Method and system for detecting corrosion state of overhead ground wire based on deep learning

Abstract

The invention discloses an overhead ground wire corrosion state detection method and system based on deep learning, and the method comprises the steps: obtaining an overhead ground wire surface image, and constructing a semantic segmentation data set comprising a background and different corrosion grade categories; constructing a deep learning semantic segmentation model based on an encoder-decoder architecture; the encoder comprises a deformable convolution attention module and a multi-scale feed-forward network so as to extract corrosion features; the decoder adopts a full convolution structure, fuses the features extracted by the encoder in each stage, and outputs a semantic segmentation result; training the deep learning semantic segmentation model by using the semantic segmentation data set; and detecting the surface of the on-site overhead ground wire by using the trained deep learning semantic segmentation model, and outputting a corrosion level semantic segmentation result of the surface image of the on-site overhead ground wire. According to the method, a deep learning semantic segmentation model based on an encoder-decoder architecture is combined, and semantic segmentation identification of the corrosion level of the corrosion area of the overhead ground wire is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of corrosion detection, in particular to an overhead ground wire corrosion state detection method and system based on deep learning. BACKGROUND

[0002] As the first lightning protection barrier of the power transmission line, the overhead ground wire is exposed to the atmospheric environment for a long time and bears multiple environmental stresses such as acid rain erosion, salt mist corrosion, chemical pollutant deposition and temperature and humidity changes. The surface coating layer of the overhead ground wire serves as a sacrificial protective layer to protect the steel substrate by corrosion, but once the coating layer is consumed, the steel substrate will be rapidly corroded, resulting in a sharp decrease in the tensile strength of the ground wire and causing strand breakage or even fracture accidents. Therefore, regular monitoring of the corrosion state of the coating layer is of great significance for preventing power accidents, optimizing maintenance decisions and prolonging the service life of the equipment.

[0003] Traditional detection methods mainly rely on manual visual inspection or contact measurement, which have the problems of low efficiency, strong subjectivity and difficulty in covering large areas. With the development of artificial intelligence deep learning, image-based detection methods have gradually attracted attention. However, the standard convolution in traditional semantic segmentation models cannot adapt to the irregular and multi-scale corrosion patterns on the metal surface, making it difficult to accurately extract corrosion features. SUMMARY

[0004] In view of the problem that the traditional semantic segmentation model in the prior art cannot accurately extract corrosion features in corrosion detection, the present application provides an overhead ground wire corrosion state detection method and system based on deep learning, which can realize corrosion level semantic segmentation and identification of the corrosion area of the overhead ground wire by constructing a semantic segmentation dataset including background and different corrosion level categories, and combining a deep learning semantic segmentation model based on an encoder-decoder architecture. The specific technical solutions are as follows: In a first aspect, the present application provides an overhead ground wire corrosion state detection method based on deep learning, comprising the following steps: Obtaining an overhead ground wire surface image and constructing a semantic segmentation dataset including background and different corrosion level categories; Constructing a deep learning semantic segmentation model based on an encoder-decoder architecture; the encoder includes a deformable convolution attention module to extract irregular corrosion area features, and the encoder also includes a multi-scale feedforward network to extract corrosion features of different scales through a parallel multi-branch convolution structure; the decoder adopts a full convolution structure to gradually upsample through bilinear interpolation, fuse the features extracted by each stage of the encoder, and output a semantic segmentation result; Training the deep learning semantic segmentation model using the semantic segmentation dataset; The trained deep learning semantic segmentation model is used for detecting the surface of the overhead ground wire on site, and outputs the corrosion grade semantic segmentation result of the overhead ground wire surface image on site.

[0005] Preferably, the encoder adopts a four-stage down-sampling structure, each stage including a down-sampling module and a plurality of basic blocks, each basic block including a deformable convolution attention module and a multi-scale feedforward network.

[0006] Preferably, the attention module of the deformable convolution generates linear mapping of the input features through two parallel 1x1 convolution layers; performs GELU activation function and deformable convolution operation on the first mapping to generate spatial attention weight; performs Hadamard product operation on the spatial attention weight and the second linear mapping, and then outputs the features through 1x1 convolution.

[0007] Preferably, the processing process of the deformable convolution attention module is described as: In the formula, represents a spatial mixer; is a Hadamard product operation; , , are three convolution layer weights, respectively; represents deformable convolution with a kernel size of ; represents GELU activation function.

[0008] Preferably, the multi-scale feedforward network splits the input features in the channel dimension and inputs them into a deep separable convolution layer branch with a convolution kernel size of 3x3, 5x5 and 7x7 and an identity mapping branch for processing, splices the outputs of each branch, and fuses them through 1x1 convolution to obtain multi-scale information features.

[0009] Preferably, the processing process of the multi-scale feedforward network is described as: In the formula, represents a channel mixer operation; is the input of MSFFN; and are the weights of the convolution layer with an expansion ratio of , represents pyramid convolution with a convolution kernel size of .

[0010] Preferably, the construction of the semantic segmentation dataset including background and different corrosion level categories adopts data augmentation operations, and the data augmentation includes one or more of random cropping, random scaling, random horizontal / vertical flipping, random rotation, Gaussian blur, color disturbance.

[0011] Preferably, the detection of the trained deep learning semantic segmentation model on the surface of the live overhead ground wire includes: deploying the trained deep learning semantic segmentation model on an unmanned aerial vehicle embedded platform, and collecting live overhead ground wire surface images through an image acquisition unit of the unmanned aerial vehicle for inference detection; or collecting live overhead ground wire surface images and corresponding positioning data through the unmanned aerial vehicle, and transmitting the images and the positioning data to a server on which the trained deep learning semantic segmentation model is deployed for inference detection.

[0012] Preferably, after the corrosion level semantic segmentation result of the live overhead ground wire surface image is output, the method further includes: based on the corrosion level semantic segmentation result, calculating an overhead ground wire surface corrosion area proportion, and performing connected domain analysis to output a corrosion risk index.

[0013] In a second aspect, the present application further provides an overhead ground wire corrosion state detection system based on deep learning, which applies the aforementioned method and includes: a data acquisition unit configured to acquire overhead ground wire surface images and construct a semantic segmentation dataset including background and different corrosion level categories; a semantic segmentation model construction unit configured to construct a deep learning semantic segmentation model based on an encoder-decoder architecture; the encoder includes a deformable convolution attention module to extract irregular corrosion region features, and the encoder further includes a multi-scale feedforward network to extract corrosion features of different scales through a parallel multi-branch convolution structure; the decoder adopts a full convolution structure to gradually upsample through bilinear interpolation, fuse features extracted by the encoder at each stage, and output a semantic segmentation result; a semantic segmentation model training unit configured to train the deep learning semantic segmentation model using the semantic segmentation dataset; a corrosion detection unit configured to detect the live overhead ground wire surface using the trained deep learning semantic segmentation model, and output a corrosion level semantic segmentation result of the live overhead ground wire surface image.

[0014] Compared with the prior art, the present application has the following beneficial effects: The deep learning-based overhead ground wire corrosion state detection method and system of the present application realizes corrosion level semantic segmentation identification of the corrosion area of the overhead ground wire by constructing a semantic segmentation data set containing background and different corrosion level categories, and combining a deep learning semantic segmentation model based on an encoder-decoder architecture. The deformable convolution attention module in the encoder can effectively extract irregular corrosion area features, the multi-scale feedforward network extracts corrosion features of different scales through a parallel multi-branch convolution structure, the decoder adopts a full convolution structure and gradually fuses features through bilinear interpolation upsampling, and an accurate overhead ground wire corrosion level semantic segmentation result is output. The deep learning semantic segmentation model based on the encoder-decoder architecture discards the dependence of the traditional Transformer structure on large-scale pre-training, fully utilizes the inductive bias of convolution, is more suitable for industrial real-time detection scenarios, and can effectively improve the automation level and accuracy of overhead ground wire corrosion detection, and ensure the safe and stable operation of the power system. BRIEF DESCRIPTION OF DRAWINGS

[0015] In order to more clearly illustrate the specific embodiments of the present application or the technical solutions in the prior art, the drawings needed to be used in the specific embodiments or the prior art description will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, each element or part is not necessarily drawn according to the actual scale.

[0016] Figure 1 A deep learning-based overhead ground wire corrosion state detection method flowchart of an embodiment of the present application.

[0017] Figure 2 A deformable convolution attention module network structure schematic diagram of an embodiment of the present application.

[0018] Figure 3 A multi-scale feedforward network structure schematic diagram of an embodiment of the present application.

[0019] Figure 4 A whole architecture schematic diagram of a model deployment background server of an embodiment of the present application.

[0020] Figure 5 A deep learning-based overhead ground wire corrosion state detection system schematic diagram of an embodiment of the present application. DETAILED DESCRIPTION

[0021] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0022] It should be understood that the terms "comprises" and "comprising," when used in this specification, indicate the presence of the stated features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0023] It should also be understood that the terms used in the present specification are only for the purpose of describing particular embodiments and are not intended to limit the present application. As used in the present specification and the appended claims, the singular forms "a," "an" and "the" are intended to include the plural forms unless the context clearly indicates otherwise.

[0024] It should be further understood that the term "and / or" used in the present specification means any combination of one or more of the associated listed items and all possible combinations, and includes these combinations.

[0025] The following examples are provided Figures 1 to 5 .

[0026] As Figure 1 shown, the embodiment of the present application provides a deep learning-based overhead ground wire corrosion state detection method, comprising the following steps: Step S1, acquiring an overhead ground wire surface image, and constructing a semantic segmentation dataset including a background and different corrosion level categories; Real overhead ground wire corrosion images can be collected by a drone as a dataset, which is divided into six corrosion level categories according to the power technical supervision regulations, and semantic segmentation annotation is performed to obtain a background and six corrosion level categories. For example, select a galvanized steel overhead ground wire of a 110kV transmission line in a coastal area as a detection object. In clear and non-fog weather (humidity <70%, light intensity >50000Lux), control the drone to cruise along the ground wire at a predetermined height and speed. The camera collects high-definition images of the top and side of the overhead ground wire at a frequency of 1 frame per second, about 2000 original images are collected for each line, and corresponding corrosion images are selected. In some embodiments, different corrosion level overhead ground wire samples can also be prepared by salt spray test, and overhead ground wire surface images are obtained by shooting overhead ground wire samples as a dataset; or by searching the Internet and other open databases, etc.

[0027] The data amount is significantly expanded through a series of data enhancement operations. Specifically, random cropping is performed on each original image and its labeled mask simultaneously, such as randomly cropping a 512x512 sub-image from a 1024x1024 image to simulate different local perspectives; or random scaling is performed on the original image, with a scale range of [0.8, 1.2] to learn multi-scale features, or random horizontal / vertical flipping is performed on the original image; or random rotation (within ±15°) is performed on the original image; or Gaussian blur is performed on the original image; or color perturbation is performed on the original image, such as adjusting brightness and contrast to cope with light changes and the like, thereby efficiently generating a large number of training samples with different shapes, scales, and environments.

[0028] Step S2, constructing a deep learning semantic segmentation model based on an encoder-decoder architecture; the encoder includes a deformable convolution attention module to extract irregular erosion region features, and the encoder also includes a multi-scale feedforward network to extract erosion features of different scales through a parallel multi-branch convolution structure; the decoder adopts a full convolution structure, gradually up-samples through bilinear interpolation, fuses the features extracted by the encoder at each stage, and outputs a semantic segmentation result. The encoder adopts a four-stage down-sampling structure, each stage including a down-sampling module and a plurality of basic blocks, and each basic block including a deformable convolution attention module and a multi-scale feedforward network.

[0029] As shown in Figure 2 , the deformable convolution attention module generates linear mapping of input features through two parallel 1x1 convolution layers; performs GELU activation function and deformable convolution operation on the first mapping to generate spatial attention weights; performs Hadamard product operation on the spatial attention weights and the second linear mapping, and then outputs features through a 1x1 convolution. The deformable convolution is decomposed into depth convolution + pointwise convolution to reduce the calculation cost of offset generation; and the softmax function (instead of sigmoid) is used to normalize the modulation scalar to improve the training stability.

[0030] The processing process of the deformable convolution attention module is described as: In the formula, represents a spatial mixer; is a Hadamard product operation; , , are three convolution layer weights, respectively; represents a kernel size Deformable convolution; This represents the GELU activation function. In this embodiment, the kernel size is set to 3×3, and the following formula is used. The formula for calculating the GELU value is as follows: like Figure 3 As shown, the multi-scale feedforward network splits the input features along the channel dimension and inputs them into depth-separable convolutional layer branches with kernel sizes of 3×3, 5×5, and 7×7, and an identity mapping branch for processing. By concatenating the outputs of each branch, the multi-scale information features are obtained by fusion through 1×1 convolution.

[0031] The processing procedure of the multi-scale feedforward network is expressed as follows: In the formula, Indicates channel mixer operation; For MSFFN input; and These represent expansion ratios respectively. of The weights of the convolutional layer Indicates the kernel size as The pyramid convolution is calculated using the following formula in this embodiment: in, , , These represent three depthwise separable convolutional layer branches; Indicates the branch of the identity mapping; Indicates the number of channels in the convolution branch; sets a ratio. The table determines the number of channels in each branch, and its calculation formula is as follows: ; Indicates the kernel size as Depth-wise convolution; set the default kernel size to... ; This indicates that the outputs of each branch are spliced ​​along the channel dimension.

[0032] In summary, the processing procedure of BasicBlock can be formulated as follows: in, are the input, output and output of a channel mixer, respectively, denotes a normalization layer operation.

[0033] In the present embodiment, StemBlock captures local details and down-samples by 4 times through a convolutional layer; four-stage encoder (output resolution from 1 / 4 to 1 / 32), each stage contains DownsamplingBlock (depthwise convolution + step 2 down-sampling) and stacked BasicBlock (deformable convolution attention module + multi-scale feedforward network).

[0034] The encoder adopts a full convolution structure instead of an MLP decoder to avoid dimension transformation overhead; the multi-scale feature maps of the four stages of the encoder are fused and up-sampled to the original resolution by bilinear interpolation; the semantic segmentation mask (semantic segmentation result) is output by 1x1 convolution, and the background and corrosion level regions are predicted.

[0035] The decoder adopts a lightweight full convolution structure to avoid using MLP with huge computational complexity. By receiving multi-scale feature maps (e.g. size 1 / 4, 1 / 8, 1 / 16, 1 / 32 of the original image) output by the four stages of the encoder. First, adjust the channel number of each feature map to a unified dimension by 1x1 convolution. Then, starting from the deepest feature, sequentially perform 2 times up-sampling by bilinear interpolation (BilinearInterpolation), and perform skip connection (SkipConnection) and addition fusion with the feature map of the corresponding stage of the encoder. The fused feature is refined by a 3x3 convolution. Repeat the up-sampling-fusion-convolution process until the feature map is restored to the original input resolution (512x512). Finally, use a 1x1 convolution layer to map the channel number to the class number, divide it into 7, and output the corrosion class probability of each pixel belonging to the 7 classes.

[0036] Step S3, training the deep learning semantic segmentation model using the semantic segmentation dataset; The labeled semantic segmentation dataset is randomly divided into training set, validation set and test set according to a certain proportion, such as 70%:15%:15%, which are respectively used for deep learning semantic segmentation model training, hyperparameter tuning and final performance evaluation.

[0037] ​​Based on the encoder-decoder based deep learning semantic segmentation model built in step S2, the training set data is input into the model for forward propagation, the loss is calculated, and then the network weight is updated through the back propagation algorithm for iterative optimization. During training, instead of scaling all images to a fixed size, multiple scales such as 256x256, 512x512, 1024x1024, etc. are fed into a batch, or the same image is randomly scaled. The model is forced to learn the corrosion features at different scales, so that it can maintain stable performance when facing input images of different resolutions or different shooting distances.

[0038] After the end of each training cycle, the current model performance is evaluated using the validation set. Core indicators such as mIoU and mPA are calculated to monitor whether the model is overfitting or underfitting. The model weight that performs best on the validation set during the entire training process is saved, rather than the model of the last Epoch. After training is complete, the final saved best model is evaluated for performance in a fair manner using a test set that has never participated in any training process, resulting in a trained deep learning semantic segmentation model that can output the corrosion level semantic segmentation result of the aerial ground wire surface image.

[0039] Step S4, the trained deep learning semantic segmentation model detects the on-site aerial ground wire surface and outputs the corrosion level semantic segmentation result of the on-site aerial ground wire surface image.

[0040] The trained model is deployed to the actual environment to automatically analyze the on-site aerial ground wire surface image and output the corrosion level semantic segmentation result of the on-site aerial ground wire surface image. For example, the system receives the on-site aerial ground wire image returned by the on-site unmanned aerial vehicle or input locally. The input image is standardized to be consistent with the training, including adjusting the image to the required input size of the model, such as 512x512. The pixel value is normalized from [0, 255] to [0, 1] or standardized using mean and standard deviation. The preprocessed image tensor is input into the loaded model, and the special deformable convolution module in the model dynamically adjusts the receptive field and focuses on irregular corrosion areas to capture their unique texture details. The multi-scale feedforward network enhances the local multi-scale information in the encoder by using different scales of depth separable convolution in parallel to aggregate fine-grained multi-scale information features. The multi-level, multi-scale information features extracted by the encoder are sent to the decoder.

[0041] The decoder fuses features from different stages of the encoder through a skip connection. Through operations such as bilinear upsampling, the feature map size is gradually restored to the original input resolution. Finally, a 1x1 convolution layer is used to output the corrosion class probability of each pixel belonging to 7 categories.

[0042] Specifically, the output live overhead ground wire surface image corrosion level semantic segmentation result also includes: Based on the corrosion level semantic segmentation result, the output overhead ground wire surface corrosion area proportion is calculated, and connected component analysis is performed to output the corrosion risk index.

[0043] Among them, the ArgMax operation is performed on the output corrosion category probability to determine the final category, and a single-channel semantic segmentation mask image is generated. Post-processing calculation is performed based on the semantic segmentation mask image.

[0044] The mean intersection over union (mIoU) is the most commonly used and most representative index in semantic segmentation tasks. It calculates the average intersection over union of all classes between the ground truth and the prediction. The calculation formula is: In the formula, is the total number of classes, background and six corrosion levels; is the current class index being calculated (for example represents the background, represents the first level of corrosion); (True Positives for class k) is the true positive, for class , the number of pixels correctly predicted by the model as class ; (False Positives for class k): is the false positive, for class , the number of pixels incorrectly predicted by the model as class ; (False Negatives for class k) is the false negative, for class , the number of pixels that the model fails to predict as class .

[0045] mPA (average pixel classification accuracy of all classes) calculates the average proportion of correctly classified pixels to the total number of pixels for all classes, and the calculation formula is: Calculate the total corrosion area proportion: Wherein, NumberofPixels(Class1to6) is the total number of pixels of all corrosion grades, representing the area of the entire corrosion region; NumberofPixels(Class0to6) is the total number of pixels of the overhead ground wire region, representing the total area of the overhead ground wire.

[0046] By performing connected component analysis, the size and number of independent corrosion regions are identified. The segmentation mask is rendered into a color map (different colors represent different corrosion grades), which is superimposed on the original image to generate an intuitive corrosion analysis effect diagram.

[0047] When the corrosion degree reaches level five or six, the overhead ground wire needs to be immediately overhauled and replaced. If the matching result is that the corrosion degree reaches level three or four, targeted maintenance recommendations are generated and the inspection and monitoring are strengthened; if it is lower than level three, the inspection is maintained, and the suggestion of increasing the inspection frequency can be made according to the aggravation of the degree.

[0048] It should be noted that the deformable convolution attention module proposed in the embodiment fuses deformable convolution and convolution attention. By dynamically generating sampling offsets and modulation weights depending on the input content through deformable convolution, the spatial receptive field is adaptively adjusted, so that the network can accurately focus on the irregularly shaped corrosion region; and then through the spatial attention mechanism constructed by Hadamard product, the features are content-aware weighted aggregation, so as to realize the adaptive feature aggregation capability similar to Transformer while maintaining the computational efficiency of convolution, and improve the modeling effect of complex corrosion morphology. It can adaptively aggregate features and explicitly model spatial interactions (similar to self-attention mechanism), while with the increase of image resolution, the computational complexity is reduced from quadratic to linear. At the same time, the multi-scale feedforward network proposed in the embodiment enhances the local multi-scale information in the encoder, which effectively aggregates the fine-grained multi-scale information features by using deep separable convolutions of different scales in parallel.

[0049] The deep learning-based overhead ground wire corrosion state detection method and system constructs a semantic segmentation dataset containing a background and different corrosion level categories, and combines a deep learning semantic segmentation model based on an encoder-decoder architecture to achieve corrosion level semantic segmentation identification of the corrosion area of the overhead ground wire. The deformable convolution attention module in the encoder can effectively extract irregular corrosion area features, the multi-scale feedforward network extracts corrosion features of different scales through a parallel multi-branch convolution structure, the decoder adopts a full convolution structure and gradually fuses features through bilinear interpolation upsampling to output accurate overhead ground wire corrosion level semantic segmentation results. The deep learning semantic segmentation model based on the encoder-decoder architecture discards the dependence of the traditional Transformer structure on large-scale pre-training, fully utilizes the inductive bias of convolution, is more suitable for industrial real-time detection scenarios, and can effectively improve the automation level and accuracy of overhead ground wire corrosion detection and ensure the safe and stable operation of the power system.

[0050] Specifically, the detection of the field overhead ground wire surface by the trained deep learning semantic segmentation model comprises: The trained deep learning semantic segmentation model is deployed on a UAV embedded platform, and field overhead ground wire surface images are collected by an image collection unit of the UAV for inference detection; In specific implementation, the trained model is optimized through quantization, pruning, etc., is converted into a format suitable for embedded hardware, and is integrated into a UAV flight control system. When the UAV is patrolling, the onboard camera of the UAV collects ground wire surface images in real time. The images are directly sent to the onboard computing unit for real-time preprocessing and model inference. The lightweight deep learning semantic segmentation model outputs segmentation results and corrosion level judgments in real time on the UAV.

[0051] As shown in Figure 4 The field overhead ground wire surface images and corresponding positioning data are collected by the UAV and transmitted to a background server on which the trained deep learning semantic segmentation model is deployed for inference detection; In specific implementation, the UAV collects high-definition ground wire images and corresponding GPS, IMU, etc. positioning and attitude data, and transmits them to an application server through a 4G / 5G network. The application server receives the data and transmits them to an AI inference server. The AI inference server runs the complete deep learning semantic segmentation model for inference analysis. Complex post-processing calculations are completed on the server side, and the original data, analysis results and reports are stored in a database. The generated corrosion report, early warning information and visual result graph are pushed to the mobile terminal or monitoring large screen of the inspection personnel.

[0052] As shown in Figure 5 The embodiment of the present application also provides a deep learning-based overhead ground wire corrosion state detection system, which applies the foregoing method and comprises: a data acquisition unit configured to acquire an overhead ground wire surface image and construct a semantic segmentation dataset including a background and different corrosion level categories; a semantic segmentation model construction unit configured to construct a deep learning semantic segmentation model based on an encoder-decoder architecture; the encoder includes a deformable convolution attention module to extract irregular corrosion region features, and the encoder further includes a multi-scale feedforward network to extract corrosion features of different scales through a parallel multi-branch convolution structure; the decoder adopts a full convolution structure to gradually upsample through bilinear interpolation and fuse features extracted by the encoder at different stages to output a semantic segmentation result; a semantic segmentation model training unit configured to train the deep learning semantic segmentation model using the semantic segmentation dataset; a corrosion detection unit configured to detect an overhead ground wire surface in the field using the trained deep learning semantic segmentation model and output a corrosion level semantic segmentation result of an overhead ground wire surface image in the field.

[0053] The functions of the units in the embodiment are explained in the same way as a deep learning-based overhead ground wire corrosion state detection method, and have the same technical effects, which are not repeated here.

[0054] Those skilled in the art can appreciate that the units of each example described in combination with the embodiments disclosed herein can be realized in electronic hardware, computer software or a combination of both. In order to clearly illustrate the interchangeability of hardware and software, the components of each example have been described in the above description in general terms. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. The skilled person can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0055] In the embodiments provided in the present application, it should be understood that the division of units is only a logical functional division, and when actually implemented, there can be another division manner, for example, multiple units can be combined into one unit, one unit can be split into multiple units, or some features can be ignored, etc.

[0056] In addition, each functional unit in each embodiment of the present application can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0057] The integrated unit, if implemented in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or say the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server or a network device, etc.) to execute all or part of the steps of the method described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a mobile hard disk, a magnetic disk or an optical disk, and various media that can store program codes.

[0058] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application, and they should be covered in the scope of the specification of the present application.

Claims

1. A deep learning-based method for detecting the corrosion status of overhead ground wires, characterized in that, Includes the following steps: Obtain surface images of overhead ground wires and construct a semantic segmentation dataset that includes background and different corrosion level categories; A deep learning semantic segmentation model based on an encoder-decoder architecture is constructed. The encoder includes a deformable convolutional attention module to extract features of irregular eroded regions. The encoder also includes a multi-scale feedforward network to extract eroded features at different scales through parallel multi-branch convolutional structures. The decoder employs a fully convolutional structure, progressively upsampling through bilinear interpolation, fusing features extracted from each stage of the encoder, and outputting semantic segmentation results. The deep learning semantic segmentation model is trained using the semantic segmentation dataset. The trained deep learning semantic segmentation model is used to detect the surface of the overhead ground wire in the field, and the semantic segmentation results of the corrosion level of the surface image of the overhead ground wire are output.

2. The method for detecting the corrosion status of overhead ground wires based on deep learning according to claim 1, characterized in that, The encoder adopts a four-stage downsampling structure, with each stage including a downsampling module and multiple base blocks. Each base block includes a deformable convolutional attention module and a multi-scale feedforward network.

3. The method for detecting the corrosion status of overhead ground wires based on deep learning according to claim 2, characterized in that, The attention module of the deformable convolution generates a linear mapping of the input features through two parallel 1×1 convolutional layers; Perform the GELU activation function and deformable convolution operation on the first path mapping to generate spatial attention weights; The spatial attention weights are subjected to a Hadamard product with the second linear mapping, and then the features are output through a 1×1 convolution.

4. The method for detecting the corrosion status of overhead ground wires based on deep learning according to claim 3, characterized in that, The processing procedure of the deformable convolutional attention module is described as follows: In the formula, Indicates a spatial mixer; It is the Hadamard product operation; , , There are three of them. Convolutional layer weights; Indicates the kernel size as Deformable convolution; This represents the GELU activation function.

5. The method for detecting the corrosion status of overhead ground wires based on deep learning according to claim 2, characterized in that, The multi-scale feedforward network splits the input features along the channel dimension and feeds them into depth-separable convolutional layer branches with kernel sizes of 3×3, 5×5, and 7×7, as well as an identity mapping branch for processing. By concatenating the outputs of each branch, the multi-scale information features are obtained through 1×1 convolution.

6. The method for detecting the corrosion status of overhead ground wires based on deep learning according to claim 5, characterized in that, The processing procedure of the multi-scale feedforward network is expressed as follows: In the formula, Indicates channel mixer operation; For MSFFN input; and These represent expansion ratios respectively. of The weights of the convolutional layer Indicates the kernel size as Pyramid convolution.

7. The method for detecting the corrosion status of overhead ground wires based on deep learning according to claim 1, characterized in that, The construction of the semantic segmentation dataset, which includes background and different corrosion level categories, employs data augmentation operations, including one or more of the following: random cropping, random scaling, random horizontal / vertical flipping, random rotation, Gaussian blur, and color perturbation.

8. The method for detecting the corrosion status of overhead ground wires based on deep learning according to claim 1, characterized in that, The step of using the trained deep learning semantic segmentation model to detect the surface of the overhead ground wire includes: The trained deep learning semantic segmentation model is deployed on the UAV embedded platform, and the UAV's image acquisition unit collects images of the surface of the overhead ground wire for inference and detection. Alternatively, drones can be used to collect images of the surface of overhead ground wires and corresponding positioning data, which can then be transmitted to a server equipped with a trained deep learning semantic segmentation model for inference and detection.

9. A method for detecting the corrosion status of overhead ground wires based on deep learning according to any one of claims 1, characterized in that, After outputting the semantic segmentation results of the corrosion level of the overhead ground wire surface image, the following is also included: Based on the semantic segmentation results of corrosion level, the percentage of corrosion area on the surface of overhead ground wire is calculated and output, and connected component analysis is performed to output the corrosion risk index.

10. A deep learning-based overhead ground wire corrosion status detection system, characterized in that, The method described by any one of claims 1 to 9 includes: The data acquisition unit is used to acquire surface images of overhead ground wires and construct a semantic segmentation dataset that includes background and different corrosion level categories; The semantic segmentation model building unit is used to construct a deep learning semantic segmentation model based on an encoder-decoder architecture. The encoder includes a deformable convolutional attention module to extract features of irregular eroded regions. The encoder also includes a multi-scale feedforward network to extract eroded features at different scales through a parallel multi-branch convolutional structure. The decoder adopts a fully convolutional structure and gradually upsamples through bilinear interpolation, fusing the features extracted at each stage of the encoder to output the semantic segmentation result. A semantic segmentation model training unit is used to train the deep learning semantic segmentation model using the semantic segmentation dataset; The corrosion detection unit is used to detect the surface of the overhead ground wire in the field using the trained deep learning semantic segmentation model, and output the semantic segmentation results of the corrosion level of the surface image of the overhead ground wire.

Images