Power transmission line hidden danger detection method

Abstract

The invention provides a power transmission line hidden danger detection method. The method comprises the following steps: S1, constructing a power transmission line hidden danger target detection data set; s2, constructing a power transmission line hidden danger detection network model; s3, training a hidden danger detection network model based on the power transmission line; and S4, detecting a to-be-detected hidden danger target of the power transmission line, inputting a picture containing the hidden danger target of the power transmission line into the trained hidden danger detection network model of the power transmission line, and outputting a detection result of the position of the hidden danger target of the power transmission line in the picture. According to the method, high-precision and rapid detection of potential hazards in the power transmission line can be realized through a deep learning technology.

Description

Technical Field

[0001] The present invention relates to the field of power system safety detection, and in particular to a method for detecting hidden dangers of power transmission lines. Background Art

[0002] As a key component of the power system, the safety and stability of transmission lines are of vital importance. The purpose of intelligent inspection of transmission lines is to identify the components and power lines in the transmission lines, as well as the location information of hidden dangers, to provide support for subsequent power maintenance. Transmission line hidden danger detection methods are mainly divided into traditional methods based on image processing and methods based on deep learning. Traditional methods require manual extraction of target features, which are applicable to a single scenario, inflexible, and consume a lot of manpower and time costs, so they cannot meet actual inspection needs. The method based on deep learning has good background adaptability, good detection effect and fast speed, and is now widely used in the task of transmission line hidden danger detection. Although the current target detection algorithm has achieved excellent detection results, it still has certain limitations in the detection of hidden danger targets in complex transmission lines.

[0003] During the inspection of transmission lines, since the lines are often built in complex environments, such as mountainous areas and urban edges, there are problems such as occlusion and small targets. This leads to the fact that traditional algorithms have low recognition accuracy in complex scenarios, large consumption of computing resources, and insufficient real-time performance when completing the task of transmission line hidden danger detection. Therefore, there is an urgent need to provide a transmission line hidden danger detection method with high accuracy and strong real-time performance. Summary of the invention

[0004] The purpose of the present invention is to solve the technical problems in the above background, and a method for detecting hidden dangers of power transmission lines is proposed, comprising the following steps:

[0005] S1. Construct a data set for detecting hidden dangers of power transmission lines;

[0006] S2. Construct a network model for detecting hidden dangers in power transmission lines;

[0007] S3, training a transmission line hidden danger detection network model;

[0008] S4. Detect the transmission line hidden danger target to be detected, input the picture containing the transmission line hidden danger target into the trained transmission line hidden danger detection network model, and output the detection result of the location of the transmission line hidden danger target in the picture.

[0009] In a preferred embodiment, step S2 further comprises the following steps:

[0010] The GhostConv module is introduced into the S21 and yolov9 network architectures;

[0011] S22. Design and add a coordinate attention module to optimize the model's ability to locate object structures by embedding position information into channel attention.

[0012] S23. The inverted residual moving module is adopted, CNN and Transformer are integrated, and the detection efficiency of the model for complex transmission line hidden danger targets is enhanced by integrating local features and global context information.

[0013] In a preferred solution, the GhostConv module in step S21 includes a traditional convolutional layer and a lightweight linear transformation layer.

[0014] In the preferred solution, the GhostConv module extracts the feature map, including the following steps: first, using part of the convolution kernel to generate part of the intrinsic feature map, then generating redundant features through linear change operation, and finally fusion of the features as the output of the GhostConv module, which is specifically expressed by the following mathematical formula:

[0015] Y′=X*f′+b;

[0016]

[0017] In the above formula, Y′∈R h′×w′×m represents the output feature, f′∈R c×k×k×m Represents the convolution kernel. The number of channels m of the output feature map of the convolution operation is smaller than the number of channels n of the traditional convolution layer, y′ i (i=1,...,m) are the m feature maps of Y′. Each feature map in Y′ must be processed by the linear operation Φ i,j (j=1,...,s) to obtain s feature maps, so m feature maps are linearly operated to obtain m·s feature maps. The final m·s feature maps are consistent with the number of feature maps output by traditional convolution, that is, n=m·s. Among the s linear transformations of each feature, the last linear transformation is the identity transformation.

[0018] In the preferred scheme, the coordinate attention module decomposes the channel attention into two one-dimensional feature encoding processes, aggregating features in the horizontal and vertical directions respectively, and retaining fine-grained position information.

[0019] In a preferred solution, the coordinate attention module includes two steps: coordinate information embedding and coordinate attention generation.

[0020] In the preferred solution, in the coordinate information embedding step, the module uses two one-dimensional global pooling operations to aggregate features in the horizontal and vertical directions respectively, thereby generating a pair of direction-aware features. For the input feature map, the pooling kernels of size (H, 1) and (1, W) are used for encoding. Therefore, the height h CAThe output formula of the cth channel at is:

[0021]

[0022] The width is w CA The output of the cth channel at is expressed as:

[0023]

[0024] In a preferred embodiment, the coordinate attention generation stage includes the following steps:

[0025] First, the coordinate information is embedded in the generated partial feature map for linking, and then sent to the shared 1×1 convolution transformation function F1. The specific operation is:

[0026]

[0027] Among them, f CA ∈R C / r×(H+W) is the intermediate feature map of spatial information in the horizontal and vertical directions, r is the downsampling ratio, and f CA Split into two independent tensors f CA h ∈R C / r×H and f CA w ∈R C / r×W ;

[0028] 1×1 convolution transformation and Used to and They are transformed into tensors with the same number of channels as the input feature map, as shown in the following formula:

[0029]

[0030] Then, and Expanded to attention weights;

[0031] Finally, the output of the coordinate attention module is expressed as:

[0032]

[0033] In the preferred scheme, the inverted residual moving module in step S23 enhances the recognition ability of the transmission line hidden danger detection network model for key hidden danger features by integrating local details and global background information, and adopts a multi-head self-attention mechanism to assign weights according to the importance of features, thereby optimizing the ability of the inverted residual moving module to capture the hidden danger target features in the transmission line.

[0034] In the preferred scheme, the specific workflow of the inverted residual moving module in step S23 includes: first, using the convolution layer to extract features from the input feature map; then, using the attention mechanism to generate an attention weight feature map, and connecting the weighted feature map residual to the input feature map; in addition, the residually connected feature map further fuses feature information of semantic levels of different scales; finally, obtaining the output of the inverted residual moving module.

[0035] Compared with the prior art, the beneficial effects of the present invention include:

[0036] (1) The present invention achieves accurate detection of hidden danger targets of complex transmission lines through strategies such as efficient feature extraction and enhancement, improvement of precise positioning capabilities, and fusion of local and global information.

[0037] (2) The present invention introduces a lightweight Ghost module, which can generate rich redundant image features without significantly increasing the amount of calculation, significantly enhancing the feature representation of the detection target, and helping the model to more accurately identify transmission line hidden danger targets in complex backgrounds.

[0038] (3) The present invention improves the detection performance of the model for complex transmission line hidden danger targets by designing and adding CA modules and iRMB modules, and enhances the adaptability of the model in complex scenarios.

[0039] (4) The algorithm framework proposed in this invention can be more easily extended to string target detection tasks in other industrial fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is the overall framework diagram of the power transmission line hidden danger detection proposed by the present invention.

[0041] Figure 2 This is the schematic diagram of conventional convolution and Ghost network.

[0042] Figure 3 Figure 2 is the structure diagram of the coordinate attention (CA) module.

[0043] Figure 4 The structural diagram of the lightweight inverted residual move (iRMB) module. DETAILED DESCRIPTION

[0044] The present invention is further described below in conjunction with the accompanying drawings.

[0045] The method for detecting hidden dangers of power transmission lines in complex backgrounds based on deep learning models includes:

[0046] Step 1: Construct a data set for detecting hidden dangers in transmission lines.

[0047] Step 2: Construct a transmission line hidden danger detection network model.

[0048] like Figure 1 As shown in the figure, the present invention proposes a transmission line hidden danger detection method based on deep learning, which aims to achieve high-precision and rapid detection of potential hidden dangers in transmission lines through deep learning technology. The method is built on the general yolov9 network architecture and combines a variety of optimization strategies to improve detection efficiency and accuracy. First, the GhostConv module is introduced, which can generate more feature maps with fewer parameters, while reducing the computational complexity and maintaining a high accuracy. In addition, a coordinate attention module (Co rdinate Attention, CA) is designed and added, which optimizes the model's positioning ability for object structures by embedding location information into channel attention. Finally, a lightweight inverted residual mobile module (iRMB) module is adopted, which combines the advantages of CNN and Transformer, and enhances the model's detection efficiency for complex transmission line hidden danger targets by integrating local features and global context information.

[0049] The Ghost network is specifically:

[0050] like Figure 2 As shown in the figure, Ghost network is a new type of convolutional neural network. Through the GhostConv module, more feature maps are generated at a low cost, which effectively reduces the amount of calculation and network parameters while ensuring the performance of the detection model, thereby improving the efficiency of the model. Compared with traditional convolution, the GhostConv module divides the original convolution layer into two parts. First, a small number of convolution kernels are used to generate a small number of intrinsic feature maps, and then redundant features are generated through simple linear change operations. Finally, the feature fusion is used as the output of the network module.

[0051] If we use X∈R c×h×w As the input feature of the network module, c is the number of channels of the input feature map, h, w represent the height and width of the input feature map. The traditional convolution operation can be expressed as;

[0052] Y=X*f+b

[0053] In the above formula, Y∈R h′×w′×n represents the output feature map containing n channels, h′ and w′ are the height and width of the output feature map, and * represents the convolution operation. f∈R c×k×k×n represents the convolution kernel, k×k is the size of the convolution kernel, and b is the bias term.

[0054] The Ghost network model is mainly composed of two parts: the traditional convolution layer and the lightweight linear transformation layer. First, the convolution layer is used to generate a small amount of image features; then, a linear transformation operation is performed on these image features to increase the redundancy of the feature map; specifically, it can be expressed by the mathematical formula:

[0055] Y′=X*f′+b

[0056]

[0057] In the above formula, Y′∈R h′×w′×m represents the output feature, f′∈R c×k×k×m Represents the convolution kernel. The number of channels m of the output feature map of the convolution operation is smaller than the number of channels n of the traditional convolution layer. i (i=1,...,m) are the m feature maps of Y′. Each feature map in Y′ must be processed by the linear operation Φ i,j (j=1,...,s) to get s feature maps. Therefore, m feature maps can be obtained through linear operations, and the final m·s feature maps are consistent with the number of feature maps output by traditional convolution, that is, n=m·s. In addition, among the s linear transformations of each feature, the last linear transformation is the identity transformation.

[0058] The coordinate attention (CA) module is as follows:

[0059] like Figure 3 As shown in the figure, the coordinate attention module decomposes the channel attention into two one-dimensional feature encoding processes, aggregating features in the horizontal and vertical directions respectively, so as to capture long-distance dependencies while maintaining precise information and retain fine-grained location information, thereby helping the model to better locate object structures and identify target areas. The CA attention module consists of two main steps: coordinate information embedding and coordinate attention generation. In the coordinate information embedding step, the module uses two one-dimensional global pooling operations to aggregate features in the horizontal and vertical directions respectively, thereby generating a pair of direction-aware features. For the input feature map, pooling kernels of size (H, 1) and (1, W) are used for encoding, so the height h CA The output of the cth channel at can be formulated as:

[0060]

[0061] The width is w CA The output of the cth channel at can be expressed as:

[0062]

[0063] The coordinate attention map generation stage is mainly divided into the following steps. First, the coordinate information is embedded into the partially generated feature map for linking, and then sent to the shared 1×1 convolution transformation function F1. The specific operation can be expressed as:

[0064]

[0065] Among them, f CA ∈R C / r×(H+W) is the intermediate feature map of spatial information in the horizontal and vertical directions, r is the downsampling ratio, and f CA Split into two independent tensors f CA h ∈R C / r×H and f CA w ∈R C / r×W The other two 1×1 convolution transformations and Used to and They are transformed into tensors with the same number of channels as the input feature map, as shown in the following formula:

[0066]

[0067] Then, and Expanded to attention weight. Finally, the output of the CA attention module can be expressed as:

[0068]

[0069] The inverted residual moving module (iRMB) is specifically:

[0070] like Figure 4 As shown in the figure, the Inverted Residual Mobile Block (iRMB) combines the efficiency of CNN in modeling short-distance dependencies and the ability of Transformer in dynamically modeling long-distance interactions. It achieves efficient use of computing resources and high accuracy while keeping the model lightweight.

[0071] The specific workflow of iRMB is as follows: first, the convolutional layer is used to extract features from the input feature map; then, the attention mechanism is used to generate the attention weight feature map, and the weighted feature map residual is connected to the input feature map; in addition, the residual connection feature map further integrates the feature information of semantic levels at different scales; finally, the output of the iRMB model is obtained. The RMB module enhances the model's ability to identify key hidden danger features by integrating local details and global background information, ensuring the integrity of hidden danger target information. At the same time, the multi-head self-attention mechanism adopted by the module can flexibly assign weights according to the importance of the features, thereby optimizing the performance of the iRMB module in capturing the characteristics of hidden danger targets in transmission lines, and further improving the model's detection accuracy and efficiency for hidden danger targets.

[0072] Step 3: Train the transmission line hidden danger detection network model.

[0073] Step 4: Detect the target of hidden dangers in the transmission line to be detected.

[0074] The image containing the transmission line hidden danger target is input into the trained network, and the detection result of each external rectangular frame position where the transmission line hidden danger target is located in the image is output.

[0075] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting hidden dangers in power transmission lines, characterized in that: The following steps are involved: S1. Construct a data set for detecting hidden dangers of power transmission lines; S2. Construct a network model for detecting hidden dangers in power transmission lines; S3, training a transmission line hidden danger detection network model; S4. Detect the transmission line hidden danger target to be detected, input the picture containing the transmission line hidden danger target into the trained transmission line hidden danger detection network model, and output the detection result of the location of the transmission line hidden danger target in the picture.

2. A method for detecting hidden dangers in power transmission lines according to claim 1, characterized in that: Step S2 also includes the following steps: The GhostConv module is introduced into the S21 and yolov9 network architectures; S22. Design and add a coordinate attention module to optimize the model's ability to locate object structures by embedding position information into channel attention. S23. The inverted residual moving module is adopted, CNN and Transformer are integrated, and the detection efficiency of the model for complex transmission line hidden danger targets is enhanced by integrating local features and global context information.

3. The method for detecting hidden dangers of power transmission lines according to claim 2 is characterized in that: The GhostConv module in step S21 includes a traditional convolutional layer and a lightweight linear transformation layer.

4. A method for detecting hidden dangers in power transmission lines according to claim 1, characterized in that: The GhostConv module extracts feature maps, including the following steps: first use some convolution kernels to generate some intrinsic feature maps, then generate redundant features through linear change operations, and finally fuse the features as the output of the GhostConv module, which is specifically expressed by the following mathematical formula: Y′=X*f′+b; and ij =Φ i,j (and i ′), In the above formula, Y′∈R h′×w′×m represents the output feature, f′∈R c×k×k×m Represents the convolution kernel. The number of channels m of the output feature map of the convolution operation is smaller than the number of channels n of the traditional convolution layer. i ′(i=1,...,m) is the m feature map of Y′, and each feature map in Y′ must be processed by the linear operation Φ i,j (j=1,...,s) to obtain s feature maps, so m feature maps are linearly operated to obtain m·s feature maps. The final m·s feature maps are consistent with the number of feature maps output by traditional convolution, that is, n=m·s. Among the s linear transformations of each feature, the last linear transformation is the identity transformation.

5. The method for detecting hidden dangers of a power transmission line according to claim 1, characterized in that: The coordinate attention module retains fine-grained position information by decomposing channel attention into two one-dimensional feature encoding processes, aggregating features in the horizontal and vertical directions respectively.

6. A method for detecting hidden dangers in power transmission lines according to claim 1, characterized in that: The coordinate attention module includes two steps: coordinate information embedding and coordinate attention generation.

7. A method for detecting hidden dangers in power transmission lines according to claim 6, characterized in that: In the coordinate information embedding step, the module uses two one-dimensional global pooling operations to aggregate features in the horizontal and vertical directions respectively, thereby generating a pair of direction-aware features. For the input feature map, the pooling kernels of size (H, 1) and (1, W) are used for encoding. Therefore, the height h CA The output formula of the cth channel at is: The width is w CA The output of the cth channel at is expressed as:

8. A method for detecting hidden dangers in power transmission lines according to claim 6, characterized in that: The coordinate attention generation phase includes the following steps: First, the coordinate information is embedded in the generated partial feature map for linking, and then sent to the shared 1×1 convolution transformation function F1. The specific operation is: Among them, f CA ∈R C / r×(H+W) is the intermediate feature map of spatial information in the horizontal and vertical directions, r is the downsampling ratio, and f CA Split into two independent tensors f CA h ∈R C / r×H and f CA w ∈R C / r×W ; 1×1 convolution transformation and Used to and They are transformed into tensors with the same number of channels as the input feature map, as shown in the following formula: Then, and Expanded to attention weights; Finally, the output of the coordinate attention module is expressed as:

9. A method for detecting hidden dangers in power transmission lines according to claim 2, characterized in that: The inverted residual moving module in step S23 enhances the recognition ability of the transmission line hidden danger detection network model for key hidden danger features by integrating local details and global background information. It adopts a multi-head self-attention mechanism and assigns weights according to the importance of features, thereby optimizing the ability of the inverted residual moving module to capture the hidden danger target features in the transmission line.

10. A method for detecting hidden dangers in power transmission lines according to claim 2, characterized in that: The specific workflow of the inverted residual moving module in step S23 includes: first, using the convolution layer to extract features from the input feature map; then, using the attention mechanism to generate an attention weight feature map, and connecting the weighted feature map residual to the input feature map; in addition, the residually connected feature map further fuses feature information of semantic levels of different scales; finally, obtaining the output of the inverted residual moving module.

Images