated through a sigmoid layer; The Ms is multiplied by the second feature to obtain the final feature map.

[0015] The technical solution provided by the embodiments of the present disclosure may have the following beneficial effects: In the embodiments of the present disclosure, through the above-mentioned switch cabinet state recognition method based on small target perception, on the one hand, SPD non-stride convolution, i.e. SPD-Conv, is introduced into the backbone network of YOLOv8 to replace the stride convolution in the original backbone network, thereby improving the feature extraction capability of the model and reducing the loss of small target features. The SPPF module in the backbone network of YOLOv8 is improved to a multi-scale feature extraction module MSFE_AC, which can realize multi-scale perception of the target, expand the receptive field, and better retain the global information of the small target of the switch cabinet indicator light. Adding a multi-scale feature extraction module and a small target detection layer enables the network to pay attention to more detailed information, thereby improving the recognition effect of the small target of the status of the substation drain line bolts and the indicator lights of the switch cabinet, which is more suitable for the application scenario of the smart substation. On the other hand, there is a significant improvement in the level of intelligent control, which can be widely used in substation switch cabinet state detection, especially for smart substations. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The accompanying drawings herein are incorporated into the specification and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the specification are used to explain the principles of the present disclosure. Obviously, the accompanying drawings described below are only some embodiments of the present disclosure, and for ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without creative work.

[0017] Figure 1 A step diagram of a switch cabinet state recognition method based on small target perception in an exemplary embodiment of the present disclosure is shown; Figure 2 A schematic diagram showing the structure of the YOLOv8 model of the present application in an exemplary embodiment of the present disclosure is shown; Figure 3 A schematic diagram showing a CBS module in an exemplary embodiment of the present disclosure; Figure 4 A schematic diagram showing an SPD-Conv (s=1) module in an exemplary embodiment of the present disclosure; Figure 5 A schematic diagram showing an SPPF module in an exemplary embodiment of the present disclosure; Figure 6 A schematic diagram showing a MSFE_AC module in an exemplary embodiment of the present disclosure; Figure 7 A schematic diagram of a CBAM attention module in an exemplary embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0018] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in a variety of forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that the disclosure will be more comprehensive and complete and to fully convey the concepts of the example embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0019] In addition, the accompanying drawings are only schematic illustrations of the embodiments of the present disclosure and are not necessarily drawn to scale. The same reference numerals in the figures represent the same or similar parts, and their repeated descriptions will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities.

[0020] This example embodiment provides a switch cabinet state recognition method based on small target perception. Figure 1 As shown in , the switch cabinet state recognition method based on small target perception may include: step S101 to step S105.

[0021] Step S101: acquiring a substation switch cabinet image, and preprocessing the substation switch cabinet image; Step S102: annotating the preprocessed substation switch cabinet image, and dividing the annotated substation switch cabinet image into a training set and a test set according to a preset ratio; Step S103: constructing a YOLOv8 model; wherein the YOLOv8 model includes an input layer, a backbone network, a neck network and a detection head; Step S104: training and testing the YOLOv8 model using the training set and the test set respectively to obtain the trained YOLOv8 model; Step S105: Use the YOLOv8 model to detect the state of the substation switch cabinet to obtain a switch cabinet state recognition result.

[0022] Through the above switch cabinet state recognition method based on small target perception, on the one hand, SPD non-strided convolution, namely SPD-Conv, is introduced into the backbone network of YOLOv8 to replace the strided convolution in the original backbone network, which improves the feature extraction ability of the model and reduces the loss of small target features. The SPPF module in the backbone network of YOLOv8 is improved to a multi-scale feature extraction module MSFE_AC, which can realize multi-scale perception of the target, expand the receptive field, and better retain the global information of the small target of the switch cabinet indicator light. Adding a multi-scale feature extraction module and a small target detection layer enables the network to pay attention to more detailed information, thereby improving the recognition effect of the small target of the substation drainage line bolts and the status of each indicator light of the switch cabinet, which is more suitable for the application scenario of smart substations. On the other hand, there is a significant improvement in the level of intelligent control, which can be widely used in substation switch cabinet state detection, especially for smart substations.

[0023] Next, we will refer to Figures 1 to 7 Each step of the switch cabinet state identification method based on small target perception in this example implementation is described in more detail.

[0024] In step S101, a camera is used to collect different images of the substation switch cabinet. Since the collected data is limited, the data set is expanded by random scaling, random cropping, adjusting contrast, etc.

[0025] In step S102, the dataset images are annotated using the Labelme annotation tool and saved as VOC format files recognizable by the YOLO target detection algorithm. The dataset is divided into a training set and a test set in a ratio of 8:2.

[0026] In step S103 and step S104, the YOLOv8 model is mainly composed of four parts, including the input, the backbone network, the neck, and the detection head. The backbone network consists of the CBS module, the C2f module, and the SPPF module. The SPD non-strided convolution layer, namely SPD-Conv (s=1), is introduced into the backbone network to replace the original CBS module, which solves the problem of loss of detail information in the original backbone network, effectively increases the network's ability to extract features of small targets of the switch cabinet indicator light status, and improves the problem of target feature loss.

[0027] The proposed multi-scale feature extraction module MFSE_AC replaces the SPPF in the backbone network. The original SPPF module uses multiple pooling operations. By extracting the maximum value in each pooling window, it can better retain the part with higher feature significance, but it discards the non-maximum information, resulting in the neglect of detail information. The improved MFSE_AC module applies dilated convolution and average pooling operations of different scales in parallel, and adds residual links and CBAM attention mechanisms. MSFE_AC can capture contextual information of different levels in the image and realize multi-scale perception of the target. In addition, using different dilation rates for dilated convolution helps to expand the receptive field, enabling the network to better understand broader contextual information, reducing the loss of small target features and improving the overall performance of the model.

[0028] The features of the backbone network that pass through the C2f module for the first time are fused with the deep features in the neck network, and further passed through the C2f module to obtain a detection head for small targets. A small target detection layer is added to the original neck network structure of YOLOv8. By reducing the size of the feature map, the network can focus on more detailed information and improve the detection effect of small targets.

[0029] YOLOv8 model selection and parameter initialization: Select YOLOv8 by modifying the configuration file. The specific setting parameters are as follows: set the network depth to 1.0, the network width to 1.0; select CIOU as the loss function; the weight decay coefficient is 0.0005; the Epoch is set to 300; the Mosaic enhancement index is set to (0.5, 1.5), turn off Mixup enhancement, and turn off Mosaic data enhancement in the last 80 rounds of training.

[0030] In step S105, after the training of the model is completed, the trained model is used to detect the indicator lights of the substation switch cabinet to determine the status of each indicator light.

[0031] In a specific embodiment, a camera is used to collect images of a substation switch cabinet, and the red indicator light, yellow indicator light, green indicator light, white indicator light, circuit on / off indicator light, energy storage status indicator light, and remote local status in the switch cabinet are selected as research objects. The initial images obtained are screened and images with low quality are removed, and finally 50 images are obtained. Due to the insufficient number of data sets, this application uses random scaling, random cropping, and contrast adjustment to enhance the data, and the total number is expanded to 300.

[0032] All images in the dataset are labeled using the Labelme tool, and the annotation information is saved as a txt file. The dataset format is set to VOC format, and the dataset is divided into a training set: test set = 8:2 ratio.

[0033] The improved YOLOv8 model is mainly composed of four parts, including the input, the backbone network, the neck, and the detection head. Figure 2 This is the YOLOv8 structure diagram and specific structural principle diagram of this application. The network structure is mainly composed of SPD-Conv (s=1) module, C2f module, MSFE_AC module, and PAN-FPN module. At the input end of the network, the data set is enhanced in combination with Mosaic and Mixup to prevent serious overfitting. Its backbone network consists of SPD-Conv (s=1) module, C2f module, and MSFE_AC module. The neck network is connected to the MSFE_AC module and consists of a PAN-FPN module. The detection layer divides regression and classification into two parts through a decoupling head, which is integrated during prediction, while improving the convergence speed and accuracy of the algorithm. Compared with other YOLOv8 series algorithms, the improved YOLOv8 model algorithm has higher accuracy and higher model generalization ability.

[0034] For the YOLOv8 initial model, the CBS module in its backbone network is as follows Figure 3 As shown in the figure, it consists of three parts: Conv2D, BatchNorm2D, and SiLU activation function. The two-dimensional convolution Conv2D uses strided convolution, that is, the convolution step length stride=2. In most scenes, there is a lot of redundant pixel information, which can be easily skipped by strided convolution, and the model can still learn features well; but for the small target of the substation switch cabinet indicator light, the redundant information is no longer valid. The use of strided convolution will cause the model to lose fine-grained information of feature extraction, which is not conducive to small target detection. Introducing SPD-Conv (s=1) to replace the CBS module can improve the problem of target feature information loss.

[0035] This application introduces SPD-Conv (s=1) to replace the original CBS module in the backbone network, solving the problem of detail information loss in the original strided convolution, enabling the model to better retain feature details and improve the performance of small target detection. SPD-Conv (s=1) consists of a space-to-depth (SPD) layer and a non-strided convolution layer. The basic principle is: the SPD layer uses a scale factor of scale=2 to Figure 4 The features shown in (a) are sliced ​​and downsampled by 2 times, as shown in Figure 4 As shown in (b), the shapes of the four sub-feature maps obtained are , then connect the four sub-feature maps along the channel dimension, such as Figure 4 As shown in (c), the final feature map is reduced by half in space and increased four times in dimension. After the SPD layer, it undergoes a non-step convolution to retain all discriminative feature information as much as possible. Through the above steps, SPD-Conv (s=1) effectively solves the problem of detail information loss in the original backbone network, increases the network's feature extraction capability for small targets of the switch cabinet indicator light status, and improves the problem of target feature loss.

[0036] For the YOLOv8 initial model, the SPPF module in the backbone network connects the backbone network and the neck network to fuse multi-scale features. Figure 5 As shown, it is a schematic diagram of the SPPF module. In the SPPF module, a compression operation is first completed on the input feature map through a Conv module, and then three identical maximum pooling operations MaxPool2d are performed in series. Finally, the concat feature fusion is performed on the four dimensions after the three pooling operations and before the pooling. After the fusion, a Conv module is performed to upgrade the dimension to obtain the final feature. This module uses the maximum pooling operation multiple times: The advantage of maximum pooling is that it can effectively extract important features from the input sequence and reduce the length of the input sequence. However, maximum pooling will also ignore some minor information, thereby affecting the performance of the small target detection model. Therefore, this application introduces a multi-scale feature extraction module MSFE_AC, such as Figure 6 As shown in Figure 3, this module reduces the loss of small target features and improves the overall performance of the model.

[0037] This module combines the dilated spatial convolutional pooling pyramid and the convolutional attention module. By applying dilated convolutions of different scales with dilation factors of 6, 12, and 18, single pooling operations, and Conv feature compression to the input features in parallel, MSFE can capture contextual information at different levels in the image and achieve multi-scale perception of the target; in addition, using different dilation rates for dilated convolutions helps to expand the receptive field, allowing the network to better understand broader contextual information; adding residual links can alleviate the problem of gradient disappearance and improve the network's feature extraction capabilities; finally, the convolutional attention module CBAM (Convolutional Block Attention Module) is introduced after the residual connection. This module focuses on the key information of small targets by emphasizing meaningful features in the channel and spatial dimensions, providing the model with a more comprehensive and effective feature extraction capability.

[0038] The structure of the convolutional attention module CBAM is as follows Figure 7 As shown. CBAM contains two independent submodules, the Channel Attention Module (CAM) and the Spatial Attention Module (SAM), which perform channel and spatial attention respectively: the input feature map F (H×W×C) is subjected to global max pooling (Global Max Pooling) and global average pooling (Global Average Pooling) based on width W and height H respectively to obtain two 1×1×C feature maps, and then they are sent to a two-layer neural network (MLP) respectively. The number of neurons in the first layer is C / r (r is the reduction rate), the activation function is Relu, and the number of neurons in the second layer is C. Then, the features output by the MLP are added based on element-wise operation, and then activated by sigmoid operation to generate the final channelattention feature, that is, Mc. Finally, Mc is multiplied by the input feature map F to generate the input features required by the spatial attention module.

[0039] The feature map output by the channel attention module is used as the input feature map of the spatial attention. First, a global maximum pooling and global average pooling based on space are performed to obtain two H×W×1 feature maps, and then these two feature maps are spliced ​​based on the channel. Then, after a 7×7 convolution operation, the dimension is reduced to 1 channel, that is, H×W×1. Then, a special attention feature (spatial attention feature), Ms, is generated through a sigmoid layer. Finally, the feature is multiplied with the input feature of the module to obtain the final generated feature.

[0040] The standard yolov8 model uses a prediction-head with three different scales for prediction. When the size of the input image is 640x640, the yolov8 model will obtain outputs of three different scales: 80x80 (640 / 8), 40x40 (640 / 16), and 20x20 (640 / 32). However, the switch cabinet indicator lights to be detected by the inspection robot are generally about 10 pixels in length and width, which is difficult to pass the standard yolov8 detection model. Therefore, this application improves the detection head and adds a higher resolution detection head. The 160x160 feature map output from the 17th layer of the network is concat-operated with the 3rd layer feature map of the network to obtain a higher resolution feature map as the input of the detection head. The output of the detection head is a 160x160 target. By increasing the model's ability to resolve fine-grained features, the model's detection effect on small targets is enhanced, greatly improving the detection accuracy of the target.

[0041] Model training. The model parameter initialization specifically includes the following parts: CIOU is selected as the loss function; the weight decay coefficient is 0.0005; the Epoch is set to 300; the Mosaic enhancement index is set to (0.5, 1.5), Mixup enhancement is turned off, and Mosaic data enhancement is turned off in the last 80 rounds of training; Batch-size is set to 4.

[0042] This model uses the mean average precision (mAP), a performance evaluation indicator commonly used in target detection algorithms, for evaluation.

[0043] The experimental environment of this application is: system: Windows 11; processor 12th Gen Intel(R) core(TM) i3-12100F 3.30 GHz; memory: 16GB; graphics card: NVIDIA GeForce GTX 1650 SUPER; deep learning framework is Pytorch version 2.1.1; compilation language is Python version 3.9.18; DUDA version is 12.2; CUDNN version is 8.9.6; After the model is trained, the trained model is used to detect the status of the indicator lights on the substation switch cabinet to determine whether the switch cabinet is operating correctly.

[0044] The switch cabinet state recognition algorithm designed in this application can be deployed in a power inspection robot for real-time detection and real-time feedback of switch cabinet state information.

[0045] Through the above switch cabinet state recognition method based on small target perception, on the one hand, SPD non-strided convolution, namely SPD-Conv, is introduced into the backbone network of YOLOv8 to replace the strided convolution in the original backbone network, which improves the feature extraction ability of the model and reduces the loss of small target features. The SPPF module in the backbone network of YOLOv8 is improved to a multi-scale feature extraction module MSFE_AC, which can realize multi-scale perception of the target, expand the receptive field, and better retain the global information of the small target of the switch cabinet indicator light. Adding a multi-scale feature extraction module and a small target detection layer enables the network to pay attention to more detailed information, thereby improving the recognition effect of the small target of the substation drainage line bolts and the status of each indicator light of the switch cabinet, which is more suitable for the application scenario of smart substations. On the other hand, there is a significant improvement in the level of intelligent control, which can be widely used in substation switch cabinet state detection, especially for smart substations.

[0046] It should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise", etc. in the above description indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the embodiments of the present disclosure and simplifying the description, and do not indicate or imply that the referred device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the embodiments of the present disclosure.

[0047] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the embodiments of the present disclosure, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined.

[0048] In the embodiments of the present disclosure, unless otherwise clearly specified and limited, the terms "installed", "connected", "connected", "fixed" and the like should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present disclosure can be understood according to specific circumstances.

[0049] In the embodiments of the present disclosure, unless otherwise clearly specified and limited, a first feature being "above" or "below" a second feature may include the first and second features being in direct contact, or may include the first and second features not being in direct contact but being in contact through another feature between them. Moreover, a first feature being "above", "above" and "above" a second feature includes the first feature being directly above and obliquely above the second feature, or simply indicates that the first feature is higher in level than the second feature. A first feature being "below", "below" and "below" a second feature includes the first feature being directly below and obliquely below the second feature, or simply indicates that the first feature is lower in level than the second feature.

[0050] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example" or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present disclosure. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification.

[0051] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the invention disclosed herein. This application is intended to cover any modification, use or adaptation of the present disclosure, which follows the general principles of the present disclosure and includes common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The specification and examples are intended to be exemplary only, and the true scope and spirit of the present disclosure are indicated by the appended claims.

Claims

1. A switch cabinet state recognition method based on small target perception, characterized in that: The method includes: Acquire a substation switch cabinet image, and preprocess the substation switch cabinet image; Annotating the preprocessed substation switch cabinet image, and dividing the annotated substation switch cabinet image into a training set and a test set according to a preset ratio; Constructing a YOLOv8 model; wherein the YOLOv8 model includes an input layer, a backbone network, a neck network and a detection head; The YOLOv8 model is trained and tested using the training set and the test set respectively to obtain the trained YOLOv8 model; The YOLOv8 model is used to detect the state of the substation switch cabinet to obtain the switch cabinet state recognition result.

2. The switch cabinet state recognition method based on small target perception according to claim 1 is characterized in that: The step of acquiring a substation switch cabinet image and preprocessing the substation switch cabinet image includes: The substation switch cabinet images corresponding to various working states of the substation switch cabinet are obtained, and random scaling, random cropping, contrast adjustment and data enhancement processing are sequentially performed on the substation switch cabinet images.

3. The switch cabinet state recognition method based on small target perception according to claim 1 is characterized in that: The step of annotating the preprocessed substation switch cabinet image and dividing the annotated substation switch cabinet image into a training set and a training set according to a preset ratio includes: The pre-processed substation switch cabinet image is labeled using the Labelme tool, and the labeled information is saved as a txt file, and the data format is set to VOC format; The labeled substation switch cabinet images are divided into the training set and the test set in a ratio of 8:

2.

4. The switch cabinet state recognition method based on small target perception according to claim 1 is characterized in that: The backbone network includes: 5 SPD-Conv modules, 4 C2f modules and 1 MSFE_AC module; among them, One of the SPD-Conv modules, one of the SPD-Conv modules, one of the C2f modules, one of the SPD-Conv modules, one of the C2f modules, the SPD-Conv module, one of the C2f modules, one of the SPD-Conv modules, one of the C2f modules, one of the SPD-Conv modules, one of the C2f modules and one of the MFSE_AC modules are connected in series in sequence.

5. The switch cabinet state recognition method based on small target perception according to claim 4 is characterized in that: The SPD-Conv module includes one SPD layer and one non-strided convolution layer connected in sequence; wherein, The SPD layer slices the initial feature map of the substation switch cabinet image using a scaling factor and performs 2-fold downsampling to obtain four sub-feature maps of the same shape; Connecting the four sub-feature maps along the channel dimension to obtain a feature map that is half the size in space and four times the size in dimension; The feature map is input into the non-stride convolution to retain the discriminative feature information of the substation switch cabinet image.

6. The switch cabinet state recognition method based on small target perception according to claim 5 is characterized in that: The MSFE_AC module includes: 3 dilated convolutions, 3 Conv modules, 1 pool layer, 1 concat module, 1 Upsample layer and 1 CBAM attention module; among them, The initial feature map is input in parallel into one of the Conv modules, three of the dilated convolutions of different scales, and one of the pool layers, wherein the Conv module compresses the initial feature map, the dilated convolutions expand the receptive field of the initial feature map, and the pool layer performs a single pooling layer operation on the initial feature map; The output of the pool layer is input to the Upsample layer for processing; Input the output of the Conv module, the three dilated convolutions and the Upsample layer into the concat module for feature fusion to obtain a fused feature map; Using one of the Conv modules to compress the fused feature map, and adding the compressed fused feature map to the initial feature map; The added fused feature map and the initial feature map are input into the CBAM attention module to obtain the final feature map.

7. The switch cabinet state recognition method based on small target perception according to claim 6 is characterized in that: The CBAM attention module includes a CAM submodule and a SAM submodule; wherein, The input fusion feature map F (H×W×C) is subjected to global maximum pooling and global average pooling based on width W and height H respectively to obtain two 1×1×C first feature maps; Sending the first feature maps to a two-layer MLP network for processing; wherein the number of neurons in the first layer of the MLP network is C / r, the activation function is Relu, and the number of neurons in the second layer of the MLP network is C; The features output by the MLP network are added based on element-wise operation, and then activated by sigmoid operation to generate the final Mc; Perform a multiplication operation on the Mc and the fused feature map F to generate a second feature map; Input the second feature into the input of the SAM submodule, perform a space-based global maximum pooling and a global average pooling, and obtain two H×W×1 third feature maps; The two third feature maps are concatenated based on the channel, and the dimension is reduced to one channel through a 7×7 convolution operation, and Ms is generated through a sigmoid layer; The Ms is multiplied by the second feature to obtain the final feature map.

Images