Target detection method for power distribution network operation safety supervision

Abstract

The invention provides a target detection method for power distribution network operation safety supervision. The method comprises the steps of firstly obtaining a distribution network operation image, then substituting the distribution network operation image into a target detection model, performing feature extraction on the distribution network operation image through the target detection model, then obtaining distribution network operation data, and performing target detection on the distribution network operation data, thereby realizing safety supervision on the distribution network operation. The method improves the capability of detecting small target parts and targets under complex background and illumination interference in the intelligent safety supervision process of power distribution network operation, greatly improves the detection precision of such conditions, and reduces the probability of missing detection and false detection of small targets in a large scene of power distribution network operation.

Description

Technical Field

[0001] The present invention relates to the field of target detection, and in particular to a target detection method for safety supervision of distribution network operations. Background Art

[0002] At present, in the existing technology, distribution network projects face challenges such as difficult safety management and control, high requirements for the standardization of the operation process, and low intelligence level. The traditional "people-manage-people" management model has many disadvantages such as low decision-making and command efficiency, blind spots in ground monitoring, and uneven operating levels of personnel. On-site violation behavior recognition technology based on artificial intelligence image recognition can usually only perform static single-element violation detection in simple scenarios, which limits the application scenarios and effectiveness of violation recognition. Summary of the invention

[0003] The present invention aims to solve at least one of the technical problems existing in the prior art or related art.

[0004] To this end, a first aspect of the present invention proposes a target detection method for safety supervision of distribution network operations.

[0005] In view of this, a first aspect of the present invention provides a target detection method for distribution network operation safety supervision, including: obtaining a distribution network operation image; inputting the distribution network operation image into a target detection model, the target detection model including a trunk structure, a neck structure and a head structure, and obtaining distribution network operation data and performing target detection through each structure in turn; wherein, the target detection model is based on the Yolov8 model as a benchmark model, and is obtained by improving the trunk structure and neck structure of the benchmark model, and the trunk structure of the target detection model includes a convolutional layer (Conv layer), a lightweight convolutional layer, a lightweight vision module and an SPPF module; the neck structure of the target detection model includes an optimized feature pyramid network fusion structure, and the optimized feature pyramid network fusion structure can improve the calculation speed of the target detection model.

[0006] In addition, the target detection method for power distribution network operation safety supervision in the above technical solution provided by the present invention may also have the following additional technical features: In some technical solutions of the present invention, optionally, the network distribution operation image is input into the target detection model, the target detection model includes a trunk structure, a neck structure and a head structure, and the network distribution operation data is obtained through each structure in turn and the target detection is performed, including: the network distribution operation image is input into three sequentially connected Conv layers for feature extraction, the output of the first Conv layer is feature spliced ​​with the output of the third Conv layer, and the spliced ​​data is input into a lightweight convolution layer for feature extraction to obtain P2 layer data; the P2 layer data is input into three sequentially connected lightweight convolution layers, a lightweight visual module and a The lightweight convolution layer performs feature extraction to obtain P3 layer data; the P3 layer data is input into a lightweight convolution layer and a lightweight vision module connected in sequence for feature extraction to obtain P4 layer data; the P4 layer data is input into the SPPF module to obtain P5 layer data; the P2 layer data, P3 layer data, P4 layer data and P5 layer data are respectively sent to the optimized feature pyramid network fusion structure for splicing and fusion to obtain P2 layer output, P3 layer output, P4 layer output and P5 layer output respectively; the P2 layer output, P3 layer output, P4 layer output and P5 layer output are sent to the head structure to obtain distribution network operation data.

[0007] In some technical solutions of the present invention, optionally, the lightweight convolution layer performs corresponding processing on the first input data based on the step size; wherein, when the lightweight convolution layer sets the step size to 1, the first input data passes through the first Conv layer, DwConv layer, and second Conv layer connected in sequence, and the first input data is feature-concatenated with the output of the second Conv layer to obtain the first output data; when the lightweight convolution layer sets the step size to 2, the first input data passes through the first Conv layer, DwConv layer, and second Conv layer connected in sequence to obtain the second output data.

[0008] In some technical solutions of the present invention, optionally, the lightweight visual module is four 3×3 DwConv layers, one 1×1 Conv layer, and a global representation module connected in sequence, and the output of the 1×1 Conv layer and the output of the global representation module are feature spliced ​​through a residual structure; wherein the first 3×3 DwConv layer and the fourth 3×3 DwConv layer are jump connected, and the second 3×3 DwConv layer and the third 3×3 DwConv layer are jump connected.

[0009] In some technical solutions of the present invention, optionally, the processing method of the global representation module is: the output of the 1×1 Conv layer passes through the Transformer structure, the data passing through the Transformer structure is shuffled, and the shuffled data passes through two sequentially connected 1×1 Conv layers to obtain the output of the global representation module.

[0010] In some technical solutions of the present invention, optionally, the target detection model is based on the Yolov8 model as a benchmark model, and is obtained by improving the trunk structure and neck structure of the benchmark model, and also includes: replacing the Conv layer in the Yolov8 model with a DwConv layer, replacing the C2f module in the Yolov8 model with an MSC2f module, and obtaining an optimized feature pyramid network fusion structure.

[0011] In some technical solutions of the present invention, optionally, the C2f module in the Yolov8 model is optimized to an MSC2f module, including: based on a Bottleneck number parameter, adding at least one Bottleneck layer to the C2f module; when the Bottleneck number parameter is 1, adding at least one Bottleneck layer to the C2f module, and the output of the first Bottleneck layer passes through a Bottleneck layer respectively to obtain a third data output and a fourth data output; wherein the input of the Concat structure is replaced by the third data output, the fourth data output and the output via the Split structure.

[0012] In some technical solutions of the present invention, optionally, the Bottleneck layer makes a judgment based on the position of the second input data in the neck structure, and performs corresponding processing on the second input data based on the judgment result; when the second input data is P2 layer data, the second input data passes through a MSConv layer, a DwConv layer, and a MSConv layer connected in sequence, and the second input data is feature-concatenated with the output of the second MSConv layer to obtain third output data; otherwise, the second input data passes through two MSConv layers connected in sequence to obtain third output data.

[0013] In some technical schemes of the present invention, optionally, the processing method of the MSConv module is: the third input data flattens the channel information through a 1×1 Conv layer, and the flattened channel information is divided into three parts and passed into three 3×3 Conv layers respectively to obtain the first channel processing data, the second channel processing data and the third channel processing data respectively; the first channel processing data, the second channel processing data and the third channel processing data are subjected to channel shuffling processing to obtain the fourth output data.

[0014] A second aspect of the present invention provides a computer-readable storage medium having a program or instruction stored thereon, which, when executed by a processor, implements a target detection method for safety supervision of distribution network operations as in any of the above-mentioned technical solutions.

[0015] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which: Figure 1 This is one of the flow charts of a target detection method for safety supervision of distribution network operations according to an embodiment of the present invention; Figure 2 is one of the schematic diagrams of a target detection model according to an embodiment of the present invention; Figure 3 FIG. 1 is a schematic diagram of a lightweight visual module according to an embodiment of the present invention. DETAILED DESCRIPTION

[0017] In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0018] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited to the specific embodiments disclosed below.

[0019] Refer to the following Figures 1 to 3 A target detection method for safety supervision of distribution network operations according to some embodiments of the present invention is described.

[0020] like Figure 1 As shown, the first aspect of the present invention provides a target detection method for power distribution network operation safety supervision, comprising: Step 102, obtaining a distribution network operation image; Step 104, input the distribution network operation image into the target detection model, the target detection model includes a trunk structure, a neck structure and a head structure, and the distribution network operation data is acquired through each structure in turn and the target detection is performed; Among them, the target detection model is based on the Yolov8 model as the benchmark model, which is obtained by improving the backbone structure and neck structure of the benchmark model. The backbone structure of the target detection model includes a convolutional layer (Conv layer), a lightweight convolutional layer, a lightweight visual module and an SPPF module; The neck structure of the target detection model includes optimizing the feature pyramid network fusion structure, and optimizing the feature pyramid network fusion structure can improve the calculation speed of the target detection model.

[0021] The present invention proposes a target detection method for the safety supervision of distribution network operations. First, the distribution network operation image is obtained, and then the distribution network operation image is substituted into the target detection model. The distribution network operation image is feature extracted through the target detection model, and then the distribution network operation data is obtained. Target detection is performed on the distribution network operation data, thereby realizing the safety supervision of distribution network operations.

[0022] Among them, the target detection model is based on the Yolov8 model as the benchmark model, and is obtained by improving the backbone structure and neck structure of the benchmark model. The distribution network operation image passes through the backbone structure, neck structure and head structure in turn. The backbone structure performs multi-scale feature extraction on the distribution network operation image, and the neck structure performs feature fusion on the data after feature extraction, and then introduces the feature fused data into the detection head (i.e., the head structure) to complete the detection.

[0023] The backbone structure of the target detection model includes Conv layer, lightweight convolution layer, lightweight vision module and SPPF module. The lightweight convolution layer can reduce the computational complexity of the target detection model, and the lightweight vision module can maintain high detection performance with fewer FLOPS and parameters. The neck structure of the target detection model includes optimizing the feature pyramid network fusion structure, which can improve the computational speed of the target detection model.

[0024] The present invention uses the target detection model improved based on Yolov8 to perform safety supervision of distribution network operations, thereby improving the detection capabilities of small target components and targets under complex backgrounds and light interference during the intelligent safety supervision of distribution network operations, greatly improving the detection accuracy of such situations, and reducing the probability of missed detection and false detection of small targets in large distribution network operation scenes.

[0025] It should be understood that since some targets (such as the second safety rope) are small in size and the downsampling rate of YOLOv8 is large, the spatial position information will gradually be lost through layer-by-layer feature extraction, and it is difficult to obtain the feature information of small targets from deeper feature maps. Therefore, the original YOLOv8 model has poor detection capabilities for small-sized targets. The original model has a detection scale of 80×80 for small targets, and the receptive field obtained by detecting each grid is 8×8. When the detection target is small in size in the original image, it is difficult for the original network to recognize the feature information of the target in the grid. Therefore, the improved model proposed in the present invention includes a small target detection layer, and this small target detection layer is on an image with a larger spatial resolution. Figure 2As shown in the figure, the present invention adds 160×160 layers to the original network, including the introduction of corresponding feature fusion layers and additional detection heads, so as to obtain more accurate small-size target feature representation and semantic information. First, the feature layer with a scale of 80×80 (P3) is continuously stacked on the neck, and a feature layer with deep semantic feature information of small targets is constructed after the upsampling process; then, it is continued to be spliced ​​with the MV2 (P2) module with a scale of 160×160 at the fourth layer of the backbone network to supplement and enhance the expression ability of the fusion feature layer for the position information and semantic features of small targets. Then, the 160×160 fusion feature layer is incorporated into the MSC2f module to improve the expression ability of the semantic features and position information of small targets; the fusion layer with a scale of 160×160 is sent to the next Conv (convolution) layer and the additional detection head after the MSC2f module. The addition of the 160×160 scale layer enables the information of small target features to continue to be transmitted to the other three scale feature layers along the downsampling path, thereby enhancing the feature fusion ability of the model and improving the accuracy of small target detection. The introduction of additional decoupling joints increases the detection range for smaller components.

[0026] Furthermore, in some embodiments of the present invention, the network distribution operation image is input into the target detection model, and the target detection model includes a trunk structure, a neck structure, and a head structure. The network distribution operation data is obtained through each structure in turn and the target detection is performed, including: the network distribution operation image is input into three sequentially connected Conv layers for feature extraction, the output of the first Conv layer is spliced ​​with the output of the third Conv layer, and the spliced ​​data is input into a lightweight convolution layer for feature extraction to obtain P2 layer data; the P2 layer data is input into three sequentially connected lightweight convolution layers, a lightweight visual module, and a The lightweight convolution layer performs feature extraction to obtain P3 layer data; the P3 layer data is input into a lightweight convolution layer and a lightweight vision module connected in sequence for feature extraction to obtain P4 layer data; the P4 layer data is input into the SPPF module to obtain P5 layer data; the P2 layer data, P3 layer data, P4 layer data and P5 layer data are respectively sent to the optimized feature pyramid network fusion structure for splicing and fusion to obtain P2 layer output, P3 layer output, P4 layer output and P5 layer output respectively; the P2 layer output, P3 layer output, P4 layer output and P5 layer output are sent to the head structure to obtain distribution network operation data.

[0027] In this embodiment, the processing flow of the network distribution operation image in the target detection model is as follows: First, the network distribution operation image is input into three sequentially connected Conv layers for feature extraction, and the output of the first Conv layer is feature spliced ​​with the output of the third Conv layer. Among them, the first Conv layer is the first Conv layer that the network distribution operation image enters, and the third Conv layer is the third Conv layer that the network distribution operation image enters, and the spliced ​​data is input into the lightweight convolution layer for feature extraction to obtain the P2 layer data.

[0028] Then, the P2 layer data is input into three sequentially connected lightweight convolutional layers, a lightweight visual module and a lightweight convolutional layer for feature extraction to obtain the P3 layer data.

[0029] Then, the P3 layer data is input into a lightweight convolutional layer and a lightweight visual module connected in sequence for feature extraction to obtain the P4 layer data.

[0030] Finally, the P4 layer data is input into the SPPF module to obtain the P5 layer data.

[0031] On this basis, the P2 layer data, P3 layer data, P4 layer data and P5 layer data are respectively sent to the optimized feature pyramid network fusion structure for splicing and fusion, and the P2 layer output, P3 layer output, P4 layer output and P5 layer output are respectively obtained, and the P2 layer output, P3 layer output, P4 layer output and P5 layer output are sent to the head structure to obtain the distribution network operation data.

[0032] By processing distribution network operation images through the above method, it is possible to improve the detection capabilities of small target components and targets under complex backgrounds and light interference during the intelligent safety supervision of distribution network operations, greatly improve the detection accuracy of such situations, and reduce the probability of missed detection and false detection of small targets in large distribution network operation scenes.

[0033] Specifically, Figure 2 As shown in the figure, the lightweight visual module is ChSp-MbSFViT (w,h=2,2), and the lightweight convolutional layer is MV2 (full name Mobilenetv2). The MV2 with a downward arrow represents a stride (Stride in the figure) of 2, and the MV2 without an arrow represents a stride of 1 Furthermore, in some embodiments of the present invention, the lightweight convolution layer performs corresponding processing on the first input data based on the step size; wherein, when the lightweight convolution layer sets the step size to 1, the first input data passes through the first Conv layer, DwConv layer, and second Conv layer connected in sequence, and the first input data is feature-concatenated with the output of the second Conv layer to obtain the first output data; when the lightweight convolution layer sets the step size to 2, the first input data passes through the first Conv layer, DwConv layer, and second Conv layer connected in sequence to obtain the second output data.

[0034] In this embodiment, the lightweight convolution layer processes the first input data accordingly based on the step size, and processes accordingly according to the step size. The above method can perform flexible feature extraction and optimize computational efficiency, thereby improving the reasoning speed of the model and making the model perform better in complex tasks. Specifically, the concatenated data is input into a lightweight convolutional layer with a step size of 1 for feature extraction to obtain P2 layer data.

[0035] Specifically, the P2 layer data is input into a lightweight convolution layer with a stride of 2, a lightweight convolution layer with a stride of 1, a lightweight convolution layer with a stride of 2, a lightweight visual module and a lightweight convolution layer with a stride of 2, which are connected in sequence for feature extraction to obtain the P3 layer data.

[0036] Specifically, the P3 layer data is input into a lightweight convolutional layer with a step size of 2 and a lightweight visual module connected in sequence for feature extraction to obtain the P4 layer data.

[0037] Furthermore, in some embodiments of the present invention, the lightweight visual module is four 3×3 DwConv layers, one 1×1 Conv layer, and a global representation module connected in sequence, and the output of the 1×1 Conv layer and the output of the global representation module are feature spliced ​​through a residual structure; wherein the first 3×3 DwConv layer and the fourth 3×3 DwConv layer are jump connected, and the second 3×3 DwConv layer and the third 3×3 DwConv layer are jump connected.

[0038] In this embodiment, the structure of the lightweight visual module is four 3×3 DwConv layers, one 1×1 Conv layer, and a global representation module connected in sequence. The output of the 1×1 Conv layer and the output of the global representation module are feature concatenated through a residual structure.

[0039] On this basis, the first 3×3 DwConv layer and the fourth 3×3 DwConv layer are skip-connected, and the second 3×3 DwConv layer and the third 3×3 DwConv layer are skip-connected.

[0040] By making the first 3×3 DwConv layer and the fourth 3×3 DwConv layer jump connected, and the second 3×3 DwConv layer and the third 3×3 DwConv layer jump connected, four DwConvs are used for symmetrical cross-connection when representing local information features, which enhances the ability to capture spatial location information, reduces the number of parameters, and improves the running speed of the model.

[0041] In addition, a residual structure is used to concatenate the features after the global representation module to retain the spatial position information extracted by the convolutional layer.

[0042] It should be noted that the first 3×3 DwConv layer, the second 3×3 DwConv layer, the third 3×3 DwConv layer, and the fourth 3×3 DwConv layer are sorted based on the passing order of the data.

[0043] It is important to understand that for a given input feature map , its local spatial information is encoded by a convolutional layer with a 3×3 convolution kernel, and then a projection is applied to map the tensor to a high d-dimensional space with a 1×1 convolutional layer. At this point, the shape of the feature map is , and then global feature modeling is performed through the Transformer structure; in order to enable MobileViT to obtain a global representation that includes spatial sensing bias, it expands arrive Non-overlapping planar patches, , , , P is the patch number, h is the patch height, and W is the patch width. P is the pixel feature of the pth position of each patch. Transformer encodes the relationship between patches to obtain a global feature sequence , the formula is expressed as follows: Unlike the original Transformer, MobileViT preserves information about the patch order and spatial order of the pixels contained in the patch. Get , It is then resized back to its original size in the low C-dimensional space using a 1×1 convolution kernel and concatenated with the input Through the shortcut branch, these concatenated feature maps are then fused with another 3×3 convolutional layer for the final output.

[0044] For the expanded feature map , which is usually expanded in the form of a sequence and then input into the Transformer. During self-attention, each pixel in the feature map is processed along with other pixels, so that the amount of calculation is processed. In MobileViT, patches are first generated based on the partitions of the input feature map, and then self-attention is calculated only for pixels at the same position in each individual patch. Figure 3 As shown in Figure 1, Transformer is used to associate the pixels around the center of the patch. Since the information of adjacent pixels has been encoded by convolution, Transformer is able to merge the pixel information of all patches in the image. At this point, the amount of calculation is reduced to ,because Use convolution to encode localization information within a 3×3 region, and Encode the global information of the p-th position of each Patch, including any pixel Contains information based on all pixels below However, the representation of local information features in MobileViT is still insufficient for spatial location information, which is crucial for downstream tasks for multiple target detection tasks at complex distribution network construction sites. Therefore, the lightweight vision module is improved on the basis of MobileViT.

[0045] Furthermore, in some embodiments of the present invention, the processing method of the global representation module is as follows: the output of the 1×1 Conv layer passes through the Transformer structure, the data passing through the Transformer structure is shuffled, and the shuffled data passes through two sequentially connected 1×1 Conv layers to obtain the output of the global representation module.

[0046] In this embodiment, the processing method of the global representation module for data is described. The processing method of the global representation module is as follows: the output of the 1×1 Conv layer passes through the Transformer structure, the data passing through the Transformer structure is shuffled, and the shuffled data passes through two sequentially connected 1×1 Conv layers to obtain the output of the global representation module.

[0047] Because the information between some different channels may not interact with each other after the Transformer structure, shuffling the data processed by the Transformer structure can increase the fusion between different channel information and improve the expression of features. In addition, using a 1×1 Conv layer can simplify the structure of the fusion block. Through the global representation module, the present invention maintains a high detection performance with fewer FLOPS and parameters.

[0048] Furthermore, in some embodiments of the present invention, the target detection model is based on the Yolov8 model as a benchmark model, and is obtained by improving the trunk structure and neck structure of the benchmark model. It also includes: replacing the Conv layer in the Yolov8 model with the DwConv layer, and replacing the C2f module in the Yolov8 model with the MSC2f module to obtain an optimized feature pyramid network fusion structure.

[0049] In this embodiment, improvements are made based on the Yolov8 model, wherein the Conv layer in the Yolov8 model is replaced by a DwConv layer, and the C2f module in the Yolov8 model is replaced by an MSC2f module, so as to obtain an optimized feature pyramid network fusion structure and thereby obtain an improved neck structure.

[0050] Furthermore, in some embodiments of the present invention, the C2f module in the Yolov8 model is optimized to an MSC2f module, including: based on a Bottleneck number parameter, adding at least one Bottleneck layer to the C2f module; when the Bottleneck number parameter is 1, adding at least one Bottleneck layer to the C2f module, and the output of the first Bottleneck layer passes through a Bottleneck layer respectively to obtain a third data output and a fourth data output; wherein the input of the Concat structure is replaced by the third data output, the fourth data output and the output via the Split structure.

[0051] In this embodiment, the C2f module is optimized by adding at least one Bottleneck layer to the C2f module according to a preset Bottleneck number parameter.

[0052] Among them, when the Bottleneck number parameter is 1, at least one Bottleneck layer is added to the C2f module, and the output of the first Bottleneck layer passes through a Bottleneck layer respectively to obtain the third data output and the fourth data output, and the input of the Concat structure is replaced by the third data output, the fourth data output and the output via the Split structure.

[0053] Changing the number of Bottlenecks through the Bottleneck number parameter can enhance the model feature extraction capability. At the same time, it is convenient for embedded deployment. Only one Bottleneck layer is added, and the output of the first Bottleneck layer passes through a Bottleneck layer respectively, thereby improving the model feature extraction capability and facilitating subsequent deployment.

[0054] Specifically, the first Bottleneck layer is the first Bottleneck layer that the input passes through into the MSC2f module.

[0055] Furthermore, in some embodiments of the present invention, the Bottleneck layer makes a judgment based on the position of the second input data in the neck structure, and performs corresponding processing on the second input data based on the judgment result; when the second input data is P2 layer data, the second input data passes through a MSConv layer, a DwConv layer, and a MSConv layer connected in sequence, and the second input data is feature-concatenated with the output of the second MSConv layer to obtain third output data; otherwise, the second input data passes through two MSConv layers connected in sequence to obtain third output data.

[0056] In this embodiment, the Bottleneck layer makes a judgment based on the position of the second input data in the neck structure, and performs corresponding processing on the second input data based on the judgment result.

[0057] Specifically, Figure 2 As shown, the P2 layer data is input to the bottom of the optimized feature pyramid network fusion structure. Then, when the second input data is the P2 layer data, the position of the second input data in the optimized feature pyramid network fusion structure is determined based on whether the second input data is the P2 layer data.

[0058] When the second input data is P2 layer data, the second input data passes through a sequentially connected MSConv layer, a DwConv layer, and an MSConv layer, and the second input data and the output of the second MSConv layer are feature concatenated to obtain the third output data. Otherwise, the second input data passes through two sequentially connected MSConv layers to obtain the third output data. By processing the second input data accordingly, the ability of feature representation is enhanced and the performance of the model is improved.

[0059] Furthermore, in some embodiments of the present invention, the processing method of the MSConv module is: the third input data flattens the channel information through a 1×1 Conv layer, and the flattened channel information is divided into three parts and passed into three 3×3 Conv layers respectively to obtain first channel processing data, second channel processing data and third channel processing data respectively; the first channel processing data, the second channel processing data and the third channel processing data are channel shuffled to obtain fourth output data.

[0060] In this embodiment, the processing method of the MSConv module is as follows: the third input data is flattened through a 1×1 Conv layer to flatten the channel information, and the flattened channel information is divided into three parts and respectively passed into three 3×3 Conv layers to obtain first channel processing data, second channel processing data, and third channel processing data; the first channel processing data, the second channel processing data, and the third channel processing data are subjected to channel shuffling processing to obtain fourth output data. Through the MSConv module, the feature fusion capability under complex light conditions can be significantly improved, and then different channel information can be rearranged and correlated to enhance feature representation.

[0061] Furthermore, in some embodiments of the present invention, a computer-readable storage medium is provided, on which a program or instruction is stored. When the program or instruction is executed by a processor, the computer-readable storage medium in any of the above-mentioned technical solutions is implemented, thereby having all the beneficial technical effects of the computer-readable storage medium in any of the above-mentioned embodiments.

[0062] The computer readable storage medium includes a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0063] A computer-readable storage medium may be a tangible device that can retain and store instructions for use by an instruction execution device, and may be an electronic storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the above devices, but is not limited thereto. A non-exhaustive list of more specific examples of computer-readable storage media includes: portable computer floppy disks, hard disks, random access memories, read-only memories, erasable programmable read-only memories, static random access memories, portable compact disk read-only memories, digital versatile disks, memory cards, floppy disks, encoding mechanical devices (such as punched cards or grooves with raised structures having instructions recorded therein), and any suitable combination of the above devices. The computer-readable storage medium used herein should not be understood as a transmission signal itself, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media, or electrical signals transmitted through wires, etc.

[0064] In the claims, specification and drawings of the present invention, the term "multiple" refers to two or more than two. Unless otherwise clearly defined, the orientation or positional relationship indicated by the terms "upper" and "lower" is based on the orientation or positional relationship shown in the drawings, which is only for the purpose of more conveniently describing the present invention and making the description process simpler, rather than indicating or implying that the device or element referred to must have the specific orientation described, be constructed and operated in a specific orientation, so these descriptions cannot be understood as limiting the present invention; the terms "connect", "install", "fix" and the like should be understood in a broad sense. For example, "connection" can be a fixed connection between multiple objects, or a detachable connection between multiple objects, or an integral connection; it can be a direct connection between multiple objects, or an indirect connection between multiple objects through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood based on the specific circumstances of the above data.

[0065] In the claims, specification and drawings of the present invention, the description of the terms "one embodiment", "some embodiments", "specific embodiments" and the like 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 invention. In the claims, specification and drawings of the present invention, 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.

[0066] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A target detection method for safety supervision of distribution network operations, characterized in that: include: Get the network distribution operation image; The distribution network operation image is input into a target detection model, wherein the target detection model includes a trunk structure, a neck structure, and a head structure, and the distribution network operation data is acquired through each structure in turn and target detection is performed; The target detection model is based on the Yolov8 model as a benchmark model, and is obtained by improving the backbone structure and neck structure of the benchmark model. The backbone structure of the target detection model includes a convolutional layer (Conv layer), a lightweight convolutional layer, a lightweight visual module and an SPPF module. The neck structure of the target detection model includes an optimized feature pyramid network fusion structure, and the optimized feature pyramid network fusion structure can improve the calculation speed of the target detection model.

2. The target detection method for power distribution network operation safety supervision according to claim 1 is characterized in that: The network distribution operation image is input into a target detection model, wherein the target detection model includes a trunk structure, a neck structure, and a head structure, and the network distribution operation data is acquired through each structure in turn and target detection is performed, including: The network distribution operation image is input into three sequentially connected Conv layers for feature extraction, the output of the first Conv layer is concatenated with the output of the third Conv layer for feature splicing, and the data after feature splicing is input into a lightweight convolution layer for feature extraction to obtain P2 layer data; Input the P2 layer data into the three lightweight convolutional layers, one lightweight visual module and one lightweight convolutional layer connected in sequence for feature extraction to obtain P3 layer data; Input the P3 layer data to one of the lightweight convolutional layers and one of the lightweight visual modules connected in sequence for feature extraction to obtain P4 layer data; Input the P4 layer data into the SPPF module to obtain the P5 layer data; The P2 layer data, the P3 layer data, the P4 layer data and the P5 layer data are respectively sent to the optimized feature pyramid network fusion structure for splicing and fusion, and the P2 layer output, the P3 layer output, the P4 layer output and the P5 layer output are respectively obtained; The P2 layer output, the P3 layer output, the P4 layer output and the P5 layer output are sent to the header structure to obtain the network distribution operation data.

3. The target detection method for power distribution network operation safety supervision according to claim 2 is characterized in that: The lightweight convolution layer performs corresponding processing on the first input data based on the step size; When the step size of the lightweight convolution layer is set to 1, the first input data passes through the first Conv layer, the DwConv layer, and the second Conv layer connected in sequence, and the first input data is feature-concatenated with the output of the second Conv layer to obtain the first output data. When the step size of the lightweight convolution layer is set to 2, the first input data obtains the second output data through the first Conv layer, the DwConv layer and the second Conv layer connected in sequence.

4. The target detection method for power distribution network operation safety supervision according to claim 3 is characterized in that: The lightweight visual module is composed of four 3×3 DwConv layers, a 1×1 Conv layer, and a global representation module connected in sequence, and the output of the 1×1 Conv layer and the output of the global representation module are feature spliced ​​through a residual structure; Among them, the first 3×3 DwConv layer and the fourth 3×3 DwConv layer are jump-connected, and the second 3×3 DwConv layer and the third 3×3 DwConv layer are jump-connected.

5. The target detection method for power distribution network operation safety supervision according to claim 4 is characterized in that: The processing method of the global representation module is as follows: the output of the 1×1 Conv layer passes through the Transformer structure, the data passing through the Transformer structure is shuffled, and the shuffled data passes through two 1×1 Conv layers connected in sequence to obtain the output of the global representation module.

6. The target detection method for power distribution network operation safety supervision according to claim 2 is characterized in that: The target detection model is based on the Yolov8 model as a benchmark model, obtained by improving the trunk structure and neck structure of the benchmark model, and also includes: The Conv layer in the Yolov8 model is replaced with the DwConv layer, and the C2f module in the Yolov8 model is replaced with the MSC2f module to obtain the optimized feature pyramid network fusion structure.

7. The target detection method for power distribution network operation safety supervision according to claim 6 is characterized in that: The optimization of the C2f module in the Yolov8 model to the MSC2f module includes: Based on the Bottleneck number parameter, at least one Bottleneck layer is added to the C2f module; When the Bottleneck number parameter is 1, at least one Bottleneck layer is added to the C2f module, and the output of the first Bottleneck layer passes through a Bottleneck layer respectively to obtain a third data output and a fourth data output; The input of the Concat structure is replaced by the third data output, the fourth data output and the output via the Split structure.

8. The target detection method for power distribution network operation safety supervision according to claim 7 is characterized in that: The Bottleneck layer determines the position of the second input data in the neck structure based on the position of the second input data, and processes the second input data accordingly based on the determination result; When the second input data is the P2 layer data, the second input data passes through a sequentially connected MSConv layer, a DwConv layer, and an MSConv layer, and the second input data is feature concatenated with the output of the second MSConv layer to obtain third output data; Otherwise, the second input data is passed through two sequentially connected MSConv layers to obtain third output data.

9. The target detection method for power distribution network operation safety supervision according to claim 8 is characterized in that: The processing method of the MSConv module is as follows: the third input data is passed through a 1×1 Conv layer to flatten the channel information, and the flattened channel information is divided into three parts and respectively passed into three 3×3 Conv layers to obtain first channel processing data, second channel processing data and third channel processing data respectively; Channel shuffling processing is performed on the first channel-processed data, the second channel-processed data, and the third channel-processed data to obtain fourth output data.

10. A computer-readable storage medium storing a program or an instruction, characterized in that: When the program or the instruction is executed by the processor, the target detection method for safety supervision of distribution network operations as described in any one of claims 1 to 9 is implemented.

Images