Deep learning-based power grid line ice melting effect detection method

Abstract

The application discloses to the technical field of power grid line, specifically is power grid line ice-melting effect detection method based on deep learning, including specific steps as follows: S1: constructing power grid line ice-melting effect detection dataset;S2: build power grid line ice-melting effect detection network GIMEDNet;S3: training power grid line ice-melting effect detection network GIMEDNet;S4: the application GIMEDNet carries out power grid line ice-melting effect detection.The GIMEDNet of the application can greatly improve the quality and efficiency of ice-melting effect detection through bidirectional attention mechanism, and ensures the high-quality detection of power grid line ice-melting effect.

Description

Technical Field

[0001] This invention relates to the field of power grid line technology, specifically to a method for detecting the de-icing effect of power grid lines based on deep learning. Background Technology

[0002] Power grid lines are prone to snow accumulation and icing during cold seasons, especially after icing weather. The melting of the ice layer has a significant impact on the safety and stability of power equipment. Monitoring the effectiveness of ice melting after icing is crucial for ensuring the normal operation of power grid lines and preventing accidents. Traditional ice melting detection methods, such as manual inspection or automated inspection based on simple image processing techniques (e.g., edge detection, threshold segmentation), typically have many limitations. Manual inspection is not only inefficient and labor-intensive but also easily affected by weather, viewing angle, and human factors, and it is difficult to achieve large-scale and real-time monitoring. Traditional image processing methods are poorly adaptable to complex environments (such as varying ice thickness, changes in lighting, and background interference), making accurate and stable detection difficult.

[0003] In summary, traditional methods for detecting de-icing have the following drawbacks:

[0004] (1) Traditional methods for detecting the melting effect rely on manual observation and analysis, which have disadvantages such as high labor costs, slow detection speed, strong subjectivity of results and poor adaptability.

[0005] (2) In solutions that have adopted deep learning technology, there are often problems such as too many parameters of deep learning networks, long training time, and low operating efficiency. Furthermore, the detection of the de-icing effect of power grid lines has not been optimized, resulting in poor detection results. Summary of the Invention

[0006] To address the aforementioned technical problems, according to one aspect of the present invention, the present invention provides the following technical solution:

[0007] A deep learning-based method for detecting the de-icing effect of power grid lines includes the following specific steps:

[0008] S1: Construct a dataset for detecting the de-icing effect of power grid lines;

[0009] S2: Construct the GIMEDNet network for detecting the de-icing effect of power grid lines;

[0010] S3: Training the GIMEDNet network for detecting the ice melting effect of power grid lines;

[0011] S4: Use GIMEDNet to detect the de-icing effect of power grid lines.

[0012] As a preferred embodiment of the deep learning-based power line de-icing effect detection method of the present invention, the specific steps of S1 are as follows:

[0013] S11: Collect image data of the de-icing results of power grid lines;

[0014] S12: The collected image data is annotated using annotation software to annotate the icy areas in the image pixel by pixel. After the annotation is completed, the melting result mask image will be obtained. The power grid line melting effect detection dataset consists of the original power grid line melting result image and its corresponding mask image.

[0015] As a preferred embodiment of the deep learning-based power line de-icing effect detection method of the present invention, the specific steps of S2 are as follows:

[0016] S21: Input the image F1 of the power grid line de-icing result into GIMEDNet. First, it is processed by the Conv(3×3)_BN_ReLU module to obtain the de-icing effect feature map F2. The Conv(3×3)_BN_ReLU module is composed of a Conv layer, a BN layer and a ReLU layer with a convolution kernel size of 3×3 and a stride of 1.

[0017] S22: After performing a Conv(7×7) downsampling operation on F2, it is input into the RMBAMB module to obtain the ice-melting effect feature map F3, which has half the resolution of F2. The RMBAMB module includes the DCMRB module, the MFEDSB module, and the BAMB module. The DCMRB module uses deformable convolution for feature extraction and spatial adaptive transformation to adapt to ice-melting effect features with different deformations. The MFEDSB module uses depthwise separable convolution, which can improve the ice-melting effect feature extraction capability while reducing computational complexity, reducing redundant computation, improving the efficiency and generalization ability of the model, and enabling the network to more fully capture and fuse key features at different scales. The BAMB module is used to implement a bidirectional attention mechanism, and the BAMB module combines dilated convolution and HCA activation function to effectively improve the ice-melting effect feature extraction capability.

[0018] S23: After F3 undergoes a Conv(3×3) downsampling operation, it is input into the RMBAMB module to extract and refine the ice melting effect features of F3, resulting in an ice melting effect feature map F4, whose resolution is half that of F3; after F4 undergoes a Conv(3×3) downsampling operation, it is input into the DCMRB module to extract and refine the ice melting effect features of F4, resulting in an ice melting effect feature map F5, whose resolution is half that of F4;

[0019] S24: After upsampling F5 through the EEAUP module, a concat operation is performed with F3. The concat feature map is then input into the RBAMB module to obtain the ice-melting effect feature map F6. The resolution of F6 is twice that of F5. The RBAMB module consists of a DCMRB module and a BAMB module connected in series. After upsampling F6 through the EEAUP module, a concat operation is performed with F2. The concat feature map is then input into the BAMB module. The result processed by the BAMB module is then convolved with Conv (3×3) to obtain the ice-melting effect feature map F7. The resolution of F7 is twice that of F6.

[0020] S25: After upsampling the feature maps F5, F6, and F7 of the ice melting effect to the resolution of F1, they are stitched together in the channel dimension. The stitched result is converted into a single channel after Conv(3×3) convolution operation, and then processed by a linear layer to obtain the detection result of the ice melting effect of the power grid line.

[0021] As a preferred embodiment of the deep learning-based power grid line de-icing effect detection method described in this invention, wherein: in the DCMRB module, the input feature map Input DCMRB The calculations are performed on both branches separately, with the specific steps as follows:

[0022] In the first branch, the input feature map is... DCMRB After performing a Conv(1x1) convolution operation, GN normalization is applied to obtain the ice-melting effect feature map D1; D1 is then activated by LeakyReLU to obtain the ice-melting effect feature map D2; D2 is then processed in the first and second sub-branches respectively:

[0023] In the first sub-branch, D2 is subjected to deformable convolution DConv(3x3) operation and then normalized by GN to obtain the ice melting effect feature map D3; D3 is activated by LeakyReLU and then subjected to Conv(1x1) operation to obtain the ice melting effect feature map D4.

[0024] In the second sub-branch, D2 is subjected to Conv(1x1) convolution operation and then channel attention is enhanced by SE Block to obtain the ice melting effect feature map D5;

[0025] In the second branch, the input feature map is... DCMRB After performing a Conv(1x1) convolution operation, channel attention is enhanced by an SE Block to obtain the ice-melting effect feature map D6;

[0026] After element-wise addition of D4, D5, and D6, the model is activated using Leaky ReLU to obtain the output feature map of the DCMRB module. DCMRB Output DCMRB The calculation process can be described by equations (1)-(7);

[0027] D1 = GN(Conv) 1×1 (Input DCMRB )) (1)

[0028] D2=LeakyReLU(D1) (2)

[0029] D3=GN(Conv 3×3 (D2)) (3)

[0030] D4 = Conv 1×1 (LeakyReLU(D3)) (4)

[0031] D5 = SE(Conv) 1×1 (D2)) (5)

[0032] D6=SE(Conv 1×1 (Input DCMRB (6)

[0033] Output DCMRB =LeakyReLU(Add(D4,D5,D6)) (7).

[0034] As a preferred embodiment of the deep learning-based power grid line de-icing effect detection method described in this invention, wherein: in the MFEDSB module, the input feature map Input MFEDSB The calculations are performed on each of the four branches, with the specific steps as follows:

[0035] In the first branch, for Input MFEDSB No action will be taken.

[0036] In the second branch, the Input MFEDSB The number of channels is reduced by using Conv(1×1) convolution operations to reduce computational complexity, and the feature map M1 of the ice melting effect is obtained.

[0037] In the third branch, the Input MFEDSB The number of channels is reduced by Conv(1×1) convolution operation, and then the depthwise separable convolution DSConv(3×3) operation is performed to extract the ice melting effect features, resulting in the ice melting effect feature map M2.

[0038] In the fourth branch, the Input MFEDSBThe number of channels is reduced by Conv(1×1) convolution operation, and then a deeper feature extraction of the ice melting effect is performed by two consecutive DSConv(3×3) convolution operations to obtain the ice melting effect feature map M3. The two DSConv(3×3) convolution operations have the same receptive field as the DSConv(5×5) convolution operation, but with fewer computational parameters, which can reduce the number of network parameters.

[0039] Connect M1, M2, and M3 with Input MFEDSB After performing the Concat operation, the DSConv(3×3) operation is used to extract the ice-melting effect features, resulting in the output feature map of the MFEDSB module. MFEDSB Output MFEDSB The calculation process can be described by equations (8)-(11);

[0040] M1 = conv 1×1 (Input MFEDSB (8)

[0041] M2 = DSConv 3×3 (conv 1×1 (M1)) (9)

[0042] M3 = DSConv 3×3 (DSConv 3×3 (conv 1×1 (Input MFEDSB ))) (10)

[0043] Output MFESB =DSConv 3×3 Concat(M1,M2,M3,Intput MFEDSB )) (11).

[0044] As a preferred embodiment of the deep learning-based power grid line de-icing effect detection method described in this invention, wherein: in the BAMB module, the input feature map Input BAMB The calculations are performed on each of the four branches, with the specific steps as follows:

[0045] In the first branch, for Input BAMB After performing the DConv(3x3) operation, the feature map B1 of the ice melting effect is obtained by activating it through the HCA activation function;

[0046] In the second branch, for Input BAMB No action will be taken.

[0047] In the third branch, for Input BAMB No processing is performed;

[0048] In the fourth branch, for Input BAMB After performing the GAP operation, the ice melting effect feature map B3 is obtained; B3 is then activated by the HCA activation function after two Conv(3x3)(2dilated) operations to obtain the ice melting effect feature map B4.

[0049] B1 and Input BAMB After matrix multiplication, the ice-melting effect feature map B2 is obtained; B4 and Input are then combined. BAMB After matrix multiplication, the feature map of the ice melting effect is obtained as B5;

[0050] Element-wise addition of B4 and B5 yields the output feature map of the BAMB module. BAMB Output BAMB The calculation process can be described by equations (12)-(17);

[0051] B1 = HCA(DConv) 3×3 (Input BAMB (12)

[0052]

[0053] B3 = GAP(Input) BAMB (14)

[0054] B4 = HCA(DConv) 3×3 (DConv 3×3 (B3))) (15)

[0055]

[0056] Output BAMB =Add(B4,B5) (17).

[0057] In a preferred embodiment of the deep learning-based power grid line de-icing effect detection method of the present invention, the HCA activation function is defined as follows:

[0058] HCA(x)=α·Cosh(β·x+γ)+δ·log(1+∈·|x|) (18)

[0059] Where α adjusts the amplitude of the activation function, controlling the maximum value of the output; β adjusts the degree of influence of the input on the activation function, controlling the steepness of the function; γ is a translation parameter, controlling the translation effect of the input; δ adjusts the amplitude of the second term, affecting the response to the absolute value of the input; ∈ adjusts the weighting degree of the absolute value of the input, increasing the smoothness to a certain extent.

[0060] The definition of the Cosh function is as follows:

[0061]

[0062] As a preferred embodiment of the deep learning-based power grid line de-icing effect detection method of the present invention, the upsampling process of the EEAUP module is as follows:

[0063] First, given the input feature map X and the target upsampling scale s, the initial upsampling result is generated using bilinear interpolation or deconvolution:

[0064] X up =U(X,s) (20)

[0065] Where U represents a conventional upsampling operation;

[0066] Then, for any pixel (i,j) in the feature map, the gradient magnitude is calculated using a Sobel filter or gradient operator:

[0067]

[0068] After normalization, the edge attention weights are obtained:

[0069]

[0070] Here, ∈ is a small constant to prevent numerical instability;

[0071] Finally, edge information is fused, and the preliminary upsampling results are adaptively adjusted to obtain the final upsampling result:

[0072] X EEAUP (i,j)=X up (i,j)+αW edge (i,j)·X up (i,j) (23)

[0073] Here, Δ is a hyperparameter used to control the intensity of edge enhancement.

[0074] As a preferred embodiment of the deep learning-based power line de-icing effect detection method of the present invention, the specific steps of S3 are as follows:

[0075] S31: Use the power grid line ice melting effect detection dataset in S1 to train the power grid line ice melting effect detection network GIMEDNet built in S2;

[0076] S32: During training, GIMEDNet optimizes the loss function by adjusting the weights in the network using the Adam algorithm, causing the loss function to converge and thus bringing the network to a balanced state.

[0077] As a preferred embodiment of the deep learning-based power grid line de-icing effect detection method of the present invention, the specific steps of S4 are as follows:

[0078] S41: After GIMEDNet training is completed, GIMEDNet is used to detect the de-icing effect of power grid lines;

[0079] S42: In practical applications, the input is the image of the de-icing result of the power grid line. After GIMEDNet processes the image of the de-icing result, it outputs the de-icing effect detection result.

[0080] Compared with existing technologies:

[0081] The GIMEDNet of this invention, through a bidirectional attention mechanism, can significantly improve the quality and efficiency of de-icing effect detection, ensuring high-quality detection of de-icing effects on power grid lines. Furthermore, GIMEDNet possesses continuous learning and self-optimization capabilities, enabling it to continuously improve the detection accuracy of de-icing effects on power grid lines through continuous training and iteration. Based on this, this invention offers faster detection speed compared to complex methods and higher detection quality compared to traditional methods. Attached Figure Description

[0082] Figure 1 This is a schematic diagram of the overall process of the present invention;

[0083] Figure 2 This is a schematic diagram of the GIMEDNet network structure of the present invention;

[0084] Figure 3 This is a schematic diagram of the RMBAMB module structure of the present invention;

[0085] Figure 4 This is a schematic diagram of the DCMRB module structure of the present invention;

[0086] Figure 5 This is a schematic diagram of the MFEDSB module structure of the present invention;

[0087] Figure 6 This is a schematic diagram of the BAMB module structure of the present invention. Detailed Implementation

[0088] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0089] This invention provides a deep learning-based method for detecting the de-icing effect of power grid lines. Please refer to [link / reference]. Figures 1-6 The specific steps are as follows:

[0090] S1: Construct a dataset for detecting the de-icing effect of power grid lines.

[0091] The specific steps of S1 are as follows:

[0092] S11: Collect image data of the de-icing results of power grid lines (e.g., by taking photos on-site or using existing public datasets);

[0093] S12: The collected image data is annotated using annotation software to annotate the icy areas in the image pixel by pixel. After the annotation is completed, the melting result mask image will be obtained. The power grid line melting effect detection dataset consists of the original power grid line melting result image and its corresponding mask image.

[0094] There are many annotation tool software programs available; in this invention, LabelMe is used as the annotation tool software.

[0095] S2: Build a network for detecting the de-icing effect of power grid lines, GIMEDNet.

[0096] A grid ice melting effect detection network, GIMEDNet (GridIce Melting EffectDetection Net), was constructed for detecting the ice melting effect of power grid lines. The GIMEDNet network structure is as follows: Figure 2 As shown in the figure, ⊕ represents the Concat operation.

[0097] Here, Concat represents the channel-dimensional concatenation operation in deep learning. When concatenating along the channel dimension, the length and width of the feature maps must be the same, but the number of channels can be different. For example, if there are two feature maps A and B, with shapes of H×W×16 and H×W×32 respectively, they can be concatenated along the channel dimension to obtain a new feature map with a shape of H×W×48.

[0098] The specific steps of S2 are as follows:

[0099] S21: Input the image F1 of the power grid line de-icing result into GIMEDNet. First, it is processed by the Conv(3×3)_BN_ReLU module to obtain the de-icing effect feature map F2. The Conv(3×3)_BN_ReLU module is composed of a Conv layer, a BN layer and a ReLU layer with a convolution kernel size of 3×3 and a stride of 1.

[0100] In the above text, the kernel size of 3×3 means that the length and width of the kernel are both 3.

[0101] S22: After performing a Conv(7×7) downsampling operation on F2, it is input into the RMBAMB (Residual Multi-scale Bilateral Attention Mechanism Block) module to obtain the ice-melting effect feature map F3. The resolution of F3 is half that of F2. The structure of the RMBAMB module is as follows: Figure 3 As shown;

[0102] This invention designs a DCMRB (Deformable Conv Multi-dimension Residual Block) module, which uses deformable convolution (DConv) for feature extraction and spatial adaptive transformation to adapt to the melting effect features of different deformations.

[0103] The advantage of this module lies in its ability to significantly expand the receptive field without reducing resolution. Furthermore, it can adapt to complex ice-melting effect structures through deformable convolutions. Combined with SEBlock's global attention mechanism, it enhances the network's focus on key features, thereby improving the ability to recognize ice-melting effect features. The structure of the DCMRB module is as follows: Figure 4 As shown.

[0104] In the DCMRB module, the input feature map is processed. DCMRB Perform calculations on both branches separately, such as... Figure 4 As shown, the specific steps are as follows:

[0105] In the first branch, the input feature map is... DCMRB After performing a Conv(1x1) convolution operation, normalization is performed using GN (Group Normalization) to obtain the ice-melting effect feature map D1; D1 is then activated by LeakyReLU to obtain the ice-melting effect feature map D2; D2 is then processed in the first and second sub-branches respectively:

[0106] In the first sub-branch, D2 is subjected to deformable convolution DConv(3x3) operation and then normalized by GN to obtain the ice melting effect feature map D3; D3 is activated by LeakyReLU and then subjected to Conv(1x1) operation to obtain the ice melting effect feature map D4.

[0107] In the second sub-branch, D2 is subjected to Conv(1x1) convolution operation and then channel attention is enhanced by SE Block to obtain the ice melting effect feature map D5;

[0108] In the second branch, the input feature map is... DCMRBAfter performing a Conv(1x1) convolution operation, channel attention is enhanced by an SE Block to obtain the ice-melting effect feature map D6;

[0109] After performing an element-wise addition operation on D4, D5, and D6, and then activating them using LeakyReLU, the output feature map of the DCMRB module is obtained. DCMRB Output DCMRB The calculation process can be described by equations (1)-(7);

[0110] D1 = GN(Conv) 1×1 (Input DCMRB )) (1)

[0111] D2=LeakyReLU(D1) (2)

[0112] D3=GN(Conv 3×3 (D2)) (3)

[0113] D4 = Conv 1×1 (LeakyReLU(D3)) (4)

[0114] D5 = SE(Conv) 1×1 (D2)) (5)

[0115] D6=SE(Conv 1×1 (Input DCMRB (6)

[0116] Output DCMRB =LeakyReLU(Add(D4,D5,D6)) (7).

[0117] This invention designs an MFEDSB (Multi-scale Feature Extraction-DepthSeparation Block) module to improve the performance of GIMEDNet. This module employs depthwise separable convolution (DSConv), which can improve the feature extraction capability for ice-melting effects while reducing computational complexity, reducing redundant computation, and improving the model's efficiency and generalization ability. This allows the network to more fully capture and fuse key features at different scales. The MFEDSB module structure is as follows: Figure 5 As shown.

[0118] In the MFEDSB module, the input feature map is processed. MFEDSB Calculations are performed on each of the four branches, as follows: Figure 5 As shown, the specific steps are as follows:

[0119] In the first branch, for Input MFEDSB No action will be taken.

[0120] In the second branch, the Input MFEDSB The number of channels is reduced by using Conv(1×1) convolution operations to reduce computational complexity, and the feature map M1 of the ice melting effect is obtained.

[0121] In the third branch, the Input MFEDSB The number of channels is reduced by Conv(1×1) convolution operation, and then the depthwise separable convolution DSConv(3×3) operation is performed to extract the ice melting effect features, resulting in the ice melting effect feature map M2.

[0122] In the fourth branch, the Input MFEDSB The number of channels is reduced by Conv(1×1) convolution operation, and then a deeper feature extraction of the ice melting effect is performed by two consecutive DSConv(3×3) convolution operations to obtain the ice melting effect feature map M3. The two DSConv(3×3) convolution operations have the same receptive field as the DSConv(5×5) convolution operation, but with fewer computational parameters, which can reduce the number of network parameters.

[0123] Connect M1, M2, and M3 with Input MFEDSB After performing the Concat operation, the DSConv(3×3) operation is used to extract the ice-melting effect features, resulting in the output feature map of the MFEDSB module. MFEDSB Output MFEDSB The calculation process can be described by equations (8)-(11);

[0124] M1 = conv 1×1 (Input MFEDSB (8)

[0125] M2 = DSConv 3×3 (conv 1×1 (M1)) (9)

[0126] M3 = DSConv 3×3 (DSConv 3×3 (conv 1×1 (Input MFEDSB ))) (10)

[0127] Output MFEDSB =DSConv 3×3 Concat(M1,M2,M3,Intput MFEDSB )) (11).

[0128] This invention designs a BAMB (Bilateral Attention Mechanism Block) module to implement a bidirectional attention mechanism. The module combines dilated convolution and the HCA (HyperLog-Cosh Activation) activation function, effectively improving the feature extraction capability for ice-melting effects. Dilated convolution expands the receptive field, capturing multi-scale features while maintaining computational efficiency. The HCA activation function optimizes training stability and accelerates convergence. The bidirectional attention mechanism allows the network to focus more on key features, further improving the detection accuracy of ice-melting effects. The structure of the BAMB module is as follows: Figure 6 As shown. Figure 6 middle This refers to matrix multiplication.

[0129] Dilated convolution, also known as dilated convolution, is a variant of standard convolution that introduces dilation. It inserts holes between kernel elements to increase the receptive field without increasing computational complexity or reducing resolution.

[0130] For dilated convolution, the distance between the convolution kernels is determined by a dilation rate.

[0131] Assume the input feature map is I, the convolution kernel is K, the kernel size is k×k, and the input feature map size is H. in ×W in The porosity is d, and the output feature map size is H. out ×W out The formula for calculating dilated convolution is as follows:

[0132]

[0133] In the BAMB module, the input feature map is processed. BAMB Calculations are performed on each of the four branches, as follows: Figure 6 As shown, the specific steps are as follows:

[0134] In the first branch, for Input BAMB After performing the DConv(3x3) operation, the feature map B1 of the ice melting effect is obtained by activating it through the HCA activation function;

[0135] In the second branch, for Input BAMB No action will be taken.

[0136] In the third branch, for Input BAMB No processing is performed;

[0137] In the fourth branch, for Input BAMB After performing the GAP (Global Average Pooling) operation, the ice melting effect feature map B3 is obtained; B3 is then activated by the HCA activation function after two Conv(3x3)(2dilated) operations to obtain the ice melting effect feature map B4.

[0138] B1 and Input BAMB After matrix multiplication, the ice-melting effect feature map B2 is obtained; B4 and Input are then combined. BAMB After matrix multiplication, the feature map of the ice melting effect is obtained as B5;

[0139] Adding B4 and B5 element-wise yields the output feature map of the BAMB module. BAMB Output BAMB The calculation process can be described by equations (12)-(17);

[0140] B1 = GCA(DConv) 3×3 (Input BAMB (12)

[0141]

[0142] B3 = GAP(Input) BAMB (14)

[0143] B4 = HCA(DConv) 3×3 (DConv 3×3 (B3))) (15)

[0144]

[0145] Output BAMB =Add(B4,B5) (17).

[0146] This invention designs a hyperlog-hyperbolic cosine activation function (HCA), which combines an exponential function and a parameterized adjustment factor, making the activation function more flexible and complex.

[0147] The HCA activation function is defined as follows:

[0148] HCA(x)=α·Cosh(β·x+γ)+δ·log(1+∈·|x|) (18)

[0149] Where α adjusts the amplitude of the activation function, controlling the maximum value of the output; β adjusts the degree of influence of the input on the activation function, controlling the steepness of the function; γ is a translation parameter, controlling the translation effect of the input; δ adjusts the amplitude of the second term, affecting the response to the absolute value of the input; ∈ adjusts the weighting degree of the absolute value of the input, increasing the smoothness to a certain extent.

[0150] The definition of the Cosh function is as follows:

[0151]

[0152] S23: After F3 undergoes a Conv(3×3) downsampling operation, it is input into the RMBAMB module to extract and refine the ice melting effect features of F3, resulting in an ice melting effect feature map F4, whose resolution is half that of F3; after F4 undergoes a Conv(3×3) downsampling operation, it is input into the DCMRB module to extract and refine the ice melting effect features of F4, resulting in an ice melting effect feature map F5, whose resolution is half that of F4;

[0153] S24: After upsampling F5 through the EEAUP module, a concat operation is performed with F3. The concat feature map is then input into the RBAMB module to obtain the ice-melting effect feature map F6. The resolution of F6 is twice that of F5. The RBAMB module consists of a DCMRB module and a BAMB module connected in series. After upsampling F6 through the EEAUP module, a concat operation is performed with F2. The concat feature map is then input into the BAMB module. The result processed by the BAMB module is then convolved with Conv (3×3) to obtain the ice-melting effect feature map F7. The resolution of F7 is twice that of F6.

[0154] This invention designs an EEAUP upsampling module that combines edge awareness and adaptive adjustment (see...). Figure 2 The proposed method, EEAUP, aims to improve the detail preservation of features during upsampling. It first performs conventional upsampling on the input feature map (such as bilinear interpolation or deconvolution), then calculates edge responses using gradient information to generate edge enhancement weights. These weights adaptively adjust the upsampled features, resulting in additional enhancement of edge regions, effectively reducing blur and improving detail clarity. Compared to traditional upsampling methods, EEAUP restores edge details more accurately while maintaining overall structural information, making it better suited for tasks involving ice melting effect detection.

[0155] The upsampling process of the EEAUP module is as follows:

[0156] First, given the input feature map X and the target upsampling scale s, the initial upsampling result is generated using bilinear interpolation or deconvolution:

[0157] X up =U(X,s) (20)

[0158] Where U represents a conventional upsampling operation;

[0159] Then, for any pixel (i,j) in the feature map, the gradient magnitude is calculated using a Sobel filter or gradient operator:

[0160]

[0161] After normalization, the edge attention weights are obtained:

[0162]

[0163] Here, ∈ is a small constant to prevent numerical instability;

[0164] Finally, edge information is fused, and the preliminary upsampling results are adaptively adjusted to obtain the final upsampling result:

[0165] X EEAUP (i,j)=X up (i,j)+αW edge (i,j)·X up (i,j) (23)

[0166] Here, α is a hyperparameter used to control the intensity of edge enhancement.

[0167] S25: After upsampling the feature maps F5, F6, and F7 of the ice melting effect to the resolution of F1, they are concatenated in the channel dimension. The concatenated result is converted into a single channel after Conv(3×3) convolution operation, and then processed by a linear layer to obtain the detection result of the ice melting effect of the power grid line.

[0168] The result of the power grid line de-icing effect test is a single value; the larger the output value, the better the de-icing effect.

[0169] S3: Training the GIMEDNet network for detecting the de-icing effect of power grid lines.

[0170] The specific steps of S3 are as follows:

[0171] S31: Use the power grid line ice melting effect detection dataset in S1 to train the power grid line ice melting effect detection network GIMEDNet built in S2; initialize the hyperparameters required for training the GIMEDNet network, such as the size of a batch of samples during training, the number of training rounds, and the initial learning rate.

[0172] S32: During training, GIMEDNet optimizes the loss function by adjusting the weights in the network using the Adam algorithm, causing the loss function to converge and thus bringing the network to a balanced state.

[0173] Among them, the Adam optimization algorithm is a common optimization algorithm in deep learning.

[0174] S4: Use GIMEDNet to detect the de-icing effect of power grid lines.

[0175] The specific steps of S4 are as follows:

[0176] S41: After GIMEDNet training is completed, GIMEDNet is used to detect the de-icing effect of power grid lines;

[0177] S42: In practical applications, the input is the image of the de-icing result of the power grid line. After GIMEDNet processes the image of the de-icing result, it outputs the de-icing effect detection result.

[0178] Example: Using GIMEDNet to detect the de-icing effect of power grid lines.

[0179] Assume the input image F1, representing the result of de-icing on the power grid line, has dimensions of 1×512×512 (channels × width × height). Inputting F1 into GIMEDNet, it first undergoes processing by the Conv(3×3)_BN_ReLU module, resulting in a de-icing effect feature map F2 with dimensions of 16×512×512. After undergoing a Conv(7×7) downsampling operation, F2 is input into the RMBAMB module, yielding a de-icing effect feature map F3 with dimensions of 32×256×256. After further downsampling by Conv(3×3), F3 is input into the RMBAMB module, where the de-icing effect features of F3 are extracted and refined, resulting in a de-icing effect feature map F4 with dimensions of 48×128×128. After downsampling F4 using Conv(3×3), it is input into the DCMRB module to extract and refine the ice-melting effect features of F4, resulting in an ice-melting effect feature map F5 with dimensions of 64×64×64. F5 is then upsampled using the EEAUP module and concatenated with F3. The concatenated feature map is then input into the RBAMB module, resulting in an ice-melting effect feature map F6 with dimensions of 48×256×256. F6 is then upsampled using the EEAUP module and concatenated with F2. The concatenated feature map is then input into the BAMB module. The result processed by the BAMB module is then convolved using Conv(3×3), resulting in an ice-melting effect feature map F7 with dimensions of 48×512×512. After upsampling the feature maps F5, F6, and F7 to a resolution of 512×512, they are concatenated along the channel dimension. The resulting (64+48+48)×512×512 feature map is then converted to a single channel (1×512×512) after a Conv(3×3) convolution operation. Finally, it is processed by a linear layer with a weight matrix of size 512×512 to obtain the final power grid line ice melting effect detection result. The power grid line ice melting effect detection result is a numerical value ranging from 0 to 100; a higher value indicates a better ice melting effect.

[0180] Although the present invention has been described above with reference to embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of the invention. In particular, as long as there is no structural conflict, the features in the disclosed embodiments can be combined with each other in any manner. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A method for detecting the effect of power grid line de-icing based on deep learning, characterized in that, The specific steps include the following: S1: Constructing a power grid line ice-melting effect detection dataset, and the specific steps include the following: S11: Collecting power grid line ice-melting result image data; S12: Labeling the collected image data by using a labeling tool software, performing pixel-by-pixel labeling on the ice-coated area in the image by using the labeling tool software, and obtaining an ice-melting result mask graph after labeling is completed, wherein the power grid line ice-melting effect detection dataset is composed of an original power grid line ice-melting result image and a corresponding mask graph; S21: a power grid line ice melting result image In the input GIMEDNet, first processed by the Conv(3×3)_BN_ReLU module to obtain the ice melting effect feature map ; wherein the Conv(3×3)_BN_ReLU module is composed of a Conv layer with a convolution kernel size of 3×3 and a step of 1, a BN layer, and a ReLU layer in series S22: Will After a Conv(7×7) downsampling operation, the data is input into the RMBAMB module to obtain the ice-melting effect feature map. , The resolution is The RMBAMB module comprises a DCMRB module, an MFEDSB module, and a BAMB module. The DCMRB module uses deformable convolution for feature extraction and spatial adaptive transformation to adapt to the melting effect features of different deformations. The MFEDSB module uses depthwise separable convolution, which can improve the feature extraction capability of melting effect while reducing computational complexity, reducing redundant computation, and improving the efficiency and generalization ability of the model, enabling the network to more fully capture and fuse key features at different scales. The BAMB module is used to implement a bidirectional attention mechanism, and the BAMB module combines dilated convolution and HCA activation function to effectively improve the feature extraction capability of melting effect. S23: After the down-sampling operation of Conv(3x3), the input is input into the RMBAMB module to extract and refine the ice-melting effect features of the input, to obtain an ice-melting effect feature map , The resolution is half of that of the input; After the down-sampling operation of Conv(3x3), the input is input into the DCMRB module to extract and refine the ice-melting effect features of the input, to obtain an ice-melting effect feature map , The resolution is half of that of the input; After the down-sampling operation of Conv(3x3), the input is input into the DCMRB module to extract and refine the ice-melting effect features of the input, to obtain an ice-melting effect feature map , The resolution is half of that of the input; ​ S24: After upsampling by the EEAUP module, the feature map is input into the RBAMB module, and the deicing effect feature map is obtained After the Concat operation, the feature map after the Concat operation is input into the RBAMB module, and the deicing effect feature map is obtained , The resolution of the feature map is twice that of the feature map The RBAMB module is composed of a DCMRB module and a BAMB module in series. After upsampling by the EEAUP module, the feature map is input into the RBAMB module, and the deicing effect feature map is obtained After the Concat operation, the feature map after the Concat operation is input into the BAMB module, and the deicing effect feature map is obtained , The resolution of the feature map is twice that of the feature map ​ S25: the ice-melting effect feature map , , are restored to the resolution of respectively through upsampling operations, spliced in the channel dimension, converted to a single channel after a convolution operation of Conv(3x3), and processed by a linear layer to obtain the ice-melting effect detection result of the power grid line. S2: Building a power grid line ice-melting effect detection network GIMEDNet, and the specific steps include the following: S3: Training the power grid line ice-melting effect detection network GIMEDNet; 2. The deep learning-based power grid line ice-melting effect detection method according to claim 1, characterized in that, In the DCMRB module, the input feature map Input DCMRB The calculation is performed on two branches respectively, and the specific steps are as follows: In the first branch, the input feature map Input DCMRB After the Conv(1x1) convolution operation is normalized by GN to obtain the ice melting effect feature map ; After LeakyReLU activation, the ice melting effect feature map is obtained ; Respectively into the first sub-branch and the second sub-branch for processing: In the first sub-branch, the following operations are performed: After the deformable convolution DConv(3x3) operation, the ice-melting effect feature map is obtained through GN normalization ; After the LeakyReLU activation, the ice-melting effect feature map is obtained through the Conv(1x1) operation ; In the second sub-branch, the After the Conv(1x1) convolution operation, the channel attention is enhanced by the SE Block to obtain the ice-melting effect feature map ; In the second branch, the input feature map Input DCMRB After the Conv(1x1) convolution operation, the channel attention is enhanced by the SE Block to obtain the ice-melting effect feature map ; After element-wise addition operation with , and , the output feature map Output DCMRB of the DCMRB module is obtained through LeakyReLU activation. DCMRB The calculation process of Output DCMRB is described by equations (1)-(7). (1) (2) (3) (4) (5) (6) (7)。 3. The deep learning-based power grid line ice-melting effect detection method according to claim 1, characterized in that, In the MFEDSB module, the input feature map Input MFEDSB The calculation is performed on the four branches respectively, and the specific steps are as follows: In the first branch, Input MFEDSB is not processed at all; In the second branch, Input MFEDSB The channel number is reduced by Conv(1x1) convolution operation to reduce the calculation complexity and obtain the ice melting effect feature map ; In the third branch, Input MFEDSB The channel number is reduced by a Conv(1x1) convolution operation, and then a depth separable convolution DSConv(3x3) operation is performed to extract the ice melting effect features to obtain an ice melting effect feature map ; In the fourth branch, Input MFEDSB The channel number is reduced by a Conv(1×1) convolution operation, and then two consecutive DSConv(3×3) convolution operations are used for deeper feature extraction of the ice-melting effect to obtain an ice-melting effect feature map The two DSConv(3×3) convolution operations have the same receptive field as the DSConv(5×5) convolution operation, and have fewer calculation parameters, thereby reducing the network parameter amount. will be described below. , , After the Concat operation with Input MFEDSB , the ice-melting effect feature extraction is performed through the DSConv (3x3) operation to obtain the output feature map Output MFEDSB of the MFEDSB module; the calculation process of Output MFEDSB is described by equations (8)-(11). (8) (9) (10) (11)。 4. The deep learning-based power grid line ice-melting effect detection method according to claim 1, characterized in that, In the BAMB module, the input feature map Input BAMB The calculation is performed on the four branches respectively, and the specific steps are as follows: In the first branch, the Input BAMB is activated by the HCA activation function after the DConv(3x3) operation to obtain the ice-melting effect feature map ; In the second branch, Input BAMB is not processed at all; In the third branch, Input BAMB is not processed at all; In the fourth branch, after the GAP operation is performed on the Input BAMB , the ice-melting effect feature map is obtained ; After two Conv(3x3)(2 dilated) operations and activation by the HCA activation function, the ice-melting effect feature map is obtained ; Will and Input BAMB After matrix multiplication, the ice-melting effect feature map is obtained ; Will and Input BAMB After matrix multiplication, the ice-melting effect feature map is obtained ; The element-wise addition operation is performed with to obtain the output feature map Output of the BAMB module BAMB ; the calculation process of Output BAMB is described by equations (12)-(17); (12) (13) (14) (15) (16) (17)。 5. The deep learning-based power grid line ice-melting effect detection method according to claim 4, characterized in that, S4: Applying GIMEDNet to perform power grid line ice-melting effect detection. (18) wherein, is a magnitude of the activation function, controlling the maximum value of the output; is a degree of influence of the input on the activation function, controlling the steepness of the function; is a shift parameter, controlling the shift effect of the input; adjusts the magnitude of the second term, affecting the absolute value response to the input; adjusts the degree of weighting of the absolute value of the input, increasing the smoothness; The The function is defined as follows: (19)。 6. The deep learning-based power grid line ice-melting effect detection method according to claim 1, characterized in that, The definition of the HCA activation function is as follows: The up-sampling process of the EEAUP module is as follows: (20) wherein represents a regular up-sampling operation; Then, for any pixel position in the feature map the gradient magnitude is calculated using a Sobel filter or gradient operator: (21) First, given the input feature map X and the target up-sampling scale s, the initial up-sampling result is generated by using bilinear interpolation or deconvolution: (22) wherein is a small constant that prevents numerical instability; After normalization, the edge attention weight is obtained: (23) wherein, is a hyperparameter used to control the strength of the edge enhancement.

7. The deep learning-based power grid line ice-melting effect detection method according to claim 1, characterized in that, Finally, the edge information is fused to adaptively adjust the preliminary up-sampling result to obtain the final up-sampling result: The specific steps of S3 are as follows: S31: Using the power grid line ice-melting effect detection dataset in S1 to train the power grid line ice-melting effect detection network GIMEDNet built in S2; 8. The deep learning-based power grid line ice-melting effect detection method according to claim 1, characterized in that, S32: During the training process, GIMEDNet adjusts the weight values in the network by using the Adam algorithm to optimize the loss function, so that the loss function converges, thereby making the network reach a balanced state. The specific steps of S4 are as follows: S41: After the training of GIMEDNet is completed, GIMEDNet is applied to perform power grid line ice-melting effect detection; S42: In actual application, the collected power grid line ice-melting result image is input, GIMEDNet processes the ice-melting result image, and outputs the ice-melting effect detection result.

Images