Inspection system and method for power transmission line in data center park

Abstract

The invention provides a data center park power transmission line inspection system and method, and relates to the technical field of computer vision, and the method comprises the steps: carrying out the multiple continuous deformable convolution of a remotely collected power transmission line image, carrying out the preliminary feature extraction, and carrying out the deep feature mining through spatial pyramid pooling and triple attention, multiple layers of feature maps with different scales are obtained through co-extraction; performing multi-stage feature fusion on the feature maps of all scales to output fusion information of four scales; and a plurality of detection heads are arranged to detect the different fusion information to obtain the target category of the fusion information and the position of the fusion information in the power transmission line image. The method has the beneficial effects that attention on fine-grained deep information and transmission of effective information are enhanced through deep feature mining, and the robustness and generalization of the model are improved; through multi-stage feature fusion and detection, interaction between deep semantic information and shallow position information is enhanced, and the detection precision and accuracy of the system are improved.

Description

A system and method for inspecting power transmission lines in a data center campus. Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to an inspection system and method for power transmission lines in a data center campus. Background Technology

[0002] With the rapid development of intelligent computing data centers, the requirements for power systems are becoming increasingly stringent. Besides stable power generation, a stable, fault-free, and trouble-free power transmission system is a crucial prerequisite for ensuring the smooth operation of data centers. Therefore, inspecting the overhead power lines of data centers is an essential means to ensure the stability of the data center's power system.

[0003] The current main inspection methods for overhead power lines already constructed for data centers include: 1. Sending personnel to conduct on-site inspections of completed and operational overhead power lines. While this method is relatively accurate, it suffers from high risk, high labor costs, and strong subjectivity. 2. Electronic measurement-based inspection methods primarily measure changes in the infrared spectrum around the equipment after insulator damage, as well as changes in electrical and optical parameters before and after damage. Although this method can effectively reduce the influence of subjectivity on the inspection results, it does not fundamentally change the requirement for on-site inspections by personnel. It still carries certain risks during signal acquisition and cannot accurately detect insulator obstacles such as bird nests and tree branches. 3. Inspection methods based on traditional image processing algorithms mainly involve using drones to collect image data, transmitting it back, and then using manual or image processing technology for direct or indirect judgment and inspection. This method reduces labor costs to some extent and has a high safety factor. However, this method suffers from poor generalization, low robustness, and computational redundancy. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention provides an inspection system for power transmission lines in a central industrial park, comprising: an image acquisition terminal for acquiring images of power transmission lines during remote inspections by inspection personnel; a feature extraction module connected to the image acquisition terminal for performing multiple consecutive deformable convolutions on the power transmission line images for preliminary feature extraction, and performing deep feature mining through spatial pyramid pooling and triple attention to extract multiple layers of feature maps at different scales; a feature fusion module connected to the feature extraction module for performing multi-stage feature fusion of the feature maps at each scale to output fused information at four scales; and a detection module connected to the feature fusion module and the image acquisition terminal for setting up multiple detection heads to detect each different fused information, determining the target category and position of each fused information in the power transmission line image, and displaying the results on the image acquisition terminal for inspection personnel to view.

[0005] Preferably, the feature map includes a shallow feature map, a shallower feature map, and a deep feature map with progressively increasing scale; the feature extraction module is configured with a backbone network model for feature extraction, the backbone network model including a first feature extraction unit, a second feature extraction unit, a third feature extraction unit, a fourth feature extraction unit, and a fifth feature extraction unit connected in sequence by input and output; the third feature extraction unit is used to output the shallow feature map; the fourth feature extraction unit is used to output the shallower feature map; and the fifth feature extraction unit is used to output the deep feature map.

[0006] Preferably, the first feature extraction unit includes a deformable convolution operation subunit; the second feature extraction unit, the third feature extraction unit, and the fourth feature extraction unit each include a deformable convolution operation subunit and a feature extraction operation subunit with their inputs and outputs connected in sequence; the fifth feature extraction unit includes a deformable convolution operation subunit, a feature extraction operation subunit, a spatial pyramid pooling subunit, and a triple attention subunit with their inputs and outputs connected in sequence.

[0007] Preferably, the spatial pyramid pooling subunit includes: a first ConvBNReLU module, the input of which is connected to the output of the feature extraction operation subunit in the fifth feature extraction unit; four max pooling layer modules with their inputs and outputs connected in sequence, the input of the first max pooling layer module being connected to the output of the first ConvBNReLU module; and a concatenation module, which concatenates the outputs of each max pooling layer module and the output of the first ConvBNReLU module and outputs them to a second ConvBNReLU module, which then processes the outputs and outputs them to the triple attention subunit.

[0008] Preferably, the fusion information includes first to fourth fusion information; the feature fusion module includes an upsampling branch and a downsampling branch; the upsampling branch includes a first upsampling unit and a second upsampling unit, the first upsampling unit is used to upsample the deep feature map and fuse it with the features of the shallower feature map to obtain the first fusion information, and the second upsampling unit is used to upsample the first fusion information and fuse it with the features of the shallow feature map to obtain the second fusion information; the downsampling branch includes a first downsampling unit and a second downsampling unit, the first downsampling unit is used to deformably convolve the second fusion information and fuse it with the features of the second fusion information to obtain the third fusion information, and the second downsampling unit is used to deformably convolve the third fusion information and fuse it with the features of the first fusion information to obtain the fourth fusion information.

[0009] Preferably, the first upsampling unit and the second upsampling unit have the same structure, each including: an upsampling layer, the output of which is connected to the input of the stitching layer, and the output of which is connected to the input of the feature extraction calculation layer; the upsampling layer of the first upsampling unit inputs the deep feature map, and the stitching layer of the first upsampling unit also inputs the shallower feature map; the upsampling layer of the second upsampling unit inputs the first fusion information, and the stitching layer of the second upsampling unit also inputs the shallow feature map.

[0010] Preferably, the first downsampling unit and the second downsampling unit have the same structure, both including: a deformable convolutional layer, the output of which is connected to the input of the stitching layer, and the output of which is connected to the input of the feature extraction calculation layer; the deformable convolutional layer of the first downsampling unit inputs the second fusion information, and the stitching layer of the first downsampling unit also inputs the second fusion information; the deformable convolutional layer of the second downsampling unit inputs the third fusion information, and the stitching layer of the second upsampling unit also inputs the first fusion information.

[0011] Preferably, the upsampling layer uses deformable convolution operation for feature extraction.

[0012] Preferably, the step size of each maximum pooling layer module increases sequentially according to the connection order.

[0013] The present invention also provides a method for inspecting power transmission lines in a central park, applied to the aforementioned inspection system, comprising: step S1, wherein the inspection system performs multiple consecutive deformable convolutions on the input power transmission line image for preliminary feature extraction, and performs deep feature mining through spatial pyramid pooling and triple attention, thereby extracting multiple feature maps of different scales; step S2, wherein the inspection system performs multi-stage feature fusion on the feature maps of each scale to output fused information at four scales; step S3, wherein the inspection system sets up multiple detection heads to detect each different fused information, thereby determining the target category and position of each fused information in the power transmission line image.

[0014] The above technical solution has the following advantages or beneficial effects: 1. The image acquisition terminal can be a remote image acquisition device such as a drone, which allows inspection personnel to conduct remote inspections and image acquisition. This eliminates the need for manual on-site testing of electrical characteristics and other aspects, thereby reducing labor costs and improving human safety.

[0015] 2. Enhanced multi-scale target detection capability and generalization: The feature extraction module performs multiple consecutive deformable convolutions on the input image, enabling more flexible extraction of features from targets of the same category at different scales, effectively solving the multi-scale problem in data center power line inspection. Furthermore, this module utilizes spatial pyramid pooling and triple attention for deep feature mining, enhancing the system's focus on fine-grained deep information and the transmission of effective information, thereby significantly improving the model's robustness and generalization, allowing it to better cope with complex campus background interference.

[0016] 3. Improved accuracy and precision of model detection: The feature fusion module, through multi-stage feature fusion, effectively couples feature maps of different scales, enhancing the interaction between deep semantic information and shallow location information. Furthermore, the detection module employs multiple detection heads to detect the fused multi-scale information, allowing shallow feature information to contribute to the final detection results, thereby significantly improving the system's detection accuracy and precision. Attached Figure Description

[0017] Figure 1 is a structural schematic diagram of a data center campus power transmission line inspection system according to a preferred embodiment of the present invention; Figure 2 is a schematic diagram of deformable convolution operation according to a preferred embodiment of the present invention; Figure 3 is a schematic diagram of multi-scale target detection according to a preferred embodiment of the present invention; Figure 4 is a schematic diagram comparing the SPPF technology in the prior art with the improved SimSPPF unit in the present invention according to a preferred embodiment of the present invention; Figure 5 is a schematic diagram of the target category and position in the power transmission line image of each fused information in the detected power transmission line image according to a preferred embodiment of the present invention; Figure 6 is a flowchart of a data center campus power transmission line inspection method according to a preferred embodiment of the present invention. Detailed Implementation

[0018] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. The present invention is not limited to this embodiment; other embodiments that conform to the spirit of the present invention may also fall within the scope of the present invention.

[0019] In a preferred embodiment of the present invention, based on the aforementioned problems existing in the prior art, an inspection system for power transmission lines in a data center campus is provided, comprising: an image acquisition terminal, used to acquire images of power transmission lines during remote inspections by inspection personnel; a feature extraction module 1, connected to the image acquisition terminal, used to perform multiple consecutive deformable convolutions on the power transmission line images input to the inspection system for preliminary feature extraction, and to perform deep feature mining through spatial pyramid pooling and triple attention, thereby extracting multiple feature maps of different scales; a feature fusion module 2, connected to the feature extraction module 1, used to perform multi-stage feature fusion of the feature maps at each scale to output fusion information at four scales; and a detection module 3, connected to the feature fusion module 2 and the image acquisition terminal, used to set up multiple detection heads to detect each different fusion information, obtain the target category and position of each fusion information in the power transmission line image, and display it on the image acquisition terminal for inspection personnel to view.

[0020] Specifically, this embodiment provides an inspection system for power transmission lines in a central park, which can achieve the following beneficial effects: 1. Enhanced multi-scale target detection capability and generalization. The image acquisition terminal of this system can be a remote image acquisition device such as a drone, which allows inspection personnel to conduct remote inspections and image acquisition. This eliminates the need for manual on-site inspections of electrical and optical characteristics, thereby reducing labor costs and improving the safety of personnel.

[0021] Furthermore, a feature extraction module processes the power transmission line images. This module employs multiple consecutive deformable convolutions for feature extraction, enabling the model to more flexibly perceive and adapt to target features of the same category at different scales, effectively addressing the multi-scale problem commonly encountered in data center power transmission line inspections. In addition, this module introduces Spatial Pyramid Pooling (SPP) and a triple attention mechanism for mining deep features. This design enhances the system's focus on fine-grained deep information and the transmission of effective information, enabling it to better cope with complex background interference in data center campuses and significantly improving the model's robustness and generalization.

[0022] 2. Improving the Accuracy and Precision of Model Detection: The feature fusion module in this system effectively couples feature maps of different scales through multi-stage feature fusion, enhancing the interaction between deep semantic information and shallow location information. The detection module employs multiple detection heads to detect the fused information at four different scales. This design allows shallow feature information to contribute to the final detection results, significantly improving the system's detection accuracy and precision, and reducing missed or false detections caused by complex backgrounds and dense target areas.

[0023] 3. Optimized Computational Efficiency and Practical Application Value: This system, through an optimized feature extraction module, effectively reduces computational redundancy while maintaining high-precision detection, overcoming the inefficiency of traditional image processing methods. This makes the entire inspection system more practical and advantageous in real-world applications, providing an efficient and reliable solution for data center power inspection.

[0024] Compared to traditional manual inspection, this system can reduce the risk factor and labor costs; compared to traditional electronic measurement methods, this system can accurately detect insulating obstacles such as bird nests and tree branches; compared to traditional image processing algorithms, this system has better generalization and robustness.

[0025] More specifically, compared to other power system transmission lines, data center power transmission lines are characterized by high system complexity, high density, complex campus environment, and compact layout. Therefore, directly using existing algorithms for data center power transmission line inspection cannot effectively solve the complex background interference and multi-scale characteristics of the same target. This invention, based on the YOLOv8 network model, addresses the multi-scale characteristics and background interference problems of existing intelligent algorithms in power transmission lines, proposing a multi-scale, highly information-interactive data center power transmission line inspection system.

[0026] Since the YOLOv8 model has been applied in object detection, this invention uses the YOLOv8n model as the main framework. The backbone network is improved based on deformable convolution (DCNV2) and residual networks, taking into account the multi-scale and complex background characteristics of data center power transmission lines. It employs deep separable convolution, which is more suitable for multi-scale targets, for feature extraction. Furthermore, to further enhance the model's ability to extract features of the same target in different dimensions, a Triplet attention module is inserted at the deepest layer of the backbone network to strengthen the model's ability to locally focus on deep, high-dimensional, effective information. In addition, a larger-scale SimSPPF module (spatial pyramid pooling) is used to extract deep semantic information, effectively mining deeper, finer-grained information.

[0027] In the feature fusion and detection stages, deformable convolution is used for feature extraction during the upsampling phase to improve the information coupling between the upper and lower stages. Furthermore, the first fused information output from the FPN network (the feature fusion module in this system, consisting of a bottom-up path, a top-down path, and lateral connections) is used for detection, strengthening the contribution of this information to the final detection result and thus improving the model's detection capability. Therefore, compared to other algorithms, this invention exhibits better application performance in data center power inspection scenarios.

[0028] In a preferred embodiment of the present invention, the feature map includes a shallow feature map, a shallower feature map, and a deep feature map with progressively increasing scale; the feature extraction module 1 is configured with a backbone network model for feature extraction, the backbone network model including a first feature extraction unit 11, a second feature extraction unit 12, a third feature extraction unit 13, a fourth feature extraction unit 14, and a fifth feature extraction unit 15 connected in sequence by input and output; the third feature extraction unit 13 is used to output the shallow feature map; the fourth feature extraction unit 14 is used to output the shallower feature map; and the fifth feature extraction unit 15 is used to output the deep feature map.

[0029] Specifically, as shown in Figure 1, the feature extraction module of this invention corresponds to the backbone network in the YOLOv8 network model. Through multiple consecutive deformable convolution extractions by the first to fourth feature extraction units, shallow feature maps and relatively shallow feature maps are extracted. Then, deep feature mining is performed by the fifth feature extraction unit to extract deep feature maps, resulting in a total of three layers of feature maps at different scales.

[0030] This invention achieves effective extraction of multi-scale features from transmission line images by configuring a backbone network model containing five feature extraction units in the feature extraction module. Its beneficial effects are reflected in the following aspects: 1. Refined extraction and hierarchical processing of multi-scale features. By configuring the first to fifth feature extraction units, this invention can extract shallow feature maps, relatively shallow feature maps, and deep feature maps sequentially according to scale. This hierarchical and refined feature extraction method enables the system to capture target information at different scales in transmission line images. The shallow feature map (output by the third feature extraction unit 13) retains more location and detail information, suitable for detecting small targets; the relatively shallow feature map (output by the fourth feature extraction unit 14) incorporates some semantic information, suitable for detecting medium-sized targets; while the deep feature map (output by the fifth feature extraction unit 15) contains rich semantic information, suitable for recognizing large targets. This design provides a solid foundation for subsequent multi-scale target detection.

[0031] 2. Enhanced Resistance to Complex Background Interference: This invention employs deformable convolution for feature extraction, enabling the backbone network model to flexibly adjust its sampling position. This allows it to better adapt to object deformation and scale changes caused by complex background interference such as lighting variations, angle differences, and trees and bird nests in power transmission line images. This adaptive feature extraction capability significantly improves the system's robustness, enabling it to maintain high detection performance even in complex and ever-changing data center environments.

[0032] 3. Ensuring the Effectiveness of Subsequent Feature Fusion: Through the hierarchical design of five feature extraction units, this invention ensures that rich information at different levels can be obtained during the feature extraction stage. The outputs of shallow, mid-shallow, and deep feature maps provide high-quality input for the subsequent feature fusion module, enabling the feature fusion process to more effectively combine semantic and positional information, thereby ultimately improving the detection accuracy of the entire system.

[0033] In a preferred embodiment of the present invention, the first feature extraction unit 11 includes a deformable convolution operation subunit; the second, third, and fourth feature extraction units each include a deformable convolution operation subunit and a feature extraction operation subunit whose inputs and outputs are connected in sequence; the fifth feature extraction unit includes a deformable convolution operation subunit, a feature extraction operation subunit, a spatial pyramid pooling subunit, and a triple attention subunit whose inputs and outputs are connected in sequence.

[0034] Specifically, as shown in Figure 1, the first feature extraction unit 11 includes a deformable convolution operation subunit (DCNV2); the second feature extraction unit 12, the third feature extraction unit 13, and the fourth feature extraction unit 14 each include a deformable convolution operation subunit (DCNV2) and a feature extraction operation subunit (C2f) with the input and output connected in sequence; the fifth feature extraction unit 15 includes a deformable convolution operation subunit (DCNV2), a feature extraction operation subunit (C2f), a spatial pyramid pooling subunit (SimSPPF), and a triple attention subunit (Trilpet Attention) with the input and output connected in sequence.

[0035] By introducing deformable convolutions to replace traditional convolutions for feature extraction in the backbone network, the model's ability to detect objects of the same class at different scales is improved (Figure 3 shows a schematic diagram of multi-scale target detection). The feature extraction expressions for each stage of the backbone network are as follows:

[0036]

[0037] out in the formula i Indicates the output of each stage; in i C represents the input at each stage, C2f represents the feature extraction learning operation, DC represents the deformable convolution operation, T represents the TripletAttention mechanism, and S represents the SimSPPF unit operation.

[0038] Figure 2 illustrates the operation of Deformable Convolution. This figure compares and contrasts Deformable Convolution with Standard Convolution: Left side - Standard Convolution: This represents the standard convolution operation. Its sampling points are a fixed, square grid (e.g., a 3x3 rectangle). Regardless of the shape or size of the target in the image, it uses this fixed-shape grid to extract features.

[0039] The right side – Deformable Convolution: This section represents deformable convolution. Unlike standard convolution, its sampling points are not fixed. By adding an extra offset to the network, these sampling points can freely move and deform according to the actual shape and pose of the object in the image. The arrows in the diagram visually represent the offset of the sampling points from their original regular positions.

[0040] The core idea of ​​Figure 2 is that standard convolution becomes rigid in its fixed sampling method when dealing with targets of varying shapes and scales (such as insulators at different angles) during power transmission line inspections. Deformable convolution, on the other hand, is more flexible; it can automatically learn to adjust the position of sampling points to better fit the actual shape of the target, thus extracting target features more effectively. The purpose of introducing deformable convolution into this system is to improve the model's ability to detect targets of different scales and shapes.

[0041] Furthermore, SimSPPF units with a larger receptive field are added to the deepest layer of the backbone network to mine more fine-grained deep information and improve the model's attention to the multi-dimensional fine-grained deep features of overhead transmission lines.

[0042] Furthermore, the spatial pyramid pooling subunit used in this system is an improved SimSPPF unit obtained by improving the spatial pyramid pooling in the existing technology. The TripletAttention attention mechanism is added after the improved SimSPPF unit to enhance the model's ability to transmit deep effective information and further optimize the model's attention to deep effective semantic information.

[0043] In a preferred embodiment of the present invention, as shown in Figure 2, the spatial pyramid pooling subunit Triplet Attention includes: a first ConvBNReLU module, the input of which is connected to the output of the feature extraction operation subunit in the fifth feature extraction unit; four inputs and outputs connected sequentially to a max pooling layer module (MAXPool, Stride is the stride), the input of the first max pooling layer module being connected to the output of the first ConvBNReLU module; and a concatenation module, which concatenates the outputs of each max pooling layer module and the output of the first ConvBNReLU module and outputs them to a second ConvBNReLU module, which then processes the outputs and outputs them to the triplet attention subunit.

[0044] In a preferred embodiment of the present invention, the step size of each maximum pooling layer module increases sequentially according to the connection order.

[0045] Specifically, the spatial pyramid pooling subunit in this embodiment differs from the spatial pyramid pooling in the prior art, as shown in Figure 4. The prior art only has three max pooling layer modules with step sizes of 5, 9, and 13, respectively. The improved SimSPPF unit of this invention has one more max pooling layer (MAXPool) with a step size of 17 compared to the existing SPPF and SinSPPF technologies. In Figure 4, Concat represents the splicing module.

[0046] This invention improves the spatial pyramid pooling subunit by combining it with a triple attention subunit, achieving significant benefits in feature extraction. Specifically: 1. Enhanced extraction and representation of deep features: This invention configures an improved SimSPPF unit and a TripletAttention mechanism in the fifth feature extraction unit, strengthening the model's ability to transmit effective deep information and further optimizing the model's attention to effective deep semantic information. This design enables the system to more effectively mine and utilize finer-grained deep features in overhead transmission lines, improving its understanding of complex scenarios.

[0047] 2. Enhancing the Model's Adaptability to Multi-Scale Targets: Compared to existing technologies, the improved SimSPPF unit in this invention adds a max-pooling layer with a stride of 17, thereby capturing information at larger scales and providing richer multi-scale information for subsequent feature fusion. This improvement enables the system to perform better in detecting targets of different sizes, enhancing the model's robustness and generalization ability.

[0048] 3. Optimized Feature Fusion and Information Transmission: The spatial pyramid pooling subunit in this invention uses a splicing module to concatenate the outputs of each max-pooling layer module and the first ConvBNReLU module, and then outputs this concatenation to the second ConvBNReLU module for processing. This design ensures the effective transmission and fusion of deep feature information, providing a more discriminative feature representation for subsequent detection modules, thereby significantly improving the system's detection accuracy.

[0049] In a preferred embodiment of the present invention, the fusion information includes first to fourth fusion information; the feature fusion module 2 includes an upsampling branch 21 and a downsampling branch 22; the upsampling branch 21 includes a first upsampling unit 211 and a second upsampling unit 212, the first upsampling unit 211 is used to upsample the deep feature map and fuse it with the features of the shallower feature map to obtain the first fusion information, and the second upsampling unit 212 is used to upsample the first fusion information and fuse it with the features of the shallow feature map to obtain the second fusion information; the downsampling branch 22 includes a first downsampling unit 221 and a second downsampling unit 222, the first downsampling unit 221 is used to deformably convolve the second fusion information and fuse it with the features of the second fusion information to obtain the third fusion information and fuse it with the features of the first fusion information to obtain the fourth fusion information.

[0050] Specifically, as shown in Figure 1 in this embodiment, the feature fusion module 2 adopts an FPN network structure, including a bottom-up line-up sampling branch 21 and a top-down line-down sampling branch 22, with the two branches connected laterally.

[0051] In a preferred embodiment of the present invention, the first upsampling unit and the second upsampling unit have the same structure, both including: an upsampling layer Upsample, the output of which is connected to the input of the concatenation layer Concat, and the output of which is connected to the input of the feature extraction calculation layer C2f; the upsampling layer of the first upsampling unit 211 inputs a deep feature map, and the concatenation layer of the first upsampling unit also inputs a shallower feature map; the upsampling layer of the second upsampling unit 212 inputs first fusion information, and the concatenation layer of the second upsampling unit also inputs a shallower feature map.

[0052] In a preferred embodiment of the present invention, the upsampling layer uses deformable convolution operation for feature extraction.

[0053] In a preferred embodiment of the present invention, the first downsampling unit 221 and the second downsampling unit 222 have the same structure, both including: a deformable convolutional layer DCNV2, the output of the deformable convolutional layer being connected to the input of the concatenation layer Concat, and the output of the concatenation layer being connected to the input of the feature extraction calculation layer C2f; the deformable convolutional layer of the first downsampling unit inputs second fusion information, and the concatenation layer of the first downsampling unit also inputs second fusion information; the deformable convolutional layer of the second downsampling unit inputs third fusion information, and the concatenation layer of the second upsampling unit also inputs first fusion information.

[0054] Specifically, this invention improves the feature fusion module by combining upsampling and downsampling branches, effectively enhancing the efficiency and quality of feature fusion. Its beneficial effects are reflected in the following aspects: 1. Strengthening the deep fusion and interaction of multi-scale information. This invention fuses deep semantic information with shallow positional information through an upsampling branch. Specifically, the first upsampling unit fuses the deep feature map with a shallower feature map to obtain first fused information, and the second upsampling unit fuses the first fused information with the shallow feature map to obtain second fused information. This bottom-up upsampling path ensures that the rich semantic information in the deep features can be gradually transmitted to the shallow layer, enhancing the model's ability to recognize targets at different scales.

[0055] 2. Enhancing the Effective Utilization of Feature Information: In the downsampling branch, this invention extracts features through deformable convolution and combines it with a splicing layer for information fusion. The first downsampling unit processes the second fused information and fuses it with itself to obtain the third fused information; the second downsampling unit then fuses the third fused information with the first fused information to obtain the fourth fused information. This top-down downsampling path achieves effective backflow and reuse of feature information, avoids potential semantic ambiguity, and ensures that feature information at different levels can interact and fuse more fully, thereby improving the detection accuracy of the entire system.

[0056] 3. Enhancing the robustness of the model in complex environments: This invention introduces deformable convolution in the downsampling branch, which allows the system to more flexibly adapt to changes in target shape and scale during feature fusion, avoiding potential semantic ambiguity issues that may arise during contextual information transmission. In complex and compact environments such as data center campuses, this design can better handle the challenges posed by target occlusion or deformation, thereby improving the system's robustness.

[0057] The first fusion information after the first stage of upsampling (denoted as ∂ in Figure 1) is also used for detection to improve the contribution of this stage information to the model detection results and realize the four-scale coupled feature fusion structure.

[0058] Specifically, this system is based on the YOLOv8 network model and has been improved. Therefore, before using this system, the power line images input to the inspection system are augmented with data. The system is then trained by dividing the training, testing and validation sets into a 9:1:1 ratio with existing open-source datasets, including CPLD datasets and a certain number of image architecture datasets collected from IDC data centers, in order to better suit the application.

[0059] After the aforementioned backbone network 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.

[0060] Figure 5 shows a schematic diagram of the target categories and positions of various fused information in the transmission line image obtained by this system.

[0061] The present invention also provides a method for inspecting power transmission lines in a central park, applied to the aforementioned inspection system, as shown in Figure 6, comprising: Step S1, the inspection system performs multiple consecutive deformable convolutions on the input power transmission line image for preliminary feature extraction, and performs deep feature mining through spatial pyramid pooling and triple attention, extracting multiple feature maps of different scales; Step S2, the inspection system performs multi-stage feature fusion on the feature maps of each scale to output fused information of four scales; Step S3, the inspection system sets up multiple detection heads to detect each different fused information, and obtains the target category and position of each fused information in the power transmission line image.

[0062] 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. An inspection system for power transmission lines in a data center campus, characterized in that, include: Image acquisition terminal, used to acquire images of power transmission lines during remote inspections by inspection personnel; The feature extraction module, connected to the image acquisition terminal, is used to perform preliminary feature extraction by performing multiple consecutive deformable convolutions on the transmission line image, and to perform deep feature mining through spatial pyramid pooling and triple attention, thereby extracting multiple feature maps of different scales. The feature fusion module, connected to the feature extraction module, is used to perform multi-stage feature fusion on the feature maps at each scale to output fusion information at four scales; The detection module, connected to the feature fusion module and the image acquisition terminal, is used to set up multiple detection heads to detect different fusion information, obtain the target category and position of each fusion information in the transmission line image, and display it on the image acquisition terminal for inspection personnel to view.

2. The inspection system according to claim 1, characterized in that, The feature map includes shallow feature maps, relatively shallow feature maps, and deep feature maps with progressively increasing scales. The feature extraction module is configured with a backbone network model for feature extraction. The backbone network model includes a first feature extraction unit, a second feature extraction unit, a third feature extraction unit, a fourth feature extraction unit, and a fifth feature extraction unit whose inputs and outputs are connected in sequence. The third feature extraction unit is used to output the shallow feature map. The fourth feature extraction unit is used to output the relatively shallow feature map. The fifth feature extraction unit is used to output the deep feature map.

3. The inspection system according to claim 2, characterized in that, The first feature extraction unit includes a deformable convolution operation subunit; the second feature extraction unit, the third feature extraction unit, and the fourth feature extraction unit each include a deformable convolution operation subunit and a feature extraction operation subunit with the input and output connected in sequence; the fifth feature extraction unit includes a deformable convolution operation subunit, a feature extraction operation subunit, a spatial pyramid pooling subunit, and a triple attention subunit with the input and output connected in sequence.

4. The inspection system according to claim 3, characterized in that, The spatial pyramid pooling subunit includes: a first ConvBNReLU module, the input of which is connected to the output of the feature extraction operation subunit in the fifth feature extraction unit; four max pooling layer modules with their inputs and outputs connected in sequence, the input of the first max pooling layer module being connected to the output of the first ConvBNReLU module; and a concatenation module, which concatenates the outputs of each max pooling layer module and the output of the first ConvBNReLU module and outputs them to a second ConvBNReLU module, which then processes the outputs and outputs them to the triple attention subunit.

5. The inspection system according to claim 2, characterized in that, The fusion information includes first to fourth fusion information; the feature fusion module includes an upsampling branch and a downsampling branch; the upsampling branch includes a first upsampling unit and a second upsampling unit, the first upsampling unit is used to upsample the deep feature map and fuse it with the features of the shallower feature map to obtain the first fusion information, and the second upsampling unit is used to upsample the first fusion information and fuse it with the features of the shallow feature map to obtain the second fusion information; the downsampling branch includes a first downsampling unit and a second downsampling unit, the first downsampling unit is used to deformably convolve the second fusion information and fuse it with the features of the second fusion information to obtain the third fusion information, and the second downsampling unit is used to deformably convolve the third fusion information and fuse it with the features of the first fusion information to obtain the fourth fusion information.

6. The inspection system according to claim 5, characterized in that, The first upsampling unit and the second upsampling unit have the same structure, both including: an upsampling layer, the output of which is connected to the input of the stitching layer, and the output of which is connected to the input of the feature extraction calculation layer; the upsampling layer of the first upsampling unit inputs the deep feature map, and the stitching layer of the first upsampling unit also inputs the shallower feature map; the upsampling layer of the second upsampling unit inputs the first fusion information, and the stitching layer of the second upsampling unit also inputs the shallow feature map.

7. The inspection system according to claim 5, characterized in that, The first downsampling unit and the second downsampling unit have the same structure, both including: a deformable convolutional layer, the output of which is connected to the input of the stitching layer, and the output of which is connected to the input of the feature extraction calculation layer; the deformable convolutional layer of the first downsampling unit inputs the second fusion information, and the stitching layer of the first downsampling unit also inputs the second fusion information; the deformable convolutional layer of the second downsampling unit inputs the third fusion information, and the stitching layer of the second upsampling unit also inputs the first fusion information.

8. The inspection system according to claim 6, characterized in that, The upsampling layer uses deformable convolution operations for feature extraction.

9. The inspection system according to claim 4, characterized in that, The step size of each maximum pooling layer module increases sequentially according to the connection order.

10. A method for inspecting power transmission lines in a central industrial park, applied to the inspection system described in any one of claims 1-9, characterized in that, include: Step S1: During the inspection of the power transmission line by the inspection personnel, the inspection system collects images of the power transmission line, performs multiple consecutive deformable convolutions on the images to perform preliminary feature extraction, and performs deep feature mining through spatial pyramid pooling and triple attention to extract multiple feature maps of different scales; Step S2: The inspection system performs multi-stage feature fusion on the feature maps of each scale to output fused information of four scales. Step S3: The inspection system is equipped with multiple detection heads to detect each different fused information, thereby determining the target category and position of each fused information in the transmission line image.

Images