Electric power fitting detection method and system for data center park power transmission line inspection

Abstract

The invention provides an electric power fitting detection method and system for data center park power transmission line inspection, and belongs to the technical field of electric power fitting detection, and the method comprises the steps: obtaining an electric power fitting image, constructing a data set, and dividing the data set; an improved YOLOv8-based electric power fitting identification model is constructed, a feature extraction network is based on a lightweight structure, an attention mechanism and a pyramid pooling structure are introduced, a feature fusion network uses a multi-scale coupling structure and up-down sampling is added, and a plurality of detection heads construct different dimension models; performing training, verification and testing based on the training set, the verification set and the test set, and determining a target electric power fitting identification model; and performing target detection on the new electric power fitting image. The method has the beneficial effects that an improved YOLOv8-based electric power fitting identification model is adopted, the limitations of high model complexity, poor side end adaptability, insufficient detection precision and the like of an existing electric power fitting detection method in a data center park power transmission line inspection scene are solved, and the detection performance is improved.

Description

Technical Field

[0001] This invention relates to the field of power fitting testing technology, and in particular to a method and system for testing power fittings during the inspection of power transmission lines in data center parks. Background Technology

[0002] A stable power transmission system is a critical prerequisite for ensuring the smooth operation of data centers. Power fittings, such as insulators, surge arresters, and tension clamps, are essential components of transmission lines, playing an irreplaceable role in connection, support, and insulation protection. However, during long-term operation, power fittings are not only subject to erosion from natural environments such as wind, frost, rain, and snow, but also to the stress of mechanical vibration, inevitably leading to aging or damage. Once some power fittings fail, it will directly trigger a power line fault in the data center campus, resulting in a power outage of the Internet Data Center (IDC) system and ultimately severely impacting business operations. Therefore, intelligent real-time monitoring of various power fittings in the data center campus transmission lines, enabling timely, accurate, and efficient fault location identification, not only provides a scientific basis for the maintenance and operation management of power fittings, helping maintenance personnel quickly locate problems and take targeted measures, but also is of great significance for ensuring the stable operation of the data center power system.

[0003] Currently, the mainstream methods for inspecting power transmission lines in data center parks mainly include manual inspection, photoelectric signal acquisition and analysis, and traditional image processing algorithms. However, these methods all have certain limitations: manual inspection requires a lot of manpower and resources and is inefficient; photoelectric signal detection suffers from problems such as a mismatch between input and output, poor robustness, and insufficient real-time performance; although traditional machine learning algorithms have achieved automation to a certain extent, they have poor robustness, high computational complexity, and relatively fixed retrieval patterns.

[0004] In contrast, deep learning algorithms, with their advantages of not relying on human experience, strong robustness, and outstanding generalization ability, have gradually become a research hotspot in the field of power fitting inspection. In recent years, many scholars have combined deep learning algorithms with power fitting inspection and achieved significant research results. However, existing power fitting inspection algorithms still have shortcomings: on the one hand, traditional deep learning algorithms rely excessively on data, have a large number of model parameters, poor edge-end adaptability, and limited practicality; on the other hand, existing lightweight inspection algorithms still need to be improved in terms of model representation ability and detection accuracy.

[0005] Existing power fitting detection algorithms can be categorized into single-stage and two-stage detection models based on whether candidate regions need to be generated in advance during the detection process. Two-stage models, such as Cascade Region-based Convolutional Neural Network (Cascade RCNN) and Faster Region-based Convolutional Neural Network (Faster RCNN), achieve high detection accuracy through candidate region pre-division, but their slow inference speed makes them difficult to meet real-time requirements. Single-stage models, such as You Only Live Once (YOLO) and Single Shot MultiBox Detector (SSD), have advantages such as fewer parameters, faster detection speed, and strong edge adaptability. However, in the scenario of power fitting detection during power line inspection in data center parks, directly using existing YOLOv8, YOLOv9, or SSD algorithms still suffers from insufficient detection accuracy and high model complexity. Summary of the Invention

[0006] To address the above technical problems, this invention provides a method for inspecting power fittings during the inspection of power transmission lines in data center parks.

[0007] The technical problem solved by this invention can be achieved by the following technical solutions:

[0008] A method for testing power fittings during the inspection of power transmission lines in a data center campus includes:

[0009] Step S1: Obtain images of power fittings, construct a dataset, and divide the dataset into a training set, a test set, and a validation set;

[0010] Step S2: Construct an electric fitting recognition model based on an improved YOLOv8, including a feature extraction network, a feature fusion network, and a detection head. The feature extraction network is based on a lightweight network structure, while introducing an attention mechanism and a pyramid pooling structure. The feature fusion network adopts a multi-scale coupled feature fusion structure, which is interconnected with the output of the feature extraction network through residual jumpers of the same scale, and an upsampling / downsampling structure is added. The detection head uses multiple detection heads to construct electric fitting recognition models of different dimensions.

[0011] Step S3: Train the power fitting recognition model based on the improved YOLOv8 based on the training set to obtain the trained power fitting recognition model, and use the validation set for validation and the test set for testing to determine the target power fitting recognition model.

[0012] Step S4: Obtain a new image of the power fittings, and perform target detection on the new image of the power fittings based on the target power fittings recognition model to obtain the detection result.

[0013] Preferably, the feature extraction network includes:

[0014] Preliminary feature extraction and dimensionality reduction are performed on the input power fitting images to obtain the first intermediate features;

[0015] The first intermediate feature is subjected to two GMBConv operations to obtain the second intermediate feature.

[0016] The second intermediate feature is subjected to two GMBConv operations to obtain the third intermediate feature;

[0017] Perform a GMBConv operation on the third intermediate feature sequentially to obtain the fourth intermediate feature;

[0018] The SPPFCSPC operation is performed on the fourth intermediate feature to obtain the fifth intermediate feature.

[0019] Preferably, the network for the GMBConv operation includes:

[0020] Convolution is performed on the input features of the GMBConv operation to obtain the first intermediate sub-feature;

[0021] The first intermediate sub-feature is subjected to a depthwise separable convolution to obtain a second intermediate sub-feature; and the first intermediate sub-feature is subjected to a depthwise separable convolution to obtain a third intermediate sub-feature.

[0022] Perform GECA attention operation on the second intermediate feature and the third intermediate feature to obtain the fourth intermediate feature;

[0023] Convolution is performed on the fourth intermediate feature to obtain the fifth intermediate feature;

[0024] Perform a Dropout operation on the fifth intermediate feature to obtain the sixth intermediate feature;

[0025] The residual between the sixth intermediate feature and the input feature of the GMBConv operation is output as the output feature of the GMBConv operation.

[0026] Preferably, in the inverted bottleneck convolutional layer, the 3×3 convolution is decomposed into a first sub-depth separable convolutional layer of 1×3 and a second sub-depth separable convolutional layer of 3×1.

[0027] Preferably, each GMBConv operation is represented by the following formula:

[0028]

[0029] Where, x 1i GMBConv represents the input features; D represents the Dropout operation; Conv represents the convolution operation; E represents the GECA attention mechanism operation; DW represents the depthwise separable convolution operation; y o1 This indicates the output characteristics of the GMBConv operation.

[0030] Preferably, the GECA attention mechanism operation is represented by the following formula:

[0031] y i =DW(Sigmoid(G(x) 2i )))

[0032] Where, x 2i The input features are represented by GECA attention mechanism; DW represents depthwise separable convolution; Sigmoid represents activation function operation; G represents global average pooling operation; y i This represents the output characteristics of the GECA attention mechanism.

[0033] Preferably, the SPPFCSPC operation is expressed using the following formula:

[0034]

[0035] Where, x 3i SPPFCSPC represents the input features for the operation; Conv represents the convolution operation; Shuffle represents the channel shuffling operation; Sigmoid represents the activation function operation; MaxPool represents the max pooling operation; y o2 This indicates the output characteristics of the SPPFCSPC operation.

[0036] Preferably, the multi-scale coupled feature fusion structure is a four-scale coupled feature fusion structure, configured as follows:

[0037] The fifth intermediate feature and the fourth intermediate feature are fused to obtain the first fused feature;

[0038] The third intermediate feature and the first fusion feature are fused together to obtain the second fusion feature;

[0039] The second intermediate feature and the second fused feature are fused to obtain the third fused feature;

[0040] The third fusion feature and the third intermediate feature are fused to obtain the fourth fusion feature;

[0041] The second fusion feature and the fourth fusion feature are fused together to obtain the fifth fusion feature;

[0042] The first fusion feature and the fifth fusion feature are fused together to obtain the sixth fusion feature.

[0043] Preferably, the detection head comprises four, and the four detection heads comprise:

[0044] The first detection head is connected to the sixth feature fusion layer and is used to output a 160×160×128 dimension detection result based on the sixth fusion feature.

[0045] The second detection head is connected to the fifth feature fusion layer and is used to output an 80×80×256-dimensional detection result based on the fifth fusion feature.

[0046] The third detection head is connected to the fourth feature fusion layer and is used to output a 40×40×512 dimension detection result based on the fourth fusion feature.

[0047] The fourth detection head is connected to the third feature fusion layer and is used to output a 20×20×1024 dimension detection result based on the third fusion feature.

[0048] On the other hand, a power fitting identification system for data center campus power transmission line inspection is also provided, for implementing the power fitting detection method for data center campus power transmission line inspection as described above, including:

[0049] The dataset construction module is used to acquire images of power fittings, construct the dataset, and divide the dataset into training, testing, and validation sets.

[0050] The model building module is used to construct an electric fitting recognition model based on an improved YOLOv8, including a feature extraction network, a feature fusion network, and a detection head. The feature extraction network is based on a lightweight network structure and incorporates an attention mechanism and a pyramid pooling structure. The feature fusion network adopts a multi-scale coupled feature fusion structure, which is interconnected with the output of the feature extraction network through residual jumpers of the same scale, and an upsampling / downsampling structure is added. The detection head uses multiple detection heads to construct electric fitting recognition models of different dimensions.

[0051] The model training module is connected to the dataset construction module and the model construction module respectively. It is used to train the power fitting recognition model based on the improved YOLOv8 based on the training set to obtain the trained power fitting recognition model. The model is then validated using the validation set and tested using the test set to determine the target power fitting recognition model.

[0052] The target detection module, connected to the model training module, is used to acquire new power fitting images and perform target detection on the new power fitting images based on the target power fitting recognition model to obtain detection results.

[0053] The advantages or beneficial effects of the technical solution of this invention are as follows:

[0054] This invention employs an improved YOLOv8-based power fitting recognition model. Based on a lightweight network structure, it reduces model complexity, avoids the problem of large parameter counts in traditional models, and improves edge-end adaptability, making the model easier to deploy and run on edge devices. The model incorporates an attention mechanism to enhance its ability to capture and analyze features. A pyramid pooling structure further strengthens the model's ability to fuse high-dimensional abstract semantic information. A multi-scale coupled feature fusion structure adds upsampling and downsampling, enabling the fusion of power fitting features from different scales and levels, enriching feature information and further improving the model's ability to recognize power fittings in complex scenarios. Furthermore, multiple detection heads construct power fitting recognition models with different dimensions, enabling detection of power fittings from multiple angles, significantly improving detection accuracy. This addresses the limitations of existing power fitting detection methods in data center campus transmission line inspection scenarios, such as high model complexity, poor edge-end adaptability, and insufficient detection accuracy, thus improving detection performance. Attached Figure Description

[0055] Figure 1 This is a flowchart illustrating the power fitting testing method in a preferred embodiment of the present invention.

[0056] Figure 2 This is a schematic diagram of the architecture of the power fitting identification model based on the improved YOLOv8 in a preferred embodiment of the present invention.

[0057] Figure 3 A schematic diagram of the GMBConv network architecture is shown in a preferred embodiment of the present invention.

[0058] Figure 4 A schematic diagram of the GECA attention mechanism network architecture is shown in a preferred embodiment of the present invention.

[0059] Figure 5 A schematic diagram of the SPPFCSPC network architecture is shown in a preferred embodiment of the present invention.

[0060] Figure 6 This is a structural block diagram of a power fitting identification system for power line inspection in a data center campus, as described in a preferred embodiment of the present invention. Detailed Implementation

[0061] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0063] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the scope of the invention.

[0064] In a preferred embodiment of the present invention, based on the above-mentioned problems existing in the prior art, a method for detecting power fittings during the inspection of power transmission lines in data center campuses is provided, such as... Figure 1 and Figure 2 As shown, the method includes:

[0065] Step S1: Obtain images of power fittings, construct a dataset, and divide the dataset into a training set, a test set, and a validation set;

[0066] Specifically, power fittings are metal accessories that connect and combine various devices in a power system to transmit mechanical loads, electrical loads, and provide certain protective functions, such as, but not limited to, insulators, surge arresters, and tension clamps.

[0067] When acquiring images of power fittings, web crawling technology can be used to crawl a certain number of power fitting images from the massive amount of information on the Internet. Alternatively, existing open-source datasets can be obtained, including but not limited to datasets such as CPLD and a certain number of images collected by Internet Data Centers (IDCs). These images can be integrated with the crawled images to construct a dataset for subsequent model training, evaluation, and validation.

[0068] After constructing the dataset, it is divided into training, testing, and validation sets according to a certain ratio. In this embodiment, the ratio is 8:1:1, meaning the training set accounts for 90% and is used for model training, allowing the model to learn the features and patterns of power fitting images; the testing set accounts for 10% and is used to initially test the model's performance after training, verifying its generalization ability on unknown data; the validation set also accounts for 10% and is used to evaluate the model's performance in real time during training, adjusting the model's hyperparameters to ensure the model reaches its optimal performance state and improve the training effect and performance.

[0069] Of course, this is not the only option; the specific proportions of the training set, test set, and validation set can be set according to actual needs.

[0070] Step S2: Construct an electric fitting recognition model based on an improved YOLOv8, including a feature extraction network 100, a feature fusion network 200, and a detection head 300. The feature extraction network 100 is based on a lightweight network structure, while introducing an attention mechanism and a pyramid pooling structure. The feature fusion network 200 adopts a multi-scale coupled feature fusion structure, which is interconnected with the output of the feature extraction network 100 through residual jumpers of the same scale, and an upsampling / downsampling structure is added. The detection head 300 uses multiple detection heads to construct electric fitting recognition models of different dimensions.

[0071] Step S3: Train the power fitting recognition model based on the improved YOLOv8 based on the training set to obtain the trained power fitting recognition model, and use the validation set for validation and the test set for testing to determine the target power fitting recognition model.

[0072] Step S4: Obtain new power fitting images and perform target detection on the new power fitting images based on the target power fitting recognition model to obtain detection results.

[0073] Specifically, in response to the limitations of existing power fitting detection methods in data center campus power transmission line inspection scenarios, such as high model complexity, poor edge-end adaptability, and insufficient detection accuracy, this invention proposes a lightweight power fitting detection and recognition algorithm based on improved YOLOv8, taking into account the actual needs of power fitting detection in data center campus power transmission line inspection scenarios. The aim is to solve the above problems and improve detection performance.

[0074] YOLOv8's overall architecture continues the single-stage detector design of the YOLO series, adopting an end-to-end approach that can directly predict the target's category and bounding box, simplifying the detection process and improving detection efficiency.

[0075] Since the YOLOv8 model can effectively balance detection speed and accuracy in the field of object detection, in this embodiment, the YOLOv8n model is preferably selected.

[0076] The YOLOv8n model serves as the main framework. YOLOv8n is a lightweight model within the YOLOv8 series, inheriting the design philosophy of YOLOv8 and offering the advantage of lightweight design while maintaining detection performance. Figure 2 As shown, its main components include:

[0077] Backbone: This refers to the feature extraction network 100 mentioned above, which is used to extract key features from the input image.

[0078] Neck network: This refers to the feature fusion network 200 mentioned above, which is used to effectively fuse the extracted features for subsequent detection.

[0079] The detection head outputs the final detection result.

[0080] In the feature extraction network 100, considering the small sample size and multi-scale characteristics of power fittings in data centers, this embodiment uses an improved backbone network based on EfficientNet and employs the lighter GMBConv (Ghost MobileInverted Bottleneck Convolution) for feature extraction. This greatly reduces the computational load and improves the model's operating efficiency while ensuring the feature extraction effect.

[0081] By using rank decomposition, the 3×3 convolution in the inverted bottleneck convolution is decomposed into a cascaded convolution structure of 1×3 and 3×1, which significantly reduces the number of model parameters by about 35%, while retaining the ability to capture multi-scale features.

[0082] After rank decomposition, a GECA (Ghost-Efficient Channel Attention) mechanism is added at the feature concatenation point to enhance the transmission of effective information at the cost of a small number of additional parameters.

[0083] Meanwhile, a pyramid pooling (SPPFCSPC) structure is inserted into the deepest layer of the backbone network to further enhance the model's ability to fuse high-dimensional abstract semantic information.

[0084] Finally, the feature extraction network 100 extracts four sets of feature maps from the 640×640×3 input image through a series of operations, and passes them to the feature fusion to build contextual information interaction.

[0085] Finally, the backbone network extracts four sets of feature maps from the 640×640×3 image through a series of complex and sophisticated operations, and then passes them to the feature fusion network 200 to build contextual information interaction.

[0086] In the feature fusion network 200, to address the poor feature fusion capability of the traditional YOLOv8 model, this embodiment proposes an improved four-scale feature fusion structure. This four-scale feature fusion structure is connected to the feature extraction network 100 via jumpers of the same scale. Simultaneously, during the feature fusion process, an additional upsampling / downsampling structure is added to construct richer information fusion interactions, enabling the full fusion of features at different scales.

[0087] Head uses four detection heads to construct power fitting detection models for data center campus power transmission line inspection scenarios from different dimensions, enabling detection of targets from multiple angles and improving the accuracy and comprehensiveness of detection.

[0088] Compared with other existing mainstream lightweight algorithms, the power fitting recognition model based on the improved YOLOv8 of this invention has better recognition performance in the scenario of power transmission line inspection in data center campuses.

[0089] In a preferred embodiment, such as Figure 2 As shown, the feature extraction network 100 includes:

[0090] Stem module 101 is used to perform preliminary feature extraction and dimensionality reduction on the input power fitting image to obtain the first intermediate feature;

[0091] The first feature extraction module 102 is used to extract the first intermediate feature twice in sequence.

[0092] The GMBConv operation is used to obtain the second intermediate feature;

[0093] The second feature extraction module 103 is used to extract the second intermediate feature twice sequentially.

[0094] The GMBConv operation yields the third intermediate feature;

[0095] The third feature extraction module 104 is used to sequentially extract the third intermediate feature.

[0096] The GMBConv operation yields the fourth intermediate feature;

[0097] The pyramid pooling module 105 is used to perform SPPFCSPC operation on the fourth intermediate feature to obtain the fifth intermediate feature.

[0098] Specifically, Stem module 101 is the initial part of the YOLOv8n main framework, used to extract low-level features from the input image.

[0099] In a preferred embodiment, such as Figure 3 As shown, the network architecture for GMBConv computation includes:

[0100] The input layer is used to receive input features for GMBConv operations;

[0101] The first convolutional layer, connected to the input layer, is used to convolve the input features of the GMBConv operation to obtain the first intermediate sub-feature;

[0102] Inverted bottleneck convolutional layers include:

[0103] The first sub-depth separable convolutional layer, connected to the first convolutional layer, is used to perform depth-separable convolution on the first intermediate sub-feature to obtain the second intermediate feature;

[0104] The second sub-depth separable convolutional layer is connected to the first convolutional layer and is used to perform depth-separable convolution on the first intermediate sub-feature to obtain the third intermediate feature;

[0105] The GECA attention layer connects to the first sub-depth separable convolutional layer and the second sub-depth separable convolutional layer, respectively, and is used to perform GECA attention operations on the second and third intermediate features to obtain the fourth intermediate feature.

[0106] The second convolutional layer, connected to the GECA attention layer, is used to convolve the fourth intermediate sub-feature to obtain the fifth intermediate sub-feature.

[0107] The regularization layer, connected to the GECA attention layer, is used to perform Dropout operation on the fifth intermediate sub-feature to obtain the sixth intermediate sub-feature.

[0108] The output layer, connected to the regularization layer, is used to output the residual between the sixth intermediate feature and the input features of the GMBConv operation as the output feature of the GMBConv operation.

[0109] In a preferred embodiment, the inverted bottleneck convolutional layer decomposes the 3×3 convolution into a first sub-depth separable convolutional layer of 1×3 and a second sub-depth separable convolutional layer of 3×1.

[0110] Specifically, the rank decomposition technique is used to decompose the 3×3 convolution in the inverted bottleneck convolution into a cascaded convolution structure of 1×3 and 3×1, which significantly reduces the number of model parameters (the number of parameters is compressed by about 35%), while retaining the ability to capture multi-scale features.

[0111] In a preferred embodiment, each GMBConv operation is represented by the following formula:

[0112]

[0113] Where, x 1iGMBConv represents the input features; D represents the Dropout operation; Conv represents the convolution operation; E represents the GECA attention mechanism operation; DW represents the depthwise separable convolution operation; y o1 This indicates the output characteristics of the GMBConv operation.

[0114] like Figure 4 As shown, the network architecture of the GECA attention mechanism includes a globally average pooling layer, an activation function, and a depthwise separable convolutional layer connected in sequence. Figure 4 In this context, G represents global average pooling, S represents the sigmoid activation function, DW represents depthwise separable convolution, and K refers to the number of attention heads. For an input W×H×C feature map, after global average pooling, a first intermediate feature map of 1×1×C is obtained; after sigmoid activation, a second intermediate feature map of 1×1×C is obtained; finally, depthwise separable convolution is performed to obtain the output features of the GECA attention mechanism. In a preferred embodiment, the GECA attention mechanism operation is represented by the following formula:

[0115] y i =DW(Sigmoid(G(x) 2i )))

[0116] Where, x 2i The input features are represented by GECA attention mechanism; DW represents depthwise separable convolution; Sigmoid represents activation function operation; G represents global average pooling operation; y i This represents the output characteristics of the GECA attention mechanism.

[0117] Specifically, a GECA (Ghost-Efficient Channel Attention) attention mechanism is added at the feature concatenation point after rank decomposition to enhance the transmission of effective information at the cost of a small number of additional parameters.

[0118] Furthermore, residual jumper connections are added to the basic modules of the backbone network to enhance the information expression capabilities of the basic modules.

[0119] In a preferred embodiment, the SPPFCSPC operation is expressed using the following formula:

[0120]

[0121] Where, x 3iSPPFCSPC represents the input features for the operation; Conv represents the convolution operation; Shuffle represents the channel shuffling operation; Sigmoid represents the activation function operation; MaxPool represents the max pooling operation; y o2 This indicates the output characteristics of the SPPFCSPC operation.

[0122] Specifically, an SPPFCSPC structure is added to the deepest layer of the feature extraction network 100. The SPPFCSPC structure is composed of Spatial Pyramid Pooling Fast (SPPF) and CrossStage Partial Channel (CSPC). The SPPF structure optimizes the parallel pooling layer of SPP into a serial structure, improving the running speed of the pooling layer; the CSPC convolutional structure accelerates the transfer speed of feature data and improves network efficiency by introducing cross-stage connections.

[0123] In the SPPFCSPC module, such as Figure 5 As shown, for the input feature map, three consecutive convolutions are first performed to extract deeper semantic information. Then, three consecutive pooling layers of different sizes are used to extract effective information from different receptive fields. Next, a channel shuffling mechanism is introduced to enhance the information interaction under different pooling receptive fields. Subsequently, feature fusion is performed through stacked convolutional layers to improve the computational efficiency of the model, ensure the accuracy of small target recognition, and reduce the amount of computation.

[0124] In a preferred embodiment, such as Figure 2 As shown, the multi-scale coupled feature fusion structure is a four-scale coupled feature fusion structure, including:

[0125] The first feature fusion layer 201 is used to fuse the fifth intermediate feature and the fourth intermediate feature to obtain the first fused feature;

[0126] The second feature fusion layer 202 is used to fuse the third intermediate feature and the first fusion feature to obtain the second fusion feature;

[0127] The third feature fusion layer 203 is used to fuse the second intermediate feature and the second fused feature to obtain the third fused feature;

[0128] The fourth feature fusion layer 204 is used to fuse the third fusion feature and the third intermediate feature to obtain the fourth fusion feature;

[0129] The fifth feature fusion layer 205 is used to fuse the second fusion feature and the fourth fusion feature to obtain the fifth fusion feature;

[0130] The sixth feature fusion layer 206 is used to fuse the first fusion feature and the fifth fusion feature to obtain the sixth fusion feature.

[0131] Specifically, the traditional YOLOv8 model uses a two-stage upsampling and downsampling feature fusion model to interact with contextual information and performs detection from three dimensions: large, medium, and small. Its ability to interact with contextual information is poor, and its ability to utilize information from the backbone network is also poor.

[0132] This invention proposes a four-scale coupled feature fusion structure, which improves the detection performance of the model without excessively increasing the number of redundant parameters.

[0133] Meanwhile, during the feature fusion process, additional interconnections of residual jumpers of the same scale are added, which to some extent strengthens the information interaction between different receptive fields in the same dimension and effectively improves the contribution of information from different dimensions to the model detection results.

[0134] In a preferred embodiment, the detection head includes four components, such as... Figure 2 As shown, the four detection heads include:

[0135] The first detection head 301 is connected to the sixth feature fusion layer 206 and is used to output a detection result of 160×160×128 dimensions based on the sixth fusion feature.

[0136] The second detection head 302 is connected to the fifth feature fusion layer 205 and is used to output a detection result of 80×80×256 dimensions based on the fifth fusion feature.

[0137] The third detection head 303 is connected to the fourth feature fusion layer 204 and is used to output a 40×40×512 dimension detection result based on the fourth fusion feature.

[0138] The fourth detection head 304 is connected to the third feature fusion layer 203 and is used to output a 20×20×1024 dimension detection result based on the third fusion feature.

[0139] Specifically, after the above feature extraction and feature fusion network, the improved model output includes four dimensions of detection heads: 160×160×128, 80×80×256, 40×40×512, and 20×20×1024.

[0140] Furthermore, the model of this invention employs a transfer learning strategy for pre-training to ensure the generalization of the model until it is iterated until the model basically converges.

[0141] Furthermore, the model that has been iterated to near convergence is trained again to achieve full convergence.

[0142] On the other hand, a power fitting identification system for data center campus power transmission line inspection is also provided, used to implement the power fitting detection method for data center campus power transmission line inspection as described above, such as... Figure 6 As shown, the system includes:

[0143] Dataset construction module 1 is used to acquire images of power fittings, construct the dataset, and divide the dataset into training set, test set, and validation set;

[0144] Model building module 2 is used to construct an electric fitting recognition model based on improved YOLOv8, including a feature extraction network 100, a feature fusion network 200, and a detection head 300. The feature extraction network 100 is based on a lightweight network structure and introduces an attention mechanism and a pyramid pooling structure. The feature fusion network 200 adopts a multi-scale coupled feature fusion structure. The output of the multi-scale coupled feature fusion structure 200 and the feature extraction network 100 are interconnected through the same-scale residual jumper, and an upsampling and downsampling structure is added. The detection head 300 uses multiple detection heads to construct electric fitting recognition models of different dimensions.

[0145] Model training module 3 is connected to dataset construction module 1 and model construction module 2 respectively. It is used to train the power fitting recognition model based on the improved YOLOv8 based on the training set, obtain the trained power fitting recognition model, and use the validation set for validation and the test set for testing to determine the target power fitting recognition model.

[0146] The target detection module 4 is connected to the model training module 3. It is used to acquire new power fitting images and perform target detection on the new power fitting images based on the target power fitting recognition model to obtain detection results.

[0147] Specifically, the traditional EffcientNet backbone network suffers from problems such as a large number of parameters, insufficient effective information mining, and weak power fitting detection capabilities when applied to power line inspection scenarios in data center campuses:

[0148] First, by introducing a rank decomposition convolutional structure into the GMBConv network architecture, the 3×3 convolution is decomposed into a cascaded convolutional structure of 1×3 and 3×1, reducing the number of parameters required by the model and compressing the model size.

[0149] Secondly, the 1×1 dimensionality reduction convolution operation in the inverted bottleneck convolution of the traditional GMBConv network architecture is replaced with the GECA attention mechanism. The GECA attention mechanism can adaptively adjust the weights of each channel in the feature map, highlighting important feature information and suppressing redundant information. This not only compresses the number of redundant parameters but also enhances the effective information interaction between contexts, enabling the model to focus more on key features and improving the detection accuracy of power fittings in the data center campus power transmission line inspection scenario.

[0150] Finally, to better mine fine-grained features and further improve model performance, this embodiment of the invention adopts an SPPFCSPC structure in the deepest layer of the feature extraction network 100. Through continuous stacked serial and parallel convolution operations, features can be extracted and fused from different scales and levels. In the scenario of power line inspection in data center parks, when dealing with targets such as power fittings that have complex structures and diverse features, the SPPFCSPC module can capture more detailed information, thereby improving the model's ability to detect small and complex targets.

[0151] Addressing the multi-scale and accuracy challenges encountered in the identification and detection of power fittings during the feature fusion stage in data center campus power line inspection scenarios:

[0152] First, this embodiment employs a same-scale coupled jumper connection structure. Compared to traditional feature fusion methods that typically focus only on the fusion of features at different scales while neglecting the connections between different features at the same scale, this invention establishes direct connections between feature maps at the same scale, enhancing the information interaction between the feature extraction network 100 and the feature fusion network 200.

[0153] Secondly, this embodiment employs a feature fusion structure with a larger receptive field. By adding an upper and lower fusion structure, the interaction and fusion of deep semantic information and shallow location information are strengthened, enabling the model to establish connections between features at different levels, thereby improving the positioning accuracy and recognition accuracy of power fittings in the data center campus power line inspection scenario.

[0154] The above are merely preferred embodiments of the present invention and are not intended to limit the implementation methods and protection scope of the present invention. Those skilled in the art should recognize that any equivalent substitutions and obvious changes made using the content of this specification and illustrations should be included within the protection scope of the present invention.

Claims

1. A method for inspecting power fittings during power line inspection in a data center campus, characterized in that, include: Step S1: Obtain images of power fittings, construct a dataset, and divide the dataset into a training set, a test set, and a validation set; Step S2: Construct an electric fitting recognition model based on an improved YOLOv8, including a feature extraction network, a feature fusion network, and a detection head. The feature extraction network is based on a lightweight network structure, while introducing an attention mechanism and a pyramid pooling structure. The feature fusion network adopts a multi-scale coupled feature fusion structure, which is interconnected with the output of the feature extraction network through residual jumpers of the same scale, and an upsampling / downsampling structure is added. The detection head uses multiple detection heads to construct electric fitting recognition models of different dimensions. Step S3: Train the power fitting recognition model based on the improved YOLOv8 based on the training set to obtain the trained power fitting recognition model, and use the validation set for validation and the test set for testing to determine the target power fitting recognition model. Step S4: Obtain a new image of the power fittings, and perform target detection on the new image of the power fittings based on the target power fittings recognition model to obtain the detection result.

2. The method for testing power fittings according to claim 1, characterized in that, The steps of the feature extraction network include: Preliminary feature extraction and dimensionality reduction are performed on the input power fitting images to obtain the first intermediate features; The first intermediate feature is subjected to two GMBConv operations to obtain the second intermediate feature. The second intermediate feature is subjected to two GMBConv operations to obtain the third intermediate feature; Perform a GMBConv operation on the third intermediate feature sequentially to obtain the fourth intermediate feature; The SPPFCSPC operation is performed on the fourth intermediate feature to obtain the fifth intermediate feature.

3. The method for testing power fittings according to claim 2, characterized in that, The steps of the GMBConv operation include: Convolution is performed on the input features of the GMBConv operation to obtain the first intermediate sub-feature; The first intermediate sub-feature is subjected to a depthwise separable convolution to obtain a second intermediate sub-feature; and the first intermediate sub-feature is subjected to a depthwise separable convolution to obtain a third intermediate sub-feature. Perform GECA attention operation on the second intermediate feature and the third intermediate feature to obtain the fourth intermediate feature; Convolution is performed on the fourth intermediate feature to obtain the fifth intermediate feature; Perform a Dropout operation on the fifth intermediate feature to obtain the sixth intermediate feature; The residual between the sixth intermediate feature and the input feature of the GMBConv operation is output as the output feature of the GMBConv operation.

4. The method for testing power fittings according to claim 3, characterized in that, In the inverted bottleneck convolutional layer, the 3×3 convolution is decomposed into a first sub-depth separable convolution of 1×3 and a second sub-depth separable convolution of 3×1.

5. The method for testing power fittings according to claim 2, characterized in that, Each GMBConv operation is represented by the following formula: Where, x 1i GMBConv represents the input features; D represents the Dropout operation; Conv represents the convolution operation; E represents the GECA attention mechanism operation; DW represents the depthwise separable convolution operation; y o1 This indicates the output characteristics of the GMBConv operation.

6. The method for testing power fittings according to claim 3, characterized in that, The GECA attention mechanism operation is represented by the following formula: y i =DW(Sigmoid(G(x 2i ))) Where, x 2i The input features are represented by GECA attention mechanism; DW represents depthwise separable convolution; Sigmoid represents activation function operation; G represents global average pooling operation; y i This represents the output characteristics of the GECA attention mechanism.

7. The method for testing power fittings according to claim 2, characterized in that, The SPPFCSPC operation is expressed by the following formula: Where, x 3i SPPFCSPC represents the input features for the operation; Conv represents the convolution operation; Shuffle represents the channel shuffling operation; Sigmoid represents the activation function operation; MaxPool represents the max pooling operation; y o2 This indicates the output characteristics of the SPPFCSPC operation.

8. The method for testing power fittings according to claim 2, characterized in that, The multi-scale coupled feature fusion structure is a four-scale coupled feature fusion structure, configured as follows: The fifth intermediate feature and the fourth intermediate feature are fused to obtain the first fused feature; The third intermediate feature and the first fusion feature are fused together to obtain the second fusion feature; The second intermediate feature and the second fused feature are fused to obtain the third fused feature; The third fusion feature and the third intermediate feature are fused to obtain the fourth fusion feature; The second fusion feature and the fourth fusion feature are fused together to obtain the fifth fusion feature; as well as The first fusion feature and the fifth fusion feature are fused together to obtain the sixth fusion feature.

9. The method for testing power fittings according to claim 8, characterized in that, The detection head includes four, and the four detection heads include: The first detection head is connected to the sixth feature fusion layer and is used to output a 160×160×128 dimension detection result based on the sixth fusion feature. The second detection head is connected to the fifth feature fusion layer and is used to output an 80×80×256-dimensional detection result based on the fifth fusion feature. The third detection head is connected to the fourth feature fusion layer and is used to output a 40×40×512 dimension detection result based on the fourth fusion feature. The fourth detection head is connected to the third feature fusion layer and is used to output a 20×20×1024 dimension detection result based on the third fusion feature.

10. A power fitting identification system for power line inspection in a data center campus, characterized in that, The method for testing power fittings for carrying out the inspection of power transmission lines in a data center campus as described in any one of claims 1-9 includes: The dataset construction module is used to acquire images of power fittings, construct the dataset, and divide the dataset into training, testing, and validation sets. The model building module is used to construct an electric fitting recognition model based on an improved YOLOv8, including a feature extraction network, a feature fusion network, and a detection head. The feature extraction network is based on a lightweight network structure and incorporates an attention mechanism and a pyramid pooling structure. The feature fusion network adopts a multi-scale coupled feature fusion structure, which is interconnected with the output of the feature extraction network through residual jumpers of the same scale, and an upsampling / downsampling structure is added. The detection head uses multiple detection heads to construct electric fitting recognition models of different dimensions. The model training module is connected to the dataset construction module and the model construction module respectively. It is used to train the power fitting recognition model based on the improved YOLOv8 based on the training set to obtain the trained power fitting recognition model. The model is then validated using the validation set and tested using the test set to determine the target power fitting recognition model. The target detection module, connected to the model training module, is used to acquire new power fitting images and perform target detection on the new power fitting images based on the target power fitting recognition model to obtain detection results.

Images